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 unsigned X86TTIImpl::getRegisterBitWidth(bool Vector) const { 133 unsigned PreferVectorWidth = ST->getPreferVectorWidth(); 134 if (Vector) { 135 if (ST->hasAVX512() && PreferVectorWidth >= 512) 136 return 512; 137 if (ST->hasAVX() && PreferVectorWidth >= 256) 138 return 256; 139 if (ST->hasSSE1() && PreferVectorWidth >= 128) 140 return 128; 141 return 0; 142 } 143 144 if (ST->is64Bit()) 145 return 64; 146 147 return 32; 148 } 149 150 unsigned X86TTIImpl::getLoadStoreVecRegBitWidth(unsigned) const { 151 return getRegisterBitWidth(true); 152 } 153 154 unsigned X86TTIImpl::getMaxInterleaveFactor(unsigned VF) { 155 // If the loop will not be vectorized, don't interleave the loop. 156 // Let regular unroll to unroll the loop, which saves the overflow 157 // check and memory check cost. 158 if (VF == 1) 159 return 1; 160 161 if (ST->isAtom()) 162 return 1; 163 164 // Sandybridge and Haswell have multiple execution ports and pipelined 165 // vector units. 166 if (ST->hasAVX()) 167 return 4; 168 169 return 2; 170 } 171 172 int X86TTIImpl::getArithmeticInstrCost(unsigned Opcode, Type *Ty, 173 TTI::TargetCostKind CostKind, 174 TTI::OperandValueKind Op1Info, 175 TTI::OperandValueKind Op2Info, 176 TTI::OperandValueProperties Opd1PropInfo, 177 TTI::OperandValueProperties Opd2PropInfo, 178 ArrayRef<const Value *> Args, 179 const Instruction *CxtI) { 180 // TODO: Handle more cost kinds. 181 if (CostKind != TTI::TCK_RecipThroughput) 182 return BaseT::getArithmeticInstrCost(Opcode, Ty, CostKind, Op1Info, 183 Op2Info, Opd1PropInfo, 184 Opd2PropInfo, Args, CxtI); 185 // Legalize the type. 186 std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, Ty); 187 188 int ISD = TLI->InstructionOpcodeToISD(Opcode); 189 assert(ISD && "Invalid opcode"); 190 191 static const CostTblEntry GLMCostTable[] = { 192 { ISD::FDIV, MVT::f32, 18 }, // divss 193 { ISD::FDIV, MVT::v4f32, 35 }, // divps 194 { ISD::FDIV, MVT::f64, 33 }, // divsd 195 { ISD::FDIV, MVT::v2f64, 65 }, // divpd 196 }; 197 198 if (ST->useGLMDivSqrtCosts()) 199 if (const auto *Entry = CostTableLookup(GLMCostTable, ISD, 200 LT.second)) 201 return LT.first * Entry->Cost; 202 203 static const CostTblEntry SLMCostTable[] = { 204 { ISD::MUL, MVT::v4i32, 11 }, // pmulld 205 { ISD::MUL, MVT::v8i16, 2 }, // pmullw 206 { ISD::MUL, MVT::v16i8, 14 }, // extend/pmullw/trunc sequence. 207 { ISD::FMUL, MVT::f64, 2 }, // mulsd 208 { ISD::FMUL, MVT::v2f64, 4 }, // mulpd 209 { ISD::FMUL, MVT::v4f32, 2 }, // mulps 210 { ISD::FDIV, MVT::f32, 17 }, // divss 211 { ISD::FDIV, MVT::v4f32, 39 }, // divps 212 { ISD::FDIV, MVT::f64, 32 }, // divsd 213 { ISD::FDIV, MVT::v2f64, 69 }, // divpd 214 { ISD::FADD, MVT::v2f64, 2 }, // addpd 215 { ISD::FSUB, MVT::v2f64, 2 }, // subpd 216 // v2i64/v4i64 mul is custom lowered as a series of long: 217 // multiplies(3), shifts(3) and adds(2) 218 // slm muldq version throughput is 2 and addq throughput 4 219 // thus: 3X2 (muldq throughput) + 3X1 (shift throughput) + 220 // 3X4 (addq throughput) = 17 221 { ISD::MUL, MVT::v2i64, 17 }, 222 // slm addq\subq throughput is 4 223 { ISD::ADD, MVT::v2i64, 4 }, 224 { ISD::SUB, MVT::v2i64, 4 }, 225 }; 226 227 if (ST->isSLM()) { 228 if (Args.size() == 2 && ISD == ISD::MUL && LT.second == MVT::v4i32) { 229 // Check if the operands can be shrinked into a smaller datatype. 230 bool Op1Signed = false; 231 unsigned Op1MinSize = BaseT::minRequiredElementSize(Args[0], Op1Signed); 232 bool Op2Signed = false; 233 unsigned Op2MinSize = BaseT::minRequiredElementSize(Args[1], Op2Signed); 234 235 bool signedMode = Op1Signed | Op2Signed; 236 unsigned OpMinSize = std::max(Op1MinSize, Op2MinSize); 237 238 if (OpMinSize <= 7) 239 return LT.first * 3; // pmullw/sext 240 if (!signedMode && OpMinSize <= 8) 241 return LT.first * 3; // pmullw/zext 242 if (OpMinSize <= 15) 243 return LT.first * 5; // pmullw/pmulhw/pshuf 244 if (!signedMode && OpMinSize <= 16) 245 return LT.first * 5; // pmullw/pmulhw/pshuf 246 } 247 248 if (const auto *Entry = CostTableLookup(SLMCostTable, ISD, 249 LT.second)) { 250 return LT.first * Entry->Cost; 251 } 252 } 253 254 if ((ISD == ISD::SDIV || ISD == ISD::SREM || ISD == ISD::UDIV || 255 ISD == ISD::UREM) && 256 (Op2Info == TargetTransformInfo::OK_UniformConstantValue || 257 Op2Info == TargetTransformInfo::OK_NonUniformConstantValue) && 258 Opd2PropInfo == TargetTransformInfo::OP_PowerOf2) { 259 if (ISD == ISD::SDIV || ISD == ISD::SREM) { 260 // On X86, vector signed division by constants power-of-two are 261 // normally expanded to the sequence SRA + SRL + ADD + SRA. 262 // The OperandValue properties may not be the same as that of the previous 263 // operation; conservatively assume OP_None. 264 int Cost = 265 2 * getArithmeticInstrCost(Instruction::AShr, Ty, CostKind, Op1Info, 266 Op2Info, 267 TargetTransformInfo::OP_None, 268 TargetTransformInfo::OP_None); 269 Cost += getArithmeticInstrCost(Instruction::LShr, Ty, CostKind, Op1Info, 270 Op2Info, 271 TargetTransformInfo::OP_None, 272 TargetTransformInfo::OP_None); 273 Cost += getArithmeticInstrCost(Instruction::Add, Ty, CostKind, Op1Info, 274 Op2Info, 275 TargetTransformInfo::OP_None, 276 TargetTransformInfo::OP_None); 277 278 if (ISD == ISD::SREM) { 279 // For SREM: (X % C) is the equivalent of (X - (X/C)*C) 280 Cost += getArithmeticInstrCost(Instruction::Mul, Ty, CostKind, Op1Info, 281 Op2Info); 282 Cost += getArithmeticInstrCost(Instruction::Sub, Ty, CostKind, Op1Info, 283 Op2Info); 284 } 285 286 return Cost; 287 } 288 289 // Vector unsigned division/remainder will be simplified to shifts/masks. 290 if (ISD == ISD::UDIV) 291 return getArithmeticInstrCost(Instruction::LShr, Ty, CostKind, 292 Op1Info, Op2Info, 293 TargetTransformInfo::OP_None, 294 TargetTransformInfo::OP_None); 295 296 else // UREM 297 return getArithmeticInstrCost(Instruction::And, Ty, CostKind, 298 Op1Info, Op2Info, 299 TargetTransformInfo::OP_None, 300 TargetTransformInfo::OP_None); 301 } 302 303 static const CostTblEntry AVX512BWUniformConstCostTable[] = { 304 { ISD::SHL, MVT::v64i8, 2 }, // psllw + pand. 305 { ISD::SRL, MVT::v64i8, 2 }, // psrlw + pand. 306 { ISD::SRA, MVT::v64i8, 4 }, // psrlw, pand, pxor, psubb. 307 }; 308 309 if (Op2Info == TargetTransformInfo::OK_UniformConstantValue && 310 ST->hasBWI()) { 311 if (const auto *Entry = CostTableLookup(AVX512BWUniformConstCostTable, ISD, 312 LT.second)) 313 return LT.first * Entry->Cost; 314 } 315 316 static const CostTblEntry AVX512UniformConstCostTable[] = { 317 { ISD::SRA, MVT::v2i64, 1 }, 318 { ISD::SRA, MVT::v4i64, 1 }, 319 { ISD::SRA, MVT::v8i64, 1 }, 320 321 { ISD::SHL, MVT::v64i8, 4 }, // psllw + pand. 322 { ISD::SRL, MVT::v64i8, 4 }, // psrlw + pand. 323 { ISD::SRA, MVT::v64i8, 8 }, // psrlw, pand, pxor, psubb. 324 325 { ISD::SDIV, MVT::v16i32, 6 }, // pmuludq sequence 326 { ISD::SREM, MVT::v16i32, 8 }, // pmuludq+mul+sub sequence 327 { ISD::UDIV, MVT::v16i32, 5 }, // pmuludq sequence 328 { ISD::UREM, MVT::v16i32, 7 }, // pmuludq+mul+sub sequence 329 }; 330 331 if (Op2Info == TargetTransformInfo::OK_UniformConstantValue && 332 ST->hasAVX512()) { 333 if (const auto *Entry = CostTableLookup(AVX512UniformConstCostTable, ISD, 334 LT.second)) 335 return LT.first * Entry->Cost; 336 } 337 338 static const CostTblEntry AVX2UniformConstCostTable[] = { 339 { ISD::SHL, MVT::v32i8, 2 }, // psllw + pand. 340 { ISD::SRL, MVT::v32i8, 2 }, // psrlw + pand. 341 { ISD::SRA, MVT::v32i8, 4 }, // psrlw, pand, pxor, psubb. 342 343 { ISD::SRA, MVT::v4i64, 4 }, // 2 x psrad + shuffle. 344 345 { ISD::SDIV, MVT::v8i32, 6 }, // pmuludq sequence 346 { ISD::SREM, MVT::v8i32, 8 }, // pmuludq+mul+sub sequence 347 { ISD::UDIV, MVT::v8i32, 5 }, // pmuludq sequence 348 { ISD::UREM, MVT::v8i32, 7 }, // pmuludq+mul+sub sequence 349 }; 350 351 if (Op2Info == TargetTransformInfo::OK_UniformConstantValue && 352 ST->hasAVX2()) { 353 if (const auto *Entry = CostTableLookup(AVX2UniformConstCostTable, ISD, 354 LT.second)) 355 return LT.first * Entry->Cost; 356 } 357 358 static const CostTblEntry SSE2UniformConstCostTable[] = { 359 { ISD::SHL, MVT::v16i8, 2 }, // psllw + pand. 360 { ISD::SRL, MVT::v16i8, 2 }, // psrlw + pand. 361 { ISD::SRA, MVT::v16i8, 4 }, // psrlw, pand, pxor, psubb. 362 363 { ISD::SHL, MVT::v32i8, 4+2 }, // 2*(psllw + pand) + split. 364 { ISD::SRL, MVT::v32i8, 4+2 }, // 2*(psrlw + pand) + split. 365 { ISD::SRA, MVT::v32i8, 8+2 }, // 2*(psrlw, pand, pxor, psubb) + split. 366 367 { ISD::SDIV, MVT::v8i32, 12+2 }, // 2*pmuludq sequence + split. 368 { ISD::SREM, MVT::v8i32, 16+2 }, // 2*pmuludq+mul+sub sequence + split. 369 { ISD::SDIV, MVT::v4i32, 6 }, // pmuludq sequence 370 { ISD::SREM, MVT::v4i32, 8 }, // pmuludq+mul+sub sequence 371 { ISD::UDIV, MVT::v8i32, 10+2 }, // 2*pmuludq sequence + split. 372 { ISD::UREM, MVT::v8i32, 14+2 }, // 2*pmuludq+mul+sub sequence + split. 373 { ISD::UDIV, MVT::v4i32, 5 }, // pmuludq sequence 374 { ISD::UREM, MVT::v4i32, 7 }, // pmuludq+mul+sub sequence 375 }; 376 377 // XOP has faster vXi8 shifts. 378 if (Op2Info == TargetTransformInfo::OK_UniformConstantValue && 379 ST->hasSSE2() && !ST->hasXOP()) { 380 if (const auto *Entry = 381 CostTableLookup(SSE2UniformConstCostTable, ISD, LT.second)) 382 return LT.first * Entry->Cost; 383 } 384 385 static const CostTblEntry AVX512BWConstCostTable[] = { 386 { ISD::SDIV, MVT::v64i8, 14 }, // 2*ext+2*pmulhw sequence 387 { ISD::SREM, MVT::v64i8, 16 }, // 2*ext+2*pmulhw+mul+sub sequence 388 { ISD::UDIV, MVT::v64i8, 14 }, // 2*ext+2*pmulhw sequence 389 { ISD::UREM, MVT::v64i8, 16 }, // 2*ext+2*pmulhw+mul+sub sequence 390 { ISD::SDIV, MVT::v32i16, 6 }, // vpmulhw sequence 391 { ISD::SREM, MVT::v32i16, 8 }, // vpmulhw+mul+sub sequence 392 { ISD::UDIV, MVT::v32i16, 6 }, // vpmulhuw sequence 393 { ISD::UREM, MVT::v32i16, 8 }, // vpmulhuw+mul+sub sequence 394 }; 395 396 if ((Op2Info == TargetTransformInfo::OK_UniformConstantValue || 397 Op2Info == TargetTransformInfo::OK_NonUniformConstantValue) && 398 ST->hasBWI()) { 399 if (const auto *Entry = 400 CostTableLookup(AVX512BWConstCostTable, ISD, LT.second)) 401 return LT.first * Entry->Cost; 402 } 403 404 static const CostTblEntry AVX512ConstCostTable[] = { 405 { ISD::SDIV, MVT::v16i32, 15 }, // vpmuldq sequence 406 { ISD::SREM, MVT::v16i32, 17 }, // vpmuldq+mul+sub sequence 407 { ISD::UDIV, MVT::v16i32, 15 }, // vpmuludq sequence 408 { ISD::UREM, MVT::v16i32, 17 }, // vpmuludq+mul+sub sequence 409 { ISD::SDIV, MVT::v64i8, 28 }, // 4*ext+4*pmulhw sequence 410 { ISD::SREM, MVT::v64i8, 32 }, // 4*ext+4*pmulhw+mul+sub sequence 411 { ISD::UDIV, MVT::v64i8, 28 }, // 4*ext+4*pmulhw sequence 412 { ISD::UREM, MVT::v64i8, 32 }, // 4*ext+4*pmulhw+mul+sub sequence 413 { ISD::SDIV, MVT::v32i16, 12 }, // 2*vpmulhw sequence 414 { ISD::SREM, MVT::v32i16, 16 }, // 2*vpmulhw+mul+sub sequence 415 { ISD::UDIV, MVT::v32i16, 12 }, // 2*vpmulhuw sequence 416 { ISD::UREM, MVT::v32i16, 16 }, // 2*vpmulhuw+mul+sub sequence 417 }; 418 419 if ((Op2Info == TargetTransformInfo::OK_UniformConstantValue || 420 Op2Info == TargetTransformInfo::OK_NonUniformConstantValue) && 421 ST->hasAVX512()) { 422 if (const auto *Entry = 423 CostTableLookup(AVX512ConstCostTable, ISD, LT.second)) 424 return LT.first * Entry->Cost; 425 } 426 427 static const CostTblEntry AVX2ConstCostTable[] = { 428 { ISD::SDIV, MVT::v32i8, 14 }, // 2*ext+2*pmulhw sequence 429 { ISD::SREM, MVT::v32i8, 16 }, // 2*ext+2*pmulhw+mul+sub sequence 430 { ISD::UDIV, MVT::v32i8, 14 }, // 2*ext+2*pmulhw sequence 431 { ISD::UREM, MVT::v32i8, 16 }, // 2*ext+2*pmulhw+mul+sub sequence 432 { ISD::SDIV, MVT::v16i16, 6 }, // vpmulhw sequence 433 { ISD::SREM, MVT::v16i16, 8 }, // vpmulhw+mul+sub sequence 434 { ISD::UDIV, MVT::v16i16, 6 }, // vpmulhuw sequence 435 { ISD::UREM, MVT::v16i16, 8 }, // vpmulhuw+mul+sub sequence 436 { ISD::SDIV, MVT::v8i32, 15 }, // vpmuldq sequence 437 { ISD::SREM, MVT::v8i32, 19 }, // vpmuldq+mul+sub sequence 438 { ISD::UDIV, MVT::v8i32, 15 }, // vpmuludq sequence 439 { ISD::UREM, MVT::v8i32, 19 }, // vpmuludq+mul+sub sequence 440 }; 441 442 if ((Op2Info == TargetTransformInfo::OK_UniformConstantValue || 443 Op2Info == TargetTransformInfo::OK_NonUniformConstantValue) && 444 ST->hasAVX2()) { 445 if (const auto *Entry = CostTableLookup(AVX2ConstCostTable, ISD, LT.second)) 446 return LT.first * Entry->Cost; 447 } 448 449 static const CostTblEntry SSE2ConstCostTable[] = { 450 { ISD::SDIV, MVT::v32i8, 28+2 }, // 4*ext+4*pmulhw sequence + split. 451 { ISD::SREM, MVT::v32i8, 32+2 }, // 4*ext+4*pmulhw+mul+sub sequence + split. 452 { ISD::SDIV, MVT::v16i8, 14 }, // 2*ext+2*pmulhw sequence 453 { ISD::SREM, MVT::v16i8, 16 }, // 2*ext+2*pmulhw+mul+sub sequence 454 { ISD::UDIV, MVT::v32i8, 28+2 }, // 4*ext+4*pmulhw sequence + split. 455 { ISD::UREM, MVT::v32i8, 32+2 }, // 4*ext+4*pmulhw+mul+sub sequence + split. 456 { ISD::UDIV, MVT::v16i8, 14 }, // 2*ext+2*pmulhw sequence 457 { ISD::UREM, MVT::v16i8, 16 }, // 2*ext+2*pmulhw+mul+sub sequence 458 { ISD::SDIV, MVT::v16i16, 12+2 }, // 2*pmulhw sequence + split. 459 { ISD::SREM, MVT::v16i16, 16+2 }, // 2*pmulhw+mul+sub sequence + split. 460 { ISD::SDIV, MVT::v8i16, 6 }, // pmulhw sequence 461 { ISD::SREM, MVT::v8i16, 8 }, // pmulhw+mul+sub sequence 462 { ISD::UDIV, MVT::v16i16, 12+2 }, // 2*pmulhuw sequence + split. 463 { ISD::UREM, MVT::v16i16, 16+2 }, // 2*pmulhuw+mul+sub sequence + split. 464 { ISD::UDIV, MVT::v8i16, 6 }, // pmulhuw sequence 465 { ISD::UREM, MVT::v8i16, 8 }, // pmulhuw+mul+sub sequence 466 { ISD::SDIV, MVT::v8i32, 38+2 }, // 2*pmuludq sequence + split. 467 { ISD::SREM, MVT::v8i32, 48+2 }, // 2*pmuludq+mul+sub sequence + split. 468 { ISD::SDIV, MVT::v4i32, 19 }, // pmuludq sequence 469 { ISD::SREM, MVT::v4i32, 24 }, // pmuludq+mul+sub sequence 470 { ISD::UDIV, MVT::v8i32, 30+2 }, // 2*pmuludq sequence + split. 471 { ISD::UREM, MVT::v8i32, 40+2 }, // 2*pmuludq+mul+sub sequence + split. 472 { ISD::UDIV, MVT::v4i32, 15 }, // pmuludq sequence 473 { ISD::UREM, MVT::v4i32, 20 }, // pmuludq+mul+sub sequence 474 }; 475 476 if ((Op2Info == TargetTransformInfo::OK_UniformConstantValue || 477 Op2Info == TargetTransformInfo::OK_NonUniformConstantValue) && 478 ST->hasSSE2()) { 479 // pmuldq sequence. 480 if (ISD == ISD::SDIV && LT.second == MVT::v8i32 && ST->hasAVX()) 481 return LT.first * 32; 482 if (ISD == ISD::SREM && LT.second == MVT::v8i32 && ST->hasAVX()) 483 return LT.first * 38; 484 if (ISD == ISD::SDIV && LT.second == MVT::v4i32 && ST->hasSSE41()) 485 return LT.first * 15; 486 if (ISD == ISD::SREM && LT.second == MVT::v4i32 && ST->hasSSE41()) 487 return LT.first * 20; 488 489 if (const auto *Entry = CostTableLookup(SSE2ConstCostTable, ISD, LT.second)) 490 return LT.first * Entry->Cost; 491 } 492 493 static const CostTblEntry AVX512BWShiftCostTable[] = { 494 { ISD::SHL, MVT::v8i16, 1 }, // vpsllvw 495 { ISD::SRL, MVT::v8i16, 1 }, // vpsrlvw 496 { ISD::SRA, MVT::v8i16, 1 }, // vpsravw 497 498 { ISD::SHL, MVT::v16i16, 1 }, // vpsllvw 499 { ISD::SRL, MVT::v16i16, 1 }, // vpsrlvw 500 { ISD::SRA, MVT::v16i16, 1 }, // vpsravw 501 502 { ISD::SHL, MVT::v32i16, 1 }, // vpsllvw 503 { ISD::SRL, MVT::v32i16, 1 }, // vpsrlvw 504 { ISD::SRA, MVT::v32i16, 1 }, // vpsravw 505 }; 506 507 if (ST->hasBWI()) 508 if (const auto *Entry = CostTableLookup(AVX512BWShiftCostTable, ISD, LT.second)) 509 return LT.first * Entry->Cost; 510 511 static const CostTblEntry AVX2UniformCostTable[] = { 512 // Uniform splats are cheaper for the following instructions. 513 { ISD::SHL, MVT::v16i16, 1 }, // psllw. 514 { ISD::SRL, MVT::v16i16, 1 }, // psrlw. 515 { ISD::SRA, MVT::v16i16, 1 }, // psraw. 516 { ISD::SHL, MVT::v32i16, 2 }, // 2*psllw. 517 { ISD::SRL, MVT::v32i16, 2 }, // 2*psrlw. 518 { ISD::SRA, MVT::v32i16, 2 }, // 2*psraw. 519 }; 520 521 if (ST->hasAVX2() && 522 ((Op2Info == TargetTransformInfo::OK_UniformConstantValue) || 523 (Op2Info == TargetTransformInfo::OK_UniformValue))) { 524 if (const auto *Entry = 525 CostTableLookup(AVX2UniformCostTable, ISD, LT.second)) 526 return LT.first * Entry->Cost; 527 } 528 529 static const CostTblEntry SSE2UniformCostTable[] = { 530 // Uniform splats are cheaper for the following instructions. 531 { ISD::SHL, MVT::v8i16, 1 }, // psllw. 532 { ISD::SHL, MVT::v4i32, 1 }, // pslld 533 { ISD::SHL, MVT::v2i64, 1 }, // psllq. 534 535 { ISD::SRL, MVT::v8i16, 1 }, // psrlw. 536 { ISD::SRL, MVT::v4i32, 1 }, // psrld. 537 { ISD::SRL, MVT::v2i64, 1 }, // psrlq. 538 539 { ISD::SRA, MVT::v8i16, 1 }, // psraw. 540 { ISD::SRA, MVT::v4i32, 1 }, // psrad. 541 }; 542 543 if (ST->hasSSE2() && 544 ((Op2Info == TargetTransformInfo::OK_UniformConstantValue) || 545 (Op2Info == TargetTransformInfo::OK_UniformValue))) { 546 if (const auto *Entry = 547 CostTableLookup(SSE2UniformCostTable, ISD, LT.second)) 548 return LT.first * Entry->Cost; 549 } 550 551 static const CostTblEntry AVX512DQCostTable[] = { 552 { ISD::MUL, MVT::v2i64, 1 }, 553 { ISD::MUL, MVT::v4i64, 1 }, 554 { ISD::MUL, MVT::v8i64, 1 } 555 }; 556 557 // Look for AVX512DQ lowering tricks for custom cases. 558 if (ST->hasDQI()) 559 if (const auto *Entry = CostTableLookup(AVX512DQCostTable, ISD, LT.second)) 560 return LT.first * Entry->Cost; 561 562 static const CostTblEntry AVX512BWCostTable[] = { 563 { ISD::SHL, MVT::v64i8, 11 }, // vpblendvb sequence. 564 { ISD::SRL, MVT::v64i8, 11 }, // vpblendvb sequence. 565 { ISD::SRA, MVT::v64i8, 24 }, // vpblendvb sequence. 566 567 { ISD::MUL, MVT::v64i8, 11 }, // extend/pmullw/trunc sequence. 568 { ISD::MUL, MVT::v32i8, 4 }, // extend/pmullw/trunc sequence. 569 { ISD::MUL, MVT::v16i8, 4 }, // extend/pmullw/trunc sequence. 570 }; 571 572 // Look for AVX512BW lowering tricks for custom cases. 573 if (ST->hasBWI()) 574 if (const auto *Entry = CostTableLookup(AVX512BWCostTable, ISD, LT.second)) 575 return LT.first * Entry->Cost; 576 577 static const CostTblEntry AVX512CostTable[] = { 578 { ISD::SHL, MVT::v16i32, 1 }, 579 { ISD::SRL, MVT::v16i32, 1 }, 580 { ISD::SRA, MVT::v16i32, 1 }, 581 582 { ISD::SHL, MVT::v8i64, 1 }, 583 { ISD::SRL, MVT::v8i64, 1 }, 584 585 { ISD::SRA, MVT::v2i64, 1 }, 586 { ISD::SRA, MVT::v4i64, 1 }, 587 { ISD::SRA, MVT::v8i64, 1 }, 588 589 { ISD::MUL, MVT::v64i8, 26 }, // extend/pmullw/trunc sequence. 590 { ISD::MUL, MVT::v32i8, 13 }, // extend/pmullw/trunc sequence. 591 { ISD::MUL, MVT::v16i8, 5 }, // extend/pmullw/trunc sequence. 592 { ISD::MUL, MVT::v16i32, 1 }, // pmulld (Skylake from agner.org) 593 { ISD::MUL, MVT::v8i32, 1 }, // pmulld (Skylake from agner.org) 594 { ISD::MUL, MVT::v4i32, 1 }, // pmulld (Skylake from agner.org) 595 { ISD::MUL, MVT::v8i64, 8 }, // 3*pmuludq/3*shift/2*add 596 597 { ISD::FADD, MVT::v8f64, 1 }, // Skylake from http://www.agner.org/ 598 { ISD::FSUB, MVT::v8f64, 1 }, // Skylake from http://www.agner.org/ 599 { ISD::FMUL, MVT::v8f64, 1 }, // Skylake from http://www.agner.org/ 600 601 { ISD::FADD, MVT::v16f32, 1 }, // Skylake from http://www.agner.org/ 602 { ISD::FSUB, MVT::v16f32, 1 }, // Skylake from http://www.agner.org/ 603 { ISD::FMUL, MVT::v16f32, 1 }, // Skylake from http://www.agner.org/ 604 }; 605 606 if (ST->hasAVX512()) 607 if (const auto *Entry = CostTableLookup(AVX512CostTable, ISD, LT.second)) 608 return LT.first * Entry->Cost; 609 610 static const CostTblEntry AVX2ShiftCostTable[] = { 611 // Shifts on v4i64/v8i32 on AVX2 is legal even though we declare to 612 // customize them to detect the cases where shift amount is a scalar one. 613 { ISD::SHL, MVT::v4i32, 1 }, 614 { ISD::SRL, MVT::v4i32, 1 }, 615 { ISD::SRA, MVT::v4i32, 1 }, 616 { ISD::SHL, MVT::v8i32, 1 }, 617 { ISD::SRL, MVT::v8i32, 1 }, 618 { ISD::SRA, MVT::v8i32, 1 }, 619 { ISD::SHL, MVT::v2i64, 1 }, 620 { ISD::SRL, MVT::v2i64, 1 }, 621 { ISD::SHL, MVT::v4i64, 1 }, 622 { ISD::SRL, MVT::v4i64, 1 }, 623 }; 624 625 if (ST->hasAVX512()) { 626 if (ISD == ISD::SHL && LT.second == MVT::v32i16 && 627 (Op2Info == TargetTransformInfo::OK_UniformConstantValue || 628 Op2Info == TargetTransformInfo::OK_NonUniformConstantValue)) 629 // On AVX512, a packed v32i16 shift left by a constant build_vector 630 // is lowered into a vector multiply (vpmullw). 631 return getArithmeticInstrCost(Instruction::Mul, Ty, CostKind, 632 Op1Info, Op2Info, 633 TargetTransformInfo::OP_None, 634 TargetTransformInfo::OP_None); 635 } 636 637 // Look for AVX2 lowering tricks. 638 if (ST->hasAVX2()) { 639 if (ISD == ISD::SHL && LT.second == MVT::v16i16 && 640 (Op2Info == TargetTransformInfo::OK_UniformConstantValue || 641 Op2Info == TargetTransformInfo::OK_NonUniformConstantValue)) 642 // On AVX2, a packed v16i16 shift left by a constant build_vector 643 // is lowered into a vector multiply (vpmullw). 644 return getArithmeticInstrCost(Instruction::Mul, Ty, CostKind, 645 Op1Info, Op2Info, 646 TargetTransformInfo::OP_None, 647 TargetTransformInfo::OP_None); 648 649 if (const auto *Entry = CostTableLookup(AVX2ShiftCostTable, ISD, LT.second)) 650 return LT.first * Entry->Cost; 651 } 652 653 static const CostTblEntry XOPShiftCostTable[] = { 654 // 128bit shifts take 1cy, but right shifts require negation beforehand. 655 { ISD::SHL, MVT::v16i8, 1 }, 656 { ISD::SRL, MVT::v16i8, 2 }, 657 { ISD::SRA, MVT::v16i8, 2 }, 658 { ISD::SHL, MVT::v8i16, 1 }, 659 { ISD::SRL, MVT::v8i16, 2 }, 660 { ISD::SRA, MVT::v8i16, 2 }, 661 { ISD::SHL, MVT::v4i32, 1 }, 662 { ISD::SRL, MVT::v4i32, 2 }, 663 { ISD::SRA, MVT::v4i32, 2 }, 664 { ISD::SHL, MVT::v2i64, 1 }, 665 { ISD::SRL, MVT::v2i64, 2 }, 666 { ISD::SRA, MVT::v2i64, 2 }, 667 // 256bit shifts require splitting if AVX2 didn't catch them above. 668 { ISD::SHL, MVT::v32i8, 2+2 }, 669 { ISD::SRL, MVT::v32i8, 4+2 }, 670 { ISD::SRA, MVT::v32i8, 4+2 }, 671 { ISD::SHL, MVT::v16i16, 2+2 }, 672 { ISD::SRL, MVT::v16i16, 4+2 }, 673 { ISD::SRA, MVT::v16i16, 4+2 }, 674 { ISD::SHL, MVT::v8i32, 2+2 }, 675 { ISD::SRL, MVT::v8i32, 4+2 }, 676 { ISD::SRA, MVT::v8i32, 4+2 }, 677 { ISD::SHL, MVT::v4i64, 2+2 }, 678 { ISD::SRL, MVT::v4i64, 4+2 }, 679 { ISD::SRA, MVT::v4i64, 4+2 }, 680 }; 681 682 // Look for XOP lowering tricks. 683 if (ST->hasXOP()) { 684 // If the right shift is constant then we'll fold the negation so 685 // it's as cheap as a left shift. 686 int ShiftISD = ISD; 687 if ((ShiftISD == ISD::SRL || ShiftISD == ISD::SRA) && 688 (Op2Info == TargetTransformInfo::OK_UniformConstantValue || 689 Op2Info == TargetTransformInfo::OK_NonUniformConstantValue)) 690 ShiftISD = ISD::SHL; 691 if (const auto *Entry = 692 CostTableLookup(XOPShiftCostTable, ShiftISD, LT.second)) 693 return LT.first * Entry->Cost; 694 } 695 696 static const CostTblEntry SSE2UniformShiftCostTable[] = { 697 // Uniform splats are cheaper for the following instructions. 698 { ISD::SHL, MVT::v16i16, 2+2 }, // 2*psllw + split. 699 { ISD::SHL, MVT::v8i32, 2+2 }, // 2*pslld + split. 700 { ISD::SHL, MVT::v4i64, 2+2 }, // 2*psllq + split. 701 702 { ISD::SRL, MVT::v16i16, 2+2 }, // 2*psrlw + split. 703 { ISD::SRL, MVT::v8i32, 2+2 }, // 2*psrld + split. 704 { ISD::SRL, MVT::v4i64, 2+2 }, // 2*psrlq + split. 705 706 { ISD::SRA, MVT::v16i16, 2+2 }, // 2*psraw + split. 707 { ISD::SRA, MVT::v8i32, 2+2 }, // 2*psrad + split. 708 { ISD::SRA, MVT::v2i64, 4 }, // 2*psrad + shuffle. 709 { ISD::SRA, MVT::v4i64, 8+2 }, // 2*(2*psrad + shuffle) + split. 710 }; 711 712 if (ST->hasSSE2() && 713 ((Op2Info == TargetTransformInfo::OK_UniformConstantValue) || 714 (Op2Info == TargetTransformInfo::OK_UniformValue))) { 715 716 // Handle AVX2 uniform v4i64 ISD::SRA, it's not worth a table. 717 if (ISD == ISD::SRA && LT.second == MVT::v4i64 && ST->hasAVX2()) 718 return LT.first * 4; // 2*psrad + shuffle. 719 720 if (const auto *Entry = 721 CostTableLookup(SSE2UniformShiftCostTable, ISD, LT.second)) 722 return LT.first * Entry->Cost; 723 } 724 725 if (ISD == ISD::SHL && 726 Op2Info == TargetTransformInfo::OK_NonUniformConstantValue) { 727 MVT VT = LT.second; 728 // Vector shift left by non uniform constant can be lowered 729 // into vector multiply. 