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 CmpInst::Predicate VecPred, 2088 TTI::TargetCostKind CostKind, 2089 const Instruction *I) { 2090 // TODO: Handle other cost kinds. 2091 if (CostKind != TTI::TCK_RecipThroughput) 2092 return BaseT::getCmpSelInstrCost(Opcode, ValTy, CondTy, VecPred, CostKind, 2093 I); 2094 2095 // Legalize the type. 2096 std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, ValTy); 2097 2098 MVT MTy = LT.second; 2099 2100 int ISD = TLI->InstructionOpcodeToISD(Opcode); 2101 assert(ISD && "Invalid opcode"); 2102 2103 unsigned ExtraCost = 0; 2104 if (I && (Opcode == Instruction::ICmp || Opcode == Instruction::FCmp)) { 2105 // Some vector comparison predicates cost extra instructions. 2106 if (MTy.isVector() && 2107 !((ST->hasXOP() && (!ST->hasAVX2() || MTy.is128BitVector())) || 2108 (ST->hasAVX512() && 32 <= MTy.getScalarSizeInBits()) || 2109 ST->hasBWI())) { 2110 switch (cast<CmpInst>(I)->getPredicate()) { 2111 case CmpInst::Predicate::ICMP_NE: 2112 // xor(cmpeq(x,y),-1) 2113 ExtraCost = 1; 2114 break; 2115 case CmpInst::Predicate::ICMP_SGE: 2116 case CmpInst::Predicate::ICMP_SLE: 2117 // xor(cmpgt(x,y),-1) 2118 ExtraCost = 1; 2119 break; 2120 case CmpInst::Predicate::ICMP_ULT: 2121 case CmpInst::Predicate::ICMP_UGT: 2122 // cmpgt(xor(x,signbit),xor(y,signbit)) 2123 // xor(cmpeq(pmaxu(x,y),x),-1) 2124 ExtraCost = 2; 2125 break; 2126 case CmpInst::Predicate::ICMP_ULE: 2127 case CmpInst::Predicate::ICMP_UGE: 2128 if ((ST->hasSSE41() && MTy.getScalarSizeInBits() == 32) || 2129 (ST->hasSSE2() && MTy.getScalarSizeInBits() < 32)) { 2130 // cmpeq(psubus(x,y),0) 2131 // cmpeq(pminu(x,y),x) 2132 ExtraCost = 1; 2133 } else { 2134 // xor(cmpgt(xor(x,signbit),xor(y,signbit)),-1) 2135 ExtraCost = 3; 2136 } 2137 break; 2138 default: 2139 break; 2140 } 2141 } 2142 } 2143 2144 static const CostTblEntry SLMCostTbl[] = { 2145 // slm pcmpeq/pcmpgt throughput is 2 2146 { ISD::SETCC, MVT::v2i64, 2 }, 2147 }; 2148 2149 static const CostTblEntry AVX512BWCostTbl[] = { 2150 { ISD::SETCC, MVT::v32i16, 1 }, 2151 { ISD::SETCC, MVT::v64i8, 1 }, 2152 2153 { ISD::SELECT, MVT::v32i16, 1 }, 2154 { ISD::SELECT, MVT::v64i8, 1 }, 2155 }; 2156 2157 static const CostTblEntry AVX512CostTbl[] = { 2158 { ISD::SETCC, MVT::v8i64, 1 }, 2159 { ISD::SETCC, MVT::v16i32, 1 }, 2160 { ISD::SETCC, MVT::v8f64, 1 }, 2161 { ISD::SETCC, MVT::v16f32, 1 }, 2162 2163 { ISD::SELECT, MVT::v8i64, 1 }, 2164 { ISD::SELECT, MVT::v16i32, 1 }, 2165 { ISD::SELECT, MVT::v8f64, 1 }, 2166 { ISD::SELECT, MVT::v16f32, 1 }, 2167 2168 { ISD::SETCC, MVT::v32i16, 2 }, // FIXME: should probably be 4 2169 { ISD::SETCC, MVT::v64i8, 2 }, // FIXME: should probably be 4 2170 2171 { ISD::SELECT, MVT::v32i16, 2 }, // FIXME: should be 3 2172 { ISD::SELECT, MVT::v64i8, 2 }, // FIXME: should be 3 2173 }; 2174 2175 static const CostTblEntry AVX2CostTbl[] = { 2176 { ISD::SETCC, MVT::v4i64, 1 }, 2177 { ISD::SETCC, MVT::v8i32, 1 }, 2178 { ISD::SETCC, MVT::v16i16, 1 }, 2179 { ISD::SETCC, MVT::v32i8, 1 }, 2180 2181 { ISD::SELECT, MVT::v4i64, 1 }, // pblendvb 2182 { ISD::SELECT, MVT::v8i32, 1 }, // pblendvb 2183 { ISD::SELECT, MVT::v16i16, 1 }, // pblendvb 2184 { ISD::SELECT, MVT::v32i8, 1 }, // pblendvb 2185 }; 2186 2187 static const CostTblEntry AVX1CostTbl[] = { 2188 { ISD::SETCC, MVT::v4f64, 1 }, 2189 { ISD::SETCC, MVT::v8f32, 1 }, 2190 // AVX1 does not support 8-wide integer compare. 2191 { ISD::SETCC, MVT::v4i64, 4 }, 2192 { ISD::SETCC, MVT::v8i32, 4 }, 2193 { ISD::SETCC, MVT::v16i16, 4 }, 2194 { ISD::SETCC, MVT::v32i8, 4 }, 2195 2196 { ISD::SELECT, MVT::v4f64, 1 }, // vblendvpd 2197 { ISD::SELECT, MVT::v8f32, 1 }, // vblendvps 2198 { ISD::SELECT, MVT::v4i64, 1 }, // vblendvpd 2199 { ISD::SELECT, MVT::v8i32, 1 }, // vblendvps 2200 { ISD::SELECT, MVT::v16i16, 3 }, // vandps + vandnps + vorps 2201 { ISD::SELECT, MVT::v32i8, 3 }, // vandps + vandnps + vorps 2202 }; 2203 2204 static const CostTblEntry SSE42CostTbl[] = { 2205 { ISD::SETCC, MVT::v2f64, 1 }, 2206 { ISD::SETCC, MVT::v4f32, 1 }, 2207 { ISD::SETCC, MVT::v2i64, 1 }, 2208 }; 2209 2210 static const CostTblEntry SSE41CostTbl[] = { 2211 { ISD::SELECT, MVT::v2f64, 1 }, // blendvpd 2212 { ISD::SELECT, MVT::v4f32, 1 }, // blendvps 2213 { ISD::SELECT, MVT::v2i64, 1 }, // pblendvb 2214 { ISD::SELECT, MVT::v4i32, 1 }, // pblendvb 2215 { ISD::SELECT, MVT::v8i16, 1 }, // pblendvb 2216 { ISD::SELECT, MVT::v16i8, 1 }, // pblendvb 2217 }; 2218 2219 static const CostTblEntry SSE2CostTbl[] = { 2220 { ISD::SETCC, MVT::v2f64, 2 }, 2221 { ISD::SETCC, MVT::f64, 1 }, 2222 { ISD::SETCC, MVT::v2i64, 8 }, 2223 { ISD::SETCC, MVT::v4i32, 1 }, 2224 { ISD::SETCC, MVT::v8i16, 1 }, 2225 { ISD::SETCC, MVT::v16i8, 1 }, 2226 2227 { ISD::SELECT, MVT::v2f64, 3 }, // andpd + andnpd + orpd 2228 { ISD::SELECT, MVT::v2i64, 3 }, // pand + pandn + por 2229 { ISD::SELECT, MVT::v4i32, 3 }, // pand + pandn + por 2230 { ISD::SELECT, MVT::v8i16, 3 }, // pand + pandn + por 2231 { ISD::SELECT, MVT::v16i8, 