730 if (((VT == MVT::v8i16 || VT == MVT::v4i32) && ST->hasSSE2()) || 731 ((VT == MVT::v16i16 || VT == MVT::v8i32) && ST->hasAVX())) 732 ISD = ISD::MUL; 733 } 734 735 static const CostTblEntry AVX2CostTable[] = { 736 { ISD::SHL, MVT::v32i8, 11 }, // vpblendvb sequence. 737 { ISD::SHL, MVT::v64i8, 22 }, // 2*vpblendvb sequence. 738 { ISD::SHL, MVT::v16i16, 10 }, // extend/vpsrlvd/pack sequence. 739 { ISD::SHL, MVT::v32i16, 20 }, // 2*extend/vpsrlvd/pack sequence. 740 741 { ISD::SRL, MVT::v32i8, 11 }, // vpblendvb sequence. 742 { ISD::SRL, MVT::v64i8, 22 }, // 2*vpblendvb sequence. 743 { ISD::SRL, MVT::v16i16, 10 }, // extend/vpsrlvd/pack sequence. 744 { ISD::SRL, MVT::v32i16, 20 }, // 2*extend/vpsrlvd/pack sequence. 745 746 { ISD::SRA, MVT::v32i8, 24 }, // vpblendvb sequence. 747 { ISD::SRA, MVT::v64i8, 48 }, // 2*vpblendvb sequence. 748 { ISD::SRA, MVT::v16i16, 10 }, // extend/vpsravd/pack sequence. 749 { ISD::SRA, MVT::v32i16, 20 }, // 2*extend/vpsravd/pack sequence. 750 { ISD::SRA, MVT::v2i64, 4 }, // srl/xor/sub sequence. 751 { ISD::SRA, MVT::v4i64, 4 }, // srl/xor/sub sequence. 752 753 { ISD::SUB, MVT::v32i8, 1 }, // psubb 754 { ISD::ADD, MVT::v32i8, 1 }, // paddb 755 { ISD::SUB, MVT::v16i16, 1 }, // psubw 756 { ISD::ADD, MVT::v16i16, 1 }, // paddw 757 { ISD::SUB, MVT::v8i32, 1 }, // psubd 758 { ISD::ADD, MVT::v8i32, 1 }, // paddd 759 { ISD::SUB, MVT::v4i64, 1 }, // psubq 760 { ISD::ADD, MVT::v4i64, 1 }, // paddq 761 762 { ISD::MUL, MVT::v32i8, 17 }, // extend/pmullw/trunc sequence. 763 { ISD::MUL, MVT::v16i8, 7 }, // extend/pmullw/trunc sequence. 764 { ISD::MUL, MVT::v16i16, 1 }, // pmullw 765 { ISD::MUL, MVT::v8i32, 2 }, // pmulld (Haswell from agner.org) 766 { ISD::MUL, MVT::v4i64, 8 }, // 3*pmuludq/3*shift/2*add 767 768 { ISD::FADD, MVT::v4f64, 1 }, // Haswell from http://www.agner.org/ 769 { ISD::FADD, MVT::v8f32, 1 }, // Haswell from http://www.agner.org/ 770 { ISD::FSUB, MVT::v4f64, 1 }, // Haswell from http://www.agner.org/ 771 { ISD::FSUB, MVT::v8f32, 1 }, // Haswell from http://www.agner.org/ 772 { ISD::FMUL, MVT::v4f64, 1 }, // Haswell from http://www.agner.org/ 773 { ISD::FMUL, MVT::v8f32, 1 }, // Haswell from http://www.agner.org/ 774 775 { ISD::FDIV, MVT::f32, 7 }, // Haswell from http://www.agner.org/ 776 { ISD::FDIV, MVT::v4f32, 7 }, // Haswell from http://www.agner.org/ 777 { ISD::FDIV, MVT::v8f32, 14 }, // Haswell from http://www.agner.org/ 778 { ISD::FDIV, MVT::f64, 14 }, // Haswell from http://www.agner.org/ 779 { ISD::FDIV, MVT::v2f64, 14 }, // Haswell from http://www.agner.org/ 780 { ISD::FDIV, MVT::v4f64, 28 }, // Haswell from http://www.agner.org/ 781 }; 782 783 // Look for AVX2 lowering tricks for custom cases. 784 if (ST->hasAVX2()) 785 if (const auto *Entry = CostTableLookup(AVX2CostTable, ISD, LT.second)) 786 return LT.first * Entry->Cost; 787 788 static const CostTblEntry AVX1CostTable[] = { 789 // We don't have to scalarize unsupported ops. We can issue two half-sized 790 // operations and we only need to extract the upper YMM half. 791 // Two ops + 1 extract + 1 insert = 4. 792 { ISD::MUL, MVT::v16i16, 4 }, 793 { ISD::MUL, MVT::v8i32, 4 }, 794 { ISD::SUB, MVT::v32i8, 4 }, 795 { ISD::ADD, MVT::v32i8, 4 }, 796 { ISD::SUB, MVT::v16i16, 4 }, 797 { ISD::ADD, MVT::v16i16, 4 }, 798 { ISD::SUB, MVT::v8i32, 4 }, 799 { ISD::ADD, MVT::v8i32, 4 }, 800 { ISD::SUB, MVT::v4i64, 4 }, 801 { ISD::ADD, MVT::v4i64, 4 }, 802 803 // A v4i64 multiply is custom lowered as two split v2i64 vectors that then 804 // are lowered as a series of long multiplies(3), shifts(3) and adds(2) 805 // Because we believe v4i64 to be a legal type, we must also include the 806 // extract+insert in the cost table. Therefore, the cost here is 18 807 // instead of 8. 808 { ISD::MUL, MVT::v4i64, 18 }, 809 810 { ISD::MUL, MVT::v32i8, 26 }, // extend/pmullw/trunc sequence. 811 812 { ISD::FDIV, MVT::f32, 14 }, // SNB from http://www.agner.org/ 813 { ISD::FDIV, MVT::v4f32, 14 }, // SNB from http://www.agner.org/ 814 { ISD::FDIV, MVT::v8f32, 28 }, // SNB from http://www.agner.org/ 815 { ISD::FDIV, MVT::f64, 22 }, // SNB from http://www.agner.org/ 816 { ISD::FDIV, MVT::v2f64, 22 }, // SNB from http://www.agner.org/ 817 { ISD::FDIV, MVT::v4f64, 44 }, // SNB from http://www.agner.org/ 818 }; 819 820 if (ST->hasAVX()) 821 if (const auto *Entry = CostTableLookup(AVX1CostTable, ISD, LT.second)) 822 return LT.first * Entry->Cost; 823 824 static const CostTblEntry SSE42CostTable[] = { 825 { ISD::FADD, MVT::f64, 1 }, // Nehalem from http://www.agner.org/ 826 { ISD::FADD, MVT::f32, 1 }, // Nehalem from http://www.agner.org/ 827 { ISD::FADD, MVT::v2f64, 1 }, // Nehalem from http://www.agner.org/ 828 { ISD::FADD, MVT::v4f32, 1 }, // Nehalem from http://www.agner.org/ 829 830 { ISD::FSUB, MVT::f64, 1 }, // Nehalem from http://www.agner.org/ 831 { ISD::FSUB, MVT::f32 , 1 }, // Nehalem from http://www.agner.org/ 832 { ISD::FSUB, MVT::v2f64, 1 }, // Nehalem from http://www.agner.org/ 833 { ISD::FSUB, MVT::v4f32, 1 }, // Nehalem from http://www.agner.org/ 834 835 { ISD::FMUL, MVT::f64, 1 }, // Nehalem from http://www.agner.org/ 836 { ISD::FMUL, MVT::f32, 1 }, // Nehalem from http://www.agner.org/ 837 { ISD::FMUL, MVT::v2f64, 1 }, // Nehalem from http://www.agner.org/ 838 { ISD::FMUL, MVT::v4f32, 1 }, // Nehalem from http://www.agner.org/ 839 840 { ISD::FDIV, MVT::f32, 14 }, // Nehalem from http://www.agner.org/ 841 { ISD::FDIV, MVT::v4f32, 14 }, // Nehalem from http://www.agner.org/ 842 { ISD::FDIV, MVT::f64, 22 }, // Nehalem from http://www.agner.org/ 843 { ISD::FDIV, MVT::v2f64, 22 }, // Nehalem from http://www.agner.org/ 844 }; 845 846 if (ST->hasSSE42()) 847 if (const auto *Entry = CostTableLookup(SSE42CostTable, ISD, LT.second)) 848 return LT.first * Entry->Cost; 849 850 static const CostTblEntry SSE41CostTable[] = { 851 { ISD::SHL, MVT::v16i8, 11 }, // pblendvb sequence. 852 { ISD::SHL, MVT::v32i8, 2*11+2 }, // pblendvb sequence + split. 853 { ISD::SHL, MVT::v8i16, 14 }, // pblendvb sequence. 854 { ISD::SHL, MVT::v16i16, 2*14+2 }, // pblendvb sequence + split. 855 { ISD::SHL, MVT::v4i32, 4 }, // pslld/paddd/cvttps2dq/pmulld 856 { ISD::SHL, MVT::v8i32, 2*4+2 }, // pslld/paddd/cvttps2dq/pmulld + split 857 858 { ISD::SRL, MVT::v16i8, 12 }, // pblendvb sequence. 859 { ISD::SRL, MVT::v32i8, 2*12+2 }, // pblendvb sequence + split. 860 { ISD::SRL, MVT::v8i16, 14 }, // pblendvb sequence. 861 { ISD::SRL, MVT::v16i16, 2*14+2 }, // pblendvb sequence + split. 862 { ISD::SRL, MVT::v4i32, 11 }, // Shift each lane + blend. 863 { ISD::SRL, MVT::v8i32, 2*11+2 }, // Shift each lane + blend + split. 864 865 { ISD::SRA, MVT::v16i8, 24 }, // pblendvb sequence. 866 { ISD::SRA, MVT::v32i8, 2*24+2 }, // pblendvb sequence + split. 867 { ISD::SRA, MVT::v8i16, 14 }, // pblendvb sequence. 868 { ISD::SRA, MVT::v16i16, 2*14+2 }, // pblendvb sequence + split. 869 { ISD::SRA, MVT::v4i32, 12 }, // Shift each lane + blend. 870 { ISD::SRA, MVT::v8i32, 2*12+2 }, // Shift each lane + blend + split. 871 872 { ISD::MUL, MVT::v4i32, 2 } // pmulld (Nehalem from agner.org) 873 }; 874 875 if (ST->hasSSE41()) 876 if (const auto *Entry = CostTableLookup(SSE41CostTable, ISD, LT.second)) 877 return LT.first * Entry->Cost; 878 879 static const CostTblEntry SSE2CostTable[] = { 880 // We don't correctly identify costs of casts because they are marked as 881 // custom. 882 { ISD::SHL, MVT::v16i8, 26 }, // cmpgtb sequence. 883 { ISD::SHL, MVT::v8i16, 32 }, // cmpgtb sequence. 884 { ISD::SHL, MVT::v4i32, 2*5 }, // We optimized this using mul. 885 { ISD::SHL, MVT::v2i64, 4 }, // splat+shuffle sequence. 886 { ISD::SHL, MVT::v4i64, 2*4+2 }, // splat+shuffle sequence + split. 887 888 { ISD::SRL, MVT::v16i8, 26 }, // cmpgtb sequence. 889 { ISD::SRL, MVT::v8i16, 32 }, // cmpgtb sequence. 890 { ISD::SRL, MVT::v4i32, 16 }, // Shift each lane + blend. 891 { ISD::SRL, MVT::v2i64, 4 }, // splat+shuffle sequence. 892 { ISD::SRL, MVT::v4i64, 2*4+2 }, // splat+shuffle sequence + split. 893 894 { ISD::SRA, MVT::v16i8, 54 }, // unpacked cmpgtb sequence. 895 { ISD::SRA, MVT::v8i16, 32 }, // cmpgtb sequence. 896 { ISD::SRA, MVT::v4i32, 16 }, // Shift each lane + blend. 897 { ISD::SRA, MVT::v2i64, 12 }, // srl/xor/sub sequence. 898 { ISD::SRA, MVT::v4i64, 2*12+2 }, // srl/xor/sub sequence+split. 899 900 { ISD::MUL, MVT::v16i8, 12 }, // extend/pmullw/trunc sequence. 901 { ISD::MUL, MVT::v8i16, 1 }, // pmullw 902 { ISD::MUL, MVT::v4i32, 6 }, // 3*pmuludq/4*shuffle 903 { ISD::MUL, MVT::v2i64, 8 }, // 3*pmuludq/3*shift/2*add 904 905 { ISD::FDIV, MVT::f32, 23 }, // Pentium IV from http://www.agner.org/ 906 { ISD::FDIV, MVT::v4f32, 39 }, // Pentium IV from http://www.agner.org/ 907 { ISD::FDIV, MVT::f64, 38 }, // Pentium IV from http://www.agner.org/ 908 { ISD::FDIV, MVT::v2f64, 69 }, // Pentium IV from http://www.agner.org/ 909 910 { ISD::FADD, MVT::f32, 2 }, // Pentium IV from http://www.agner.org/ 911 { ISD::FADD, MVT::f64, 2 }, // Pentium IV from http://www.agner.org/ 912 913 { ISD::FSUB, MVT::f32, 2 }, // Pentium IV from http://www.agner.org/ 914 { ISD::FSUB, MVT::f64, 2 }, // Pentium IV from http://www.agner.org/ 915 }; 916 917 if (ST->hasSSE2()) 918 if (const auto *Entry = CostTableLookup(SSE2CostTable, ISD, LT.second)) 919 return LT.first * Entry->Cost; 920 921 static const CostTblEntry SSE1CostTable[] = { 922 { ISD::FDIV, MVT::f32, 17 }, // Pentium III from http://www.agner.org/ 923 { ISD::FDIV, MVT::v4f32, 34 }, // Pentium III from http://www.agner.org/ 924 925 { ISD::FADD, MVT::f32, 1 }, // Pentium III from http://www.agner.org/ 926 { ISD::FADD, MVT::v4f32, 2 }, // Pentium III from http://www.agner.org/ 927 928 { ISD::FSUB, MVT::f32, 1 }, // Pentium III from http://www.agner.org/ 929 { ISD::FSUB, MVT::v4f32, 2 }, // Pentium III from http://www.agner.org/ 930 931 { ISD::ADD, MVT::i8, 1 }, // Pentium III from http://www.agner.org/ 932 { ISD::ADD, MVT::i16, 1 }, // Pentium III from http://www.agner.org/ 933 { ISD::ADD, MVT::i32, 1 }, // Pentium III from http://www.agner.org/ 934 935 { ISD::SUB, MVT::i8, 1 }, // Pentium III from http://www.agner.org/ 936 { ISD::SUB, MVT::i16, 1 }, // Pentium III from http://www.agner.org/ 937 { ISD::SUB, MVT::i32, 1 }, // Pentium III from http://www.agner.org/ 938 }; 939 940 if (ST->hasSSE1()) 941 if (const auto *Entry = CostTableLookup(SSE1CostTable, ISD, LT.second)) 942 return LT.first * Entry->Cost; 943 944 // It is not a good idea to vectorize division. We have to scalarize it and 945 // in the process we will often end up having to spilling regular 946 // registers. The overhead of division is going to dominate most kernels 947 // anyways so try hard to prevent vectorization of division - it is 948 // generally a bad idea. Assume somewhat arbitrarily that we have to be able 949 // to hide "20 cycles" for each lane. 950 if (LT.second.isVector() && (ISD == ISD::SDIV || ISD == ISD::SREM || 951 ISD == ISD::UDIV || ISD == ISD::UREM)) { 952 int ScalarCost = getArithmeticInstrCost( 953 Opcode, Ty->getScalarType(), CostKind, Op1Info, Op2Info, 954 TargetTransformInfo::OP_None, TargetTransformInfo::OP_None); 955 return 20 * LT.first * LT.second.getVectorNumElements() * ScalarCost; 956 } 957 958 // Fallback to the default implementation. 959 return BaseT::getArithmeticInstrCost(Opcode, Ty, CostKind, Op1Info, Op2Info); 960 } 961 962 int X86TTIImpl::getShuffleCost(TTI::ShuffleKind Kind, VectorType *BaseTp, 963 int Index, VectorType *SubTp) { 964 // 64-bit packed float vectors (v2f32) are widened to type v4f32. 965 // 64-bit packed integer vectors (v2i32) are widened to type v4i32. 966 std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, BaseTp); 967 968 // Treat Transpose as 2-op shuffles - there's no difference in lowering. 969 if (Kind == TTI::SK_Transpose) 970 Kind = TTI::SK_PermuteTwoSrc; 971 972 // For Broadcasts we are splatting the first element from the first input 973 // register, so only need to reference that input and all the output 974 // registers are the same. 975 if (Kind == TTI::SK_Broadcast) 976 LT.first = 1; 977 978 // Subvector extractions are free if they start at the beginning of a 979 // vector and cheap if the subvectors are aligned. 980 if (Kind == TTI::SK_ExtractSubvector && LT.second.isVector()) { 981 int NumElts = LT.second.getVectorNumElements(); 982 if ((Index % NumElts) == 0) 983 return 0; 984 std::pair<int, MVT> SubLT = TLI->getTypeLegalizationCost(DL, SubTp); 985 if (SubLT.second.isVector()) { 986 int NumSubElts = SubLT.second.getVectorNumElements(); 987 if ((Index % NumSubElts) == 0 && (NumElts % NumSubElts) == 0) 988 return SubLT.first; 989 // Handle some cases for widening legalization. For now we only handle 990 // cases where the original subvector was naturally aligned and evenly 991 // fit in its legalized subvector type. 992 // FIXME: Remove some of the alignment restrictions. 993 // FIXME: We can use permq for 64-bit or larger extracts from 256-bit 994 // vectors. 995 int OrigSubElts = cast<FixedVectorType>(SubTp)->getNumElements(); 996 if (NumSubElts > OrigSubElts && (Index % OrigSubElts) == 0 && 997 (NumSubElts % OrigSubElts) == 0 && 998 LT.second.getVectorElementType() == 999 SubLT.second.getVectorElementType() && 1000 LT.second.getVectorElementType().getSizeInBits() == 1001 BaseTp->getElementType()->getPrimitiveSizeInBits()) { 1002 assert(NumElts >= NumSubElts && NumElts > OrigSubElts && 1003 "Unexpected number of elements!"); 1004 auto *VecTy = FixedVectorType::get(BaseTp->getElementType(), 1005 LT.second.getVectorNumElements()); 1006 auto *SubTy = FixedVectorType::get(BaseTp->getElementType(), 1007 SubLT.second.getVectorNumElements()); 1008 int ExtractIndex = alignDown((Index % NumElts), NumSubElts); 1009 int ExtractCost = getShuffleCost(TTI::SK_ExtractSubvector, VecTy, 1010 ExtractIndex, SubTy); 1011 1012 // If the original size is 32-bits or more, we can use pshufd. Otherwise 1013 // if we have SSSE3 we can use pshufb. 1014 if (SubTp->getPrimitiveSizeInBits() >= 32 || ST->hasSSSE3()) 1015 return ExtractCost + 1; // pshufd or pshufb 1016 1017 assert(SubTp->getPrimitiveSizeInBits() == 16 && 1018 "Unexpected vector size"); 1019 1020 return ExtractCost + 2; // worst case pshufhw + pshufd 1021 } 1022 } 1023 } 1024 1025 // Handle some common (illegal) sub-vector types as they are often very cheap 1026 // to shuffle even on targets without PSHUFB. 1027 EVT VT = TLI->getValueType(DL, BaseTp); 1028 if (VT.isSimple() && VT.isVector() && VT.getSizeInBits() < 128 && 1029 !ST->hasSSSE3()) { 1030 static const CostTblEntry SSE2SubVectorShuffleTbl[] = { 1031 {TTI::SK_Broadcast, MVT::v4i16, 1}, // pshuflw 1032 {TTI::SK_Broadcast, MVT::v2i16, 1}, // pshuflw 1033 {TTI::SK_Broadcast, MVT::v8i8, 2}, // punpck/pshuflw 1034 {TTI::SK_Broadcast, MVT::v4i8, 2}, // punpck/pshuflw 1035 {TTI::SK_Broadcast, MVT::v2i8, 1}, // punpck 1036 1037 {TTI::SK_Reverse, MVT::v4i16, 1}, // pshuflw 1038 {TTI::SK_Reverse, MVT::v2i16, 1}, // pshuflw 1039 {TTI::SK_Reverse, MVT::v4i8, 3}, // punpck/pshuflw/packus 1040 {TTI::SK_Reverse, MVT::v2i8, 1}, // punpck 1041 1042 {TTI::SK_PermuteTwoSrc, MVT::v4i16, 2}, // punpck/pshuflw 1043 {TTI::SK_PermuteTwoSrc, MVT::v2i16, 2}, // punpck/pshuflw 1044 {TTI::SK_PermuteTwoSrc, MVT::v8i8, 7}, // punpck/pshuflw 1045 {TTI::SK_PermuteTwoSrc, MVT::v4i8, 4}, // punpck/pshuflw 1046 {TTI::SK_PermuteTwoSrc, MVT::v2i8, 2}, // punpck 1047 1048 {TTI::SK_PermuteSingleSrc, MVT::v4i16, 1}, // pshuflw 1049 {TTI::SK_PermuteSingleSrc, MVT::v2i16, 1}, // pshuflw 1050 {TTI::SK_PermuteSingleSrc, MVT::v8i8, 5}, // punpck/pshuflw 1051 {TTI::SK_PermuteSingleSrc, MVT::v4i8, 3}, // punpck/pshuflw 1052 {TTI::SK_PermuteSingleSrc, MVT::v2i8, 1}, // punpck 1053 }; 1054 1055 if (ST->hasSSE2()) 1056 if (const auto *Entry = 1057 CostTableLookup(SSE2SubVectorShuffleTbl, Kind, VT.getSimpleVT())) 1058 return Entry->Cost; 1059 } 1060 1061 // We are going to permute multiple sources and the result will be in multiple 1062 // destinations. Providing an accurate cost only for splits where the element 1063 // type remains the same. 1064 if (Kind == TTI::SK_PermuteSingleSrc && LT.first != 1) { 1065 MVT LegalVT = LT.second; 1066 if (LegalVT.isVector() && 1067 LegalVT.getVectorElementType().getSizeInBits() == 1068 BaseTp->getElementType()->getPrimitiveSizeInBits() && 1069 LegalVT.getVectorNumElements() < 1070 cast<FixedVectorType>(BaseTp)->getNumElements()) { 1071 1072 unsigned VecTySize = DL.getTypeStoreSize(BaseTp); 1073 unsigned LegalVTSize = LegalVT.getStoreSize(); 1074 // Number of source vectors after legalization: 1075 unsigned NumOfSrcs = (VecTySize + LegalVTSize - 1) / LegalVTSize; 1076 // Number of destination vectors after legalization: 1077 unsigned NumOfDests = LT.first; 1078 1079 auto *SingleOpTy = FixedVectorType::get(BaseTp->getElementType(), 1080 LegalVT.getVectorNumElements()); 1081 1082 unsigned NumOfShuffles = (NumOfSrcs - 1) * NumOfDests; 1083 return NumOfShuffles * 1084 getShuffleCost(TTI::SK_PermuteTwoSrc, SingleOpTy, 0, nullptr); 1085 } 1086 1087 return BaseT::getShuffleCost(Kind, BaseTp, Index, SubTp); 1088 } 1089 1090 // For 2-input shuffles, we must account for splitting the 2 inputs into many. 1091 if (Kind == TTI::SK_PermuteTwoSrc && LT.first != 1) { 1092 // We assume that source and destination have the same vector type. 1093 int NumOfDests = LT.first; 1094 int NumOfShufflesPerDest = LT.first * 2 - 1; 1095 LT.first = NumOfDests * NumOfShufflesPerDest; 1096 } 1097 1098 static const CostTblEntry AVX512VBMIShuffleTbl[] = { 1099 {TTI::SK_Reverse, MVT::v64i8, 1}, // vpermb 1100 {TTI::SK_Reverse, MVT::v32i8, 1}, // vpermb 1101 1102 {TTI::SK_PermuteSingleSrc, MVT::v64i8, 1}, // vpermb 1103 {TTI::SK_PermuteSingleSrc, MVT::v32i8, 1}, // vpermb 1104 1105 {TTI::SK_PermuteTwoSrc, MVT::v64i8, 2}, // vpermt2b 1106 {TTI::SK_PermuteTwoSrc, MVT::v32i8, 2}, // vpermt2b 1107 {TTI::SK_PermuteTwoSrc, MVT::v16i8, 2} // vpermt2b 1108 }; 1109 1110 if (ST->hasVBMI()) 1111 if (const auto *Entry = 1112 CostTableLookup(AVX512VBMIShuffleTbl, Kind, LT.second)) 1113 return LT.first * Entry->Cost; 1114 1115 static const CostTblEntry AVX512BWShuffleTbl[] = { 1116 {TTI::SK_Broadcast, MVT::v32i16, 1}, // vpbroadcastw 1117 {TTI::SK_Broadcast, MVT::v64i8, 1}, // vpbroadcastb 1118 1119 {TTI::SK_Reverse, MVT::v32i16, 2}, // vpermw 1120 {TTI::SK_Reverse, MVT::v16i16, 2}, // vpermw 1121 {TTI::SK_Reverse, MVT::v64i8, 2}, // pshufb + vshufi64x2 1122 1123 {TTI::SK_PermuteSingleSrc, MVT::v32i16, 2}, // vpermw 1124 {TTI::SK_PermuteSingleSrc, MVT::v16i16, 2}, // vpermw 1125 {TTI::SK_PermuteSingleSrc, MVT::v64i8, 8}, // extend to v32i16 1126 1127 {TTI::SK_PermuteTwoSrc, MVT::v32i16, 2}, // vpermt2w 1128 {TTI::SK_PermuteTwoSrc, MVT::v16i16, 2}, // vpermt2w 1129 {TTI::SK_PermuteTwoSrc, MVT::v8i16, 2}, // vpermt2w 1130 {TTI::SK_PermuteTwoSrc, MVT::v64i8, 19}, // 6 * v32i8 + 1 1131 1132 {TTI::SK_Select, MVT::v32i16, 1}, // vblendmw 1133 {TTI::SK_Select, MVT::v64i8, 1}, // vblendmb 1134 }; 1135 1136 if (ST->hasBWI()) 1137 if (const auto *Entry = 1138 CostTableLookup(AVX512BWShuffleTbl, Kind, LT.second)) 1139 return LT.first * Entry->Cost; 1140 1141 static const CostTblEntry AVX512ShuffleTbl[] = { 1142 {TTI::SK_Broadcast, MVT::v8f64, 1}, // vbroadcastpd 1143 {TTI::SK_Broadcast, MVT::v16f32, 1}, // vbroadcastps 1144 {TTI::SK_Broadcast, MVT::v8i64, 1}, // vpbroadcastq 1145 {TTI::SK_Broadcast, MVT::v16i32, 1}, // vpbroadcastd 1146 {TTI::SK_Broadcast, MVT::v32i16, 1}, // vpbroadcastw 1147 {TTI::SK_Broadcast, MVT::v64i8, 1}, // vpbroadcastb 1148 1149 {TTI::SK_Reverse, MVT::v8f64, 1}, // vpermpd 1150 {TTI::SK_Reverse, MVT::v16f32, 1}, // vpermps 1151 {TTI::SK_Reverse, MVT::v8i64, 1}, // vpermq 1152 {TTI::SK_Reverse, MVT::v16i32, 1}, // vpermd 1153 1154 {TTI::SK_PermuteSingleSrc, MVT::v8f64, 1}, // vpermpd 1155 {TTI::SK_PermuteSingleSrc, MVT::v4f64, 1}, // vpermpd 1156 {TTI::SK_PermuteSingleSrc, MVT::v2f64, 1}, // vpermpd 1157 {TTI::SK_PermuteSingleSrc, MVT::v16f32, 1}, // vpermps 1158 {TTI::SK_PermuteSingleSrc, MVT::v8f32, 1}, // vpermps 1159 {TTI::SK_PermuteSingleSrc, MVT::v4f32, 1}, // vpermps 1160 {TTI::SK_PermuteSingleSrc, MVT::v8i64, 1}, // vpermq 1161 {TTI::SK_PermuteSingleSrc, MVT::v4i64, 1}, // vpermq 1162 {TTI::SK_PermuteSingleSrc, MVT::v2i64, 1}, // vpermq 1163 {TTI::SK_PermuteSingleSrc, MVT::v16i32, 1}, // vpermd 1164 {TTI::SK_PermuteSingleSrc, MVT::v8i32, 1}, // vpermd 1165 {TTI::SK_PermuteSingleSrc, MVT::v4i32, 1}, // vpermd 1166 {TTI::SK_PermuteSingleSrc, MVT::v16i8, 1}, // pshufb 1167 1168 {TTI::SK_PermuteTwoSrc, MVT::v8f64, 1}, // vpermt2pd 1169 {TTI::SK_PermuteTwoSrc, MVT::v16f32, 1}, // vpermt2ps 1170 {TTI::SK_PermuteTwoSrc, MVT::v8i64, 1}, // vpermt2q 1171 {TTI::SK_PermuteTwoSrc, MVT::v16i32, 1}, // vpermt2d 1172 {TTI::SK_PermuteTwoSrc, MVT::v4f64, 1}, // vpermt2pd 1173 {TTI::SK_PermuteTwoSrc, MVT::v8f32, 1}, // vpermt2ps 1174 {TTI::SK_PermuteTwoSrc, MVT::v4i64, 1}, // vpermt2q 1175 {TTI::SK_PermuteTwoSrc, MVT::v8i32, 1}, // vpermt2d 1176 {TTI::SK_PermuteTwoSrc, MVT::v2f64, 1}, // vpermt2pd 1177 {TTI::SK_PermuteTwoSrc, MVT::v4f32, 1}, // vpermt2ps 1178 {TTI::SK_PermuteTwoSrc, MVT::v2i64, 1}, // vpermt2q 1179 {TTI::SK_PermuteTwoSrc, MVT::v4i32, 1}, // vpermt2d 1180 1181 // FIXME: This just applies the type legalization cost rules above 1182 // assuming these completely split. 1183 {TTI::SK_PermuteSingleSrc, MVT::v32i16, 14}, 1184 {TTI::SK_PermuteSingleSrc, MVT::v64i8, 14}, 1185 {TTI::SK_PermuteTwoSrc, MVT::v32i16, 42}, 1186 {TTI::SK_PermuteTwoSrc, MVT::v64i8, 42}, 1187 1188 {TTI::SK_Select, MVT::v32i16, 1}, // vpternlogq 1189 {TTI::SK_Select, MVT::v64i8, 1}, // vpternlogq 1190 {TTI::SK_Select, MVT::v8f64, 1}, // vblendmpd 1191 {TTI::SK_Select, MVT::v16f32, 1}, // vblendmps 1192 {TTI::SK_Select, MVT::v8i64, 1}, // vblendmq 1193 {TTI::SK_Select, MVT::v16i32, 1}, // vblendmd 1194 }; 1195 1196 if (ST->hasAVX512()) 1197 if (const auto *Entry = CostTableLookup(AVX512ShuffleTbl, Kind, LT.second)) 1198 return LT.first * Entry->Cost; 1199 1200 static const CostTblEntry AVX2ShuffleTbl[] = { 1201 {TTI::SK_Broadcast, MVT::v4f64, 1}, // vbroadcastpd 1202 {TTI::SK_Broadcast, MVT::v8f32, 1}, // vbroadcastps 1203 {TTI::SK_Broadcast, MVT::v4i64, 1}, // vpbroadcastq 1204 {TTI::SK_Broadcast, MVT::v8i32, 1}, // vpbroadcastd 1205 {TTI::SK_Broadcast, MVT::v16i16, 1}, // vpbroadcastw 1206 {TTI::SK_Broadcast, MVT::v32i8, 1}, // vpbroadcastb 1207 1208 {TTI::SK_Reverse, MVT::v4f64, 1}, // vpermpd 1209 {TTI::SK_Reverse, MVT::v8f32, 1}, // vpermps 1210 {TTI::SK_Reverse, MVT::v4i64, 1}, // vpermq 1211 {TTI::SK_Reverse, MVT::v8i32, 1}, // vpermd 1212 {TTI::SK_Reverse, MVT::v16i16, 2}, // vperm2i128 + pshufb 1213 {TTI::SK_Reverse, MVT::v32i8, 2}, // vperm2i128 + pshufb 1214 1215 {TTI::SK_Select, MVT::v16i16, 1}, // vpblendvb 1216 {TTI::SK_Select, MVT::v32i8, 1}, // vpblendvb 1217 1218 {TTI::SK_PermuteSingleSrc, MVT::v4f64, 1}, // vpermpd 1219 {TTI::SK_PermuteSingleSrc, MVT::v8f32, 1}, // vpermps 1220 {TTI::SK_PermuteSingleSrc, MVT::v4i64, 1}, // vpermq 1221 {TTI::SK_PermuteSingleSrc, MVT::v8i32, 1}, // vpermd 1222 {TTI::SK_PermuteSingleSrc, MVT::v16i16, 4}, // vperm2i128 + 2*vpshufb 1223 // + vpblendvb 1224 {TTI::SK_PermuteSingleSrc, MVT::v32i8, 4}, // vperm2i128 + 2*vpshufb 1225 // + vpblendvb 1226 1227 {TTI::SK_PermuteTwoSrc, MVT::v4f64, 3}, // 2*vpermpd + vblendpd 1228 {TTI::SK_PermuteTwoSrc, MVT::v8f32, 3}, // 2*vpermps + vblendps 1229 {TTI::SK_PermuteTwoSrc, MVT::v4i64, 3}, // 2*vpermq + vpblendd 1230 {TTI::SK_PermuteTwoSrc, MVT::v8i32, 3}, // 2*vpermd + vpblendd 1231 {TTI::SK_PermuteTwoSrc, MVT::v16i16, 7}, // 2*vperm2i128 + 4*vpshufb 1232 // + vpblendvb 1233 {TTI::SK_PermuteTwoSrc, MVT::v32i8, 7}, // 2*vperm2i128 + 4*vpshufb 1234 // + vpblendvb 1235 }; 1236 1237 if (ST->hasAVX2()) 1238 if (const auto *Entry = CostTableLookup(AVX2ShuffleTbl, Kind, LT.second)) 1239 return LT.first * Entry->Cost; 1240 1241 static const CostTblEntry XOPShuffleTbl[] = { 1242 {TTI::SK_PermuteSingleSrc, MVT::v4f64, 2}, // vperm2f128 + vpermil2pd 1243 {TTI::SK_PermuteSingleSrc, MVT::v8f32, 2}, // vperm2f128 + vpermil2ps 1244 {TTI::SK_PermuteSingleSrc, MVT::v4i64, 2}, // vperm2f128 + vpermil2pd 1245 {TTI::SK_PermuteSingleSrc, MVT::v8i32, 2}, // vperm2f128 + vpermil2ps 1246 {TTI::SK_PermuteSingleSrc, MVT::v16i16, 4}, // vextractf128 + 2*vpperm 1247 // + vinsertf128 1248 {TTI::SK_PermuteSingleSrc, MVT::v32i8, 4}, // vextractf128 + 2*vpperm 1249 // + vinsertf128 1250 1251 {TTI::SK_PermuteTwoSrc, MVT::v16i16, 9}, // 2*vextractf128 + 6*vpperm 1252 // + vinsertf128 1253 {TTI::SK_PermuteTwoSrc, MVT::v8i16, 1}, // vpperm 1254 {TTI::SK_PermuteTwoSrc, MVT::v32i8, 9}, // 2*vextractf128 + 6*vpperm 1255 // + vinsertf128 1256 {TTI::SK_PermuteTwoSrc, MVT::v16i8, 1}, // vpperm 1257 }; 1258 1259 if (ST->hasXOP()) 1260 if (const auto *Entry = CostTableLookup(XOPShuffleTbl, Kind, LT.second)) 1261 return LT.first * Entry->Cost; 1262 1263 static const CostTblEntry AVX1ShuffleTbl[] = { 1264 {TTI::SK_Broadcast, MVT::v4f64, 2}, // vperm2f128 + vpermilpd 1265 {TTI::SK_Broadcast, MVT::v8f32, 2}, // vperm2f128 + vpermilps 1266 {TTI::SK_Broadcast, MVT::v4i64, 2}, // vperm2f128 + vpermilpd 1267 {TTI::SK_Broadcast, MVT::v8i32, 2}, // vperm2f128 + vpermilps 1268 {TTI::SK_Broadcast, MVT::v16i16, 3}, // vpshuflw + vpshufd + vinsertf128 1269 {TTI::SK_Broadcast, MVT::v32i8, 2}, // vpshufb + vinsertf128 1270 1271 {TTI::SK_Reverse, MVT::v4f64, 2}, // vperm2f128 + vpermilpd 1272 {TTI::SK_Reverse, MVT::v8f32, 2}, // vperm2f128 + vpermilps 1273 {TTI::SK_Reverse, MVT::v4i64, 2}, // vperm2f128 + vpermilpd 1274 {TTI::SK_Reverse, MVT::v8i32, 2}, // vperm2f128 + vpermilps 1275 {TTI::SK_Reverse, MVT::v16i16, 4}, // vextractf128 + 2*pshufb 1276 // + vinsertf128 1277 {TTI::SK_Reverse, MVT::v32i8, 4}, // vextractf128 + 2*pshufb 1278 // + vinsertf128 1279 1280 {TTI::SK_Select, MVT::v4i64, 1}, // vblendpd 1281 {TTI::SK_Select, MVT::v4f64, 1}, // vblendpd 1282 {TTI::SK_Select, MVT::v8i32, 1}, // vblendps 1283 {TTI::SK_Select, MVT::v8f32, 1}, // vblendps 1284 {TTI::SK_Select, MVT::v16i16, 3}, // vpand + vpandn + vpor 1285 {TTI::SK_Select, MVT::v32i8, 3}, // vpand + vpandn + vpor 1286 1287 {TTI::SK_PermuteSingleSrc, MVT::v4f64, 2}, // vperm2f128 + vshufpd 1288 {TTI::SK_PermuteSingleSrc, MVT::v4i64, 2}, // vperm2f128 + vshufpd 1289 {TTI::SK_PermuteSingleSrc, MVT::v8f32, 4}, // 2*vperm2f128 + 2*vshufps 1290 {TTI::SK_PermuteSingleSrc, MVT::v8i32, 4}, // 2*vperm2f128 + 2*vshufps 1291 {TTI::SK_PermuteSingleSrc, MVT::v16i16, 8}, // vextractf128 + 4*pshufb 1292 // + 2*por + vinsertf128 1293 {TTI::SK_PermuteSingleSrc, MVT::v32i8, 8}, // vextractf128 + 4*pshufb 1294 // + 2*por + vinsertf128 1295 1296 {TTI::SK_PermuteTwoSrc, MVT::v4f64, 3}, // 2*vperm2f128 + vshufpd 1297 {TTI::SK_PermuteTwoSrc, MVT::v4i64, 3}, // 2*vperm2f128 + vshufpd 1298 {TTI::SK_PermuteTwoSrc, MVT::v8f32, 4}, // 2*vperm2f128 + 2*vshufps 1299 {TTI::SK_PermuteTwoSrc, MVT::v8i32, 4}, // 2*vperm2f128 + 2*vshufps 1300 {TTI::SK_PermuteTwoSrc, MVT::v16i16, 15}, // 2*vextractf128 + 8*pshufb 1301 // + 4*por + vinsertf128 1302 {TTI::SK_PermuteTwoSrc, MVT::v32i8, 15}, // 2*vextractf128 + 8*pshufb 1303 // + 4*por + vinsertf128 1304 }; 1305 1306 if (ST->hasAVX()) 1307 if (const auto *Entry = CostTableLookup(AVX1ShuffleTbl, Kind, LT.second)) 1308 return LT.first * Entry->Cost; 1309 1310 static const CostTblEntry SSE41ShuffleTbl[] = { 1311 {TTI::SK_Select, MVT::v2i64, 1}, // pblendw 1312 {TTI::SK_Select, MVT::v2f64, 1}, // movsd 1313 {TTI::SK_Select, MVT::v4i32, 1}, // pblendw 1314 {TTI::SK_Select, MVT::v4f32, 1}, // blendps 1315 {TTI::SK_Select, MVT::v8i16, 1}, // pblendw 1316 {TTI::SK_Select, MVT::v16i8, 1} // pblendvb 1317 }; 1318 1319 if (ST->hasSSE41()) 1320 if (const auto *Entry = CostTableLookup(SSE41ShuffleTbl, Kind, LT.second)) 1321 return LT.first * Entry->Cost; 1322 1323 static const CostTblEntry SSSE3ShuffleTbl[] = { 1324 {TTI::SK_Broadcast, MVT::v8i16, 1}, // pshufb 1325 {TTI::SK_Broadcast, MVT::v16i8, 1}, // pshufb 1326 1327 {TTI::SK_Reverse, MVT::v8i16, 1}, // pshufb 1328 {TTI::SK_Reverse, MVT::v16i8, 1}, // pshufb 1329 1330 {TTI::SK_Select, MVT::v8i16, 3}, // 2*pshufb + por 1331 {TTI::SK_Select, MVT::v16i8, 3}, // 2*pshufb + por 1332 1333 {TTI::SK_PermuteSingleSrc, MVT::v8i16, 1}, // pshufb 1334 {TTI::SK_PermuteSingleSrc, MVT::v16i8, 1}, // pshufb 1335 1336 {TTI::SK_PermuteTwoSrc, MVT::v8i16, 3}, // 2*pshufb + por 1337 {TTI::SK_PermuteTwoSrc, MVT::v16i8, 3}, // 2*pshufb + por 1338 }; 1339 1340 if (ST->hasSSSE3()) 1341 if (const auto *Entry = CostTableLookup(SSSE3ShuffleTbl, Kind, LT.second)) 1342 return LT.first * Entry->Cost; 1343 1344 static const CostTblEntry SSE2ShuffleTbl[] = { 1345 {TTI::SK_Broadcast, MVT::v2f64, 1}, // shufpd 1346 {TTI::SK_Broadcast, MVT::v2i64, 1}, // pshufd 1347 {TTI::SK_Broadcast, MVT::v4i32, 1}, // pshufd 1348 {TTI::SK_Broadcast, MVT::v8i16, 2}, // pshuflw + pshufd 1349 {TTI::SK_Broadcast, MVT::v16i8, 3}, // unpck + pshuflw + pshufd 1350 1351 {TTI::SK_Reverse, MVT::v2f64, 1}, // shufpd 1352 {TTI::SK_Reverse, MVT::v2i64, 1}, // pshufd 1353 {TTI::SK_Reverse, MVT::v4i32, 1}, // pshufd 1354 {TTI::SK_Reverse, MVT::v8i16, 3}, // pshuflw + pshufhw + pshufd 1355 {TTI::SK_Reverse, MVT::v16i8, 9}, // 2*pshuflw + 2*pshufhw 1356 // + 2*pshufd + 2*unpck + packus 1357 1358 {TTI::SK_Select, MVT::v2i64, 1}, // movsd 1359 {TTI::SK_Select, MVT::v2f64, 1}, // movsd 1360 {TTI::SK_Select, MVT::v4i32, 2}, // 2*shufps 1361 {TTI::SK_Select, MVT::v8i16, 3}, // pand + pandn + por 1362 {TTI::SK_Select, MVT::v16i8, 3}, // pand + pandn + por 1363 1364 {TTI::SK_PermuteSingleSrc, MVT::v2f64, 1}, // shufpd 1365 {TTI::SK_PermuteSingleSrc, MVT::v2i64, 1}, // pshufd 1366 {TTI::SK_PermuteSingleSrc, MVT::v4i32, 1}, // pshufd 1367 {TTI::SK_PermuteSingleSrc, MVT::v8i16, 5}, // 2*pshuflw + 2*pshufhw 1368 // + pshufd/unpck 1369 { TTI::SK_PermuteSingleSrc, MVT::v16i8, 10 }, // 2*pshuflw + 2*pshufhw 1370 // + 2*pshufd + 2*unpck + 2*packus 1371 1372 { TTI::SK_PermuteTwoSrc, MVT::v2f64, 1 }, // shufpd 1373 { TTI::SK_PermuteTwoSrc, MVT::v2i64, 1 }, // shufpd 1374 { TTI::SK_PermuteTwoSrc, MVT::v4i32, 2 }, // 2*{unpck,movsd,pshufd} 1375 { TTI::SK_PermuteTwoSrc, MVT::v8i16, 8 }, // blend+permute 1376 { TTI::SK_PermuteTwoSrc, MVT::v16i8, 13 }, // blend+permute 1377 }; 1378 1379 if (ST->hasSSE2()) 1380 if (const auto *Entry = CostTableLookup(SSE2ShuffleTbl, Kind, LT.second)) 1381 return LT.first * Entry->Cost; 1382 1383 static const CostTblEntry SSE1ShuffleTbl[] = { 1384 { TTI::SK_Broadcast, MVT::v4f32, 1 }, // shufps 1385 { TTI::SK_Reverse, MVT::v4f32, 1 }, // shufps 1386 { TTI::SK_Select, MVT::v4f32, 2 }, // 2*shufps 1387 { TTI::SK_PermuteSingleSrc, MVT::v4f32, 1 }, // shufps 1388 { TTI::SK_PermuteTwoSrc, MVT::v4f32, 2 }, // 2*shufps 1389 }; 1390 1391 if (ST->hasSSE1()) 1392 if (const auto *Entry = CostTableLookup(SSE1ShuffleTbl, Kind, LT.second)) 1393 return LT.first * Entry->Cost; 1394 1395 return BaseT::getShuffleCost(Kind, BaseTp, Index, SubTp); 1396 } 1397 1398 int X86TTIImpl::getCastInstrCost(unsigned Opcode, Type *Dst, Type *Src, 1399 TTI::CastContextHint CCH, 1400 TTI::TargetCostKind CostKind, 1401 const Instruction *I) { 1402 int ISD = TLI->InstructionOpcodeToISD(Opcode); 1403 assert(ISD && "Invalid opcode"); 1404 1405 // TODO: Allow non-throughput costs that aren't binary. 1406 auto AdjustCost = [&CostKind](int Cost) { 1407 if (CostKind != TTI::TCK_RecipThroughput) 1408 return Cost == 0 ? 0 : 1; 1409 return Cost; 1410 }; 1411 1412 // FIXME: Need a better design of the cost table to handle non-simple types of 1413 // potential massive combinations (elem_num x src_type x dst_type). 1414 1415 static const TypeConversionCostTblEntry AVX512BWConversionTbl[] { 1416 { ISD::SIGN_EXTEND, MVT::v32i16, MVT::v32i8, 1 }, 1417 { ISD::ZERO_EXTEND, MVT::v32i16, MVT::v32i8, 1 }, 1418 1419 // Mask sign extend has an instruction. 1420 { ISD::SIGN_EXTEND, MVT::v2i8, MVT::v2i1, 1 }, 1421 { ISD::SIGN_EXTEND, MVT::v2i16, MVT::v2i1, 1 }, 1422 { ISD::SIGN_EXTEND, MVT::v4i8, MVT::v4i1, 1 }, 1423 { ISD::SIGN_EXTEND, MVT::v4i16, MVT::v4i1, 1 }, 1424 { ISD::SIGN_EXTEND, MVT::v8i8, MVT::v8i1, 1 }, 1425 { ISD::SIGN_EXTEND, MVT::v8i16, MVT::v8i1, 1 }, 1426 { ISD::SIGN_EXTEND, MVT::v16i8, MVT::v16i1, 1 }, 1427 { ISD::SIGN_EXTEND, MVT::v16i16, MVT::v16i1, 1 }, 1428 { ISD::SIGN_EXTEND, MVT::v32i8, MVT::v32i1, 1 }, 1429 { ISD::SIGN_EXTEND, MVT::v32i16, MVT::v32i1, 1 }, 1430 { ISD::SIGN_EXTEND, MVT::v64i8, MVT::v64i1, 1 }, 1431 1432 // Mask zero extend is a sext + shift. 1433 { ISD::ZERO_EXTEND, MVT::v2i8, MVT::v2i1, 2 }, 1434 { ISD::ZERO_EXTEND, MVT::v2i16, MVT::v2i1, 2 }, 1435 { ISD::ZERO_EXTEND, MVT::v4i8, MVT::v4i1, 2 }, 1436 { ISD::ZERO_EXTEND, MVT::v4i16, MVT::v4i1, 2 }, 1437 { ISD::ZERO_EXTEND, MVT::v8i8, MVT::v8i1, 2 }, 1438 { ISD::ZERO_EXTEND, MVT::v8i16, MVT::v8i1, 2 }, 1439 { ISD::ZERO_EXTEND, MVT::v16i8, MVT::v16i1, 2 }, 1440 { ISD::ZERO_EXTEND, MVT::v16i16, MVT::v16i1, 2 }, 1441 { ISD::ZERO_EXTEND, MVT::v32i8, MVT::v32i1, 2 }, 1442 { ISD::ZERO_EXTEND, MVT::v32i16, MVT::v32i1, 2 }, 1443 { ISD::ZERO_EXTEND, MVT::v64i8, MVT::v64i1, 2 }, 1444 1445 { ISD::TRUNCATE, MVT::v32i8, MVT::v32i16, 2 }, 1446 { ISD::TRUNCATE, MVT::v16i8, MVT::v16i16, 2 }, // widen to zmm 1447 { ISD::TRUNCATE, MVT::v2i1, MVT::v2i8, 2 }, // widen to zmm 1448 { ISD::TRUNCATE, MVT::v2i1, MVT::v2i16, 2 }, // widen to zmm 1449 { ISD::TRUNCATE, MVT::v4i1, MVT::v4i8, 2 }, // widen to zmm 1450 { ISD::TRUNCATE, MVT::v4i1, MVT::v4i16, 2 }, // widen to zmm 1451 { ISD::TRUNCATE, MVT::v8i1, MVT::v8i8, 2 }, // widen to zmm 1452 { ISD::TRUNCATE, MVT::v8i1, MVT::v8i16, 2 }, // widen to zmm 1453 { ISD::TRUNCATE, MVT::v16i1, MVT::v16i8, 2 }, // widen to zmm 1454 { ISD::TRUNCATE, MVT::v16i1, MVT::v16i16, 2 }, // widen to zmm 1455 { ISD::TRUNCATE, MVT::v32i1, MVT::v32i8, 2 }, // widen to zmm 1456 { ISD::TRUNCATE, MVT::v32i1, MVT::v32i16, 2 }, 1457 { ISD::TRUNCATE, MVT::v64i1, MVT::v64i8, 2 }, 1458 }; 1459 1460 static const TypeConversionCostTblEntry AVX512DQConversionTbl[] = { 1461 { ISD::SINT_TO_FP, MVT::v8f32, MVT::v8i64, 1 }, 1462 { ISD::SINT_TO_FP, MVT::v8f64, MVT::v8i64, 1 }, 1463 1464 { ISD::UINT_TO_FP, MVT::v8f32, MVT::v8i64, 1 }, 1465 { ISD::UINT_TO_FP, MVT::v8f64, MVT::v8i64, 1 }, 1466 1467 { ISD::FP_TO_SINT, MVT::v8i64, MVT::v8f32, 1 }, 1468 { ISD::FP_TO_SINT, MVT::v8i64, MVT::v8f64, 1 }, 1469 1470 { ISD::FP_TO_UINT, MVT::v8i64, MVT::v8f32, 1 }, 1471 { ISD::FP_TO_UINT, MVT::v8i64, MVT::v8f64, 1 }, 1472 }; 1473 1474 // TODO: For AVX512DQ + AVX512VL, we also have cheap casts for 128-bit and 1475 // 256-bit wide vectors. 1476 1477 static const TypeConversionCostTblEntry AVX512FConversionTbl[] = { 1478 { ISD::FP_EXTEND, MVT::v8f64, MVT::v8f32, 1 }, 1479 { ISD::FP_EXTEND, MVT::v8f64, MVT::v16f32, 3 }, 1480 { ISD::FP_ROUND, MVT::v8f32, MVT::v8f64, 1 }, 1481 1482 { ISD::TRUNCATE, MVT::v2i1, MVT::v2i8, 3 }, // sext+vpslld+vptestmd 1483 { ISD::TRUNCATE, MVT::v4i1, MVT::v4i8, 3 }, // sext+vpslld+vptestmd 1484 { ISD::TRUNCATE, MVT::v8i1, MVT::v8i8, 3 }, // sext+vpslld+vptestmd 1485 { ISD::TRUNCATE, MVT::v16i1, MVT::v16i8, 3 }, // sext+vpslld+vptestmd 1486 { ISD::TRUNCATE, MVT::v2i1, MVT::v2i16, 3 }, // sext+vpsllq+vptestmq 1487 { ISD::TRUNCATE, MVT::v4i1, MVT::v4i16, 3 }, // sext+vpsllq+vptestmq 1488 { ISD::TRUNCATE, MVT::v8i1, MVT::v8i16, 3 }, // sext+vpsllq+vptestmq 1489 { ISD::TRUNCATE, MVT::v16i1, MVT::v16i16, 3 }, // sext+vpslld+vptestmd 1490 { ISD::TRUNCATE, MVT::v2i1, MVT::v2i32, 2 }, // zmm vpslld+vptestmd 1491 { ISD::TRUNCATE, MVT::v4i1, MVT::v4i32, 2 }, // zmm vpslld+vptestmd 1492 { ISD::TRUNCATE, MVT::v8i1, MVT::v8i32, 2 }, // zmm vpslld+vptestmd 1493 { ISD::TRUNCATE, MVT::v16i1, MVT::v16i32, 2 }, // vpslld+vptestmd 1494 { ISD::TRUNCATE, MVT::v2i1, MVT::v2i64, 2 }, // zmm vpsllq+vptestmq 1495 { ISD::TRUNCATE, MVT::v4i1, MVT::v4i64, 2 }, // zmm vpsllq+vptestmq 1496 { ISD::TRUNCATE, MVT::v8i1, MVT::v8i64, 2 }, // vpsllq+vptestmq 1497 { ISD::TRUNCATE, MVT::v16i8, MVT::v16i32, 2 }, 1498 { ISD::TRUNCATE, MVT::v16i16, MVT::v16i32, 2 }, 1499 { ISD::TRUNCATE, MVT::v8i8, MVT::v8i64, 2 }, 1500 { ISD::TRUNCATE, MVT::v8i16, MVT::v8i64, 2 }, 1501 { ISD::TRUNCATE, MVT::v8i32, MVT::v8i64, 1 }, 1502 { ISD::TRUNCATE, MVT::v4i32, MVT::v4i64, 1 }, // zmm vpmovqd 1503 { ISD::TRUNCATE, MVT::v16i8, MVT::v16i64, 5 },// 2*vpmovqd+concat+vpmovdb 1504 1505 { ISD::TRUNCATE, MVT::v16i8, MVT::v16i16, 3 }, // extend to v16i32 1506 { ISD::TRUNCATE, MVT::v32i8, MVT::v32i16, 8 }, 1507 1508 // Sign extend is zmm vpternlogd+vptruncdb. 1509 // Zero extend is zmm broadcast load+vptruncdw. 1510 { ISD::SIGN_EXTEND, MVT::v2i8, MVT::v2i1, 3 }, 1511 { ISD::ZERO_EXTEND, MVT::v2i8, MVT::v2i1, 4 }, 1512 { ISD::SIGN_EXTEND, MVT::v4i8, MVT::v4i1, 3 }, 1513 { ISD::ZERO_EXTEND, MVT::v4i8, MVT::v4i1, 4 }, 1514 { ISD::SIGN_EXTEND, MVT::v8i8, MVT::v8i1, 3 }, 1515 { ISD::ZERO_EXTEND, MVT::v8i8, MVT::v8i1, 4 }, 1516 { ISD::SIGN_EXTEND, MVT::v16i8, MVT::v16i1, 3 }, 1517 { ISD::ZERO_EXTEND, MVT::v16i8, MVT::v16i1, 4 }, 1518 1519 // Sign extend is zmm vpternlogd+vptruncdw. 1520 // Zero extend is zmm vpternlogd+vptruncdw+vpsrlw. 1521 { ISD::SIGN_EXTEND, MVT::v2i16, MVT::v2i1, 3 }, 1522 { ISD::ZERO_EXTEND, MVT::v2i16, MVT::v2i1, 4 }, 1523 { ISD::SIGN_EXTEND, MVT::v4i16, MVT::v4i1, 3 }, 1524 { ISD::ZERO_EXTEND, MVT::v4i16, MVT::v4i1, 4 }, 1525 { ISD::SIGN_EXTEND, MVT::v8i16, MVT::v8i1, 3 }, 1526 { ISD::ZERO_EXTEND, MVT::v8i16, MVT::v8i1, 4 }, 1527 { ISD::SIGN_EXTEND, MVT::v16i16, MVT::v16i1, 3 }, 1528 { ISD::ZERO_EXTEND, MVT::v16i16, MVT::v16i1, 4 }, 1529 1530 { ISD::SIGN_EXTEND, MVT::v2i32, MVT::v2i1, 1 }, // zmm vpternlogd 1531 { ISD::ZERO_EXTEND, MVT::v2i32, MVT::v2i1, 2 }, // zmm vpternlogd+psrld 1532 { ISD::SIGN_EXTEND, MVT::v4i32, MVT::v4i1, 1 }, // zmm vpternlogd 1533 { ISD::ZERO_EXTEND, MVT::v4i32, MVT::v4i1, 2 }, // zmm vpternlogd+psrld 1534 { ISD::SIGN_EXTEND, MVT::v8i32, MVT::v8i1, 1 }, // zmm vpternlogd 1535 { ISD::ZERO_EXTEND, MVT::v8i32, MVT::v8i1, 2 }, // zmm vpternlogd+psrld 1536 { ISD::SIGN_EXTEND, MVT::v2i64, MVT::v2i1, 1 }, // zmm vpternlogq 1537 { ISD::ZERO_EXTEND, MVT::v2i64, MVT::v2i1, 2 }, // zmm vpternlogq+psrlq 1538 { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v4i1, 1 }, // zmm vpternlogq 1539 { ISD::ZERO_EXTEND, MVT::v4i64, MVT::v4i1, 2 }, // zmm vpternlogq+psrlq 1540 1541 { ISD::SIGN_EXTEND, MVT::v16i32, MVT::v16i1, 1 }, // vpternlogd 1542 { ISD::ZERO_EXTEND, MVT::v16i32, MVT::v16i1, 2 }, // vpternlogd+psrld 1543 { ISD::SIGN_EXTEND, MVT::v8i64, MVT::v8i1, 1 }, // vpternlogq 1544 { ISD::ZERO_EXTEND, MVT::v8i64, MVT::v8i1, 2 }, // vpternlogq+psrlq 1545 1546 { ISD::SIGN_EXTEND, MVT::v16i32, MVT::v16i8, 1 }, 1547 { ISD::ZERO_EXTEND, MVT::v16i32, MVT::v16i8, 1 }, 1548 { ISD::SIGN_EXTEND, MVT::v16i32, MVT::v16i16, 1 }, 1549 { ISD::ZERO_EXTEND, MVT::v16i32, MVT::v16i16, 1 }, 1550 { ISD::SIGN_EXTEND, MVT::v8i64, MVT::v8i8, 1 }, 1551 { ISD::ZERO_EXTEND, MVT::v8i64, MVT::v8i8, 1 }, 1552 { ISD::SIGN_EXTEND, MVT::v8i64, MVT::v8i16, 1 }, 1553 { ISD::ZERO_EXTEND, MVT::v8i64, MVT::v8i16, 1 }, 1554 { ISD::SIGN_EXTEND, MVT::v8i64, MVT::v8i32, 1 }, 1555 { ISD::ZERO_EXTEND, MVT::v8i64, MVT::v8i32, 1 }, 1556 1557 { ISD::SIGN_EXTEND, MVT::v32i16, MVT::v32i8, 3 }, // FIXME: May not be right 1558 { ISD::ZERO_EXTEND, MVT::v32i16, MVT::v32i8, 3 }, // FIXME: May not be right 1559 1560 { ISD::SINT_TO_FP, MVT::v8f64, MVT::v8i1, 4 }, 1561 { ISD::SINT_TO_FP, MVT::v16f32, MVT::v16i1, 3 }, 1562 { ISD::SINT_TO_FP, MVT::v8f64, MVT::v8i8, 2 }, 1563 { ISD::SINT_TO_FP, MVT::v16f32, MVT::v16i8, 2 }, 1564 { ISD::SINT_TO_FP, MVT::v8f64, MVT::v8i16, 2 }, 1565 { ISD::SINT_TO_FP, MVT::v16f32, MVT::v16i16, 2 }, 1566 { ISD::SINT_TO_FP, MVT::v16f32, MVT::v16i32, 1 }, 1567 { ISD::SINT_TO_FP, MVT::v8f64, MVT::v8i32, 1 }, 1568 1569 { ISD::UINT_TO_FP, MVT::v8f64, MVT::v8i1, 4 }, 1570 { ISD::UINT_TO_FP, MVT::v16f32, MVT::v16i1, 3 }, 1571 { ISD::UINT_TO_FP, MVT::v8f64, MVT::v8i8, 2 }, 1572 { ISD::UINT_TO_FP, MVT::v16f32, MVT::v16i8, 2 }, 1573 { ISD::UINT_TO_FP, MVT::v8f64, MVT::v8i16, 2 }, 1574 { ISD::UINT_TO_FP, MVT::v16f32, MVT::v16i16, 2 }, 1575 { ISD::UINT_TO_FP, MVT::v8f64, MVT::v8i32, 1 }, 1576 { ISD::UINT_TO_FP, MVT::v16f32, MVT::v16i32, 1 }, 1577 { ISD::UINT_TO_FP, MVT::v8f32, MVT::v8i64, 26 }, 1578 { ISD::UINT_TO_FP, MVT::v8f64, MVT::v8i64, 5 }, 1579 1580 { ISD::FP_TO_SINT, MVT::v8i8, MVT::v8f64, 3 }, 1581 { ISD::FP_TO_SINT, MVT::v8i16, MVT::v8f64, 3 }, 1582 { ISD::FP_TO_SINT, MVT::v16i8, MVT::v16f32, 3 }, 1583 { ISD::FP_TO_SINT, MVT::v16i16, MVT::v16f32, 3 }, 1584 1585 { ISD::FP_TO_UINT, MVT::v8i32, MVT::v8f64, 1 }, 1586 { ISD::FP_TO_UINT, MVT::v8i16, MVT::v8f64, 3 }, 1587 { ISD::FP_TO_UINT, MVT::v8i8, MVT::v8f64, 3 }, 1588 { ISD::FP_TO_UINT, MVT::v16i32, MVT::v16f32, 1 }, 1589 { ISD::FP_TO_UINT, MVT::v16i16, MVT::v16f32, 3 }, 1590 { ISD::FP_TO_UINT, MVT::v16i8, MVT::v16f32, 3 }, 1591 }; 1592 1593 static const TypeConversionCostTblEntry AVX512BWVLConversionTbl[] { 1594 // Mask sign extend has an instruction. 1595 { ISD::SIGN_EXTEND, MVT::v2i8, MVT::v2i1, 1 }, 1596 { ISD::SIGN_EXTEND, MVT::v2i16, MVT::v2i1, 1 }, 1597 { ISD::SIGN_EXTEND, MVT::v4i8, MVT::v4i1, 1 }, 1598 { ISD::SIGN_EXTEND, MVT::v4i16, MVT::v4i1, 1 }, 1599 { ISD::SIGN_EXTEND, MVT::v8i8, MVT::v8i1, 1 }, 1600 { ISD::SIGN_EXTEND, MVT::v8i16, MVT::v8i1, 1 }, 1601 { ISD::SIGN_EXTEND, MVT::v16i8, MVT::v16i1, 1 }, 1602 { ISD::SIGN_EXTEND, MVT::v16i16, MVT::v16i1, 1 }, 1603 { ISD::SIGN_EXTEND, MVT::v32i8, MVT::v32i1, 1 }, 1604 1605 // Mask zero extend is a sext + shift. 1606 { ISD::ZERO_EXTEND, MVT::v2i8, MVT::v2i1, 2 }, 1607 { ISD::ZERO_EXTEND, MVT::v2i16, MVT::v2i1, 2 }, 1608 { ISD::ZERO_EXTEND, MVT::v4i8, MVT::v4i1, 2 }, 1609 { ISD::ZERO_EXTEND, MVT::v4i16, MVT::v4i1, 2 }, 1610 { ISD::ZERO_EXTEND, MVT::v8i8, MVT::v8i1, 2 }, 1611 { ISD::ZERO_EXTEND, MVT::v8i16, MVT::v8i1, 2 }, 1612 { ISD::ZERO_EXTEND, MVT::v16i8, MVT::v16i1, 2 }, 1613 { ISD::ZERO_EXTEND, MVT::v16i16, MVT::v16i1, 2 }, 1614 { ISD::ZERO_EXTEND, MVT::v32i8, MVT::v32i1, 2 }, 1615 1616 { ISD::TRUNCATE, MVT::v16i8, MVT::v16i16, 2 }, 1617 { ISD::TRUNCATE, MVT::v2i1, MVT::v2i8, 2 }, // vpsllw+vptestmb 1618 { ISD::TRUNCATE, MVT::v2i1, MVT::v2i16, 2 }, // vpsllw+vptestmw 1619 { ISD::TRUNCATE, MVT::v4i1, MVT::v4i8, 2 }, // vpsllw+vptestmb 1620 { ISD::TRUNCATE, MVT::v4i1, MVT::v4i16, 2 }, // vpsllw+vptestmw 1621 { ISD::TRUNCATE, MVT::v8i1, MVT::v8i8, 2 }, // vpsllw+vptestmb 1622 { ISD::TRUNCATE, MVT::v8i1, MVT::v8i16, 2 }, // vpsllw+vptestmw 1623 { ISD::TRUNCATE, MVT::v16i1, MVT::v16i8, 2 }, // vpsllw+vptestmb 1624 { ISD::TRUNCATE, MVT::v16i1, MVT::v16i16, 2 }, // vpsllw+vptestmw 1625 { ISD::TRUNCATE, MVT::v32i1, MVT::v32i8, 2 }, // vpsllw+vptestmb 1626 }; 1627 1628 static const TypeConversionCostTblEntry AVX512DQVLConversionTbl[] = { 1629 { ISD::SINT_TO_FP, MVT::v2f32, MVT::v2i64, 1 }, 1630 { ISD::SINT_TO_FP, MVT::v2f64, MVT::v2i64, 1 }, 1631 { ISD::SINT_TO_FP, MVT::v4f32, MVT::v4i64, 1 }, 1632 { ISD::SINT_TO_FP, MVT::v4f64, MVT::v4i64, 1 }, 1633 1634 { ISD::UINT_TO_FP, MVT::v2f32, MVT::v2i64, 1 }, 1635 { ISD::UINT_TO_FP, MVT::v2f64, MVT::v2i64, 1 }, 1636 { ISD::UINT_TO_FP, MVT::v4f32, MVT::v4i64, 1 }, 1637 { ISD::UINT_TO_FP, MVT::v4f64, MVT::v4i64, 1 }, 1638 1639 { ISD::FP_TO_SINT, MVT::v2i64, MVT::v2f32, 1 }, 1640 { ISD::FP_TO_SINT, MVT::v4i64, MVT::v4f32, 1 }, 1641 { ISD::FP_TO_SINT, MVT::v2i64, MVT::v2f64, 1 }, 1642 { ISD::FP_TO_SINT, MVT::v4i64, MVT::v4f64, 1 }, 1643 1644 { ISD::FP_TO_UINT, MVT::v2i64, MVT::v2f32, 1 }, 1645 { ISD::FP_TO_UINT, MVT::v4i64, MVT::v4f32, 1 }, 1646 { ISD::FP_TO_UINT, MVT::v2i64, MVT::v2f64, 1 }, 1647 { ISD::FP_TO_UINT, MVT::v4i64, MVT::v4f64, 1 }, 1648 }; 1649 1650 static const TypeConversionCostTblEntry AVX512VLConversionTbl[] = { 1651 { ISD::TRUNCATE, MVT::v2i1, MVT::v2i8, 3 }, // sext+vpslld+vptestmd 1652 { ISD::TRUNCATE, MVT::v4i1, MVT::v4i8, 3 }, // sext+vpslld+vptestmd 1653 { ISD::TRUNCATE, MVT::v8i1, MVT::v8i8, 3 }, // sext+vpslld+vptestmd 1654 { ISD::TRUNCATE, MVT::v16i1, MVT::v16i8, 8 }, // split+2*v8i8 1655 { ISD::TRUNCATE, MVT::v2i1, MVT::v2i16, 3 }, // sext+vpsllq+vptestmq 1656 { ISD::TRUNCATE, MVT::v4i1, MVT::v4i16, 3 }, // sext+vpsllq+vptestmq 1657 { ISD::TRUNCATE, MVT::v8i1, MVT::v8i16, 3 }, // sext+vpsllq+vptestmq 1658 { ISD::TRUNCATE, MVT::v16i1, MVT::v16i16, 8 }, // split+2*v8i16 1659 { ISD::TRUNCATE, MVT::v2i1, MVT::v2i32, 2 }, // vpslld+vptestmd 1660 { ISD::TRUNCATE, MVT::v4i1, MVT::v4i32, 2 }, // vpslld+vptestmd 1661 { ISD::TRUNCATE, MVT::v8i1, MVT::v8i32, 2 }, // vpslld+vptestmd 1662 { ISD::TRUNCATE, MVT::v2i1, MVT::v2i64, 2 }, // vpsllq+vptestmq 1663 { ISD::TRUNCATE, MVT::v4i1, MVT::v4i64, 2 }, // vpsllq+vptestmq 1664 { ISD::TRUNCATE, MVT::v4i32, MVT::v4i64, 1 }, // vpmovqd 1665 1666 // sign extend is vpcmpeq+maskedmove+vpmovdw+vpacksswb 1667 // zero extend is vpcmpeq+maskedmove+vpmovdw+vpsrlw+vpackuswb 1668 { ISD::SIGN_EXTEND, MVT::v2i8, MVT::v2i1, 5 }, 1669 { ISD::ZERO_EXTEND, MVT::v2i8, MVT::v2i1, 6 }, 1670 { ISD::SIGN_EXTEND, MVT::v4i8, MVT::v4i1, 5 }, 1671 { ISD::ZERO_EXTEND, MVT::v4i8, MVT::v4i1, 6 }, 1672 { ISD::SIGN_EXTEND, MVT::v8i8, MVT::v8i1, 5 }, 1673 { ISD::ZERO_EXTEND, MVT::v8i8, MVT::v8i1, 6 }, 1674 { ISD::SIGN_EXTEND, MVT::v16i8, MVT::v16i1, 10 }, 1675 { ISD::ZERO_EXTEND, MVT::v16i8, MVT::v16i1, 12 }, 1676 1677 // sign extend is vpcmpeq+maskedmove+vpmovdw 1678 // zero extend is vpcmpeq+maskedmove+vpmovdw+vpsrlw 1679 { ISD::SIGN_EXTEND, MVT::v2i16, MVT::v2i1, 4 }, 1680 { ISD::ZERO_EXTEND, MVT::v2i16, MVT::v2i1, 5 }, 1681 { ISD::SIGN_EXTEND, MVT::v4i16, MVT::v4i1, 4 }, 1682 { ISD::ZERO_EXTEND, MVT::v4i16, MVT::v4i1, 5 }, 1683 { ISD::SIGN_EXTEND, MVT::v8i16, MVT::v8i1, 4 }, 1684 { ISD::ZERO_EXTEND, MVT::v8i16, MVT::v8i1, 5 }, 1685 { ISD::SIGN_EXTEND, MVT::v16i16, MVT::v16i1, 10 }, 1686 { ISD::ZERO_EXTEND, MVT::v16i16, MVT::v16i1, 12 }, 1687 1688 { ISD::SIGN_EXTEND, MVT::v2i32, MVT::v2i1, 1 }, // vpternlogd 1689 { ISD::ZERO_EXTEND, MVT::v2i32, MVT::v2i1, 2 }, // vpternlogd+psrld 1690 { ISD::SIGN_EXTEND, MVT::v4i32, MVT::v4i1, 1 }, // vpternlogd 1691 { ISD::ZERO_EXTEND, MVT::v4i32, MVT::v4i1, 2 }, // vpternlogd+psrld 1692 { ISD::SIGN_EXTEND, MVT::v8i32, MVT::v8i1, 1 }, // vpternlogd 1693 { ISD::ZERO_EXTEND, MVT::v8i32, MVT::v8i1, 2 }, // vpternlogd+psrld 1694 { ISD::SIGN_EXTEND, MVT::v2i64, MVT::v2i1, 1 }, // vpternlogq 1695 { ISD::ZERO_EXTEND, MVT::v2i64, MVT::v2i1, 2 }, // vpternlogq+psrlq 1696 { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v4i1, 1 }, // vpternlogq 1697 { ISD::ZERO_EXTEND, MVT::v4i64, MVT::v4i1, 2 }, // vpternlogq+psrlq 1698 1699 { ISD::UINT_TO_FP, MVT::v2f64, MVT::v2i8, 2 }, 1700 { ISD::UINT_TO_FP, MVT::v4f64, MVT::v4i8, 2 }, 1701 { ISD::UINT_TO_FP, MVT::v8f32, MVT::v8i8, 2 }, 1702 { ISD::UINT_TO_FP, MVT::v2f64, MVT::v2i16, 5 }, 1703 { ISD::UINT_TO_FP, MVT::v4f64, MVT::v4i16, 2 }, 1704 { ISD::UINT_TO_FP, MVT::v8f32, MVT::v8i16, 2 }, 1705 { ISD::UINT_TO_FP, MVT::v2f32, MVT::v2i32, 2 }, 1706 { ISD::UINT_TO_FP, MVT::v2f64, MVT::v2i32, 1 }, 1707 { ISD::UINT_TO_FP, MVT::v4f32, MVT::v4i32, 1 }, 1708 { ISD::UINT_TO_FP, MVT::v4f64, MVT::v4i32, 1 }, 1709 { ISD::UINT_TO_FP, MVT::v8f32, MVT::v8i32, 1 }, 1710 { ISD::UINT_TO_FP, MVT::v2f32, MVT::v2i64, 5 }, 1711 { ISD::UINT_TO_FP, MVT::v2f64, MVT::v2i64, 5 }, 1712 { ISD::UINT_TO_FP, MVT::v4f64, MVT::v4i64, 5 }, 1713 1714 { ISD::UINT_TO_FP, MVT::f32, MVT::i64, 1 }, 1715 { ISD::UINT_TO_FP, MVT::f64, MVT::i64, 1 }, 1716 1717 { ISD::FP_TO_SINT, MVT::v8i8, MVT::v8f32, 3 }, 1718 { ISD::FP_TO_UINT, MVT::v8i8, MVT::v8f32, 3 }, 1719 1720 { ISD::FP_TO_UINT, MVT::i64, MVT::f32, 1 }, 1721 { ISD::FP_TO_UINT, MVT::i64, MVT::f64, 1 }, 1722 1723 { ISD::FP_TO_UINT, MVT::v2i32, MVT::v2f32, 1 }, 1724 { ISD::FP_TO_UINT, MVT::v4i32, MVT::v4f32, 1 }, 1725 { ISD::FP_TO_UINT, MVT::v2i32, MVT::v2f64, 1 }, 1726 { ISD::FP_TO_UINT, MVT::v4i32, MVT::v4f64, 1 }, 1727 { ISD::FP_TO_UINT, MVT::v8i32, MVT::v8f32, 1 }, 1728 }; 1729 1730 static const TypeConversionCostTblEntry AVX2ConversionTbl[] = { 1731 { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v4i1, 3 }, 1732 { ISD::ZERO_EXTEND, MVT::v4i64, MVT::v4i1, 3 }, 1733 { ISD::SIGN_EXTEND, MVT::v8i32, MVT::v8i1, 3 }, 1734 { ISD::ZERO_EXTEND, MVT::v8i32, MVT::v8i1, 3 }, 1735 { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v4i8, 1 }, 1736 { ISD::ZERO_EXTEND, MVT::v4i64, MVT::v4i8, 1 }, 1737 { ISD::SIGN_EXTEND, MVT::v8i32, MVT::v8i8, 