3 }, // pand + pandn + por 2232 }; 2233 2234 static const CostTblEntry SSE1CostTbl[] = { 2235 { ISD::SETCC, MVT::v4f32, 2 }, 2236 { ISD::SETCC, MVT::f32, 1 }, 2237 2238 { ISD::SELECT, MVT::v4f32, 3 }, // andps + andnps + orps 2239 }; 2240 2241 if (ST->isSLM()) 2242 if (const auto *Entry = CostTableLookup(SLMCostTbl, ISD, MTy)) 2243 return LT.first * (ExtraCost + Entry->Cost); 2244 2245 if (ST->hasBWI()) 2246 if (const auto *Entry = CostTableLookup(AVX512BWCostTbl, ISD, MTy)) 2247 return LT.first * (ExtraCost + Entry->Cost); 2248 2249 if (ST->hasAVX512()) 2250 if (const auto *Entry = CostTableLookup(AVX512CostTbl, ISD, MTy)) 2251 return LT.first * (ExtraCost + Entry->Cost); 2252 2253 if (ST->hasAVX2()) 2254 if (const auto *Entry = CostTableLookup(AVX2CostTbl, ISD, MTy)) 2255 return LT.first * (ExtraCost + Entry->Cost); 2256 2257 if (ST->hasAVX()) 2258 if (const auto *Entry = CostTableLookup(AVX1CostTbl, ISD, MTy)) 2259 return LT.first * (ExtraCost + Entry->Cost); 2260 2261 if (ST->hasSSE42()) 2262 if (const auto *Entry = CostTableLookup(SSE42CostTbl, ISD, MTy)) 2263 return LT.first * (ExtraCost + Entry->Cost); 2264 2265 if (ST->hasSSE41()) 2266 if (const auto *Entry = CostTableLookup(SSE41CostTbl, ISD, MTy)) 2267 return LT.first * (ExtraCost + Entry->Cost); 2268 2269 if (ST->hasSSE2()) 2270 if (const auto *Entry = CostTableLookup(SSE2CostTbl, ISD, MTy)) 2271 return LT.first * (ExtraCost + Entry->Cost); 2272 2273 if (ST->hasSSE1()) 2274 if (const auto *Entry = CostTableLookup(SSE1CostTbl, ISD, MTy)) 2275 return LT.first * (ExtraCost + Entry->Cost); 2276 2277 return BaseT::getCmpSelInstrCost(Opcode, ValTy, CondTy, VecPred, CostKind, I); 2278 } 2279 2280 unsigned X86TTIImpl::getAtomicMemIntrinsicMaxElementSize() const { return 16; } 2281 2282 int X86TTIImpl::getTypeBasedIntrinsicInstrCost( 2283 const IntrinsicCostAttributes &ICA, TTI::TargetCostKind CostKind) { 2284 2285 // Costs should match the codegen from: 2286 // BITREVERSE: llvm\test\CodeGen\X86\vector-bitreverse.ll 2287 // BSWAP: llvm\test\CodeGen\X86\bswap-vector.ll 2288 // CTLZ: llvm\test\CodeGen\X86\vector-lzcnt-*.ll 2289 // CTPOP: llvm\test\CodeGen\X86\vector-popcnt-*.ll 2290 // CTTZ: llvm\test\CodeGen\X86\vector-tzcnt-*.ll 2291 static const CostTblEntry AVX512CDCostTbl[] = { 2292 { ISD::CTLZ, MVT::v8i64, 1 }, 2293 { ISD::CTLZ, MVT::v16i32, 1 }, 2294 { ISD::CTLZ, MVT::v32i16, 8 }, 2295 { ISD::CTLZ, MVT::v64i8, 20 }, 2296 { ISD::CTLZ, MVT::v4i64, 1 }, 2297 { ISD::CTLZ, MVT::v8i32, 1 }, 2298 { ISD::CTLZ, MVT::v16i16, 4 }, 2299 { ISD::CTLZ, MVT::v32i8, 10 }, 2300 { ISD::CTLZ, MVT::v2i64, 1 }, 2301 { ISD::CTLZ, MVT::v4i32, 1 }, 2302 { ISD::CTLZ, MVT::v8i16, 4 }, 2303 { ISD::CTLZ, MVT::v16i8, 4 }, 2304 }; 2305 static const CostTblEntry AVX512BWCostTbl[] = { 2306 { ISD::ABS, MVT::v32i16, 1 }, 2307 { ISD::ABS, MVT::v64i8, 1 }, 2308 { ISD::BITREVERSE, MVT::v8i64, 5 }, 2309 { ISD::BITREVERSE, MVT::v16i32, 5 }, 2310 { ISD::BITREVERSE, MVT::v32i16, 5 }, 2311 { ISD::BITREVERSE, MVT::v64i8, 5 }, 2312 { ISD::CTLZ, MVT::v8i64, 23 }, 2313 { ISD::CTLZ, MVT::v16i32, 22 }, 2314 { ISD::CTLZ, MVT::v32i16, 18 }, 2315 { ISD::CTLZ, MVT::v64i8, 17 }, 2316 { ISD::CTPOP, MVT::v8i64, 7 }, 2317 { ISD::CTPOP, MVT::v16i32, 11 }, 2318 { ISD::CTPOP, MVT::v32i16, 9 }, 2319 { ISD::CTPOP, MVT::v64i8, 6 }, 2320 { ISD::CTTZ, MVT::v8i64, 10 }, 2321 { ISD::CTTZ, MVT::v16i32, 14 }, 2322 { ISD::CTTZ, MVT::v32i16, 12 }, 2323 { ISD::CTTZ, MVT::v64i8, 9 }, 2324 { ISD::SADDSAT, MVT::v32i16, 1 }, 2325 { ISD::SADDSAT, MVT::v64i8, 1 }, 2326 { ISD::SMAX, MVT::v32i16, 1 }, 2327 { ISD::SMAX, MVT::v64i8, 1 }, 2328 { ISD::SMIN, MVT::v32i16, 1 }, 2329 { ISD::SMIN, MVT::v64i8, 1 }, 2330 { ISD::SSUBSAT, MVT::v32i16, 1 }, 2331 { ISD::SSUBSAT, MVT::v64i8, 1 }, 2332 { ISD::UADDSAT, MVT::v32i16, 1 }, 2333 { ISD::UADDSAT, MVT::v64i8, 1 }, 2334 { ISD::UMAX, MVT::v32i16, 1 }, 2335 { ISD::UMAX, MVT::v64i8, 1 }, 2336 { ISD::UMIN, MVT::v32i16, 1 }, 2337 { ISD::UMIN, MVT::v64i8, 1 }, 2338 { ISD::USUBSAT, MVT::v32i16, 1 }, 2339 { ISD::USUBSAT, MVT::v64i8, 1 }, 2340 }; 2341 static const CostTblEntry AVX512CostTbl[] = { 2342 { ISD::ABS, MVT::v8i64, 1 }, 2343 { ISD::ABS, MVT::v16i32, 1 }, 2344 { ISD::ABS, MVT::v32i16, 2 }, // FIXME: include split 2345 { ISD::ABS, MVT::v64i8, 2 }, // FIXME: include split 2346 { ISD::ABS, MVT::v4i64, 1 }, 2347 { ISD::ABS, MVT::v2i64, 1 }, 2348 { ISD::BITREVERSE, MVT::v8i64, 36 }, 2349 { ISD::BITREVERSE, MVT::v16i32, 24 }, 2350 { ISD::BITREVERSE, MVT::v32i16, 10 }, 2351 { ISD::BITREVERSE, MVT::v64i8, 10 }, 2352 { ISD::CTLZ, MVT::v8i64, 29 }, 2353 { ISD::CTLZ, MVT::v16i32, 35 }, 2354 { ISD::CTLZ, MVT::v32i16, 28 }, 2355 { ISD::CTLZ, MVT::v64i8, 18 }, 2356 { ISD::CTPOP, MVT::v8i64, 16 }, 2357 { ISD::CTPOP, MVT::v16i32, 24 }, 2358 { ISD::CTPOP, MVT::v32i16, 18 }, 2359 { ISD::CTPOP, MVT::v64i8, 12 }, 2360 { ISD::CTTZ, MVT::v8i64, 20 }, 2361 { ISD::CTTZ, MVT::v16i32, 28 }, 2362 { ISD::CTTZ, MVT::v32i16, 24 }, 2363 { ISD::CTTZ, MVT::v64i8, 