1 }, 1738 { ISD::ZERO_EXTEND, MVT::v8i32, MVT::v8i8, 1 }, 1739 { ISD::SIGN_EXTEND, MVT::v16i16, MVT::v16i1, 1 }, 1740 { ISD::ZERO_EXTEND, MVT::v16i16, MVT::v16i1, 1 }, 1741 { ISD::SIGN_EXTEND, MVT::v16i16, MVT::v16i8, 1 }, 1742 { ISD::ZERO_EXTEND, MVT::v16i16, MVT::v16i8, 1 }, 1743 { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v4i16, 1 }, 1744 { ISD::ZERO_EXTEND, MVT::v4i64, MVT::v4i16, 1 }, 1745 { ISD::SIGN_EXTEND, MVT::v8i32, MVT::v8i16, 1 }, 1746 { ISD::ZERO_EXTEND, MVT::v8i32, MVT::v8i16, 1 }, 1747 { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v4i32, 1 }, 1748 { ISD::ZERO_EXTEND, MVT::v4i64, MVT::v4i32, 1 }, 1749 { ISD::ZERO_EXTEND, MVT::v16i32, MVT::v16i16, 3 }, 1750 { ISD::SIGN_EXTEND, MVT::v16i32, MVT::v16i16, 3 }, 1751 1752 { ISD::TRUNCATE, MVT::v4i32, MVT::v4i64, 2 }, 1753 { ISD::TRUNCATE, MVT::v8i1, MVT::v8i32, 2 }, 1754 1755 { ISD::TRUNCATE, MVT::v4i8, MVT::v4i64, 2 }, 1756 { ISD::TRUNCATE, MVT::v4i16, MVT::v4i64, 2 }, 1757 { ISD::TRUNCATE, MVT::v8i8, MVT::v8i32, 2 }, 1758 { ISD::TRUNCATE, MVT::v8i16, MVT::v8i32, 2 }, 1759 1760 { ISD::FP_EXTEND, MVT::v8f64, MVT::v8f32, 3 }, 1761 { ISD::FP_ROUND, MVT::v8f32, MVT::v8f64, 3 }, 1762 1763 { ISD::UINT_TO_FP, MVT::v8f32, MVT::v8i32, 8 }, 1764 }; 1765 1766 static const TypeConversionCostTblEntry AVXConversionTbl[] = { 1767 { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v4i1, 6 }, 1768 { ISD::ZERO_EXTEND, MVT::v4i64, MVT::v4i1, 4 }, 1769 { ISD::SIGN_EXTEND, MVT::v8i32, MVT::v8i1, 7 }, 1770 { ISD::ZERO_EXTEND, MVT::v8i32, MVT::v8i1, 4 }, 1771 { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v4i8, 4 }, 1772 { ISD::ZERO_EXTEND, MVT::v4i64, MVT::v4i8, 4 }, 1773 { ISD::SIGN_EXTEND, MVT::v8i32, MVT::v8i8, 4 }, 1774 { ISD::ZERO_EXTEND, MVT::v8i32, MVT::v8i8, 4 }, 1775 { ISD::SIGN_EXTEND, MVT::v16i16, MVT::v16i1, 4 }, 1776 { ISD::ZERO_EXTEND, MVT::v16i16, MVT::v16i1, 4 }, 1777 { ISD::SIGN_EXTEND, MVT::v16i16, MVT::v16i8, 4 }, 1778 { ISD::ZERO_EXTEND, MVT::v16i16, MVT::v16i8, 4 }, 1779 { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v4i16, 4 }, 1780 { ISD::ZERO_EXTEND, MVT::v4i64, MVT::v4i16, 3 }, 1781 { ISD::SIGN_EXTEND, MVT::v8i32, MVT::v8i16, 4 }, 1782 { ISD::ZERO_EXTEND, MVT::v8i32, MVT::v8i16, 4 }, 1783 { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v4i32, 4 }, 1784 { ISD::ZERO_EXTEND, MVT::v4i64, MVT::v4i32, 4 }, 1785 1786 { ISD::TRUNCATE, MVT::v4i1, MVT::v4i64, 4 }, 1787 { ISD::TRUNCATE, MVT::v8i1, MVT::v8i32, 5 }, 1788 { ISD::TRUNCATE, MVT::v16i1, MVT::v16i16, 4 }, 1789 { ISD::TRUNCATE, MVT::v8i1, MVT::v8i64, 9 }, 1790 { ISD::TRUNCATE, MVT::v16i1, MVT::v16i64, 11 }, 1791 1792 { ISD::TRUNCATE, MVT::v16i8, MVT::v16i16, 4 }, 1793 { ISD::TRUNCATE, MVT::v8i8, MVT::v8i32, 4 }, 1794 { ISD::TRUNCATE, MVT::v8i16, MVT::v8i32, 5 }, 1795 { ISD::TRUNCATE, MVT::v4i8, MVT::v4i64, 4 }, 1796 { ISD::TRUNCATE, MVT::v4i16, MVT::v4i64, 4 }, 1797 { ISD::TRUNCATE, MVT::v4i32, MVT::v4i64, 2 }, 1798 { ISD::TRUNCATE, MVT::v8i8, MVT::v8i64, 11 }, 1799 { ISD::TRUNCATE, MVT::v8i16, MVT::v8i64, 9 }, 1800 { ISD::TRUNCATE, MVT::v8i32, MVT::v8i64, 3 }, 1801 { ISD::TRUNCATE, MVT::v16i8, MVT::v16i64, 11 }, 1802 1803 { ISD::SINT_TO_FP, MVT::v4f32, MVT::v4i1, 3 }, 1804 { ISD::SINT_TO_FP, MVT::v4f64, MVT::v4i1, 3 }, 1805 { ISD::SINT_TO_FP, MVT::v8f32, MVT::v8i1, 8 }, 1806 { ISD::SINT_TO_FP, MVT::v4f32, MVT::v4i8, 3 }, 1807 { ISD::SINT_TO_FP, MVT::v4f64, MVT::v4i8, 3 }, 1808 { ISD::SINT_TO_FP, MVT::v8f32, MVT::v8i8, 8 }, 1809 { ISD::SINT_TO_FP, MVT::v4f32, MVT::v4i16, 3 }, 1810 { ISD::SINT_TO_FP, MVT::v4f64, MVT::v4i16, 3 }, 1811 { ISD::SINT_TO_FP, MVT::v8f32, MVT::v8i16, 5 }, 1812 { ISD::SINT_TO_FP, MVT::v4f32, MVT::v4i32, 1 }, 1813 { ISD::SINT_TO_FP, MVT::v4f64, MVT::v4i32, 1 }, 1814 { ISD::SINT_TO_FP, MVT::v8f32, MVT::v8i32, 1 }, 1815 1816 { ISD::UINT_TO_FP, MVT::v4f32, MVT::v4i1, 7 }, 1817 { ISD::UINT_TO_FP, MVT::v4f64, MVT::v4i1, 7 }, 1818 { ISD::UINT_TO_FP, MVT::v8f32, MVT::v8i1, 6 }, 1819 { ISD::UINT_TO_FP, MVT::v4f32, MVT::v4i8, 2 }, 1820 { ISD::UINT_TO_FP, MVT::v4f64, MVT::v4i8, 2 }, 1821 { ISD::UINT_TO_FP, MVT::v8f32, MVT::v8i8, 5 }, 1822 { ISD::UINT_TO_FP, MVT::v4f32, MVT::v4i16, 2 }, 1823 { ISD::UINT_TO_FP, MVT::v4f64, MVT::v4i16, 2 }, 1824 { ISD::UINT_TO_FP, MVT::v8f32, MVT::v8i16, 5 }, 1825 { ISD::UINT_TO_FP, MVT::v2f64, MVT::v2i32, 6 }, 1826 { ISD::UINT_TO_FP, MVT::v4f32, MVT::v4i32, 6 }, 1827 { ISD::UINT_TO_FP, MVT::v4f64, MVT::v4i32, 6 }, 1828 { ISD::UINT_TO_FP, MVT::v8f32, MVT::v8i32, 9 }, 1829 { ISD::UINT_TO_FP, MVT::v2f64, MVT::v2i64, 5 }, 1830 { ISD::UINT_TO_FP, MVT::v4f64, MVT::v4i64, 6 }, 1831 // The generic code to compute the scalar overhead is currently broken. 1832 // Workaround this limitation by estimating the scalarization overhead 1833 // here. We have roughly 10 instructions per scalar element. 1834 // Multiply that by the vector width. 1835 // FIXME: remove that when PR19268 is fixed. 1836 { ISD::SINT_TO_FP, MVT::v4f64, MVT::v4i64, 13 }, 1837 { ISD::SINT_TO_FP, MVT::v4f64, MVT::v4i64, 13 }, 1838 1839 { ISD::FP_TO_SINT, MVT::v8i8, MVT::v8f32, 4 }, 1840 { ISD::FP_TO_SINT, MVT::v4i8, MVT::v4f64, 3 }, 1841 { ISD::FP_TO_SINT, MVT::v4i16, MVT::v4f64, 2 }, 1842 { ISD::FP_TO_SINT, MVT::v8i16, MVT::v8f32, 3 }, 1843 1844 { ISD::FP_TO_UINT, MVT::v4i8, MVT::v4f64, 3 }, 1845 { ISD::FP_TO_UINT, MVT::v4i16, MVT::v4f64, 2 }, 1846 { ISD::FP_TO_UINT, MVT::v8i8, MVT::v8f32, 4 }, 1847 { ISD::FP_TO_UINT, MVT::v8i16, MVT::v8f32, 3 }, 1848 // This node is expanded into scalarized operations but BasicTTI is overly 1849 // optimistic estimating its cost. It computes 3 per element (one 1850 // vector-extract, one scalar conversion and one vector-insert). The 1851 // problem is that the inserts form a read-modify-write chain so latency 1852 // should be factored in too. Inflating the cost per element by 1. 1853 { ISD::FP_TO_UINT, MVT::v8i32, MVT::v8f32, 8*4 }, 1854 { ISD::FP_TO_UINT, MVT::v4i32, MVT::v4f64, 4*4 }, 1855 1856 { ISD::FP_EXTEND, MVT::v4f64, MVT::v4f32, 1 }, 1857 { ISD::FP_ROUND, MVT::v4f32, MVT::v4f64, 1 }, 1858 }; 1859 1860 static const TypeConversionCostTblEntry SSE41ConversionTbl[] = { 1861 { ISD::ZERO_EXTEND, MVT::v4i64, MVT::v4i8, 2 }, 1862 { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v4i8, 2 }, 1863 { ISD::ZERO_EXTEND, MVT::v4i64, MVT::v4i16, 2 }, 1864 { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v4i16, 2 }, 1865 { ISD::ZERO_EXTEND, MVT::v4i64, MVT::v4i32, 2 }, 1866 { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v4i32, 2 }, 1867 1868 { ISD::ZERO_EXTEND, MVT::v4i16, MVT::v4i8, 1 }, 1869 { ISD::SIGN_EXTEND, MVT::v4i16, MVT::v4i8, 2 }, 1870 { ISD::ZERO_EXTEND, MVT::v4i32, MVT::v4i8, 1 }, 1871 { ISD::SIGN_EXTEND, MVT::v4i32, MVT::v4i8, 1 }, 1872 { ISD::ZERO_EXTEND, MVT::v8i16, MVT::v8i8, 1 }, 1873 { ISD::SIGN_EXTEND, MVT::v8i16, MVT::v8i8, 1 }, 1874 { ISD::ZERO_EXTEND, MVT::v8i32, MVT::v8i8, 2 }, 1875 { ISD::SIGN_EXTEND, MVT::v8i32, MVT::v8i8, 2 }, 1876 { ISD::ZERO_EXTEND, MVT::v16i16, MVT::v16i8, 2 }, 1877 { ISD::SIGN_EXTEND, MVT::v16i16, MVT::v16i8, 2 }, 1878 { ISD::ZERO_EXTEND, MVT::v16i32, MVT::v16i8, 4 }, 1879 { ISD::SIGN_EXTEND, MVT::v16i32, MVT::v16i8, 4 }, 1880 { ISD::ZERO_EXTEND, MVT::v4i32, MVT::v4i16, 1 }, 1881 { ISD::SIGN_EXTEND, MVT::v4i32, MVT::v4i16, 1 }, 1882 { ISD::ZERO_EXTEND, MVT::v8i32, MVT::v8i16, 2 }, 1883 { ISD::SIGN_EXTEND, MVT::v8i32, MVT::v8i16, 2 }, 1884 { ISD::ZERO_EXTEND, MVT::v16i32, MVT::v16i16, 4 }, 1885 { ISD::SIGN_EXTEND, MVT::v16i32, MVT::v16i16, 4 }, 1886 1887 // These truncates end up widening elements. 1888 { ISD::TRUNCATE, MVT::v2i1, MVT::v2i8, 1 }, // PMOVXZBQ 1889 { ISD::TRUNCATE, MVT::v2i1, MVT::v2i16, 1 }, // PMOVXZWQ 1890 { ISD::TRUNCATE, MVT::v4i1, MVT::v4i8, 1 }, // PMOVXZBD 1891 1892 { ISD::TRUNCATE, MVT::v2i8, MVT::v2i16, 1 }, 1893 { ISD::TRUNCATE, MVT::v4i8, MVT::v4i16, 1 }, 1894 { ISD::TRUNCATE, MVT::v8i8, MVT::v8i16, 1 }, 1895 { ISD::TRUNCATE, MVT::v4i8, MVT::v4i32, 1 }, 1896 { ISD::TRUNCATE, MVT::v4i16, MVT::v4i32, 1 }, 1897 { ISD::TRUNCATE, MVT::v8i8, MVT::v8i32, 3 }, 1898 { ISD::TRUNCATE, MVT::v8i16, MVT::v8i32, 3 }, 1899 { ISD::TRUNCATE, MVT::v16i16, MVT::v16i32, 6 }, 1900 { ISD::TRUNCATE, MVT::v2i8, MVT::v2i64, 1 }, // PSHUFB 1901 1902 { ISD::UINT_TO_FP, MVT::f32, MVT::i64, 4 }, 1903 { ISD::UINT_TO_FP, MVT::f64, MVT::i64, 4 }, 1904 1905 { ISD::FP_TO_SINT, MVT::v2i8, MVT::v2f32, 3 }, 1906 { ISD::FP_TO_SINT, MVT::v2i8, MVT::v2f64, 3 }, 1907 1908 { ISD::FP_TO_UINT, MVT::v2i8, MVT::v2f32, 3 }, 1909 { ISD::FP_TO_UINT, MVT::v2i8, MVT::v2f64, 3 }, 1910 { ISD::FP_TO_UINT, MVT::v4i16, MVT::v4f32, 2 }, 1911 }; 1912 1913 static const TypeConversionCostTblEntry SSE2ConversionTbl[] = { 1914 // These are somewhat magic numbers justified by looking at the output of 1915 // Intel's IACA, running some kernels and making sure when we take 1916 // legalization into account the throughput will be overestimated. 1917 { ISD::SINT_TO_FP, MVT::v4f32, MVT::v16i8, 8 }, 1918 { ISD::SINT_TO_FP, MVT::v2f64, MVT::v16i8, 16*10 }, 1919 { ISD::SINT_TO_FP, MVT::v4f32, MVT::v8i16, 15 }, 1920 { ISD::SINT_TO_FP, MVT::v2f64, MVT::v8i16, 8*10 }, 1921 { ISD::SINT_TO_FP, MVT::v4f32, MVT::v4i32, 5 }, 1922 { ISD::SINT_TO_FP, MVT::v2f64, MVT::v4i32, 2*10 }, 1923 { ISD::SINT_TO_FP, MVT::v2f64, MVT::v2i32, 2*10 }, 1924 { ISD::SINT_TO_FP, MVT::v4f32, MVT::v2i64, 15 }, 1925 { ISD::SINT_TO_FP, MVT::v2f64, MVT::v2i64, 2*10 }, 1926 1927 { ISD::UINT_TO_FP, MVT::v2f64, MVT::v16i8, 16*10 }, 1928 { ISD::UINT_TO_FP, MVT::v4f32, MVT::v16i8, 8 }, 1929 { ISD::UINT_TO_FP, MVT::v4f32, MVT::v8i16, 15 }, 1930 { ISD::UINT_TO_FP, MVT::v2f64, MVT::v8i16, 8*10 }, 1931 { ISD::UINT_TO_FP, MVT::v2f64, MVT::v4i32, 4*10 }, 1932 { ISD::UINT_TO_FP, MVT::v4f32, MVT::v4i32, 8 }, 1933 { ISD::UINT_TO_FP, MVT::v2f64, MVT::v2i64, 6 }, 1934 { ISD::UINT_TO_FP, MVT::v4f32, MVT::v2i64, 15 }, 1935 1936 { ISD::FP_TO_SINT, MVT::v2i8, MVT::v2f32, 4 }, 1937 { ISD::FP_TO_SINT, MVT::v2i16, MVT::v2f32, 2 }, 1938 { ISD::FP_TO_SINT, MVT::v4i8, MVT::v4f32, 3 }, 1939 { ISD::FP_TO_SINT, MVT::v4i16, MVT::v4f32, 2 }, 1940 { ISD::FP_TO_SINT, MVT::v2i16, MVT::v2f64, 2 }, 1941 { ISD::FP_TO_SINT, MVT::v2i8, MVT::v2f64, 4 }, 1942 1943 { ISD::FP_TO_SINT, MVT::v2i32, MVT::v2f64, 1 }, 1944 1945 { ISD::UINT_TO_FP, MVT::f32, MVT::i64, 6 }, 1946 { ISD::UINT_TO_FP, MVT::f64, MVT::i64, 6 }, 1947 1948 { ISD::FP_TO_UINT, MVT::i64, MVT::f32, 4 }, 1949 { ISD::FP_TO_UINT, MVT::i64, MVT::f64, 4 }, 1950 { ISD::FP_TO_UINT, MVT::v2i8, MVT::v2f32, 4 }, 1951 { ISD::FP_TO_UINT, MVT::v2i8, MVT::v2f64, 4 }, 1952 { ISD::FP_TO_UINT, MVT::v4i8, MVT::v4f32, 3 }, 1953 { ISD::FP_TO_UINT, MVT::v2i16, MVT::v2f32, 2 }, 1954 { ISD::FP_TO_UINT, MVT::v2i16, MVT::v2f64, 2 }, 1955 { ISD::FP_TO_UINT, MVT::v4i16, MVT::v4f32, 4 }, 1956 1957 { ISD::ZERO_EXTEND, MVT::v4i16, MVT::v4i8, 1 }, 1958 { ISD::SIGN_EXTEND, MVT::v4i16, MVT::v4i8, 6 }, 1959 { ISD::ZERO_EXTEND, MVT::v4i32, MVT::v4i8, 2 }, 1960 { ISD::SIGN_EXTEND, MVT::v4i32, MVT::v4i8, 3 }, 1961 { ISD::ZERO_EXTEND, MVT::v4i64, MVT::v4i8, 4 }, 1962 { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v4i8, 8 }, 1963 { ISD::ZERO_EXTEND, MVT::v8i16, MVT::v8i8, 1 }, 1964 { ISD::SIGN_EXTEND, MVT::v8i16, MVT::v8i8, 2 }, 1965 { ISD::ZERO_EXTEND, MVT::v8i32, MVT::v8i8, 6 }, 1966 { ISD::SIGN_EXTEND, MVT::v8i32, MVT::v8i8, 6 }, 1967 { ISD::ZERO_EXTEND, MVT::v16i16, MVT::v16i8, 3 }, 1968 { ISD::SIGN_EXTEND, MVT::v16i16, MVT::v16i8, 4 }, 1969 { ISD::ZERO_EXTEND, MVT::v16i32, MVT::v16i8, 9 }, 1970 { ISD::SIGN_EXTEND, MVT::v16i32, MVT::v16i8, 12 }, 1971 { ISD::ZERO_EXTEND, MVT::v4i32, MVT::v4i16, 1 }, 1972 { ISD::SIGN_EXTEND, MVT::v4i32, MVT::v4i16, 2 }, 1973 { ISD::ZERO_EXTEND, MVT::v4i64, MVT::v4i16, 3 }, 1974 { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v4i16, 10 }, 1975 { ISD::ZERO_EXTEND, MVT::v8i32, MVT::v8i16, 3 }, 1976 { ISD::SIGN_EXTEND, MVT::v8i32, MVT::v8i16, 4 }, 1977 { ISD::ZERO_EXTEND, MVT::v16i32, MVT::v16i16, 6 }, 1978 { ISD::SIGN_EXTEND, MVT::v16i32, MVT::v16i16, 8 }, 1979 { ISD::ZERO_EXTEND, MVT::v4i64, MVT::v4i32, 3 }, 1980 { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v4i32, 5 }, 1981 1982 // These truncates are really widening elements. 1983 { ISD::TRUNCATE, MVT::v2i1, MVT::v2i32, 1 }, // PSHUFD 1984 { ISD::TRUNCATE, MVT::v2i1, MVT::v2i16, 2 }, // PUNPCKLWD+DQ 1985 { ISD::TRUNCATE, MVT::v2i1, MVT::v2i8, 3 }, // PUNPCKLBW+WD+PSHUFD 1986 { ISD::TRUNCATE, MVT::v4i1, MVT::v4i16, 1 }, // PUNPCKLWD 1987 { ISD::TRUNCATE, MVT::v4i1, MVT::v4i8, 2 }, // PUNPCKLBW+WD 1988 { ISD::TRUNCATE, MVT::v8i1, MVT::v8i8, 1 }, // PUNPCKLBW 1989 1990 { ISD::TRUNCATE, MVT::v2i8, MVT::v2i16, 2 }, // PAND+PACKUSWB 1991 { ISD::TRUNCATE, MVT::v4i8, MVT::v4i16, 2 }, // PAND+PACKUSWB 1992 { ISD::TRUNCATE, MVT::v8i8, MVT::v8i16, 2 }, // PAND+PACKUSWB 1993 { ISD::TRUNCATE, MVT::v16i8, MVT::v16i16, 3 }, 1994 { ISD::TRUNCATE, MVT::v2i8, MVT::v2i32, 3 }, // PAND+2*PACKUSWB 1995 { ISD::TRUNCATE, MVT::v2i16, MVT::v2i32, 1 }, 1996 { ISD::TRUNCATE, MVT::v4i8, MVT::v4i32, 3 }, 1997 { ISD::TRUNCATE, MVT::v4i16, MVT::v4i32, 3 }, 1998 { ISD::TRUNCATE, MVT::v8i8, MVT::v8i32, 4 }, 1999 { ISD::TRUNCATE, MVT::v16i8, MVT::v16i32, 7 }, 2000 { ISD::TRUNCATE, MVT::v8i16, MVT::v8i32, 5 }, 2001 { ISD::TRUNCATE, MVT::v16i16, MVT::v16i32, 10 }, 2002 { ISD::TRUNCATE, MVT::v2i8, MVT::v2i64, 4 }, // PAND+3*PACKUSWB 2003 { ISD::TRUNCATE, MVT::v2i16, MVT::v2i64, 2 }, // PSHUFD+PSHUFLW 2004 { ISD::TRUNCATE, MVT::v2i32, MVT::v2i64, 1 }, // PSHUFD 2005 }; 2006 2007 std::pair<int, MVT> LTSrc = TLI->getTypeLegalizationCost(DL, Src); 2008 std::pair<int, MVT> LTDest = TLI->getTypeLegalizationCost(DL, Dst); 2009 2010 if (ST->hasSSE2() && !ST->hasAVX()) { 2011 if (const auto *Entry = ConvertCostTableLookup(SSE2ConversionTbl, ISD, 2012 LTDest.second, LTSrc.second)) 2013 return AdjustCost(LTSrc.first * Entry->Cost); 2014 } 2015 2016 EVT SrcTy = TLI->getValueType(DL, Src); 2017 EVT DstTy = TLI->getValueType(DL, Dst); 2018 2019 // The function getSimpleVT only handles simple value types. 2020 if (!SrcTy.isSimple() || !DstTy.isSimple()) 2021 return AdjustCost(BaseT::getCastInstrCost(Opcode, Dst, Src, CCH, CostKind)); 2022 2023 MVT SimpleSrcTy = SrcTy.getSimpleVT(); 2024 MVT SimpleDstTy = DstTy.getSimpleVT(); 2025 2026 if (ST->useAVX512Regs()) { 2027 if (ST->hasBWI()) 2028 if (const auto *Entry = ConvertCostTableLookup(AVX512BWConversionTbl, ISD, 2029 SimpleDstTy, SimpleSrcTy)) 2030 return AdjustCost(Entry->Cost); 2031 2032 if (ST->hasDQI()) 2033 if (const auto *Entry = ConvertCostTableLookup(AVX512DQConversionTbl, ISD, 2034 SimpleDstTy, SimpleSrcTy)) 2035 return AdjustCost(Entry->Cost); 2036 2037 if (ST->hasAVX512()) 2038 if (const auto *Entry = ConvertCostTableLookup(AVX512FConversionTbl, ISD, 2039 SimpleDstTy, SimpleSrcTy)) 2040 return AdjustCost(Entry->Cost); 2041 } 2042 2043 if (ST->hasBWI()) 2044 if (const auto *Entry = ConvertCostTableLookup(AVX512BWVLConversionTbl, ISD, 2045 SimpleDstTy, SimpleSrcTy)) 2046 return AdjustCost(Entry->Cost); 2047 2048 if (ST->hasDQI()) 2049 if (const auto *Entry = ConvertCostTableLookup(AVX512DQVLConversionTbl, ISD, 2050 SimpleDstTy, SimpleSrcTy)) 2051 return AdjustCost(Entry->Cost); 2052 2053 if (ST->hasAVX512()) 2054 if (const auto *Entry = ConvertCostTableLookup(AVX512VLConversionTbl, ISD, 2055 SimpleDstTy, SimpleSrcTy)) 2056 return AdjustCost(Entry->Cost); 2057 2058 if (ST->hasAVX2()) { 2059 if (const auto *Entry = ConvertCostTableLookup(AVX2ConversionTbl, ISD, 2060 SimpleDstTy, SimpleSrcTy)) 2061 return AdjustCost(Entry->Cost); 2062 } 2063 2064 if (ST->hasAVX()) { 2065 if (const auto *Entry = ConvertCostTableLookup(AVXConversionTbl, ISD, 2066 SimpleDstTy, SimpleSrcTy)) 2067 return AdjustCost(Entry->Cost); 2068 } 2069 2070 if (ST->hasSSE41()) { 2071 if (const auto *Entry = ConvertCostTableLookup(SSE41ConversionTbl, ISD, 2072 SimpleDstTy, SimpleSrcTy)) 2073 return AdjustCost(Entry->Cost); 2074 } 2075 2076 if (ST->hasSSE2()) { 2077 if (const auto *Entry = ConvertCostTableLookup(SSE2ConversionTbl, ISD, 2078 SimpleDstTy, SimpleSrcTy)) 2079 return AdjustCost(Entry->Cost); 2080 } 2081 2082 return AdjustCost( 2083 BaseT::getCastInstrCost(Opcode, Dst, Src, CCH, CostKind, I)); 2084 } 2085 2086 int X86TTIImpl::getCmpSelInstrCost(unsigned Opcode, Type *ValTy, Type *CondTy, 2087 TTI::TargetCostKind CostKind, 2088 const Instruction *I) { 2089 // TODO: Handle other cost kinds. 2090 if (CostKind != TTI::TCK_RecipThroughput) 2091 return BaseT::getCmpSelInstrCost(Opcode, ValTy, CondTy, CostKind, I); 2092 2093 // Legalize the type. 2094 std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, ValTy); 2095 2096 MVT MTy = LT.second; 2097 2098 int ISD = TLI->InstructionOpcodeToISD(Opcode); 2099 assert(ISD && "Invalid opcode"); 2100 2101 unsigned ExtraCost = 0; 2102 if (I && (Opcode == Instruction::ICmp || Opcode == Instruction::FCmp)) { 2103 // Some vector comparison predicates cost extra instructions. 2104 if (MTy.isVector() && 2105 !((ST->hasXOP() && (!ST->hasAVX2() || MTy.is128BitVector())) || 2106 (ST->hasAVX512() && 32 <= MTy.getScalarSizeInBits()) || 2107 ST->hasBWI())) { 2108 switch (cast<CmpInst>(I)->getPredicate()) { 2109 case CmpInst::Predicate::ICMP_NE: 2110 // xor(cmpeq(x,y),-1) 2111 ExtraCost = 1; 2112 break; 2113 case CmpInst::Predicate::ICMP_SGE: 2114 case CmpInst::Predicate::ICMP_SLE: 2115 // xor(cmpgt(x,y),-1) 2116 ExtraCost = 1; 2117 break; 2118 case CmpInst::Predicate::ICMP_ULT: 2119 case CmpInst::Predicate::ICMP_UGT: 2120 // cmpgt(xor(x,signbit),xor(y,signbit)) 2121 // xor(cmpeq(pmaxu(x,y),x),-1) 2122 ExtraCost = 2; 2123 break; 2124 case CmpInst::Predicate::ICMP_ULE: 2125 case CmpInst::Predicate::ICMP_UGE: 2126 if ((ST->hasSSE41() && MTy.getScalarSizeInBits() == 32) || 2127 (ST->hasSSE2() && MTy.getScalarSizeInBits() < 32)) { 2128 // cmpeq(psubus(x,y),0) 2129 // cmpeq(pminu(x,y),x) 2130 ExtraCost = 1; 2131 } else { 2132 // xor(cmpgt(xor(x,signbit),xor(y,signbit)),-1) 2133 ExtraCost = 3; 2134 } 2135 break; 2136 default: 2137 break; 2138 } 2139 } 2140 } 2141 2142 static const CostTblEntry SLMCostTbl[] = { 2143 // slm pcmpeq/pcmpgt throughput is 2 2144 { ISD::SETCC, MVT::v2i64, 2 }, 2145 }; 2146 2147 static const CostTblEntry AVX512BWCostTbl[] = { 2148 { ISD::SETCC, MVT::v32i16, 1 }, 2149 { ISD::SETCC, MVT::v64i8, 1 }, 2150 2151 { ISD::SELECT, MVT::v32i16, 1 }, 2152 { ISD::SELECT, MVT::v64i8, 1 }, 2153 }; 2154 2155 static const CostTblEntry AVX512CostTbl[] = { 2156 { ISD::SETCC, MVT::v8i64, 1 }, 2157 { ISD::SETCC, MVT::v16i32, 1 }, 2158 { ISD::SETCC, MVT::v8f64, 1 }, 2159 { ISD::SETCC, MVT::v16f32, 1 }, 2160 2161 { ISD::SELECT, MVT::v8i64, 1 }, 2162 { ISD::SELECT, MVT::v16i32, 1 }, 2163 { ISD::SELECT, MVT::v8f64, 1 }, 2164 { ISD::SELECT, MVT::v16f32, 1 }, 2165 2166 { ISD::SETCC, MVT::v32i16, 2 }, // FIXME: should probably be 4 2167 { ISD::SETCC, MVT::v64i8, 2 }, // FIXME: should probably be 4 2168 2169 { ISD::SELECT, MVT::v32i16, 2 }, // FIXME: should be 3 2170 { ISD::SELECT, MVT::v64i8, 2 }, // FIXME: should be 3 2171 }; 2172 2173 static const CostTblEntry AVX2CostTbl[] = { 2174 { ISD::SETCC, MVT::v4i64, 1 }, 2175 { ISD::SETCC, MVT::v8i32, 1 }, 2176 { ISD::SETCC, MVT::v16i16, 1 }, 2177 { ISD::SETCC, MVT::v32i8, 1 }, 2178 2179 { ISD::SELECT, MVT::v4i64, 1 }, // pblendvb 2180 { ISD::SELECT, MVT::v8i32, 1 }, // pblendvb 2181 { ISD::SELECT, MVT::v16i16, 1 }, // pblendvb 2182 { ISD::SELECT, MVT::v32i8, 1 }, // pblendvb 2183 }; 2184 2185 static const CostTblEntry AVX1CostTbl[] = { 2186 { ISD::SETCC, MVT::v4f64, 1 }, 2187 { ISD::SETCC, MVT::v8f32, 1 }, 2188 // AVX1 does not support 8-wide integer compare. 2189 { ISD::SETCC, MVT::v4i64, 4 }, 2190 { ISD::SETCC, MVT::v8i32, 4 }, 2191 { ISD::SETCC, MVT::v16i16, 4 }, 2192 { ISD::SETCC, MVT::v32i8, 4 }, 2193 2194 { ISD::SELECT, MVT::v4f64, 1 }, // vblendvpd 2195 { ISD::SELECT, MVT::v8f32, 1 }, // vblendvps 2196 { ISD::SELECT, MVT::v4i64, 1 }, // vblendvpd 2197 { ISD::SELECT, MVT::v8i32, 1 }, // vblendvps 2198 { ISD::SELECT, MVT::v16i16, 3 }, // vandps + vandnps + vorps 2199 { ISD::SELECT, MVT::v32i8, 3 }, // vandps + vandnps + vorps 2200 }; 2201 2202 static const CostTblEntry SSE42CostTbl[] = { 2203 { ISD::SETCC, MVT::v2f64, 1 }, 2204 { ISD::SETCC, MVT::v4f32, 1 }, 2205 { ISD::SETCC, MVT::v2i64, 1 }, 2206 }; 2207 2208 static const CostTblEntry SSE41CostTbl[] = { 2209 { ISD::SELECT, MVT::v2f64, 1 }, // blendvpd 2210 { ISD::SELECT, MVT::v4f32, 1 }, // blendvps 2211 { ISD::SELECT, MVT::v2i64, 1 }, // pblendvb 2212 { ISD::SELECT, MVT::v4i32, 1 }, // pblendvb 2213 { ISD::SELECT, MVT::v8i16, 1 }, // pblendvb 2214 { ISD::SELECT, MVT::v16i8, 1 }, // pblendvb 2215 }; 2216 2217 static const CostTblEntry SSE2CostTbl[] = { 2218 { ISD::SETCC, MVT::v2f64, 2 }, 2219 { ISD::SETCC, MVT::f64, 1 }, 2220 { ISD::SETCC, MVT::v2i64, 8 }, 2221 { ISD::SETCC, MVT::v4i32, 1 }, 2222 { ISD::SETCC, MVT::v8i16, 1 }, 2223 { ISD::SETCC, MVT::v16i8, 1 }, 2224 2225 { ISD::SELECT, MVT::v2f64, 3 }, // andpd + andnpd + orpd 2226 { ISD::SELECT, MVT::v2i64, 3 }, // pand + pandn + por 2227 { ISD::SELECT, MVT::v4i32, 3 }, // pand + pandn + por 2228 { ISD::SELECT, MVT::v8i16, 3 }, // pand + pandn + por 2229 { ISD::SELECT, MVT::v16i8, 3 }, // pand + pandn + por 2230 }; 2231 2232 static const CostTblEntry SSE1CostTbl[] = { 2233 { ISD::SETCC, MVT::v4f32, 2 }, 2234 { ISD::SETCC, MVT::f32, 1 }, 2235 2236 { ISD::SELECT, MVT::v4f32, 3 }, // andps + andnps + orps 2237 }; 2238 2239 if (ST->isSLM()) 2240 if (const auto *Entry = CostTableLookup(SLMCostTbl, ISD, MTy)) 2241 return LT.first * (ExtraCost + Entry->Cost); 2242 2243 if (ST->hasBWI()) 2244 if (const auto *Entry = CostTableLookup(AVX512BWCostTbl, ISD, MTy)) 2245 return LT.first * (ExtraCost + Entry->Cost); 2246 2247 if (ST->hasAVX512()) 2248 if (const auto *Entry = CostTableLookup(AVX512CostTbl, ISD, MTy)) 2249 return LT.first * (ExtraCost + Entry->Cost); 2250 2251 if (ST->hasAVX2()) 2252 if (const auto *Entry = CostTableLookup(AVX2CostTbl, ISD, MTy)) 2253 return LT.first * (ExtraCost + Entry->Cost); 2254 2255 if (ST->hasAVX()) 2256 if (const auto *Entry = CostTableLookup(AVX1CostTbl, ISD, MTy)) 2257 return LT.first * (ExtraCost + Entry->Cost); 2258 2259 if (ST->hasSSE42()) 2260 if (const auto *Entry = CostTableLookup(SSE42CostTbl, ISD, MTy)) 2261 return LT.first * (ExtraCost + Entry->Cost); 2262 2263 if (ST->hasSSE41()) 2264 if (const auto *Entry = CostTableLookup(SSE41CostTbl, ISD, MTy)) 2265 return LT.first * (ExtraCost + Entry->Cost); 2266 2267 if (ST->hasSSE2()) 2268 if (const auto *Entry = CostTableLookup(SSE2CostTbl, ISD, MTy)) 2269 return LT.first * (ExtraCost + Entry->Cost); 2270 2271 if (ST->hasSSE1()) 2272 if (const auto *Entry = CostTableLookup(SSE1CostTbl, ISD, MTy)) 2273 return LT.first * (ExtraCost + Entry->Cost); 2274 2275 return BaseT::getCmpSelInstrCost(Opcode, ValTy, CondTy, CostKind, I); 2276 } 2277 2278 unsigned X86TTIImpl::getAtomicMemIntrinsicMaxElementSize() const { return 16; } 2279 2280 int X86TTIImpl::getTypeBasedIntrinsicInstrCost( 2281 const IntrinsicCostAttributes &ICA, TTI::TargetCostKind CostKind) { 2282 2283 // Costs should match the codegen from: 2284 // BITREVERSE: llvm\test\CodeGen\X86\vector-bitreverse.ll 2285 // BSWAP: llvm\test\CodeGen\X86\bswap-vector.ll 2286 // CTLZ: llvm\test\CodeGen\X86\vector-lzcnt-*.ll 2287 // CTPOP: llvm\test\CodeGen\X86\vector-popcnt-*.ll 2288 // CTTZ: llvm\test\CodeGen\X86\vector-tzcnt-*.ll 2289 static const CostTblEntry AVX512CDCostTbl[] = { 2290 { ISD::CTLZ, MVT::v8i64, 1 }, 2291 { ISD::CTLZ, MVT::v16i32, 1 }, 2292 { ISD::CTLZ, MVT::v32i16, 8 }, 2293 { ISD::CTLZ, MVT::v64i8, 20 }, 2294 { ISD::CTLZ, MVT::v4i64, 1 }, 2295 { ISD::CTLZ, MVT::v8i32, 1 }, 2296 { ISD::CTLZ, MVT::v16i16, 4 }, 2297 { ISD::CTLZ, MVT::v32i8, 10 }, 2298 { ISD::CTLZ, MVT::v2i64, 1 }, 2299 { ISD::CTLZ, MVT::v4i32, 1 }, 2300 { ISD::CTLZ, MVT::v8i16, 4 }, 2301 { ISD::CTLZ, MVT::v16i8, 4 }, 2302 }; 2303 static const CostTblEntry AVX512BWCostTbl[] = { 2304 { ISD::ABS, MVT::v32i16, 1 }, 2305 { ISD::ABS, MVT::v64i8, 1 }, 2306 { ISD::BITREVERSE, MVT::v8i64, 5 }, 2307 { ISD::BITREVERSE, MVT::v16i32, 5 }, 2308 { ISD::BITREVERSE, MVT::v32i16, 5 }, 2309 { ISD::BITREVERSE, MVT::v64i8, 5 }, 2310 { ISD::CTLZ, MVT::v8i64, 23 }, 2311 { ISD::CTLZ, MVT::v16i32, 22 }, 2312 { ISD::CTLZ, MVT::v32i16, 18 }, 2313 { ISD::CTLZ, MVT::v64i8, 17 }, 2314 { ISD::CTPOP, MVT::v8i64, 7 }, 2315 { ISD::CTPOP, MVT::v16i32, 11 }, 2316 { ISD::CTPOP, MVT::v32i16, 9 }, 2317 { ISD::CTPOP, MVT::v64i8, 6 }, 2318 { ISD::CTTZ, MVT::v8i64, 10 }, 2319 { ISD::CTTZ, MVT::v16i32, 14 }, 2320 { ISD::CTTZ, MVT::v32i16, 12 }, 2321 { ISD::CTTZ, MVT::v64i8, 9 }, 2322 { ISD::SADDSAT, MVT::v32i16, 1 }, 2323 { ISD::SADDSAT, MVT::v64i8, 1 }, 2324 { ISD::SMAX, MVT::v32i16, 1 }, 2325 { ISD::SMAX, MVT::v64i8, 1 }, 2326 { ISD::SMIN, MVT::v32i16, 1 }, 2327 { ISD::SMIN, MVT::v64i8, 1 }, 2328 { ISD::SSUBSAT, MVT::v32i16, 1 }, 2329 { ISD::SSUBSAT, MVT::v64i8, 1 }, 2330 { ISD::UADDSAT, MVT::v32i16, 1 }, 2331 { ISD::UADDSAT, MVT::v64i8, 1 }, 2332 { ISD::UMAX, MVT::v32i16, 1 }, 2333 { ISD::UMAX, MVT::v64i8, 1 }, 2334 { ISD::UMIN, MVT::v32i16, 1 }, 2335 { ISD::UMIN, MVT::v64i8, 1 }, 2336 { ISD::USUBSAT, MVT::v32i16, 1 }, 2337 { ISD::USUBSAT, MVT::v64i8, 1 }, 2338 }; 2339 static const CostTblEntry AVX512CostTbl[] = { 2340 { ISD::ABS, MVT::v8i64, 1 }, 2341 { ISD::ABS, MVT::v16i32, 1 }, 2342 { ISD::ABS, MVT::v32i16, 2 }, // FIXME: include split 2343 { ISD::ABS, MVT::v64i8, 2 }, // FIXME: include split 2344 { ISD::ABS, MVT::v4i64, 1 }, 2345 { ISD::ABS, MVT::v2i64, 1 }, 2346 { ISD::BITREVERSE, MVT::v8i64, 36 }, 2347 { ISD::BITREVERSE, MVT::v16i32, 24 }, 2348 { ISD::BITREVERSE, MVT::v32i16, 10 }, 2349 { ISD::BITREVERSE, MVT::v64i8, 10 }, 2350 { ISD::CTLZ, MVT::v8i64, 29 }, 2351 { ISD::CTLZ, MVT::v16i32, 35 }, 2352 { ISD::CTLZ, MVT::v32i16, 28 }, 2353 { ISD::CTLZ, MVT::v64i8, 18 }, 2354 { ISD::CTPOP, MVT::v8i64, 16 }, 2355 { ISD::CTPOP, MVT::v16i32, 24 }, 2356 { ISD::CTPOP, MVT::v32i16, 18 }, 2357 { ISD::CTPOP, MVT::v64i8, 12 }, 2358 { ISD::CTTZ, MVT::v8i64, 20 }, 2359 { ISD::CTTZ, MVT::v16i32, 28 }, 2360 { ISD::CTTZ, MVT::v32i16, 24 }, 2361 { ISD::CTTZ, MVT::v64i8, 18 }, 2362 { ISD::SMAX, MVT::v8i64, 1 }, 2363 { ISD::SMAX, MVT::v16i32, 1 }, 2364 { ISD::SMAX, MVT::v32i16, 2 }, // FIXME: include split 2365 { ISD::SMAX, MVT::v64i8, 2 }, // FIXME: include split 2366 { ISD::SMAX, MVT::v4i64, 1 }, 2367 { ISD::SMAX, MVT::v2i64, 1 }, 2368 { ISD::SMIN, MVT::v8i64, 1 }, 2369 { ISD::SMIN, MVT::v16i32, 1 }, 2370 { ISD::SMIN, MVT::v32i16, 2 }, // FIXME: include split 2371 { ISD::SMIN, MVT::v64i8, 2 }, // FIXME: include split 2372 { ISD::SMIN, MVT::v4i64, 1 }, 2373 { ISD::SMIN, MVT::v2i64, 1 }, 2374 { ISD::UMAX, MVT::v8i64, 1 }, 2375 { ISD::UMAX, MVT::v16i32, 1 }, 2376 { ISD::UMAX, MVT::v32i16, 2 }, // FIXME: include split 2377 { ISD::UMAX, MVT::v64i8, 2 }, // FIXME: include split 2378 { ISD::UMAX, MVT::v4i64, 1 }, 2379 { ISD::UMAX, MVT::v2i64, 1 }, 2380 { ISD::UMIN, MVT::v8i64, 1 }, 2381 { ISD::UMIN, MVT::v16i32, 1 }, 2382 { ISD::UMIN, MVT::v32i16, 2 }, // FIXME: include split 2383 { ISD::UMIN, MVT::v64i8, 2 }, // FIXME: include split 2384 { ISD::UMIN, MVT::v4i64, 1 }, 2385 { ISD::UMIN, MVT::v2i64, 1 }, 2386 { ISD::USUBSAT, MVT::v16i32, 2 }, // pmaxud + psubd 2387 { ISD::USUBSAT, MVT::v2i64, 2 }, // pmaxuq + psubq 2388 { ISD::USUBSAT, MVT::v4i64, 2 }, // pmaxuq + psubq 2389 { ISD::USUBSAT, MVT::v8i64, 2 }, // pmaxuq + psubq 2390 { ISD::UADDSAT, MVT::v16i32, 3 }, // not + pminud + paddd 2391 { ISD::UADDSAT, MVT::v2i64, 3 }, // not + pminuq + paddq 2392 { ISD::UADDSAT, MVT::v4i64, 3 }, // not + pminuq + paddq 2393 { ISD::UADDSAT, MVT::v8i64, 3 }, // not + pminuq + paddq 2394 { ISD::SADDSAT, MVT::v32i16, 2 }, // FIXME: include split 2395 { ISD::SADDSAT, MVT::v64i8, 2 }, // FIXME: include split 2396 { ISD::SSUBSAT, MVT::v32i16, 2 }, // FIXME: include split 2397 { ISD::SSUBSAT, MVT::v64i8, 2 }, // FIXME: include split 2398 { ISD::UADDSAT, MVT::v32i16, 2 }, // FIXME: include split 2399 { ISD::UADDSAT, MVT::v64i8, 2 }, // FIXME: include split 2400 { ISD::USUBSAT, MVT::v32i16, 2 }, // FIXME: include split 2401 { ISD::USUBSAT, MVT::v64i8, 2 }, // FIXME: include split 2402 { ISD::FMAXNUM, MVT::f32, 2 }, 2403 { ISD::FMAXNUM, MVT::v4f32, 2 }, 2404 { ISD::FMAXNUM, MVT::v8f32, 2 }, 2405 { ISD::FMAXNUM, MVT::v16f32, 2 }, 2406 { ISD::FMAXNUM, MVT::f64, 2 }, 2407 { ISD::FMAXNUM, MVT::v2f64, 2 }, 2408 { ISD::FMAXNUM, MVT::v4f64, 2 }, 2409 { ISD::FMAXNUM, MVT::v8f64, 2 }, 2410 }; 2411 static const CostTblEntry XOPCostTbl[] = { 2412 { ISD::BITREVERSE, MVT::v4i64, 4 }, 2413 { ISD::BITREVERSE, MVT::v8i32, 4 }, 2414 { ISD::BITREVERSE, MVT::v16i16, 4 }, 2415 { ISD::BITREVERSE, MVT::v32i8, 4 }, 2416 { ISD::BITREVERSE, MVT::v2i64, 1 }, 2417 { ISD::BITREVERSE, MVT::v4i32, 1 }, 2418 { ISD::BITREVERSE, MVT::v8i16, 1 }, 2419 { ISD::BITREVERSE, MVT::v16i8, 1 }, 2420 { ISD::BITREVERSE, MVT::i64, 3 }, 2421 { ISD::BITREVERSE, MVT::i32, 3 }, 2422 { ISD::BITREVERSE, MVT::i16, 3 }, 2423 { ISD::BITREVERSE, MVT::i8, 3 } 2424 }; 2425 static const CostTblEntry AVX2CostTbl[] = { 2426 { ISD::ABS, MVT::v4i64, 2 }, // VBLENDVPD(X,VPSUBQ(0,X),X) 2427 { ISD::ABS, MVT::v8i32, 1 }, 2428 { ISD::ABS, MVT::v16i16, 1 }, 2429 { ISD::ABS, MVT::v32i8, 1 }, 2430 { ISD::BITREVERSE, MVT::v4i64, 5 }, 2431 { ISD::BITREVERSE, MVT::v8i32, 5 }, 2432 { ISD::BITREVERSE, MVT::v16i16, 5 }, 2433 { ISD::BITREVERSE, MVT::v32i8, 5 }, 2434 { ISD::BSWAP, MVT::v4i64, 1 }, 2435 { ISD::BSWAP, MVT::v8i32, 1 }, 2436 { ISD::BSWAP, MVT::v16i16, 1 }, 2437 { ISD::CTLZ, MVT::v4i64, 23 }, 2438 { ISD::CTLZ, MVT::v8i32, 18 }, 2439 { ISD::CTLZ, MVT::v16i16, 14 }, 2440 { ISD::CTLZ, MVT::v32i8, 9 }, 2441 { ISD::CTPOP, MVT::v4i64, 7 }, 2442 { ISD::CTPOP, MVT::v8i32, 11 }, 2443 { ISD::CTPOP, MVT::v16i16, 9 }, 2444 { ISD::CTPOP, MVT::v32i8, 6 }, 2445 { ISD::CTTZ, MVT::v4i64, 10 }, 2446 { ISD::CTTZ, MVT::v8i32, 14 }, 2447 { ISD::CTTZ, MVT::v16i16, 12 }, 2448 { ISD::CTTZ, MVT::v32i8, 9 }, 2449 { ISD::SADDSAT, MVT::v16i16, 1 }, 2450 { ISD::SADDSAT, MVT::v32i8, 1 }, 2451 { ISD::SMAX, MVT::v8i32, 1 }, 2452 { ISD::SMAX, MVT::v16i16, 1 }, 2453 { ISD::SMAX, MVT::v32i8, 1 }, 2454 { ISD::SMIN, MVT::v8i32, 1 }, 2455 { ISD::SMIN, MVT::v16i16, 1 }, 2456 { ISD::SMIN, MVT::v32i8, 1 }, 2457 { ISD::SSUBSAT, MVT::v16i16, 1 }, 2458 { ISD::SSUBSAT, MVT::v32i8, 1 }, 2459 { ISD::UADDSAT, MVT::v16i16, 1 }, 2460 { ISD::UADDSAT, MVT::v32i8, 1 }, 2461 { ISD::UADDSAT, MVT::v8i32, 3 }, // not + pminud + paddd 2462 { ISD::UMAX, MVT::v8i32, 1 }, 2463 { ISD::UMAX, MVT::v16i16, 1 }, 2464 { ISD::UMAX, MVT::v32i8, 1 }, 2465 { ISD::UMIN, MVT::v8i32, 1 }, 2466 { ISD::UMIN, MVT::v16i16, 1 }, 2467 { ISD::UMIN, MVT::v32i8, 1 }, 2468 { ISD::USUBSAT, MVT::v16i16, 1 }, 2469 { ISD::USUBSAT, MVT::v32i8, 1 }, 2470 { ISD::USUBSAT, MVT::v8i32, 2 }, // pmaxud + psubd 2471 { ISD::FSQRT, MVT::f32, 7 }, // Haswell from http://www.agner.org/ 2472 { ISD::FSQRT, MVT::v4f32, 7 }, // Haswell from http://www.agner.org/ 2473 { ISD::FSQRT, MVT::v8f32, 14 }, // Haswell from http://www.agner.org/ 2474 { ISD::FSQRT, MVT::f64, 14 }, // Haswell from http://www.agner.org/ 2475 { ISD::FSQRT, MVT::v2f64, 14 }, // Haswell from http://www.agner.org/ 2476 { ISD::FSQRT, MVT::v4f64, 28 }, // Haswell from http://www.agner.org/ 2477 }; 2478 static const CostTblEntry AVX1CostTbl[] = { 2479 { ISD::ABS, MVT::v4i64, 6 }, // VBLENDVPD(X,VPSUBQ(0,X),X) 2480 { ISD::ABS, MVT::v8i32, 3 }, 2481 { ISD::ABS, MVT::v16i16, 3 }, 2482 { ISD::ABS, MVT::v32i8, 3 }, 2483 { ISD::BITREVERSE, MVT::v4i64, 12 }, // 2 x 128-bit Op + extract/insert 2484 { ISD::BITREVERSE, MVT::v8i32, 12 }, // 2 x 128-bit Op + extract/insert 2485 { ISD::BITREVERSE, MVT::v16i16, 12 }, // 2 x 128-bit Op + extract/insert 2486 { ISD::BITREVERSE, MVT::v32i8, 12 }, // 2 x 128-bit Op + extract/insert 2487 { ISD::BSWAP, MVT::v4i64, 4 }, 2488 { ISD::BSWAP, MVT::v8i32, 4 }, 2489 { ISD::BSWAP, MVT::v16i16, 4 }, 2490 { ISD::CTLZ, MVT::v4i64, 48 }, // 2 x 128-bit Op + extract/insert 2491 { ISD::CTLZ, MVT::v8i32, 38 }, // 2 x 128-bit Op + extract/insert 2492 { ISD::CTLZ, MVT::v16i16, 30 }, // 2 x 128-bit Op + extract/insert 2493 { ISD::CTLZ, MVT::v32i8, 20 }, // 2 x 128-bit Op + extract/insert 2494 { ISD::CTPOP, MVT::v4i64, 16 }, // 2 x 128-bit Op + extract/insert 2495 { ISD::CTPOP, MVT::v8i32, 24 }, // 2 x 128-bit Op + extract/insert 2496 { ISD::CTPOP, MVT::v16i16, 20 }, // 2 x 128-bit Op + extract/insert 2497 { ISD::CTPOP, MVT::v32i8, 14 }, // 2 x 128-bit Op + extract/insert 2498 { ISD::CTTZ, MVT::v4i64, 22 }, // 2 x 128-bit Op + extract/insert 2499 { ISD::CTTZ, MVT::v8i32, 30 }, // 2 x 128-bit Op + extract/insert 2500 { ISD::CTTZ, MVT::v16i16, 26 }, // 2 x 128-bit Op + extract/insert 2501 { ISD::CTTZ, MVT::v32i8, 20 }, // 2 x 128-bit Op + extract/insert 2502 { ISD::SADDSAT, MVT::v16i16, 4 }, // 2 x 128-bit Op + extract/insert 2503 { ISD::SADDSAT, MVT::v32i8, 4 }, // 2 x 128-bit Op + extract/insert 2504 { ISD::SMAX, MVT::v8i32, 4 }, // 2 x 128-bit Op + extract/insert 2505 { ISD::SMAX, MVT::v16i16, 4 }, // 2 x 128-bit Op + extract/insert 2506 { ISD::SMAX, MVT::v32i8, 4 }, // 2 x 128-bit Op + extract/insert 2507 { ISD::SMIN, MVT::v8i32, 4 }, // 2 x 128-bit Op + extract/insert 2508 { ISD::SMIN, MVT::v16i16, 4 }, // 2 x 128-bit Op + extract/insert 2509 { ISD::SMIN, MVT::v32i8, 4 }, // 2 x 128-bit Op + extract/insert 2510 { ISD::SSUBSAT, MVT::v16i16, 4 }, // 2 x 128-bit Op + extract/insert 2511 { ISD::SSUBSAT, MVT::v32i8, 4 }, // 2 x 128-bit Op + extract/insert 2512 { ISD::UADDSAT, MVT::v16i16, 4 }, // 2 x 128-bit Op + extract/insert 2513 { ISD::UADDSAT, MVT::v32i8, 4 }, // 2 x 128-bit Op + extract/insert 2514 { ISD::UADDSAT, MVT::v8i32, 8 }, // 2 x 128-bit Op + extract/insert 2515 { ISD::UMAX, MVT::v8i32, 4 }, // 2 x 128-bit Op + extract/insert 2516 { ISD::UMAX, MVT::v16i16, 4 }, // 2 x 128-bit Op + extract/insert 2517 { ISD::UMAX, MVT::v32i8, 4 }, // 2 x 128-bit Op + extract/insert 2518 { ISD::UMIN, MVT::v8i32, 4 }, // 2 x 128-bit Op + extract/insert 2519 { ISD::UMIN, MVT::v16i16, 4 }, // 2 x 128-bit Op + extract/insert 2520 { ISD::UMIN, MVT::v32i8, 4 }, // 2 x 128-bit Op + extract/insert 2521 { ISD::USUBSAT, MVT::v16i16, 4 }, // 2 x 128-bit Op + extract/insert 2522 { ISD::USUBSAT, MVT::v32i8, 4 }, // 2 x 128-bit Op + extract/insert 2523 { ISD::USUBSAT, MVT::v8i32, 6 }, // 2 x 128-bit Op + extract/insert 2524 { ISD::FMAXNUM, MVT::f32, 3 }, 2525 { ISD::FMAXNUM, MVT::v4f32, 3 }, 2526 { ISD::FMAXNUM, MVT::v8f32, 5 }, 2527 { ISD::FMAXNUM, MVT::f64, 3 }, 2528 { ISD::FMAXNUM, MVT::v2f64, 3 }, 2529 { ISD::FMAXNUM, MVT::v4f64, 5 }, 2530 { ISD::FSQRT, MVT::f32, 14 }, // SNB from http://www.agner.org/ 2531 { ISD::FSQRT, MVT::v4f32, 14 }, // SNB from http://www.agner.org/ 2532 { ISD::FSQRT, MVT::v8f32, 28 }, // SNB from http://www.agner.org/ 2533 { ISD::FSQRT, MVT::f64, 21 }, // SNB from http://www.agner.org/ 2534 { ISD::FSQRT, MVT::v2f64, 21 }, // SNB from http://www.agner.org/ 2535 { ISD::FSQRT, MVT::v4f64, 43 }, // SNB from http://www.agner.org/ 2536 }; 2537 static const CostTblEntry GLMCostTbl[] = { 2538 { ISD::FSQRT, MVT::f32, 19 }, // sqrtss 2539 { ISD::FSQRT, MVT::v4f32, 37 }, // sqrtps 2540 { ISD::FSQRT, MVT::f64, 34 }, // sqrtsd 2541 { ISD::FSQRT, MVT::v2f64, 67 }, // sqrtpd 2542 }; 2543 static const CostTblEntry SLMCostTbl[] = { 2544 { ISD::FSQRT, MVT::f32, 20 }, // sqrtss 2545 { ISD::FSQRT, MVT::v4f32, 40 }, // sqrtps 2546 { ISD::FSQRT, MVT::f64, 35 }, // sqrtsd 2547 { ISD::FSQRT, MVT::v2f64, 70 }, // sqrtpd 2548 }; 2549 static const CostTblEntry SSE42CostTbl[] = { 2550 { ISD::ABS, MVT::v2i64, 3 }, // BLENDVPD(X,PSUBQ(0,X),X) 2551 { ISD::USUBSAT, MVT::v4i32, 2 }, // pmaxud + psubd 2552 { ISD::UADDSAT, MVT::v4i32, 3 }, // not + pminud + paddd 2553 { ISD::FSQRT, MVT::f32, 18 }, // Nehalem from http://www.agner.org/ 2554 { ISD::FSQRT, MVT::v4f32, 18 }, // Nehalem from http://www.agner.org/ 2555 }; 2556 static const CostTblEntry SSE41CostTbl[] = { 2557 { ISD::SMAX, MVT::v4i32, 1 }, 2558 { ISD::SMAX, MVT::v16i8, 1 }, 2559 { ISD::SMIN, MVT::v4i32, 1 }, 2560 { ISD::SMIN, MVT::v16i8, 1 }, 2561 { ISD::UMAX, MVT::v4i32, 1 }, 2562 { ISD::UMAX, MVT::v8i16, 1 }, 2563 { ISD::UMIN, MVT::v4i32, 1 }, 2564 { ISD::UMIN, MVT::v8i16, 1 }, 2565 }; 2566 static const CostTblEntry SSSE3CostTbl[] = { 2567 { ISD::ABS, MVT::v4i32, 1 }, 2568 { ISD::ABS, MVT::v8i16, 1 }, 2569 { ISD::ABS, MVT::v16i8, 1 }, 2570 { ISD::BITREVERSE, MVT::v2i64, 5 }, 2571 { ISD::BITREVERSE, MVT::v4i32, 5 }, 2572 { ISD::BITREVERSE, MVT::v8i16, 5 }, 2573 { ISD::BITREVERSE, MVT::v16i8, 5 }, 2574 { ISD::BSWAP, MVT::v2i64, 1 }, 2575 { ISD::BSWAP, MVT::v4i32, 1 }, 2576 { ISD::BSWAP, MVT::v8i16, 1 }, 2577 { ISD::CTLZ, MVT::v2i64, 23 }, 2578 { ISD::CTLZ, MVT::v4i32, 18 }, 2579 { ISD::CTLZ, MVT::v8i16, 14 }, 2580 { ISD::CTLZ, MVT::v16i8, 9 }, 2581 { ISD::CTPOP, MVT::v2i64, 7 }, 2582 { ISD::CTPOP, MVT::v4i32, 11 }, 2583 { ISD::CTPOP, MVT::v8i16, 9 }, 2584 { ISD::CTPOP, MVT::v16i8, 6 }, 2585 { ISD::CTTZ, MVT::v2i64, 10 }, 2586 { ISD::CTTZ, MVT::v4i32, 14 }, 2587 { ISD::CTTZ, MVT::v8i16, 12 }, 2588 { ISD::CTTZ, MVT::v16i8, 9 } 2589 }; 2590 static const CostTblEntry SSE2CostTbl[] = { 2591 { ISD::ABS, MVT::v2i64, 4 }, 2592 { ISD::ABS, MVT::v4i32, 3 }, 2593 { ISD::ABS, MVT::v8i16, 3 }, 2594 { ISD::ABS, MVT::v16i8, 3 }, 2595 { ISD::BITREVERSE, MVT::v2i64, 29 }, 2596 { ISD::BITREVERSE, MVT::v4i32, 27 }, 2597 { ISD::BITREVERSE, MVT::v8i16, 27 }, 2598 { ISD::BITREVERSE, MVT::v16i8, 20 }, 2599 { ISD::BSWAP, MVT::v2i64, 7 }, 2600 { ISD::BSWAP, MVT::v4i32, 7 }, 2601 { ISD::BSWAP, MVT::v8i16, 7 }, 2602 { ISD::CTLZ, MVT::v2i64, 25 }, 2603 { ISD::CTLZ, MVT::v4i32, 26 }, 2604 { ISD::CTLZ, MVT::v8i16, 20 }, 2605 { ISD::CTLZ, MVT::v16i8, 17 }, 2606 { ISD::CTPOP, MVT::v2i64, 12 }, 2607 { ISD::CTPOP, MVT::v4i32, 15 }, 2608 { ISD::CTPOP, MVT::v8i16, 13 }, 2609 { ISD::CTPOP, MVT::v16i8, 10 }, 2610 { ISD::CTTZ, MVT::v2i64, 14 }, 2611 { ISD::CTTZ, MVT::v4i32, 18 }, 2612 { ISD::CTTZ, MVT::v8i16, 16 }, 2613 { ISD::CTTZ, MVT::v16i8, 13 }, 2614 { ISD::SADDSAT, MVT::v8i16, 1 }, 2615 { ISD::SADDSAT, MVT::v16i8, 1 }, 2616 { ISD::SMAX, MVT::v8i16, 1 }, 2617 { ISD::SMIN, MVT::v8i16, 1 }, 2618 { ISD::SSUBSAT, MVT::v8i16, 1 }, 2619 { ISD::SSUBSAT, MVT::v16i8, 1 }, 2620 { ISD::UADDSAT, MVT::v8i16, 1 }, 2621 { ISD::UADDSAT, MVT::v16i8, 1 }, 2622 { ISD::UMAX, MVT::v8i16, 2 }, 2623 { ISD::UMAX, MVT::v16i8, 1 }, 2624 { ISD::UMIN, MVT::v8i16, 2 }, 2625 { ISD::UMIN, MVT::v16i8, 1 }, 2626 { ISD::USUBSAT, MVT::v8i16, 1 }, 2627 { ISD::USUBSAT, MVT::v16i8, 1 }, 2628 { ISD::FMAXNUM, MVT::f64, 4 }, 2629 { ISD::FMAXNUM, MVT::v2f64, 4 }, 2630 { ISD::FSQRT, MVT::f64, 32 }, // Nehalem from http://www.agner.org/ 2631 { ISD::FSQRT, MVT::v2f64, 32 }, // Nehalem from http://www.agner.org/ 2632 }; 2633 static const CostTblEntry SSE1CostTbl[] = { 2634 { ISD::FMAXNUM, MVT::f32, 4 }, 2635 { ISD::FMAXNUM, MVT::v4f32, 4 }, 2636 { ISD::FSQRT, MVT::f32, 28 }, // Pentium III from http://www.agner.org/ 2637 { ISD::FSQRT, MVT::v4f32, 56 }, // Pentium III from http://www.agner.org/ 2638 }; 2639 static const CostTblEntry BMI64CostTbl[] = { // 64-bit targets 2640 { ISD::CTTZ, MVT::i64, 1 }, 2641 }; 2642 static const CostTblEntry BMI32CostTbl[] = { // 32 or 64-bit targets 2643 { ISD::CTTZ, MVT::i32, 1 }, 2644 { ISD::CTTZ, MVT::i16, 1 }, 2645 { ISD::CTTZ, MVT::i8, 1 }, 2646 }; 2647 static const CostTblEntry LZCNT64CostTbl[] = { // 64-bit targets 2648 { ISD::CTLZ, MVT::i64, 1 }, 2649 }; 2650 static const CostTblEntry LZCNT32CostTbl[] = { // 32 or 64-bit targets 2651 { ISD::CTLZ, MVT::i32, 1 }, 2652 { ISD::CTLZ, MVT::i16, 1 }, 2653 { ISD::CTLZ, MVT::i8, 1 }, 2654 }; 2655 static const CostTblEntry POPCNT64CostTbl[] = { // 64-bit targets 2656 { ISD::CTPOP, MVT::i64, 1 }, 2657 }; 2658 static const CostTblEntry POPCNT32CostTbl[] = { // 32 or 64-bit targets 2659 { ISD::CTPOP, MVT::i32, 1 }, 2660 { ISD::CTPOP, MVT::i16, 1 }, 2661 { ISD::CTPOP, MVT::i8, 1 }, 2662 }; 2663 static const CostTblEntry X64CostTbl[] = { // 64-bit targets 2664 { ISD::BITREVERSE, MVT::i64, 14 }, 2665 { ISD::CTLZ, MVT::i64, 4 }, // BSR+XOR or BSR+XOR+CMOV 2666 { ISD::CTTZ, MVT::i64, 3 }, // TEST+BSF+CMOV/BRANCH 2667 { ISD::CTPOP, MVT::i64, 10 }, 2668 { ISD::SADDO, MVT::i64, 1 }, 2669 { ISD::UADDO, MVT::i64, 1 }, 2670 }; 2671 static const CostTblEntry X86CostTbl[] = { // 32 or 64-bit targets 2672 { ISD::BITREVERSE, MVT::i32, 14 }, 2673 { ISD::BITREVERSE, MVT::i16, 14 }, 2674 { ISD::BITREVERSE, MVT::i8, 11 }, 2675 { ISD::CTLZ, MVT::i32, 4 }, // BSR+XOR or BSR+XOR+CMOV 2676 { ISD::CTLZ, MVT::i16, 4 }, // BSR+XOR or BSR+XOR+CMOV 2677 { ISD::CTLZ, MVT::i8, 4 }, // BSR+XOR or BSR+XOR+CMOV 2678 { ISD::CTTZ, MVT::i32, 3 }, // TEST+BSF+CMOV/BRANCH 2679 { ISD::CTTZ, MVT::i16, 3 }, // TEST+BSF+CMOV/BRANCH 2680 { ISD::CTTZ, MVT::i8, 3 }, // TEST+BSF+CMOV/BRANCH 2681 { ISD::CTPOP, MVT::i32, 8 }, 2682 { ISD::CTPOP, MVT::i16, 9 }, 2683 { ISD::CTPOP, MVT::i8, 7 }, 2684 { ISD::SADDO, MVT::i32, 1 }, 2685 { ISD::SADDO, MVT::i16, 1 }, 2686 { ISD::SADDO, MVT::i8, 1 }, 2687 { ISD::UADDO, MVT::i32, 1 }, 2688 { ISD::UADDO, MVT::i16, 1 }, 2689 { ISD::UADDO, MVT::i8, 1 }, 2690 }; 2691 2692 Type *RetTy = ICA.getReturnType(); 2693 Type *OpTy = RetTy; 2694 Intrinsic::ID IID = ICA.getID(); 2695 unsigned ISD = ISD::DELETED_NODE; 2696 switch (IID) { 2697 default: 2698 break; 2699 case Intrinsic::abs: 2700 ISD = ISD::ABS; 2701 break; 2702 case Intrinsic::bitreverse: 2703 ISD = ISD::BITREVERSE; 2704 break; 2705 case Intrinsic::bswap: 2706 ISD = ISD::BSWAP; 2707 break; 2708 case Intrinsic::ctlz: 2709 ISD = ISD::CTLZ; 2710 break; 2711 case Intrinsic::ctpop: 2712 ISD = ISD::CTPOP; 2713 break; 2714 case Intrinsic::cttz: 2715 ISD = ISD::CTTZ; 2716 break; 2717 case Intrinsic::maxnum: 2718 case Intrinsic::minnum: 2719 // FMINNUM has same costs so don't duplicate. 2720 ISD = ISD::FMAXNUM; 2721 break; 2722 case Intrinsic::sadd_sat: 2723 ISD = ISD::SADDSAT; 2724 break; 2725 case Intrinsic::smax: 2726 ISD = ISD::SMAX; 2727 break; 2728 case Intrinsic::smin: 2729 ISD = ISD::SMIN; 2730 break; 2731 case Intrinsic::ssub_sat: 2732 ISD = ISD::SSUBSAT; 2733 break; 2734 case Intrinsic::uadd_sat: 2735 ISD = ISD::UADDSAT; 2736 break; 2737 case Intrinsic::umax: 2738 ISD = ISD::UMAX; 2739 break; 2740 case Intrinsic::umin: 2741 ISD = ISD::UMIN; 2742 break; 2743 case Intrinsic::usub_sat: 2744 ISD = ISD::USUBSAT; 2745 break; 2746 case Intrinsic::sqrt: 2747 ISD = ISD::FSQRT; 2748 break; 2749 case Intrinsic::sadd_with_overflow: 2750 case Intrinsic::ssub_with_overflow: 2751 // SSUBO has same costs so don't duplicate. 2752 ISD = ISD::SADDO; 2753 OpTy = RetTy->getContainedType(0); 2754 break; 2755 case Intrinsic::uadd_with_overflow: 2756 case Intrinsic::usub_with_overflow: 2757 // USUBO has same costs so don't duplicate. 2758 ISD = ISD::UADDO; 2759 OpTy = RetTy->getContainedType(0); 2760 break; 2761 } 2762 2763 if (ISD != ISD::DELETED_NODE) { 2764 // Legalize the type. 2765 std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, OpTy); 2766 MVT MTy = LT.second; 2767 2768 // Attempt to lookup cost. 2769 if (ST->useGLMDivSqrtCosts()) 2770 if (const auto *Entry = CostTableLookup(GLMCostTbl, ISD, MTy)) 2771 return LT.first * Entry->Cost; 2772 2773 if (ST->isSLM()) 2774 if (const auto *Entry = CostTableLookup(SLMCostTbl, ISD, MTy)) 2775 return LT.first * Entry->Cost; 2776 2777 if (ST->hasCDI()) 2778 if (const auto *Entry = CostTableLookup(AVX512CDCostTbl, ISD, MTy)) 2779 return LT.first * Entry->Cost; 2780 2781 if (ST->hasBWI()) 2782 if (const auto *Entry = CostTableLookup(AVX512BWCostTbl, ISD, MTy)) 2783 return LT.first * Entry->Cost; 2784 2785 if (ST->hasAVX512()) 2786 if (const auto *Entry = CostTableLookup(AVX512CostTbl, ISD, MTy)) 2787 return LT.first * Entry->Cost; 2788 2789 if (ST->hasXOP()) 2790 if (const auto *Entry = CostTableLookup(XOPCostTbl, ISD, MTy)) 2791 return LT.first * Entry->Cost; 2792 2793 if (ST->hasAVX2()) 2794 if (const auto *Entry = CostTableLookup(AVX2CostTbl, ISD, MTy)) 2795 return LT.first * Entry->Cost; 2796 2797 if (ST->hasAVX()) 2798 if (const auto *Entry = CostTableLookup(AVX1CostTbl, ISD, MTy)) 2799 return LT.first * Entry->Cost; 2800 2801 if (ST->hasSSE42()) 2802 if (const auto *Entry = CostTableLookup(SSE42CostTbl, ISD, MTy)) 2803 return LT.first * Entry->Cost; 2804 2805 if (ST->hasSSE41()) 2806 if (const auto *Entry = CostTableLookup(SSE41CostTbl, ISD, MTy)) 2807 return LT.first * Entry->Cost; 2808 2809 if (ST->hasSSSE3()) 2810 if (const auto *Entry = CostTableLookup(SSSE3CostTbl, ISD, MTy)) 2811 return LT.first * Entry->Cost; 2812 2813 if (ST->hasSSE2()) 2814 if (const auto *Entry = CostTableLookup(SSE2CostTbl, ISD, MTy)) 2815 return LT.first * Entry->Cost; 2816 2817 if (ST->hasSSE1()) 2818 if (const auto *Entry = CostTableLookup(SSE1CostTbl, ISD, MTy)) 2819 return LT.first * Entry->Cost; 2820 2821 if (ST->hasBMI()) { 2822 if (ST->is64Bit()) 2823 if (const auto *Entry = CostTableLookup(BMI64CostTbl, ISD, MTy)) 2824 return LT.first * Entry->Cost; 2825 2826 if (const auto *Entry = CostTableLookup(BMI32CostTbl, ISD, MTy)) 2827 return LT.first * Entry->Cost; 2828 } 2829 2830 if (ST->hasLZCNT()) { 2831 if (ST->is64Bit()) 2832 if (const auto *Entry = CostTableLookup(LZCNT64CostTbl, ISD, MTy)) 2833 return LT.first * Entry->Cost; 2834 2835 if (const auto *Entry = CostTableLookup(LZCNT32CostTbl, ISD, MTy)) 2836 return LT.first * Entry->Cost; 2837 } 2838 2839 if (ST->hasPOPCNT()) { 2840 if (ST->is64Bit()) 2841 if (const auto *Entry = CostTableLookup(POPCNT64CostTbl, ISD, MTy)) 2842 return LT.first * Entry->Cost; 2843 2844 if (const auto *Entry = CostTableLookup(POPCNT32CostTbl, ISD, MTy)) 2845 return LT.first * Entry->Cost; 2846 } 2847 2848 // TODO - add BMI (TZCNT) scalar handling 2849 2850 if (ST->is64Bit()) 2851 if (const auto *Entry = CostTableLookup(X64CostTbl, ISD, MTy)) 2852 return LT.first * Entry->Cost; 2853 2854 if (const auto *Entry = CostTableLookup(X86CostTbl, ISD, MTy)) 2855 return LT.first * Entry->Cost; 2856 } 2857 2858 return BaseT::getIntrinsicInstrCost(ICA, CostKind); 2859 } 2860 2861 int X86TTIImpl::getIntrinsicInstrCost(const IntrinsicCostAttributes &ICA, 2862 TTI::TargetCostKind CostKind) { 2863 if (CostKind != TTI::TCK_RecipThroughput) 2864 return BaseT::getIntrinsicInstrCost(ICA, CostKind); 2865 2866 if (ICA.isTypeBasedOnly()) 2867 return getTypeBasedIntrinsicInstrCost(ICA, CostKind); 2868 2869 static const CostTblEntry AVX512CostTbl[] = { 2870 { ISD::ROTL, MVT::v8i64, 1 }, 2871 { ISD::ROTL, MVT::v4i64, 1 }, 2872 { ISD::ROTL, MVT::v2i64, 1 }, 2873 { ISD::ROTL, MVT::v16i32, 1 }, 2874 { ISD::ROTL, MVT::v8i32, 1 }, 2875 { ISD::ROTL, MVT::v4i32, 1 }, 2876 { ISD::ROTR, MVT::v8i64, 1 }, 2877 { ISD::ROTR, MVT::v4i64, 1 }, 2878 { ISD::ROTR, MVT::v2i64, 1 }, 2879 { ISD::ROTR, MVT::v16i32, 1 }, 2880 { ISD::ROTR, MVT::v8i32, 1 }, 2881 { ISD::ROTR, MVT::v4i32, 1 } 2882 }; 2883 // XOP: ROTL = VPROT(X,Y), ROTR = VPROT(X,SUB(0,Y)) 2884 static const CostTblEntry XOPCostTbl[] = { 2885 { ISD::ROTL, MVT::v4i64, 4 }, 2886 { ISD::ROTL, MVT::v8i32, 4 }, 2887 { ISD::ROTL, MVT::v16i16, 4 }, 2888 { ISD::ROTL, MVT::v32i8, 4 }, 2889 { ISD::ROTL, MVT::v2i64, 1 }, 2890 { ISD::ROTL, MVT::v4i32, 1 }, 2891 { ISD::ROTL, MVT::v8i16, 1 }, 2892 { ISD::ROTL, MVT::v16i8, 1 }, 2893 { ISD::ROTR, MVT::v4i64, 6 }, 2894 { ISD::ROTR, MVT::v8i32, 6 }, 2895 { ISD::ROTR, MVT::v16i16, 6 }, 2896 { ISD::ROTR, MVT::v32i8, 6 }, 2897 { ISD::ROTR, MVT::v2i64, 2 }, 2898 { ISD::ROTR, MVT::v4i32, 2 }, 2899 { ISD::ROTR, MVT::v8i16, 2 }, 2900 { ISD::ROTR, MVT::v16i8, 2 } 2901 }; 2902 static const CostTblEntry X64CostTbl[] = { // 64-bit targets 2903 { ISD::ROTL, MVT::i64, 1 }, 2904 { ISD::ROTR, MVT::i64, 1 }, 2905 { ISD::FSHL, MVT::i64, 4 } 2906 }; 2907 static const CostTblEntry X86CostTbl[] = { // 32 or 64-bit targets 2908 { ISD::ROTL, MVT::i32, 1 }, 2909 { ISD::ROTL, MVT::i16, 1 }, 2910 { ISD::ROTL, MVT::i8, 1 }, 2911 { ISD::ROTR, MVT::i32, 1 }, 2912 { ISD::ROTR, MVT::i16, 1 }, 2913 { ISD::ROTR, MVT::i8, 1 }, 2914 { ISD::FSHL, MVT::i32, 4 }, 2915 { ISD::FSHL, MVT::i16, 4 }, 2916 { ISD::FSHL, MVT::i8, 4 } 2917 }; 2918 2919 Intrinsic::ID IID = ICA.getID(); 2920 Type *RetTy = ICA.getReturnType(); 2921 const SmallVectorImpl<const Value *> &Args = ICA.getArgs(); 2922 unsigned ISD = ISD::DELETED_NODE; 2923 switch (IID) { 2924 default: 2925 break; 2926 case Intrinsic::fshl: 2927 ISD = ISD::FSHL; 2928 if (Args[0] == Args[1]) 2929 ISD = ISD::ROTL; 2930 break; 2931 case Intrinsic::fshr: 2932 // FSHR has same costs so don't duplicate. 2933 ISD = ISD::FSHL; 2934 if (Args[0] == Args[1]) 2935 ISD = ISD::ROTR; 2936 break; 2937 } 2938 2939 if (ISD != ISD::DELETED_NODE) { 2940 // Legalize the type. 2941 std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, RetTy); 2942 MVT MTy = LT.second; 2943 2944 // Attempt to lookup cost. 2945 if (ST->hasAVX512()) 2946 if (const auto *Entry = CostTableLookup(AVX512CostTbl, ISD, MTy)) 2947 return LT.first * Entry->Cost; 2948 2949 if (ST->hasXOP()) 2950 if (const auto *Entry = CostTableLookup(XOPCostTbl, ISD, MTy)) 2951 return LT.first * Entry->Cost; 2952 2953 if (ST->is64Bit()) 2954 if (const auto *Entry = CostTableLookup(X64CostTbl, ISD, MTy)) 2955 return LT.first * Entry->Cost; 2956 2957 if (const auto *Entry = CostTableLookup(X86CostTbl, ISD, MTy)) 2958 return LT.first * Entry->Cost; 2959 } 2960 2961 return BaseT::getIntrinsicInstrCost(ICA, CostKind); 2962 } 2963 2964 int X86TTIImpl::getVectorInstrCost(unsigned Opcode, Type *Val, unsigned Index) { 2965 static const CostTblEntry SLMCostTbl[] = { 2966 { ISD::EXTRACT_VECTOR_ELT, MVT::i8, 4 }, 2967 { ISD::EXTRACT_VECTOR_ELT, MVT::i16, 4 }, 2968 { ISD::EXTRACT_VECTOR_ELT, MVT::i32, 4 }, 2969 { ISD::EXTRACT_VECTOR_ELT, MVT::i64, 7 } 2970 }; 2971 2972 assert(Val->isVectorTy() && "This must be a vector type"); 2973 Type *ScalarType = Val->getScalarType(); 2974 int RegisterFileMoveCost = 0; 2975 2976 if (Index != -1U && (Opcode == Instruction::ExtractElement || 2977 Opcode == Instruction::InsertElement)) { 2978 // Legalize the type. 2979 std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, Val); 2980 2981 // This type is legalized to a scalar type. 2982 if (!LT.second.isVector()) 2983 return 0; 2984 2985 // The type may be split. Normalize the index to the new type. 2986 unsigned NumElts = LT.second.getVectorNumElements(); 2987 unsigned SubNumElts = NumElts; 2988 Index = Index % NumElts; 2989 2990 // For >128-bit vectors, we need to extract higher 128-bit subvectors. 2991 // For inserts, we also need to insert the subvector back. 2992 if (LT.second.getSizeInBits() > 128) { 2993 assert((LT.second.getSizeInBits() % 128) == 0 && "Illegal vector"); 2994 unsigned NumSubVecs = LT.second.getSizeInBits() / 128; 2995 SubNumElts = NumElts / NumSubVecs; 2996 if (SubNumElts <= Index) { 2997 RegisterFileMoveCost += (Opcode == Instruction::InsertElement ? 2 : 1); 2998 Index %= SubNumElts; 2999 } 3000 } 3001 3002 if (Index == 0) { 3003 // Floating point scalars are already located in index #0. 3004 // Many insertions to #0 can fold away for scalar fp-ops, so let's assume 3005 // true for all. 3006 if (ScalarType->isFloatingPointTy()) 3007 return RegisterFileMoveCost; 3008 3009 // Assume movd/movq XMM -> GPR is relatively cheap on all targets. 3010 if (ScalarType->isIntegerTy() && Opcode == Instruction::ExtractElement) 3011 return 1 + RegisterFileMoveCost; 3012 } 3013 3014 int ISD = TLI->InstructionOpcodeToISD(Opcode); 3015 assert(ISD && "Unexpected vector opcode"); 3016 MVT MScalarTy = LT.second.getScalarType(); 3017 if (ST->isSLM()) 3018 if (auto *Entry = CostTableLookup(SLMCostTbl, ISD, MScalarTy)) 3019 return Entry->Cost + RegisterFileMoveCost; 3020 3021 // Assume pinsr/pextr XMM <-> GPR is relatively cheap on all targets. 3022 if ((MScalarTy == MVT::i16 && ST->hasSSE2()) || 3023 (MScalarTy.isInteger() && ST->hasSSE41())) 3024 return 1 + RegisterFileMoveCost; 3025 3026 // Assume insertps is relatively cheap on all targets. 3027 if (MScalarTy == MVT::f32 && ST->hasSSE41() && 3028 Opcode == Instruction::InsertElement) 3029 return 1 + RegisterFileMoveCost; 3030 3031 // For extractions we just need to shuffle the element to index 0, which 3032 // should be very cheap (assume cost = 1). For insertions we need to shuffle 3033 // the elements to its destination. In both cases we must handle the 3034 // subvector move(s). 3035 // If the vector type is already less than 128-bits then don't reduce it. 3036 // TODO: Under what circumstances should we shuffle using the full width? 3037 int ShuffleCost = 1; 3038 if (Opcode == Instruction::InsertElement) { 3039 auto *SubTy = cast<VectorType>(Val); 3040 EVT VT = TLI->getValueType(DL, Val); 3041 if (VT.getScalarType() != MScalarTy || VT.getSizeInBits() >= 128) 3042 SubTy = FixedVectorType::get(ScalarType, SubNumElts); 3043 ShuffleCost = getShuffleCost(TTI::SK_PermuteTwoSrc, SubTy, 0, SubTy); 3044 } 3045 int IntOrFpCost = ScalarType->isFloatingPointTy() ? 0 : 1; 3046 return ShuffleCost + IntOrFpCost + RegisterFileMoveCost; 3047 } 3048 3049 // Add to the base cost if we know that the extracted element of a vector is 3050 // destined to be moved to and used in the integer register file. 3051 if (Opcode == Instruction::ExtractElement && ScalarType->isPointerTy()) 3052 RegisterFileMoveCost += 1; 3053 3054 return BaseT::getVectorInstrCost(Opcode, Val, Index) + RegisterFileMoveCost; 3055 } 3056 3057 unsigned X86TTIImpl::getScalarizationOverhead(VectorType *Ty, 3058 const APInt &DemandedElts, 3059 bool Insert, bool Extract) { 3060 unsigned Cost = 0; 3061 3062 // For insertions, a ISD::BUILD_VECTOR style vector initialization can be much 3063 // cheaper than an accumulation of ISD::INSERT_VECTOR_ELT. 3064 if (Insert) { 3065 std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, Ty); 3066 MVT MScalarTy = LT.second.getScalarType(); 3067 3068 if ((MScalarTy == MVT::i16 && ST->hasSSE2()) || 3069 (MScalarTy.isInteger() && ST->hasSSE41()) || 3070 (MScalarTy == MVT::f32 && ST->hasSSE41())) { 3071 // For types we can insert directly, insertion into 128-bit sub vectors is 3072 // cheap, followed by a cheap chain of concatenations. 3073 if (LT.second.getSizeInBits() <= 128) { 3074 Cost += 3075 BaseT::getScalarizationOverhead(Ty, DemandedElts, Insert, false); 3076 } else { 3077 // In each 128-lane, if at least one index is demanded but not all 3078 // indices are demanded and this 128-lane is not the first 128-lane of 3079 // the legalized-vector, then this 128-lane needs a extracti128; If in 3080 // each 128-lane, there is at least one demanded index, this 128-lane 3081 // needs a inserti128. 3082 3083 // The following cases will help you build a better understanding: 3084 // Assume we insert several elements into a v8i32 vector in avx2, 3085 // Case#1: inserting into 1th index needs vpinsrd + inserti128. 3086 // Case#2: inserting into 5th index needs extracti128 + vpinsrd + 3087 // inserti128. 3088 // Case#3: inserting into 4,5,6,7 index needs 4*vpinsrd + inserti128. 3089 unsigned Num128Lanes = LT.second.getSizeInBits() / 128 * LT.first; 3090 unsigned NumElts = LT.second.getVectorNumElements() * LT.first; 3091 APInt WidenedDemandedElts = DemandedElts.zextOrSelf(NumElts); 3092 unsigned Scale = NumElts / Num128Lanes; 3093 // We iterate each 128-lane, and check if we need a 3094 // extracti128/inserti128 for this 128-lane. 3095 for (unsigned I = 0; I < NumElts; I += Scale) { 3096 APInt Mask = WidenedDemandedElts.getBitsSet(NumElts, I, I + Scale); 3097 APInt MaskedDE = Mask & WidenedDemandedElts; 3098 unsigned Population = MaskedDE.countPopulation(); 3099 Cost += (Population > 0 && Population != Scale && 3100 I % LT.second.getVectorNumElements() != 0); 3101 Cost += Population > 0; 3102 } 3103 Cost += DemandedElts.countPopulation(); 3104 3105 // For vXf32 cases, insertion into the 0'th index in each v4f32 3106 // 128-bit vector is free. 3107 // NOTE: This assumes legalization widens vXf32 vectors. 3108 if (MScalarTy == MVT::f32) 3109 for (unsigned i = 0, e = cast<FixedVectorType>(Ty)->getNumElements(); 3110 i < e; i += 4) 3111 if (DemandedElts[i]) 3112 Cost--; 3113 } 3114 } else if (LT.second.isVector()) { 3115 // Without fast insertion, we need to use MOVD/MOVQ to pass each demanded 3116 // integer element as a SCALAR_TO_VECTOR, then we build the vector as a 3117 // series of UNPCK followed by CONCAT_VECTORS - all of these can be 3118 // considered cheap. 3119 if (Ty->isIntOrIntVectorTy()) 3120 Cost += DemandedElts.countPopulation(); 3121 3122 // Get the smaller of the legalized or original pow2-extended number of 3123 // vector elements, which represents the number of unpacks we'll end up 3124 // performing. 3125 unsigned NumElts = LT.second.getVectorNumElements(); 3126 unsigned Pow2Elts = 3127 PowerOf2Ceil(cast<FixedVectorType>(Ty)->getNumElements()); 3128 Cost += (std::min<unsigned>(NumElts, Pow2Elts) - 1) * LT.first; 3129 } 3130 } 3131 3132 // TODO: Use default extraction for now, but we should investigate extending this 3133 // to handle repeated subvector extraction. 3134 if (Extract) 3135 Cost += BaseT::getScalarizationOverhead(Ty, DemandedElts, false, Extract); 3136 3137 return Cost; 3138 } 3139 3140 int X86TTIImpl::getMemoryOpCost(unsigned Opcode, Type *Src, 3141 MaybeAlign Alignment, unsigned AddressSpace, 3142 TTI::TargetCostKind CostKind, 3143 const Instruction *I) { 3144 // TODO: Handle other cost kinds. 3145 if (CostKind != TTI::TCK_RecipThroughput) { 3146 if (isa_and_nonnull<StoreInst>(I)) { 3147 Value *Ptr = I->getOperand(1); 3148 // Store instruction with index and scale costs 2 Uops. 3149 // Check the preceding GEP to identify non-const indices. 3150 if (auto *GEP = dyn_cast<GetElementPtrInst>(Ptr)) { 3151 if (!all_of(GEP->indices(), [](Value *V) { return isa<Constant>(V); })) 3152 return TTI::TCC_Basic * 2; 3153 } 3154 } 3155 return TTI::TCC_Basic; 3156 } 3157 3158 // Handle non-power-of-two vectors such as <3 x float> 3159 if (auto *VTy = dyn_cast<FixedVectorType>(Src)) { 3160 unsigned NumElem = VTy->getNumElements(); 3161 3162 // Handle a few common cases: 3163 // <3 x float> 3164 if (NumElem == 3 && VTy->getScalarSizeInBits() == 32) 3165 // Cost = 64 bit store + extract + 32 bit store. 3166 return 3; 3167 3168 // <3 x double> 3169 if (NumElem == 3 && VTy->getScalarSizeInBits() == 64) 3170 // Cost = 128 bit store + unpack + 64 bit store. 3171 return 3; 3172 3173 // Assume that all other non-power-of-two numbers are scalarized. 3174 if (!isPowerOf2_32(NumElem)) { 3175 APInt DemandedElts = APInt::getAllOnesValue(NumElem); 3176 int Cost = BaseT::getMemoryOpCost(Opcode, VTy->getScalarType(), Alignment, 3177 AddressSpace, CostKind); 3178 int SplitCost = getScalarizationOverhead(VTy, DemandedElts, 3179 Opcode == Instruction::Load, 3180 Opcode == Instruction::Store); 3181 return NumElem * Cost + SplitCost; 3182 } 3183 } 3184 3185 // Type legalization can't handle structs 3186 if (TLI->getValueType(DL, Src, true) == MVT::Other) 3187 return BaseT::getMemoryOpCost(Opcode, Src, Alignment, AddressSpace, 3188 CostKind); 3189 3190 // Legalize the type. 3191 std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, Src); 3192 assert((Opcode == Instruction::Load || Opcode == Instruction::Store) && 3193 "Invalid Opcode"); 3194 3195 // Each load/store unit costs 1. 3196 int Cost = LT.first * 1; 3197 3198 // This isn't exactly right. We're using slow unaligned 32-byte accesses as a 3199 // proxy for a double-pumped AVX memory interface such as on Sandybridge. 3200 if (LT.second.getStoreSize() == 32 && ST->isUnalignedMem32Slow()) 3201 Cost *= 2; 3202 3203 return Cost; 3204 } 3205 3206 int X86TTIImpl::getMaskedMemoryOpCost(unsigned Opcode, Type *SrcTy, 3207 Align Alignment, unsigned AddressSpace, 3208 TTI::TargetCostKind CostKind) { 3209 bool IsLoad = (Instruction::Load == Opcode); 3210 bool IsStore = (Instruction::Store == Opcode); 3211 3212 auto *SrcVTy = dyn_cast<FixedVectorType>(SrcTy); 3213 if (!SrcVTy) 3214 // To calculate scalar take the regular cost, without mask 3215 return getMemoryOpCost(Opcode, SrcTy, Alignment, AddressSpace, CostKind); 3216 3217 unsigned NumElem = SrcVTy->getNumElements(); 3218 auto *MaskTy = 3219 FixedVectorType::get(Type::getInt8Ty(SrcVTy->getContext()), NumElem); 3220 if ((IsLoad && !isLegalMaskedLoad(SrcVTy, Alignment)) || 3221 (IsStore && !isLegalMaskedStore(SrcVTy, Alignment)) || 3222 !isPowerOf2_32(NumElem)) { 3223 // Scalarization 3224 APInt DemandedElts = APInt::getAllOnesValue(NumElem); 3225 int MaskSplitCost = 3226 getScalarizationOverhead(MaskTy, DemandedElts, false, true); 3227 int ScalarCompareCost = getCmpSelInstrCost( 3228 Instruction::ICmp, Type::getInt8Ty(SrcVTy->getContext()), nullptr, 3229 CostKind); 3230 int BranchCost = getCFInstrCost(Instruction::Br, CostKind); 3231 int MaskCmpCost = NumElem * (BranchCost + ScalarCompareCost); 3232 int ValueSplitCost = 3233 getScalarizationOverhead(SrcVTy, DemandedElts, IsLoad, IsStore); 3234 int MemopCost = 3235 NumElem * BaseT::getMemoryOpCost(Opcode, SrcVTy->getScalarType(), 3236 Alignment, AddressSpace, CostKind); 3237 return MemopCost + ValueSplitCost + MaskSplitCost + MaskCmpCost; 3238 } 3239 3240 // Legalize the type. 3241 std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, SrcVTy); 3242 auto VT = TLI->getValueType(DL, SrcVTy); 3243 int Cost = 0; 3244 if (VT.isSimple() && LT.second != VT.getSimpleVT() && 3245 LT.second.getVectorNumElements() == NumElem) 3246 // Promotion requires expand/truncate for data and a shuffle for mask. 3247 Cost += getShuffleCost(TTI::SK_PermuteTwoSrc, SrcVTy, 0, nullptr) + 3248 getShuffleCost(TTI::SK_PermuteTwoSrc, MaskTy, 0, nullptr); 3249 3250 else if (LT.second.getVectorNumElements() > NumElem) { 3251 auto *NewMaskTy = FixedVectorType::get(MaskTy->getElementType(), 3252 LT.second.getVectorNumElements()); 3253 // Expanding requires fill mask with zeroes 3254 Cost += getShuffleCost(TTI::SK_InsertSubvector, NewMaskTy, 0, MaskTy); 3255 } 3256 3257 // Pre-AVX512 - each maskmov load costs 2 + store costs ~8. 3258 if (!ST->hasAVX512()) 3259 return Cost + LT.first * (IsLoad ? 2 : 8); 3260 3261 // AVX-512 masked load/store is cheapper 3262 return Cost + LT.first; 3263 } 3264 3265 int X86TTIImpl::getAddressComputationCost(Type *Ty, ScalarEvolution *SE, 3266 const SCEV *Ptr) { 3267 // Address computations in vectorized code with non-consecutive addresses will 3268 // likely result in more instructions compared to scalar code where the 3269 // computation can more often be merged into the index mode. The resulting 3270 // extra micro-ops can significantly decrease throughput. 3271 const unsigned NumVectorInstToHideOverhead = 10; 3272 3273 // Cost modeling of Strided Access Computation is hidden by the indexing 3274 // modes of X86 regardless of the stride value. We dont believe that there 3275 // is a difference between constant strided access in gerenal and constant 3276 // strided value which is less than or equal to 64. 3277 // Even in the case of (loop invariant) stride whose value is not known at 3278 // compile time, the address computation will not incur more than one extra 3279 // ADD instruction. 3280 if (Ty->isVectorTy() && SE) { 3281 if (!BaseT::isStridedAccess(Ptr)) 3282 return NumVectorInstToHideOverhead; 3283 if (!BaseT::getConstantStrideStep(SE, Ptr)) 3284 return 1; 3285 } 3286 3287 return BaseT::getAddressComputationCost(Ty, SE, Ptr); 3288 } 3289 3290 int X86TTIImpl::getArithmeticReductionCost(unsigned Opcode, VectorType *ValTy, 3291 bool IsPairwise, 3292 TTI::TargetCostKind CostKind) { 3293 // Just use the default implementation for pair reductions. 3294 if (IsPairwise) 3295 return BaseT::getArithmeticReductionCost(Opcode, ValTy, IsPairwise, CostKind); 3296 3297 // We use the Intel Architecture Code Analyzer(IACA) to measure the throughput 3298 // and make it as the cost. 3299 3300 static const CostTblEntry SLMCostTblNoPairWise[] = { 3301 { ISD::FADD, MVT::v2f64, 3 }, 3302 { ISD::ADD, MVT::v2i64, 5 }, 3303 }; 3304 3305 static const CostTblEntry SSE2CostTblNoPairWise[] = { 3306 { ISD::FADD, MVT::v2f64, 2 }, 3307 { ISD::FADD, MVT::v4f32, 4 }, 3308 { ISD::ADD, MVT::v2i64, 2 }, // The data reported by the IACA tool is "1.6". 3309 { ISD::ADD, MVT::v2i32, 2 }, // FIXME: chosen to be less than v4i32 3310 { ISD::ADD, MVT::v4i32, 3 }, // The data reported by the IACA tool is "3.3". 3311 { ISD::ADD, MVT::v2i16, 2 }, // The data reported by the IACA tool is "4.3". 3312 { ISD::ADD, MVT::v4i16, 3 }, // The data reported by the IACA tool is "4.3". 3313 { ISD::ADD, MVT::v8i16, 4 }, // The data reported by the IACA tool is "4.3". 3314 { ISD::ADD, MVT::v2i8, 2 }, 3315 { ISD::ADD, MVT::v4i8, 2 }, 3316 { ISD::ADD, MVT::v8i8, 2 }, 3317 { ISD::ADD, MVT::v16i8, 3 }, 3318 }; 3319 3320 static const CostTblEntry AVX1CostTblNoPairWise[] = { 3321 { ISD::FADD, MVT::v4f64, 3 }, 3322 { ISD::FADD, MVT::v4f32, 3 }, 3323 { ISD::FADD, MVT::v8f32, 4 }, 3324 { ISD::ADD, MVT::v2i64, 1 }, // The data reported by the IACA tool is "1.5". 3325 { ISD::ADD, MVT::v4i64, 3 }, 3326 { ISD::ADD, MVT::v8i32, 5 }, 3327 { ISD::ADD, MVT::v16i16, 5 }, 3328 { ISD::ADD, MVT::v32i8, 4 }, 3329 }; 3330 3331 int ISD = TLI->InstructionOpcodeToISD(Opcode); 3332 assert(ISD && "Invalid opcode"); 3333 3334 // Before legalizing the type, give a chance to look up illegal narrow types 3335 // in the table. 3336 // FIXME: Is there a better way to do this? 3337 EVT VT = TLI->getValueType(DL, ValTy); 3338 if (VT.isSimple()) { 3339 MVT MTy = VT.getSimpleVT(); 3340 if (ST->isSLM()) 3341 if (const auto *Entry = CostTableLookup(SLMCostTblNoPairWise, ISD, MTy)) 3342 return Entry->Cost; 3343 3344 if (ST->hasAVX()) 3345 if (const auto *Entry = CostTableLookup(AVX1CostTblNoPairWise, ISD, MTy)) 3346 return Entry->Cost; 3347 3348 if (ST->hasSSE2()) 3349 if (const auto *Entry = CostTableLookup(SSE2CostTblNoPairWise, ISD, MTy)) 3350 return Entry->Cost; 3351 } 3352 3353 std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, ValTy); 3354 3355 MVT MTy = LT.second; 3356 3357 auto *ValVTy = cast<FixedVectorType>(ValTy); 3358 3359 unsigned ArithmeticCost = 0; 3360 if (LT.first != 1 && MTy.isVector() && 3361 MTy.getVectorNumElements() < ValVTy->getNumElements()) { 3362 // Type needs to be split. We need LT.first - 1 arithmetic ops. 3363 auto *SingleOpTy = FixedVectorType::get(ValVTy->getElementType(), 3364 MTy.getVectorNumElements()); 3365 ArithmeticCost = getArithmeticInstrCost(Opcode, SingleOpTy, CostKind); 3366 ArithmeticCost *= LT.first - 1; 3367 } 3368 3369 if (ST->isSLM()) 3370 if (const auto *Entry = CostTableLookup(SLMCostTblNoPairWise, ISD, MTy)) 3371 return ArithmeticCost + Entry->Cost; 3372 3373 if (ST->hasAVX()) 3374 if (const auto *Entry = CostTableLookup(AVX1CostTblNoPairWise, ISD, MTy)) 3375 return ArithmeticCost + Entry->Cost; 3376 3377 if (ST->hasSSE2()) 3378 if (const auto *Entry = CostTableLookup(SSE2CostTblNoPairWise, ISD, MTy)) 3379 return ArithmeticCost + Entry->Cost; 3380 3381 // FIXME: These assume a naive kshift+binop lowering, which is probably 3382 // conservative in most cases. 3383 static const CostTblEntry AVX512BoolReduction[] = { 3384 { ISD::AND, MVT::v2i1, 3 }, 3385 { ISD::AND, MVT::v4i1, 5 }, 3386 { ISD::AND, MVT::v8i1, 7 }, 3387 { ISD::AND, MVT::v16i1, 9 }, 3388 { ISD::AND, MVT::v32i1, 11 }, 3389 { ISD::AND, MVT::v64i1, 13 }, 3390 { ISD::OR, MVT::v2i1, 3 }, 3391 { ISD::OR, MVT::v4i1, 5 }, 3392 { ISD::OR, MVT::v8i1, 7 }, 3393 { ISD::OR, MVT::v16i1, 9 }, 3394 { ISD::OR, MVT::v32i1, 11 }, 3395 { ISD::OR, MVT::v64i1, 13 }, 3396 }; 3397 3398 static const CostTblEntry AVX2BoolReduction[] = { 3399 { ISD::AND, MVT::v16i16, 2 }, // vpmovmskb + cmp 3400 { ISD::AND, MVT::v32i8, 2 }, // vpmovmskb + cmp 3401 { ISD::OR, MVT::v16i16, 2 }, // vpmovmskb + cmp 3402 { ISD::OR, MVT::v32i8, 2 }, // vpmovmskb + cmp 3403 }; 3404 3405 static const CostTblEntry AVX1BoolReduction[] = { 3406 { ISD::AND, MVT::v4i64, 2 }, // vmovmskpd + cmp 3407 { ISD::AND, MVT::v8i32, 2 }, // vmovmskps + cmp 3408 { ISD::AND, MVT::v16i16, 4 }, // vextractf128 + vpand + vpmovmskb + cmp 3409 { ISD::AND, MVT::v32i8, 4 }, // vextractf128 + vpand + vpmovmskb + cmp 3410 { ISD::OR, MVT::v4i64, 2 }, // vmovmskpd + cmp 3411 { ISD::OR, MVT::v8i32, 2 }, // vmovmskps + cmp 3412 { ISD::OR, MVT::v16i16, 4 }, // vextractf128 + vpor + vpmovmskb + cmp 3413 { ISD::OR, MVT::v32i8, 4 }, // vextractf128 + vpor + vpmovmskb + cmp 3414 }; 3415 3416 static const CostTblEntry SSE2BoolReduction[] = { 3417 { ISD::AND, MVT::v2i64, 2 }, // movmskpd + cmp 3418 { ISD::AND, MVT::v4i32, 2 }, // movmskps + cmp 3419 { ISD::AND, MVT::v8i16, 2 }, // pmovmskb + cmp 3420 { ISD::AND, MVT::v16i8, 2 }, // pmovmskb + cmp 3421 { ISD::OR, MVT::v2i64, 2 }, // movmskpd + cmp 3422 { ISD::OR, MVT::v4i32, 2 }, // movmskps + cmp 3423 { ISD::OR, MVT::v8i16, 2 }, // pmovmskb + cmp 3424 { ISD::OR, MVT::v16i8, 2 }, // pmovmskb + cmp 3425 }; 3426 3427 // Handle bool allof/anyof patterns. 3428 if (ValVTy->getElementType()->isIntegerTy(1)) { 3429 unsigned ArithmeticCost = 0; 3430 if (LT.first != 1 && MTy.isVector() && 3431 MTy.getVectorNumElements() < ValVTy->getNumElements()) { 3432 // Type needs to be split. We need LT.first - 1 arithmetic ops. 3433 auto *SingleOpTy = FixedVectorType::get(ValVTy->getElementType(), 3434 MTy.getVectorNumElements()); 3435 ArithmeticCost = getArithmeticInstrCost(Opcode, SingleOpTy, CostKind); 3436 ArithmeticCost *= LT.first - 1; 3437 } 3438 3439 if (ST->hasAVX512()) 3440 if (const auto *Entry = CostTableLookup(AVX512BoolReduction, ISD, MTy)) 3441 return ArithmeticCost + Entry->Cost; 3442 if (ST->hasAVX2()) 3443 if (const auto *Entry = CostTableLookup(AVX2BoolReduction, ISD, MTy)) 3444 return ArithmeticCost + Entry->Cost; 3445 if (ST->hasAVX()) 3446 if (const auto *Entry = CostTableLookup(AVX1BoolReduction, ISD, MTy)) 3447 return ArithmeticCost + Entry->Cost; 3448 if (ST->hasSSE2()) 3449 if (const auto *Entry = CostTableLookup(SSE2BoolReduction, ISD, MTy)) 3450 return ArithmeticCost + Entry->Cost; 3451 3452 return BaseT::getArithmeticReductionCost(Opcode, ValVTy, IsPairwise, 3453 CostKind); 3454 } 3455 3456 unsigned NumVecElts = ValVTy->getNumElements(); 3457 unsigned ScalarSize = ValVTy->getScalarSizeInBits(); 3458 3459 // Special case power of 2 reductions where the scalar type isn't changed 3460 // by type legalization. 3461 if (!isPowerOf2_32(NumVecElts) || ScalarSize != MTy.getScalarSizeInBits()) 3462 return BaseT::getArithmeticReductionCost(Opcode, ValVTy, IsPairwise, 3463 CostKind); 3464 3465 unsigned ReductionCost = 0; 3466 3467 auto *Ty = ValVTy; 3468 if (LT.first != 1 && MTy.isVector() && 3469 MTy.getVectorNumElements() < ValVTy->getNumElements()) { 3470 // Type needs to be split. We need LT.first - 1 arithmetic ops. 3471 Ty = FixedVectorType::get(ValVTy->getElementType(), 3472 MTy.getVectorNumElements()); 3473 ReductionCost = getArithmeticInstrCost(Opcode, Ty, CostKind); 3474 ReductionCost *= LT.first - 1; 3475 NumVecElts = MTy.getVectorNumElements(); 3476 } 3477 3478 // Now handle reduction with the legal type, taking into account size changes 3479 // at each level. 3480 while (NumVecElts > 1) { 3481 // Determine the size of the remaining vector we need to reduce. 3482 unsigned Size = NumVecElts * ScalarSize; 3483 NumVecElts /= 2; 3484 // If we're reducing from 256/512 bits, use an extract_subvector. 3485 if (Size > 128) { 3486 auto *SubTy = FixedVectorType::get(ValVTy->getElementType(), NumVecElts); 3487 ReductionCost += 3488 getShuffleCost(TTI::SK_ExtractSubvector, Ty, NumVecElts, SubTy); 3489 Ty = SubTy; 3490 } else if (Size == 128) { 3491 // Reducing from 128 bits is a permute of v2f64/v2i64. 3492 FixedVectorType *ShufTy; 3493 if (ValVTy->isFloatingPointTy()) 3494 ShufTy = 3495 FixedVectorType::get(Type::getDoubleTy(ValVTy->getContext()), 2); 3496 else 3497 ShufTy = 3498 FixedVectorType::get(Type::getInt64Ty(ValVTy->getContext()), 2); 3499 ReductionCost += 3500 getShuffleCost(TTI::SK_PermuteSingleSrc, ShufTy, 0, nullptr); 3501 } else if (Size == 64) { 3502 // Reducing from 64 bits is a shuffle of v4f32/v4i32. 3503 FixedVectorType *ShufTy; 3504 if (ValVTy->isFloatingPointTy()) 3505 ShufTy = 3506 FixedVectorType::get(Type::getFloatTy(ValVTy->getContext()), 4); 3507 else 3508 ShufTy = 3509 FixedVectorType::get(Type::getInt32Ty(ValVTy->getContext()), 4); 3510 ReductionCost += 3511 getShuffleCost(TTI::SK_PermuteSingleSrc, ShufTy, 0, nullptr); 3512 } else { 3513 // Reducing from smaller size is a shift by immediate. 3514 auto *ShiftTy = FixedVectorType::get( 3515 Type::getIntNTy(ValVTy->getContext(), Size), 128 / Size); 3516 ReductionCost += getArithmeticInstrCost( 3517 Instruction::LShr, ShiftTy, CostKind, 3518 TargetTransformInfo::OK_AnyValue, 3519 TargetTransformInfo::OK_UniformConstantValue, 3520 TargetTransformInfo::OP_None, TargetTransformInfo::OP_None); 3521 } 3522 3523 // Add the arithmetic op for this level. 3524 ReductionCost += getArithmeticInstrCost(Opcode, Ty, CostKind); 3525 } 3526 3527 // Add the final extract element to the cost. 3528 return ReductionCost + getVectorInstrCost(Instruction::ExtractElement, Ty, 0); 3529 } 3530 3531 int X86TTIImpl::getMinMaxCost(Type *Ty, Type *CondTy, bool IsUnsigned) { 3532 std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, Ty); 3533 3534 MVT MTy = LT.second; 3535 3536 int ISD; 3537 if (Ty->isIntOrIntVectorTy()) { 3538 ISD = IsUnsigned ? ISD::UMIN : ISD::SMIN; 3539 } else { 3540 assert(Ty->isFPOrFPVectorTy() && 3541 "Expected float point or integer vector type."); 3542 ISD = ISD::FMINNUM; 3543 } 3544 3545 static const CostTblEntry SSE1CostTbl[] = { 3546 {ISD::FMINNUM, MVT::v4f32, 1}, 3547 }; 3548 3549 static const CostTblEntry SSE2CostTbl[] = { 3550 {ISD::FMINNUM, MVT::v2f64, 1}, 3551 {ISD::SMIN, MVT::v8i16, 1}, 3552 {ISD::UMIN, MVT::v16i8, 1}, 3553 }; 3554 3555 static const CostTblEntry SSE41CostTbl[] = { 3556 {ISD::SMIN, MVT::v4i32, 1}, 3557 {ISD::UMIN, MVT::v4i32, 1}, 3558 {ISD::UMIN, MVT::v8i16, 1}, 3559 {ISD::SMIN, MVT::v16i8, 1}, 3560 }; 3561 3562 static const CostTblEntry SSE42CostTbl[] = { 3563 {ISD::UMIN, MVT::v2i64, 3}, // xor+pcmpgtq+blendvpd 3564 }; 3565 3566 static const CostTblEntry AVX1CostTbl[] = { 3567 {ISD::FMINNUM, MVT::v8f32, 1}, 3568 {ISD::FMINNUM, MVT::v4f64, 1}, 3569 {ISD::SMIN, MVT::v8i32, 3}, 3570 {ISD::UMIN, MVT::v8i32, 3}, 3571 {ISD::SMIN, MVT::v16i16, 3}, 3572 {ISD::UMIN, MVT::v16i16, 3}, 3573 {ISD::SMIN, MVT::v32i8, 3}, 3574 {ISD::UMIN, MVT::v32i8, 3}, 3575 }; 3576 3577 static const CostTblEntry AVX2CostTbl[] = { 3578 {ISD::SMIN, MVT::v8i32, 1}, 3579 {ISD::UMIN, MVT::v8i32, 1}, 3580 {ISD::SMIN, MVT::v16i16, 1}, 3581 {ISD::UMIN, MVT::v16i16, 1}, 3582 {ISD::SMIN, MVT::v32i8, 1}, 3583 {ISD::UMIN, MVT::v32i8, 1}, 3584 }; 3585 3586 static const CostTblEntry AVX512CostTbl[] = { 3587 {ISD::FMINNUM, MVT::v16f32, 1}, 3588 {ISD::FMINNUM, MVT::v8f64, 1}, 3589 {ISD::SMIN, MVT::v2i64, 1}, 3590 {ISD::UMIN, MVT::v2i64, 1}, 3591 {ISD::SMIN, MVT::v4i64, 1}, 3592 {ISD::UMIN, MVT::v4i64, 1}, 3593 {ISD::SMIN, MVT::v8i64, 1}, 3594 {ISD::UMIN, MVT::v8i64, 1}, 3595 {ISD::SMIN, MVT::v16i32, 1}, 3596 {ISD::UMIN, MVT::v16i32, 1}, 3597 }; 3598 3599 static const CostTblEntry AVX512BWCostTbl[] = { 3600 {ISD::SMIN, MVT::v32i16, 1}, 3601 {ISD::UMIN, MVT::v32i16, 1}, 3602 {ISD::SMIN, MVT::v64i8, 1}, 3603 {ISD::UMIN, MVT::v64i8, 1}, 3604 }; 3605 3606 // If we have a native MIN/MAX instruction for this type, use it. 3607 if (ST->hasBWI()) 3608 if (const auto *Entry = CostTableLookup(AVX512BWCostTbl, ISD, MTy)) 3609 return LT.first * Entry->Cost; 3610 3611 if (ST->hasAVX512()) 3612 if (const auto *Entry = CostTableLookup(AVX512CostTbl, ISD, MTy)) 3613 return LT.first * Entry->Cost; 3614 3615 if (ST->hasAVX2()) 3616 if (const auto *Entry = CostTableLookup(AVX2CostTbl, ISD, MTy)) 3617 return LT.first * Entry->Cost; 3618 3619 if (ST->hasAVX()) 3620 if (const auto *Entry = CostTableLookup(AVX1CostTbl, ISD, MTy)) 3621 return LT.first * Entry->Cost; 3622 3623 if (ST->hasSSE42()) 3624 if (const auto *Entry = CostTableLookup(SSE42CostTbl, ISD, MTy)) 3625 return LT.first * Entry->Cost; 3626 3627 if (ST->hasSSE41()) 3628 if (const auto *Entry = CostTableLookup(SSE41CostTbl, ISD, MTy)) 3629 return LT.first * Entry->Cost; 3630 3631 if (ST->hasSSE2()) 3632 if (const auto *Entry = CostTableLookup(SSE2CostTbl, ISD, MTy)) 3633 return LT.first * Entry->Cost; 3634 3635 if (ST->hasSSE1()) 3636 if (const auto *Entry = CostTableLookup(SSE1CostTbl, ISD, MTy)) 3637 return LT.first * Entry->Cost; 3638 3639 unsigned CmpOpcode; 3640 if (Ty->isFPOrFPVectorTy()) { 3641 CmpOpcode = Instruction::FCmp; 3642 } else { 3643 assert(Ty->isIntOrIntVectorTy() && 3644 "expecting floating point or integer type for min/max reduction"); 3645 CmpOpcode = Instruction::ICmp; 3646 } 3647 3648 TTI::TargetCostKind CostKind = TTI::TCK_RecipThroughput; 3649 // Otherwise fall back to cmp+select. 3650 return getCmpSelInstrCost(CmpOpcode, Ty, CondTy, CostKind) + 3651 getCmpSelInstrCost(Instruction::Select, Ty, CondTy, CostKind); 3652 } 3653 3654 int X86TTIImpl::getMinMaxReductionCost(VectorType *ValTy, VectorType *CondTy, 3655 bool IsPairwise, bool IsUnsigned, 3656 TTI::TargetCostKind CostKind) { 3657 // Just use the default implementation for pair reductions. 3658 if (IsPairwise) 3659 return BaseT::getMinMaxReductionCost(ValTy, CondTy, IsPairwise, IsUnsigned, 3660 CostKind); 3661 3662 std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, ValTy); 3663 3664 MVT MTy = LT.second; 3665 3666 int ISD; 3667 if (ValTy->isIntOrIntVectorTy()) { 3668 ISD = IsUnsigned ? ISD::UMIN : ISD::SMIN; 3669 } else { 3670 assert(ValTy->isFPOrFPVectorTy() && 3671 "Expected float point or integer vector type."); 3672 ISD = ISD::FMINNUM; 3673 } 3674 3675 // We use the Intel Architecture Code Analyzer(IACA) to measure the throughput 3676 // and make it as the cost. 3677 3678 static const CostTblEntry SSE2CostTblNoPairWise[] = { 3679 {ISD::UMIN, MVT::v2i16, 5}, // need pxors to use pminsw/pmaxsw 3680 {ISD::UMIN, MVT::v4i16, 7}, // need pxors to use pminsw/pmaxsw 3681 {ISD::UMIN, MVT::v8i16, 9}, // need pxors to use pminsw/pmaxsw 3682 }; 3683 3684 static const CostTblEntry SSE41CostTblNoPairWise[] = { 3685 {ISD::SMIN, MVT::v2i16, 3}, // same as sse2 3686 {ISD::SMIN, MVT::v4i16, 5}, // same as sse2 3687 {ISD::UMIN, MVT::v2i16, 5}, // same as sse2 3688 {ISD::UMIN, MVT::v4i16, 7}, // same as sse2 3689 {ISD::SMIN, MVT::v8i16, 4}, // phminposuw+xor 3690 {ISD::UMIN, MVT::v8i16, 4}, // FIXME: umin is cheaper than umax 3691 {ISD::SMIN, MVT::v2i8, 3}, // pminsb 3692 {ISD::SMIN, MVT::v4i8, 5}, // pminsb 3693 {ISD::SMIN, MVT::v8i8, 7}, // pminsb 3694 {ISD::SMIN, MVT::v16i8, 6}, 3695 {ISD::UMIN, MVT::v2i8, 3}, // same as sse2 3696 {ISD::UMIN, MVT::v4i8, 5}, // same as sse2 3697 {ISD::UMIN, MVT::v8i8, 7}, // same as sse2 3698 {ISD::UMIN, MVT::v16i8, 6}, // FIXME: umin is cheaper than umax 3699 }; 3700 3701 static const CostTblEntry AVX1CostTblNoPairWise[] = { 3702 {ISD::SMIN, MVT::v16i16, 6}, 3703 {ISD::UMIN, MVT::v16i16, 6}, // FIXME: umin is cheaper than umax 3704 {ISD::SMIN, MVT::v32i8, 8}, 3705 {ISD::UMIN, MVT::v32i8, 8}, 3706 }; 3707 3708 static const CostTblEntry AVX512BWCostTblNoPairWise[] = { 3709 {ISD::SMIN, MVT::v32i16, 8}, 3710 {ISD::UMIN, MVT::v32i16, 8}, // FIXME: umin is cheaper than umax 3711 {ISD::SMIN, MVT::v64i8, 10}, 3712 {ISD::UMIN, MVT::v64i8, 10}, 3713 }; 3714 3715 // Before legalizing the type, give a chance to look up illegal narrow types 3716 // in the table. 3717 // FIXME: Is there a better way to do this? 3718 EVT VT = TLI->getValueType(DL, ValTy); 3719 if (VT.isSimple()) { 3720 MVT MTy = VT.getSimpleVT(); 3721 if (ST->hasBWI()) 3722 if (const auto *Entry = CostTableLookup(AVX512BWCostTblNoPairWise, ISD, MTy)) 3723 return Entry->Cost; 3724 3725 if (ST->hasAVX()) 3726 if (const auto *Entry = CostTableLookup(AVX1CostTblNoPairWise, ISD, MTy)) 3727 return Entry->Cost; 3728 3729 if (ST->hasSSE41()) 3730 if (const auto *Entry = CostTableLookup(SSE41CostTblNoPairWise, ISD, MTy)) 3731 return Entry->Cost; 3732 3733 if (ST->hasSSE2()) 3734 if (const auto *Entry = CostTableLookup(SSE2CostTblNoPairWise, ISD, MTy)) 3735 return Entry->Cost; 3736 } 3737 3738 auto *ValVTy = cast<FixedVectorType>(ValTy); 3739 unsigned NumVecElts = ValVTy->getNumElements(); 3740 3741 auto *Ty = ValVTy; 3742 unsigned MinMaxCost = 0; 3743 if (LT.first != 1 && MTy.isVector() && 3744 MTy.getVectorNumElements() < ValVTy->getNumElements()) { 3745 // Type needs to be split. We need LT.first - 1 operations ops. 3746 Ty = FixedVectorType::get(ValVTy->getElementType(), 3747 MTy.getVectorNumElements()); 3748 auto *SubCondTy = FixedVectorType::get(CondTy->getElementType(), 3749 MTy.getVectorNumElements()); 3750 MinMaxCost = getMinMaxCost(Ty, SubCondTy, IsUnsigned); 3751 MinMaxCost *= LT.first - 1; 3752 NumVecElts = MTy.getVectorNumElements(); 3753 } 3754 3755 if (ST->hasBWI()) 3756 if (const auto *Entry = CostTableLookup(AVX512BWCostTblNoPairWise, ISD, MTy)) 3757 return MinMaxCost + Entry->Cost; 3758 3759 if (ST->hasAVX()) 3760 if (const auto *Entry = CostTableLookup(AVX1CostTblNoPairWise, ISD, MTy)) 3761 return MinMaxCost + Entry->Cost; 3762 3763 if (ST->hasSSE41()) 3764 if (const auto *Entry = CostTableLookup(SSE41CostTblNoPairWise, ISD, MTy)) 3765 return MinMaxCost + Entry->Cost; 3766 3767 if (ST->hasSSE2()) 3768 if (const auto *Entry = CostTableLookup(SSE2CostTblNoPairWise, ISD, MTy)) 3769 return MinMaxCost + Entry->Cost; 3770 3771 unsigned ScalarSize = ValTy->getScalarSizeInBits(); 3772 3773 // Special case power of 2 reductions where the scalar type isn't changed 3774 // by type legalization. 3775 if (!isPowerOf2_32(ValVTy->getNumElements()) || 3776 ScalarSize != MTy.getScalarSizeInBits()) 3777 return BaseT::getMinMaxReductionCost(ValTy, CondTy, IsPairwise, IsUnsigned, 3778 CostKind); 3779 3780 // Now handle reduction with the legal type, taking into account size changes 3781 // at each level. 3782 while (NumVecElts > 1) { 3783 // Determine the size of the remaining vector we need to reduce. 3784 unsigned Size = NumVecElts * ScalarSize; 3785 NumVecElts /= 2; 3786 // If we're reducing from 256/512 bits, use an extract_subvector. 3787 if (Size > 128) { 3788 auto *SubTy = FixedVectorType::get(ValVTy->getElementType(), NumVecElts); 3789 MinMaxCost += 3790 getShuffleCost(TTI::SK_ExtractSubvector, Ty, NumVecElts, SubTy); 3791 Ty = SubTy; 3792 } else if (Size == 128) { 3793 // Reducing from 128 bits is a permute of v2f64/v2i64. 3794 VectorType *ShufTy; 3795 if (ValTy->isFloatingPointTy()) 3796 ShufTy = 3797 FixedVectorType::get(Type::getDoubleTy(ValTy->getContext()), 2); 3798 else 3799 ShufTy = FixedVectorType::get(Type::getInt64Ty(ValTy->getContext()), 2); 3800 MinMaxCost += 3801 getShuffleCost(TTI::SK_PermuteSingleSrc, ShufTy, 0, nullptr); 3802 } else if (Size == 64) { 3803 // Reducing from 64 bits is a shuffle of v4f32/v4i32. 3804 FixedVectorType *ShufTy; 3805 if (ValTy->isFloatingPointTy()) 3806 ShufTy = FixedVectorType::get(Type::getFloatTy(ValTy->getContext()), 4); 3807 else 3808 ShufTy = FixedVectorType::get(Type::getInt32Ty(ValTy->getContext()), 4); 3809 MinMaxCost += 3810 getShuffleCost(TTI::SK_PermuteSingleSrc, ShufTy, 0, nullptr); 3811 } else { 3812 // Reducing from smaller size is a shift by immediate. 3813 auto *ShiftTy = FixedVectorType::get( 3814 Type::getIntNTy(ValTy->getContext(), Size), 128 / Size); 3815 MinMaxCost += getArithmeticInstrCost( 3816 Instruction::LShr, ShiftTy, TTI::TCK_RecipThroughput, 3817 TargetTransformInfo::OK_AnyValue, 3818 TargetTransformInfo::OK_UniformConstantValue, 3819 TargetTransformInfo::OP_None, TargetTransformInfo::OP_None); 3820 } 3821 3822 // Add the arithmetic op for this level. 3823 auto *SubCondTy = 3824 FixedVectorType::get(CondTy->getElementType(), Ty->getNumElements()); 3825 MinMaxCost += getMinMaxCost(Ty, SubCondTy, IsUnsigned); 3826 } 3827 3828 // Add the final extract element to the cost. 3829 return MinMaxCost + getVectorInstrCost(Instruction::ExtractElement, Ty, 0); 3830 } 3831 3832 /// Calculate the cost of materializing a 64-bit value. This helper 3833 /// method might only calculate a fraction of a larger immediate. Therefore it 3834 /// is valid to return a cost of ZERO. 3835 int X86TTIImpl::getIntImmCost(int64_t Val) { 3836 if (Val == 0) 3837 return TTI::TCC_Free; 3838 3839 if (isInt<32>(Val)) 3840 return TTI::TCC_Basic; 3841 3842 return 2 * TTI::TCC_Basic; 3843 } 3844 3845 int X86TTIImpl::getIntImmCost(const APInt &Imm, Type *Ty, 3846 TTI::TargetCostKind CostKind) { 3847 assert(Ty->isIntegerTy()); 3848 3849 unsigned BitSize = Ty->getPrimitiveSizeInBits(); 3850 if (BitSize == 0) 3851 return ~0U; 3852 3853 // Never hoist constants larger than 128bit, because this might lead to 3854 // incorrect code generation or assertions in codegen. 3855 // Fixme: Create a cost model for types larger than i128 once the codegen 3856 // issues have been fixed. 3857 if (BitSize > 128) 3858 return TTI::TCC_Free; 3859 3860 if (Imm == 0) 3861 return TTI::TCC_Free; 3862 3863 // Sign-extend all constants to a multiple of 64-bit. 3864 APInt ImmVal = Imm; 3865 if (BitSize % 64 != 0) 3866 ImmVal = Imm.sext(alignTo(BitSize, 64)); 3867 3868 // Split the constant into 64-bit chunks and calculate the cost for each 3869 // chunk. 3870 int Cost = 0; 3871 for (unsigned ShiftVal = 0; ShiftVal < BitSize; ShiftVal += 64) { 3872 APInt Tmp = ImmVal.ashr(ShiftVal).sextOrTrunc(64); 3873 int64_t Val = Tmp.getSExtValue(); 3874 Cost += getIntImmCost(Val); 3875 } 3876 // We need at least one instruction to materialize the constant. 3877 return std::max(1, Cost); 3878 } 3879 3880 int X86TTIImpl::getIntImmCostInst(unsigned Opcode, unsigned Idx, 3881 const APInt &Imm, Type *Ty, 3882 TTI::TargetCostKind CostKind, 3883 Instruction *Inst) { 3884 assert(Ty->isIntegerTy()); 3885 3886 unsigned BitSize = Ty->getPrimitiveSizeInBits(); 3887 // There is no cost model for constants with a bit size of 0. Return TCC_Free 3888 // here, so that constant hoisting will ignore this constant. 3889 if (BitSize == 0) 3890 return TTI::TCC_Free; 3891 3892 unsigned ImmIdx = ~0U; 3893 switch (Opcode) { 3894 default: 3895 return TTI::TCC_Free; 3896 case Instruction::GetElementPtr: 3897 // Always hoist the base address of a GetElementPtr. This prevents the 3898 // creation of new constants for every base constant that gets constant 3899 // folded with the offset. 3900 if (Idx == 0) 3901 return 2 * TTI::TCC_Basic; 3902 return TTI::TCC_Free; 3903 case Instruction::Store: 3904 ImmIdx = 0; 3905 break; 3906 case Instruction::ICmp: 3907 // This is an imperfect hack to prevent constant hoisting of 3908 // compares that might be trying to check if a 64-bit value fits in 3909 // 32-bits. The backend can optimize these cases using a right shift by 32. 3910 // Ideally we would check the compare predicate here. There also other 3911 // similar immediates the backend can use shifts for. 3912 if (Idx == 1 && Imm.getBitWidth() == 64) { 3913 uint64_t ImmVal = Imm.getZExtValue(); 3914 if (ImmVal == 0x100000000ULL || ImmVal == 0xffffffff) 3915 return TTI::TCC_Free; 3916 } 3917 ImmIdx = 1; 3918 break; 3919 case Instruction::And: 3920 // We support 64-bit ANDs with immediates with 32-bits of leading zeroes 3921 // by using a 32-bit operation with implicit zero extension. Detect such 3922 // immediates here as the normal path expects bit 31 to be sign extended. 3923 if (Idx == 1 && Imm.getBitWidth() == 64 && isUInt<32>(Imm.getZExtValue())) 3924 return TTI::TCC_Free; 3925 ImmIdx = 1; 3926 break; 3927 case Instruction::Add: 3928 case Instruction::Sub: 3929 // For add/sub, we can use the opposite instruction for INT32_MIN. 3930 if (Idx == 1 && Imm.getBitWidth() == 64 && Imm.getZExtValue() == 0x80000000) 3931 return TTI::TCC_Free; 3932 ImmIdx = 1; 3933 break; 3934 case Instruction::UDiv: 3935 case Instruction::SDiv: 3936 case Instruction::URem: 3937 case Instruction::SRem: 3938 // Division by constant is typically expanded later into a different 3939 // instruction sequence. This completely changes the constants. 3940 // Report them as "free" to stop ConstantHoist from marking them as opaque. 3941 return TTI::TCC_Free; 3942 case Instruction::Mul: 3943 case Instruction::Or: 3944 case Instruction::Xor: 3945 ImmIdx = 1; 3946 break; 3947 // Always return TCC_Free for the shift value of a shift instruction. 3948 case Instruction::Shl: 3949 case Instruction::LShr: 3950 case Instruction::AShr: 3951 if (Idx == 1) 3952 return TTI::TCC_Free; 3953 break; 3954 case Instruction::Trunc: 3955 case Instruction::ZExt: 3956 case Instruction::SExt: 3957 case Instruction::IntToPtr: 3958 case Instruction::PtrToInt: 3959 case Instruction::BitCast: 3960 case Instruction::PHI: 3961 case Instruction::Call: 3962 case Instruction::Select: 3963 case Instruction::Ret: 3964 case Instruction::Load: 3965 break; 3966 } 3967 3968 if (Idx == ImmIdx) { 3969 int NumConstants = divideCeil(BitSize, 64); 3970 int Cost = X86TTIImpl::getIntImmCost(Imm, Ty, CostKind); 3971 return (Cost <= NumConstants * TTI::TCC_Basic) 3972 ? static_cast<int>(TTI::TCC_Free) 3973 : Cost; 3974 } 3975 3976 return X86TTIImpl::getIntImmCost(Imm, Ty, CostKind); 3977 } 3978 3979 int X86TTIImpl::getIntImmCostIntrin(Intrinsic::ID IID, unsigned Idx, 3980 const APInt &Imm, Type *Ty, 3981 TTI::TargetCostKind CostKind) { 3982 assert(Ty->isIntegerTy()); 3983 3984 unsigned BitSize = Ty->getPrimitiveSizeInBits(); 3985 // There is no cost model for constants with a bit size of 0. Return TCC_Free 3986 // here, so that constant hoisting will ignore this constant. 3987 if (BitSize == 0) 3988 return TTI::TCC_Free; 3989 3990 switch (IID) { 3991 default: 3992 return TTI::TCC_Free; 3993 case Intrinsic::sadd_with_overflow: 3994 case Intrinsic::uadd_with_overflow: 3995 case Intrinsic::ssub_with_overflow: 3996 case Intrinsic::usub_with_overflow: 3997 case Intrinsic::smul_with_overflow: 3998 case Intrinsic::umul_with_overflow: 3999 if ((Idx == 1) && Imm.getBitWidth() <= 64 && isInt<32>(Imm.getSExtValue())) 4000 return TTI::TCC_Free; 4001 break; 4002 case Intrinsic::experimental_stackmap: 4003 if ((Idx < 2) || (Imm.getBitWidth() <= 64 && isInt<64>(Imm.getSExtValue()))) 4004 return TTI::TCC_Free; 4005 break; 4006 case Intrinsic::experimental_patchpoint_void: 4007 case Intrinsic::experimental_patchpoint_i64: 4008 if ((Idx < 4) || (Imm.getBitWidth() <= 64 && isInt<64>(Imm.getSExtValue()))) 4009 return TTI::TCC_Free; 4010 break; 4011 } 4012 return X86TTIImpl::getIntImmCost(Imm, Ty, CostKind); 4013 } 4014 4015 unsigned 4016 X86TTIImpl::getCFInstrCost(unsigned Opcode, TTI::TargetCostKind CostKind) { 4017 if (CostKind != TTI::TCK_RecipThroughput) 4018 return Opcode == Instruction::PHI ? 0 : 1; 4019 // Branches are assumed to be predicted. 4020 return CostKind == TTI::TCK_RecipThroughput ? 0 : 1; 4021 } 4022 4023 int X86TTIImpl::getGatherOverhead() const { 4024 // Some CPUs have more overhead for gather. The specified overhead is relative 4025 // to the Load operation. "2" is the number provided by Intel architects. This 4026 // parameter is used for cost estimation of Gather Op and comparison with 4027 // other alternatives. 4028 // TODO: Remove the explicit hasAVX512()?, That would mean we would only 4029 // enable gather with a -march. 4030 if (ST->hasAVX512() || (ST->hasAVX2() && ST->hasFastGather())) 4031 return 2; 4032 4033 return 1024; 4034 } 4035 4036 int X86TTIImpl::getScatterOverhead() const { 4037 if (ST->hasAVX512()) 4038 return 2; 4039 4040 return 1024; 4041 } 4042 4043 // Return an average cost of Gather / Scatter instruction, maybe improved later 4044 int X86TTIImpl::getGSVectorCost(unsigned Opcode, Type *SrcVTy, const Value *Ptr, 4045 Align Alignment, unsigned AddressSpace) { 4046 4047 assert(isa<VectorType>(SrcVTy) && "Unexpected type in getGSVectorCost"); 4048 unsigned VF = cast<FixedVectorType>(SrcVTy)->getNumElements(); 4049 4050 // Try to reduce index size from 64 bit (default for GEP) 4051 // to 32. It is essential for VF 16. If the index can't be reduced to 32, the 4052 // operation will use 16 x 64 indices which do not fit in a zmm and needs 4053 // to split. Also check that the base pointer is the same for all lanes, 4054 // and that there's at most one variable index. 4055 auto getIndexSizeInBits = [](const Value *Ptr, const DataLayout &DL) { 4056 unsigned IndexSize = DL.getPointerSizeInBits(); 4057 const GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr); 4058 if (IndexSize < 64 || !GEP) 4059 return IndexSize; 4060 4061 unsigned NumOfVarIndices = 0; 4062 const Value *Ptrs = GEP->getPointerOperand(); 4063 if (Ptrs->getType()->isVectorTy() && !getSplatValue(Ptrs)) 4064 return IndexSize; 4065 for (unsigned i = 1; i < GEP->getNumOperands(); ++i) { 4066 if (isa<Constant>(GEP->getOperand(i))) 4067 continue; 4068 Type *IndxTy = GEP->getOperand(i)->getType(); 4069 if (auto *IndexVTy = dyn_cast<VectorType>(IndxTy)) 4070 IndxTy = IndexVTy->getElementType(); 4071 if ((IndxTy->getPrimitiveSizeInBits() == 64 && 4072 !isa<SExtInst>(GEP->getOperand(i))) || 4073 ++NumOfVarIndices > 1) 4074 return IndexSize; // 64 4075 } 4076 return (unsigned)32; 4077 }; 4078 4079 // Trying to reduce IndexSize to 32 bits for vector 16. 4080 // By default the IndexSize is equal to pointer size. 4081 unsigned IndexSize = (ST->hasAVX512() && VF >= 16) 4082 ? getIndexSizeInBits(Ptr, DL) 4083 : DL.getPointerSizeInBits(); 4084 4085 auto *IndexVTy = FixedVectorType::get( 4086 IntegerType::get(SrcVTy->getContext(), IndexSize), VF); 4087 std::pair<int, MVT> IdxsLT = TLI->getTypeLegalizationCost(DL, IndexVTy); 4088 std::pair<int, MVT> SrcLT = TLI->getTypeLegalizationCost(DL, SrcVTy); 4089 int SplitFactor = std::max(IdxsLT.first, SrcLT.first); 4090 if (SplitFactor > 1) { 4091 // Handle splitting of vector of pointers 4092 auto *SplitSrcTy = 4093 FixedVectorType::get(SrcVTy->getScalarType(), VF / SplitFactor); 4094 return SplitFactor * getGSVectorCost(Opcode, SplitSrcTy, Ptr, Alignment, 4095 AddressSpace); 4096 } 4097 4098 // The gather / scatter cost is given by Intel architects. It is a rough 4099 // number since we are looking at one instruction in a time. 4100 const int GSOverhead = (Opcode == Instruction::Load) 4101 ? getGatherOverhead() 4102 : getScatterOverhead(); 4103 return GSOverhead + VF * getMemoryOpCost(Opcode, SrcVTy->getScalarType(), 4104 MaybeAlign(Alignment), AddressSpace, 4105 TTI::TCK_RecipThroughput); 4106 } 4107 4108 /// Return the cost of full scalarization of gather / scatter operation. 4109 /// 4110 /// Opcode - Load or Store instruction. 4111 /// SrcVTy - The type of the data vector that should be gathered or scattered. 4112 /// VariableMask - The mask is non-constant at compile time. 4113 /// Alignment - Alignment for one element. 4114 /// AddressSpace - pointer[s] address space. 4115 /// 4116 int X86TTIImpl::getGSScalarCost(unsigned Opcode, Type *SrcVTy, 4117 bool VariableMask, Align Alignment, 4118 unsigned AddressSpace) { 4119 unsigned VF = cast<FixedVectorType>(SrcVTy)->getNumElements(); 4120 APInt DemandedElts = APInt::getAllOnesValue(VF); 4121 TTI::TargetCostKind CostKind = TTI::TCK_RecipThroughput; 4122 4123 int MaskUnpackCost = 0; 4124 if (VariableMask) { 4125 auto *MaskTy = 4126 FixedVectorType::get(Type::getInt1Ty(SrcVTy->getContext()), VF); 4127 MaskUnpackCost = 4128 getScalarizationOverhead(MaskTy, DemandedElts, false, true); 4129 int ScalarCompareCost = 4130 getCmpSelInstrCost(Instruction::ICmp, Type::getInt1Ty(SrcVTy->getContext()), 4131 nullptr, CostKind); 4132 int BranchCost = getCFInstrCost(Instruction::Br, CostKind); 4133 MaskUnpackCost += VF * (BranchCost + ScalarCompareCost); 4134 } 4135 4136 // The cost of the scalar loads/stores. 4137 int MemoryOpCost = VF * getMemoryOpCost(Opcode, SrcVTy->getScalarType(), 4138 MaybeAlign(Alignment), AddressSpace, 4139 CostKind); 4140 4141 int InsertExtractCost = 0; 4142 if (Opcode == Instruction::Load) 4143 for (unsigned i = 0; i < VF; ++i) 4144 // Add the cost of inserting each scalar load into the vector 4145 InsertExtractCost += 4146 getVectorInstrCost(Instruction::InsertElement, SrcVTy, i); 4147 else 4148 for (unsigned i = 0; i < VF; ++i) 4149 // Add the cost of extracting each element out of the data vector 4150 InsertExtractCost += 4151 getVectorInstrCost(Instruction::ExtractElement, SrcVTy, i); 4152 4153 return MemoryOpCost + MaskUnpackCost + InsertExtractCost; 4154 } 4155 4156 /// Calculate the cost of Gather / Scatter operation 4157 int X86TTIImpl::getGatherScatterOpCost(unsigned Opcode, Type *SrcVTy, 4158 const Value *Ptr, bool VariableMask, 4159 Align Alignment, 4160 TTI::TargetCostKind CostKind, 4161 const Instruction *I = nullptr) { 4162 4163 if (CostKind != TTI::TCK_RecipThroughput) 4164 return 1; 4165 4166 assert(SrcVTy->isVectorTy() && "Unexpected data type for Gather/Scatter"); 4167 unsigned VF = cast<FixedVectorType>(SrcVTy)->getNumElements(); 4168 PointerType *PtrTy = dyn_cast<PointerType>(Ptr->getType()); 4169 if (!PtrTy && Ptr->getType()->isVectorTy()) 4170 PtrTy = dyn_cast<PointerType>( 4171 cast<VectorType>(Ptr->getType())->getElementType()); 4172 assert(PtrTy && "Unexpected type for Ptr argument"); 4173 unsigned AddressSpace = PtrTy->getAddressSpace(); 4174 4175 bool Scalarize = false; 4176 if ((Opcode == Instruction::Load && 4177 !isLegalMaskedGather(SrcVTy, Align(Alignment))) || 4178 (Opcode == Instruction::Store && 4179 !isLegalMaskedScatter(SrcVTy, Align(Alignment)))) 4180 Scalarize = true; 4181 // Gather / Scatter for vector 2 is not profitable on KNL / SKX 4182 // Vector-4 of gather/scatter instruction does not exist on KNL. 4183 // We can extend it to 8 elements, but zeroing upper bits of 4184 // the mask vector will add more instructions. Right now we give the scalar 4185 // cost of vector-4 for KNL. TODO: Check, maybe the gather/scatter instruction 4186 // is better in the VariableMask case. 4187 if (ST->hasAVX512() && (VF == 2 || (VF == 4 && !ST->hasVLX()))) 4188 Scalarize = true; 4189 4190 if (Scalarize) 4191 return getGSScalarCost(Opcode, SrcVTy, VariableMask, Alignment, 4192 AddressSpace); 4193 4194 return getGSVectorCost(Opcode, SrcVTy, Ptr, Alignment, AddressSpace); 4195 } 4196 4197 bool X86TTIImpl::isLSRCostLess(TargetTransformInfo::LSRCost &C1, 4198 TargetTransformInfo::LSRCost &C2) { 4199 // X86 specific here are "instruction number 1st priority". 4200 return std::tie(C1.Insns, C1.NumRegs, C1.AddRecCost, 4201 C1.NumIVMuls, C1.NumBaseAdds, 4202 C1.ScaleCost, C1.ImmCost, C1.SetupCost) < 4203 std::tie(C2.Insns, C2.NumRegs, C2.AddRecCost, 4204 C2.NumIVMuls, C2.NumBaseAdds, 4205 C2.ScaleCost, C2.ImmCost, C2.SetupCost); 4206 } 4207 4208 bool X86TTIImpl::canMacroFuseCmp() { 4209 return ST->hasMacroFusion() || ST->hasBranchFusion(); 4210 } 4211 4212 bool X86TTIImpl::isLegalMaskedLoad(Type *DataTy, Align Alignment) { 4213 if (!ST->hasAVX()) 4214 return false; 4215 4216 // The backend can't handle a single element vector. 4217 if (isa<VectorType>(DataTy) && 4218 cast<FixedVectorType>(DataTy)->getNumElements() == 1) 4219 return false; 4220 Type *ScalarTy = DataTy->getScalarType(); 4221 4222 if (ScalarTy->isPointerTy()) 4223 return true; 4224 4225 if (ScalarTy->isFloatTy() || ScalarTy->isDoubleTy()) 4226 return true; 4227 4228 if (!ScalarTy->isIntegerTy()) 4229 return false; 4230 4231 unsigned IntWidth = ScalarTy->getIntegerBitWidth(); 4232 return IntWidth == 32 || IntWidth == 64 || 4233 ((IntWidth == 8 || IntWidth == 16) && ST->hasBWI()); 4234 } 4235 4236 bool X86TTIImpl::isLegalMaskedStore(Type *DataType, Align Alignment) { 4237 return isLegalMaskedLoad(DataType, Alignment); 4238 } 4239 4240 bool X86TTIImpl::isLegalNTLoad(Type *DataType, Align Alignment) { 4241 unsigned DataSize = DL.getTypeStoreSize(DataType); 4242 // The only supported nontemporal loads are for aligned vectors of 16 or 32 4243 // bytes. Note that 32-byte nontemporal vector loads are supported by AVX2 4244 // (the equivalent stores only require AVX). 4245 if (Alignment >= DataSize && (DataSize == 16 || DataSize == 32)) 4246 return DataSize == 16 ? ST->hasSSE1() : ST->hasAVX2(); 4247 4248 return false; 4249 } 4250 4251 bool X86TTIImpl::isLegalNTStore(Type *DataType, Align Alignment) { 4252 unsigned DataSize = DL.getTypeStoreSize(DataType); 4253 4254 // SSE4A supports nontemporal stores of float and double at arbitrary 4255 // alignment. 4256 if (ST->hasSSE4A() && (DataType->isFloatTy() || DataType->isDoubleTy())) 4257 return true; 4258 4259 // Besides the SSE4A subtarget exception above, only aligned stores are 4260 // available nontemporaly on any other subtarget. And only stores with a size 4261 // of 4..32 bytes (powers of 2, only) are permitted. 4262 if (Alignment < DataSize || DataSize < 4 || DataSize > 32 || 4263 !isPowerOf2_32(DataSize)) 4264 return false; 4265 4266 // 32-byte vector nontemporal stores are supported by AVX (the equivalent 4267 // loads require AVX2). 4268 if (DataSize == 32) 4269 return ST->hasAVX(); 4270 else if (DataSize == 16) 4271 return ST->hasSSE1(); 4272 return true; 4273 } 4274 4275 bool X86TTIImpl::isLegalMaskedExpandLoad(Type *DataTy) { 4276 if (!isa<VectorType>(DataTy)) 4277 return false; 4278 4279 if (!ST->hasAVX512()) 4280 return false; 4281 4282 // The backend can't handle a single element vector. 4283 if (cast<FixedVectorType>(DataTy)->getNumElements() == 1) 4284 return false; 4285 4286 Type *ScalarTy = cast<VectorType>(DataTy)->getElementType(); 4287 4288 if (ScalarTy->isFloatTy() || ScalarTy->isDoubleTy()) 4289 return true; 4290 4291 if (!ScalarTy->isIntegerTy()) 4292 return false; 4293 4294 unsigned IntWidth = ScalarTy->getIntegerBitWidth(); 4295 return IntWidth == 32 || IntWidth == 64 || 4296 ((IntWidth == 8 || IntWidth == 16) && ST->hasVBMI2()); 4297 } 4298 4299 bool X86TTIImpl::isLegalMaskedCompressStore(Type *DataTy) { 4300 return isLegalMaskedExpandLoad(DataTy); 4301 } 4302 4303 bool X86TTIImpl::isLegalMaskedGather(Type *DataTy, Align Alignment) { 4304 // Some CPUs have better gather performance than others. 4305 // TODO: Remove the explicit ST->hasAVX512()?, That would mean we would only 4306 // enable gather with a -march. 4307 if (!(ST->hasAVX512() || (ST->hasFastGather() && ST->hasAVX2()))) 4308 return false; 4309 4310 // This function is called now in two cases: from the Loop Vectorizer 4311 // and from the Scalarizer. 4312 // When the Loop Vectorizer asks about legality of the feature, 4313 // the vectorization factor is not calculated yet. The Loop Vectorizer 4314 // sends a scalar type and the decision is based on the width of the 4315 // scalar element. 4316 // Later on, the cost model will estimate usage this intrinsic based on 4317 // the vector type. 4318 // The Scalarizer asks again about legality. It sends a vector type. 4319 // In this case we can reject non-power-of-2 vectors. 4320 // We also reject single element vectors as the type legalizer can't 4321 // scalarize it. 4322 if (auto *DataVTy = dyn_cast<FixedVectorType>(DataTy)) { 4323 unsigned NumElts = DataVTy->getNumElements(); 4324 if (NumElts == 1) 4325 return false; 4326 } 4327 Type *ScalarTy = DataTy->getScalarType(); 4328 if (ScalarTy->isPointerTy()) 4329 return true; 4330 4331 if (ScalarTy->isFloatTy() || ScalarTy->isDoubleTy()) 4332 return true; 4333 4334 if (!ScalarTy->isIntegerTy()) 4335 return false; 4336 4337 unsigned IntWidth = ScalarTy->getIntegerBitWidth(); 4338 return IntWidth == 32 || IntWidth == 64; 4339 } 4340 4341 bool X86TTIImpl::isLegalMaskedScatter(Type *DataType, Align Alignment) { 4342 // AVX2 doesn't support scatter 4343 if (!ST->hasAVX512()) 4344 return false; 4345 return isLegalMaskedGather(DataType, Alignment); 4346 } 4347 4348 bool X86TTIImpl::hasDivRemOp(Type *DataType, bool IsSigned) { 4349 EVT VT = TLI->getValueType(DL, DataType); 4350 return TLI->isOperationLegal(IsSigned ? ISD::SDIVREM : ISD::UDIVREM, VT); 4351 } 4352 4353 bool X86TTIImpl::isFCmpOrdCheaperThanFCmpZero(Type *Ty) { 4354 return false; 4355 } 4356 4357 bool X86TTIImpl::areInlineCompatible(const Function *Caller, 4358 const Function *Callee) const { 4359 const TargetMachine &TM = getTLI()->getTargetMachine(); 4360 4361 // Work this as a subsetting of subtarget features. 4362 const FeatureBitset &CallerBits = 4363 TM.getSubtargetImpl(*Caller)->getFeatureBits(); 4364 const FeatureBitset &CalleeBits = 4365 TM.getSubtargetImpl(*Callee)->getFeatureBits(); 4366 4367 FeatureBitset RealCallerBits = CallerBits & ~InlineFeatureIgnoreList; 4368 FeatureBitset RealCalleeBits = CalleeBits & ~InlineFeatureIgnoreList; 4369 return (RealCallerBits & RealCalleeBits) == RealCalleeBits; 4370 } 4371 4372 bool X86TTIImpl::areFunctionArgsABICompatible( 4373 const Function *Caller, const Function *Callee, 4374 SmallPtrSetImpl<Argument *> &Args) const { 4375 if (!BaseT::areFunctionArgsABICompatible(Caller, Callee, Args)) 4376 return false; 4377 4378 // If we get here, we know the target features match. If one function 4379 // considers 512-bit vectors legal and the other does not, consider them 4380 // incompatible. 4381 const TargetMachine &TM = getTLI()->getTargetMachine(); 4382 4383 if (TM.getSubtarget<X86Subtarget>(*Caller).useAVX512Regs() == 4384 TM.getSubtarget<X86Subtarget>(*Callee).useAVX512Regs()) 4385 return true; 4386 4387 // Consider the arguments compatible if they aren't vectors or aggregates. 4388 // FIXME: Look at the size of vectors. 4389 // FIXME: Look at the element types of aggregates to see if there are vectors. 4390 // FIXME: The API of this function seems intended to allow arguments 4391 // to be removed from the set, but the caller doesn't check if the set 4392 // becomes empty so that may not work in practice. 4393 return llvm::none_of(Args, [](Argument *A) { 4394 auto *EltTy = cast<PointerType>(A->getType())->getElementType(); 4395 return EltTy->isVectorTy() || EltTy->isAggregateType(); 4396 }); 4397 } 4398 4399 X86TTIImpl::TTI::MemCmpExpansionOptions 4400 X86TTIImpl::enableMemCmpExpansion(bool OptSize, bool IsZeroCmp) const { 4401 TTI::MemCmpExpansionOptions Options; 4402 Options.MaxNumLoads = TLI->getMaxExpandSizeMemcmp(OptSize); 4403 Options.NumLoadsPerBlock = 2; 4404 // All GPR and vector loads can be unaligned. 4405 Options.AllowOverlappingLoads = true; 4406 if (IsZeroCmp) { 4407 // Only enable vector loads for equality comparison. Right now the vector 4408 // version is not as fast for three way compare (see #33329). 4409 const unsigned PreferredWidth = ST->getPreferVectorWidth(); 4410 if (PreferredWidth >= 512 && ST->hasAVX512()) Options.LoadSizes.push_back(64); 4411 if (PreferredWidth >= 256 && ST->hasAVX()) Options.LoadSizes.push_back(32); 4412 if (PreferredWidth >= 128 && ST->hasSSE2()) Options.LoadSizes.push_back(16); 4413 } 4414 if (ST->is64Bit()) { 4415 Options.LoadSizes.push_back(8); 4416 } 4417 Options.LoadSizes.push_back(4); 4418 Options.LoadSizes.push_back(2); 4419 Options.LoadSizes.push_back(1); 4420 return Options; 4421 } 4422 4423 bool X86TTIImpl::enableInterleavedAccessVectorization() { 4424 // TODO: We expect this to be beneficial regardless of arch, 4425 // but there are currently some unexplained performance artifacts on Atom. 4426 // As a temporary solution, disable on Atom. 4427 return !(ST->isAtom()); 4428 } 4429 4430 // Get estimation for interleaved load/store operations for AVX2. 4431 // \p Factor is the interleaved-access factor (stride) - number of 4432 // (interleaved) elements in the group. 4433 // \p Indices contains the indices for a strided load: when the 4434 // interleaved load has gaps they indicate which elements are used. 4435 // If Indices is empty (or if the number of indices is equal to the size 4436 // of the interleaved-access as given in \p Factor) the access has no gaps. 4437 // 4438 // As opposed to AVX-512, AVX2 does not have generic shuffles that allow 4439 // computing the cost using a generic formula as a function of generic 4440 // shuffles. We therefore use a lookup table instead, filled according to 4441 // the instruction sequences that codegen currently generates. 4442 int X86TTIImpl::getInterleavedMemoryOpCostAVX2( 4443 unsigned Opcode, FixedVectorType *VecTy, unsigned Factor, 4444 ArrayRef<unsigned> Indices, Align Alignment, unsigned AddressSpace, 4445 TTI::TargetCostKind CostKind, bool UseMaskForCond, bool UseMaskForGaps) { 4446 4447 if (UseMaskForCond || UseMaskForGaps) 4448 return BaseT::getInterleavedMemoryOpCost(Opcode, VecTy, Factor, Indices, 4449 Alignment, AddressSpace, CostKind, 4450 UseMaskForCond, UseMaskForGaps); 4451 4452 // We currently Support only fully-interleaved groups, with no gaps. 4453 // TODO: Support also strided loads (interleaved-groups with gaps). 4454 if (Indices.size() && Indices.size() != Factor) 4455 return BaseT::getInterleavedMemoryOpCost(Opcode, VecTy, Factor, Indices, 4456 Alignment, AddressSpace, 4457 CostKind); 4458 4459 // VecTy for interleave memop is <VF*Factor x Elt>. 4460 // So, for VF=4, Interleave Factor = 3, Element type = i32 we have 4461 // VecTy = <12 x i32>. 4462 MVT LegalVT = getTLI()->getTypeLegalizationCost(DL, VecTy).second; 4463 4464 // This function can be called with VecTy=<6xi128>, Factor=3, in which case 4465 // the VF=2, while v2i128 is an unsupported MVT vector type 4466 // (see MachineValueType.h::getVectorVT()). 4467 if (!LegalVT.isVector()) 4468 return BaseT::getInterleavedMemoryOpCost(Opcode, VecTy, Factor, Indices, 4469 Alignment, AddressSpace, 4470 CostKind); 4471 4472 unsigned VF = VecTy->getNumElements() / Factor; 4473 Type *ScalarTy = VecTy->getElementType(); 4474 4475 // Calculate the number of memory operations (NumOfMemOps), required 4476 // for load/store the VecTy. 4477 unsigned VecTySize = DL.getTypeStoreSize(VecTy); 4478 unsigned LegalVTSize = LegalVT.getStoreSize(); 4479 unsigned NumOfMemOps = (VecTySize + LegalVTSize - 1) / LegalVTSize; 4480 4481 // Get the cost of one memory operation. 4482 auto *SingleMemOpTy = FixedVectorType::get(VecTy->getElementType(), 4483 LegalVT.getVectorNumElements()); 4484 unsigned MemOpCost = getMemoryOpCost(Opcode, SingleMemOpTy, 4485 MaybeAlign(Alignment), AddressSpace, 4486 CostKind); 4487 4488 auto *VT = FixedVectorType::get(ScalarTy, VF); 4489 EVT ETy = TLI->getValueType(DL, VT); 4490 if (!ETy.isSimple()) 4491 return BaseT::getInterleavedMemoryOpCost(Opcode, VecTy, Factor, Indices, 4492 Alignment, AddressSpace, 4493 CostKind); 4494 4495 // TODO: Complete for other data-types and strides. 4496 // Each combination of Stride, ElementTy and VF results in a different 4497 // sequence; The cost tables are therefore accessed with: 4498 // Factor (stride) and VectorType=VFxElemType. 4499 // The Cost accounts only for the shuffle sequence; 4500 // The cost of the loads/stores is accounted for separately. 4501 // 4502 static const CostTblEntry AVX2InterleavedLoadTbl[] = { 4503 { 2, MVT::v4i64, 6 }, //(load 8i64 and) deinterleave into 2 x 4i64 4504 { 2, MVT::v4f64, 6 }, //(load 8f64 and) deinterleave into 2 x 4f64 4505 4506 { 3, MVT::v2i8, 10 }, //(load 6i8 and) deinterleave into 3 x 2i8 4507 { 3, MVT::v4i8, 4 }, //(load 12i8 and) deinterleave into 3 x 4i8 4508 { 3, MVT::v8i8, 9 }, //(load 24i8 and) deinterleave into 3 x 8i8 4509 { 3, MVT::v16i8, 11}, //(load 48i8 and) deinterleave into 3 x 16i8 4510 { 3, MVT::v32i8, 13}, //(load 96i8 and) deinterleave into 3 x 32i8 4511 { 3, MVT::v8f32, 17 }, //(load 24f32 and)deinterleave into 3 x 8f32 4512 4513 { 4, MVT::v2i8, 12 }, //(load 8i8 and) deinterleave into 4 x 2i8 4514 { 4, MVT::v4i8, 4 }, //(load 16i8 and) deinterleave into 4 x 4i8 4515 { 4, MVT::v8i8, 20 }, //(load 32i8 and) deinterleave into 4 x 8i8 4516 { 4, MVT::v16i8, 39 }, //(load 64i8 and) deinterleave into 4 x 16i8 4517 { 4, MVT::v32i8, 80 }, //(load 128i8 and) deinterleave into 4 x 32i8 4518 4519 { 8, MVT::v8f32, 40 } //(load 64f32 and)deinterleave into 8 x 8f32 4520 }; 4521 4522 static const CostTblEntry AVX2InterleavedStoreTbl[] = { 4523 { 2, MVT::v4i64, 6 }, //interleave into 2 x 4i64 into 8i64 (and store) 4524 { 2, MVT::v4f64, 6 }, //interleave into 2 x 4f64 into 8f64 (and store) 4525 4526 { 3, MVT::v2i8, 7 }, //interleave 3 x 2i8 into 6i8 (and store) 4527 { 3, MVT::v4i8, 8 }, //interleave 3 x 4i8 into 12i8 (and store) 4528 { 3, MVT::v8i8, 11 }, //interleave 3 x 8i8 into 24i8 (and store) 4529 { 3, MVT::v16i8, 11 }, //interleave 3 x 16i8 into 48i8 (and store) 4530 { 3, MVT::v32i8, 13 }, //interleave 3 x 32i8 into 96i8 (and store) 4531 4532 { 4, MVT::v2i8, 12 }, //interleave 4 x 2i8 into 8i8 (and store) 4533 { 4, MVT::v4i8, 9 }, //interleave 4 x 4i8 into 16i8 (and store) 4534 { 4, MVT::v8i8, 10 }, //interleave 4 x 8i8 into 32i8 (and store) 4535 { 4, MVT::v16i8, 10 }, //interleave 4 x 16i8 into 64i8 (and store) 4536 { 4, MVT::v32i8, 12 } //interleave 4 x 32i8 into 128i8 (and store) 4537 }; 4538 4539 if (Opcode == Instruction::Load) { 4540 if (const auto *Entry = 4541 CostTableLookup(AVX2InterleavedLoadTbl, Factor, ETy.getSimpleVT())) 4542 return NumOfMemOps * MemOpCost + Entry->Cost; 4543 } else { 4544 assert(Opcode == Instruction::Store && 4545 "Expected Store Instruction at this point"); 4546 if (const auto *Entry = 4547 CostTableLookup(AVX2InterleavedStoreTbl, Factor, ETy.getSimpleVT())) 4548 return NumOfMemOps * MemOpCost + Entry->Cost; 4549 } 4550 4551 return BaseT::getInterleavedMemoryOpCost(Opcode, VecTy, Factor, Indices, 4552 Alignment, AddressSpace, CostKind); 4553 } 4554 4555 // Get estimation for interleaved load/store operations and strided load. 4556 // \p Indices contains indices for strided load. 4557 // \p Factor - the factor of interleaving. 4558 // AVX-512 provides 3-src shuffles that significantly reduces the cost. 4559 int X86TTIImpl::getInterleavedMemoryOpCostAVX512( 4560 unsigned Opcode, FixedVectorType *VecTy, unsigned Factor, 4561 ArrayRef<unsigned> Indices, Align Alignment, unsigned AddressSpace, 4562 TTI::TargetCostKind CostKind, bool UseMaskForCond, bool UseMaskForGaps) { 4563 4564 if (UseMaskForCond || UseMaskForGaps) 4565 return BaseT::getInterleavedMemoryOpCost(Opcode, VecTy, Factor, Indices, 4566 Alignment, AddressSpace, CostKind, 4567 UseMaskForCond, UseMaskForGaps); 4568 4569 // VecTy for interleave memop is <VF*Factor x Elt>. 4570 // So, for VF=4, Interleave Factor = 3, Element type = i32 we have 4571 // VecTy = <12 x i32>. 4572 4573 // Calculate the number of memory operations (NumOfMemOps), required 4574 // for load/store the VecTy. 4575 MVT LegalVT = getTLI()->getTypeLegalizationCost(DL, VecTy).second; 4576 unsigned VecTySize = DL.getTypeStoreSize(VecTy); 4577 unsigned LegalVTSize = LegalVT.getStoreSize(); 4578 unsigned NumOfMemOps = (VecTySize + LegalVTSize - 1) / LegalVTSize; 4579 4580 // Get the cost of one memory operation. 4581 auto *SingleMemOpTy = FixedVectorType::get(VecTy->getElementType(), 4582 LegalVT.getVectorNumElements()); 4583 unsigned MemOpCost = getMemoryOpCost(Opcode, SingleMemOpTy, 4584 MaybeAlign(Alignment), AddressSpace, 4585 CostKind); 4586 4587 unsigned VF = VecTy->getNumElements() / Factor; 4588 MVT VT = MVT::getVectorVT(MVT::getVT(VecTy->getScalarType()), VF); 4589 4590 if (Opcode == Instruction::Load) { 4591 // The tables (AVX512InterleavedLoadTbl and AVX512InterleavedStoreTbl) 4592 // contain the cost of the optimized shuffle sequence that the 4593 // X86InterleavedAccess pass will generate. 4594 // The cost of loads and stores are computed separately from the table. 4595 4596 // X86InterleavedAccess support only the following interleaved-access group. 4597 static const CostTblEntry AVX512InterleavedLoadTbl[] = { 4598 {3, MVT::v16i8, 12}, //(load 48i8 and) deinterleave into 3 x 16i8 4599 {3, MVT::v32i8, 14}, //(load 96i8 and) deinterleave into 3 x 32i8 4600 {3, MVT::v64i8, 22}, //(load 96i8 and) deinterleave into 3 x 32i8 4601 }; 4602 4603 if (const auto *Entry = 4604 CostTableLookup(AVX512InterleavedLoadTbl, Factor, VT)) 4605 return NumOfMemOps * MemOpCost + Entry->Cost; 4606 //If an entry does not exist, fallback to the default implementation. 4607 4608 // Kind of shuffle depends on number of loaded values. 4609 // If we load the entire data in one register, we can use a 1-src shuffle. 4610 // Otherwise, we'll merge 2 sources in each operation. 4611 TTI::ShuffleKind ShuffleKind = 4612 (NumOfMemOps > 1) ? TTI::SK_PermuteTwoSrc : TTI::SK_PermuteSingleSrc; 4613 4614 unsigned ShuffleCost = 4615 getShuffleCost(ShuffleKind, SingleMemOpTy, 0, nullptr); 4616 4617 unsigned NumOfLoadsInInterleaveGrp = 4618 Indices.size() ? Indices.size() : Factor; 4619 auto *ResultTy = FixedVectorType::get(VecTy->getElementType(), 4620 VecTy->getNumElements() / Factor); 4621 unsigned NumOfResults = 4622 getTLI()->getTypeLegalizationCost(DL, ResultTy).first * 4623 NumOfLoadsInInterleaveGrp; 4624 4625 // About a half of the loads may be folded in shuffles when we have only 4626 // one result. If we have more than one result, we do not fold loads at all. 4627 unsigned NumOfUnfoldedLoads = 4628 NumOfResults > 1 ? NumOfMemOps : NumOfMemOps / 2; 4629 4630 // Get a number of shuffle operations per result. 4631 unsigned NumOfShufflesPerResult = 4632 std::max((unsigned)1, (unsigned)(NumOfMemOps - 1)); 4633 4634 // The SK_MergeTwoSrc shuffle clobbers one of src operands. 4635 // When we have more than one destination, we need additional instructions 4636 // to keep sources. 4637 unsigned NumOfMoves = 0; 4638 if (NumOfResults > 1 && ShuffleKind == TTI::SK_PermuteTwoSrc) 4639 NumOfMoves = NumOfResults * NumOfShufflesPerResult / 2; 4640 4641 int Cost = NumOfResults * NumOfShufflesPerResult * ShuffleCost + 4642 NumOfUnfoldedLoads * MemOpCost + NumOfMoves; 4643 4644 return Cost; 4645 } 4646 4647 // Store. 4648 assert(Opcode == Instruction::Store && 4649 "Expected Store Instruction at this point"); 4650 // X86InterleavedAccess support only the following interleaved-access group. 4651 static const CostTblEntry AVX512InterleavedStoreTbl[] = { 4652 {3, MVT::v16i8, 12}, // interleave 3 x 16i8 into 48i8 (and store) 4653 {3, MVT::v32i8, 14}, // interleave 3 x 32i8 into 96i8 (and store) 4654 {3, MVT::v64i8, 26}, // interleave 3 x 64i8 into 96i8 (and store) 4655 4656 {4, MVT::v8i8, 10}, // interleave 4 x 8i8 into 32i8 (and store) 4657 {4, MVT::v16i8, 11}, // interleave 4 x 16i8 into 64i8 (and store) 4658 {4, MVT::v32i8, 14}, // interleave 4 x 32i8 into 128i8 (and store) 4659 {4, MVT::v64i8, 24} // interleave 4 x 32i8 into 256i8 (and store) 4660 }; 4661 4662 if (const auto *Entry = 4663 CostTableLookup(AVX512InterleavedStoreTbl, Factor, VT)) 4664 return NumOfMemOps * MemOpCost + Entry->Cost; 4665 //If an entry does not exist, fallback to the default implementation. 4666 4667 // There is no strided stores meanwhile. And store can't be folded in 4668 // shuffle. 4669 unsigned NumOfSources = Factor; // The number of values to be merged. 4670 unsigned ShuffleCost = 4671 getShuffleCost(TTI::SK_PermuteTwoSrc, SingleMemOpTy, 0, nullptr); 4672 unsigned NumOfShufflesPerStore = NumOfSources - 1; 4673 4674 // The SK_MergeTwoSrc shuffle clobbers one of src operands. 4675 // We need additional instructions to keep sources. 4676 unsigned NumOfMoves = NumOfMemOps * NumOfShufflesPerStore / 2; 4677 int Cost = NumOfMemOps * (MemOpCost + NumOfShufflesPerStore * ShuffleCost) + 4678 NumOfMoves; 4679 return Cost; 4680 } 4681 4682 int X86TTIImpl::getInterleavedMemoryOpCost( 4683 unsigned Opcode, Type *VecTy, unsigned Factor, ArrayRef<unsigned> Indices, 4684 Align Alignment, unsigned AddressSpace, TTI::TargetCostKind CostKind, 4685 bool UseMaskForCond, bool UseMaskForGaps) { 4686 auto isSupportedOnAVX512 = [](Type *VecTy, bool HasBW) { 4687 Type *EltTy = cast<VectorType>(VecTy)->getElementType(); 4688 if (EltTy->isFloatTy() || EltTy->isDoubleTy() || EltTy->isIntegerTy(64) || 4689 EltTy->isIntegerTy(32) || EltTy->isPointerTy()) 4690 return true; 4691 if (EltTy->isIntegerTy(16) || EltTy->isIntegerTy(8)) 4692 return HasBW; 4693 return false; 4694 }; 4695 if (ST->hasAVX512() && isSupportedOnAVX512(VecTy, ST->hasBWI())) 4696 return getInterleavedMemoryOpCostAVX512( 4697 Opcode, cast<FixedVectorType>(VecTy), Factor, Indices, Alignment, 4698 AddressSpace, CostKind, UseMaskForCond, UseMaskForGaps); 4699 if (ST->hasAVX2()) 4700 return getInterleavedMemoryOpCostAVX2( 4701 Opcode, cast<FixedVectorType>(VecTy), Factor, Indices, Alignment, 4702 AddressSpace, CostKind, UseMaskForCond, UseMaskForGaps); 4703 4704 return BaseT::getInterleavedMemoryOpCost(Opcode, VecTy, Factor, Indices, 4705 Alignment, AddressSpace, CostKind, 4706 UseMaskForCond, UseMaskForGaps); 4707 } 4708