18 }, 2364 { ISD::SMAX, MVT::v8i64, 1 }, 2365 { ISD::SMAX, MVT::v16i32, 1 }, 2366 { ISD::SMAX, MVT::v32i16, 2 }, // FIXME: include split 2367 { ISD::SMAX, MVT::v64i8, 2 }, // FIXME: include split 2368 { ISD::SMAX, MVT::v4i64, 1 }, 2369 { ISD::SMAX, MVT::v2i64, 1 }, 2370 { ISD::SMIN, MVT::v8i64, 1 }, 2371 { ISD::SMIN, MVT::v16i32, 1 }, 2372 { ISD::SMIN, MVT::v32i16, 2 }, // FIXME: include split 2373 { ISD::SMIN, MVT::v64i8, 2 }, // FIXME: include split 2374 { ISD::SMIN, MVT::v4i64, 1 }, 2375 { ISD::SMIN, MVT::v2i64, 1 }, 2376 { ISD::UMAX, MVT::v8i64, 1 }, 2377 { ISD::UMAX, MVT::v16i32, 1 }, 2378 { ISD::UMAX, MVT::v32i16, 2 }, // FIXME: include split 2379 { ISD::UMAX, MVT::v64i8, 2 }, // FIXME: include split 2380 { ISD::UMAX, MVT::v4i64, 1 }, 2381 { ISD::UMAX, MVT::v2i64, 1 }, 2382 { ISD::UMIN, MVT::v8i64, 1 }, 2383 { ISD::UMIN, MVT::v16i32, 1 }, 2384 { ISD::UMIN, MVT::v32i16, 2 }, // FIXME: include split 2385 { ISD::UMIN, MVT::v64i8, 2 }, // FIXME: include split 2386 { ISD::UMIN, MVT::v4i64, 1 }, 2387 { ISD::UMIN, MVT::v2i64, 1 }, 2388 { ISD::USUBSAT, MVT::v16i32, 2 }, // pmaxud + psubd 2389 { ISD::USUBSAT, MVT::v2i64, 2 }, // pmaxuq + psubq 2390 { ISD::USUBSAT, MVT::v4i64, 2 }, // pmaxuq + psubq 2391 { ISD::USUBSAT, MVT::v8i64, 2 }, // pmaxuq + psubq 2392 { ISD::UADDSAT, MVT::v16i32, 3 }, // not + pminud + paddd 2393 { ISD::UADDSAT, MVT::v2i64, 3 }, // not + pminuq + paddq 2394 { ISD::UADDSAT, MVT::v4i64, 3 }, // not + pminuq + paddq 2395 { ISD::UADDSAT, MVT::v8i64, 3 }, // not + pminuq + paddq 2396 { ISD::SADDSAT, MVT::v32i16, 2 }, // FIXME: include split 2397 { ISD::SADDSAT, MVT::v64i8, 2 }, // FIXME: include split 2398 { ISD::SSUBSAT, MVT::v32i16, 2 }, // FIXME: include split 2399 { ISD::SSUBSAT, MVT::v64i8, 2 }, // FIXME: include split 2400 { ISD::UADDSAT, MVT::v32i16, 2 }, // FIXME: include split 2401 { ISD::UADDSAT, MVT::v64i8, 2 }, // FIXME: include split 2402 { ISD::USUBSAT, MVT::v32i16, 2 }, // FIXME: include split 2403 { ISD::USUBSAT, MVT::v64i8, 2 }, // FIXME: include split 2404 { ISD::FMAXNUM, MVT::f32, 2 }, 2405 { ISD::FMAXNUM, MVT::v4f32, 2 }, 2406 { ISD::FMAXNUM, MVT::v8f32, 2 }, 2407 { ISD::FMAXNUM, MVT::v16f32, 2 }, 2408 { ISD::FMAXNUM, MVT::f64, 2 }, 2409 { ISD::FMAXNUM, MVT::v2f64, 2 }, 2410 { ISD::FMAXNUM, MVT::v4f64, 2 }, 2411 { ISD::FMAXNUM, MVT::v8f64, 2 }, 2412 }; 2413 static const CostTblEntry XOPCostTbl[] = { 2414 { ISD::BITREVERSE, MVT::v4i64, 4 }, 2415 { ISD::BITREVERSE, MVT::v8i32, 4 }, 2416 { ISD::BITREVERSE, MVT::v16i16, 4 }, 2417 { ISD::BITREVERSE, MVT::v32i8, 4 }, 2418 { ISD::BITREVERSE, MVT::v2i64, 1 }, 2419 { ISD::BITREVERSE, MVT::v4i32, 1 }, 2420 { ISD::BITREVERSE, MVT::v8i16, 1 }, 2421 { ISD::BITREVERSE, MVT::v16i8, 1 }, 2422 { ISD::BITREVERSE, MVT::i64, 3 }, 2423 { ISD::BITREVERSE, MVT::i32, 3 }, 2424 { ISD::BITREVERSE, MVT::i16, 3 }, 2425 { ISD::BITREVERSE, MVT::i8, 3 } 2426 }; 2427 static const CostTblEntry AVX2CostTbl[] = { 2428 { ISD::ABS, MVT::v4i64, 2 }, // VBLENDVPD(X,VPSUBQ(0,X),X) 2429 { ISD::ABS, MVT::v8i32, 1 }, 2430 { ISD::ABS, MVT::v16i16, 1 }, 2431 { ISD::ABS, MVT::v32i8, 1 }, 2432 { ISD::BITREVERSE, MVT::v4i64, 5 }, 2433 { ISD::BITREVERSE, MVT::v8i32, 5 }, 2434 { ISD::BITREVERSE, MVT::v16i16, 5 }, 2435 { ISD::BITREVERSE, MVT::v32i8, 5 }, 2436 { ISD::BSWAP, MVT::v4i64, 1 }, 2437 { ISD::BSWAP, MVT::v8i32, 1 }, 2438 { ISD::BSWAP, MVT::v16i16, 1 }, 2439 { ISD::CTLZ, MVT::v4i64, 23 }, 2440 { ISD::CTLZ, MVT::v8i32, 18 }, 2441 { ISD::CTLZ, MVT::v16i16, 14 }, 2442 { ISD::CTLZ, MVT::v32i8, 9 }, 2443 { ISD::CTPOP, MVT::v4i64, 7 }, 2444 { ISD::CTPOP, MVT::v8i32, 11 }, 2445 { ISD::CTPOP, MVT::v16i16, 9 }, 2446 { ISD::CTPOP, MVT::v32i8, 6 }, 2447 { ISD::CTTZ, MVT::v4i64, 10 }, 2448 { ISD::CTTZ, MVT::v8i32, 14 }, 2449 { ISD::CTTZ, MVT::v16i16, 12 }, 2450 { ISD::CTTZ, MVT::v32i8, 9 }, 2451 { ISD::SADDSAT, MVT::v16i16, 1 }, 2452 { ISD::SADDSAT, MVT::v32i8, 1 }, 2453 { ISD::SMAX, MVT::v8i32, 1 }, 2454 { ISD::SMAX, MVT::v16i16, 1 }, 2455 { ISD::SMAX, MVT::v32i8, 1 }, 2456 { ISD::SMIN, MVT::v8i32, 1 }, 2457 { ISD::SMIN, MVT::v16i16, 1 }, 2458 { ISD::SMIN, MVT::v32i8, 1 }, 2459 { ISD::SSUBSAT, MVT::v16i16, 1 }, 2460 { ISD::SSUBSAT, MVT::v32i8, 1 }, 2461 { ISD::UADDSAT, MVT::v16i16, 1 }, 2462 { ISD::UADDSAT, MVT::v32i8, 1 }, 2463 { ISD::UADDSAT, MVT::v8i32, 3 }, // not + pminud + paddd 2464 { ISD::UMAX, MVT::v8i32, 1 }, 2465 { ISD::UMAX, MVT::v16i16, 1 }, 2466 { ISD::UMAX, MVT::v32i8, 1 }, 2467 { ISD::UMIN, MVT::v8i32, 1 }, 2468 { ISD::UMIN, MVT::v16i16, 1 }, 2469 { ISD::UMIN, MVT::v32i8, 1 }, 2470 { ISD::USUBSAT, MVT::v16i16, 1 }, 2471 { ISD::USUBSAT, MVT::v32i8, 1 }, 2472 { ISD::USUBSAT, MVT::v8i32, 2 }, // pmaxud + psubd 2473 { ISD::FSQRT, MVT::f32, 7 }, // Haswell from http://www.agner.org/ 2474 { ISD::FSQRT, MVT::v4f32, 7 }, // Haswell from http://www.agner.org/ 2475 { ISD::FSQRT, MVT::v8f32, 14 }, // Haswell from http://www.agner.org/ 2476 { ISD::FSQRT, MVT::f64, 14 }, // Haswell from http://www.agner.org/ 2477 { ISD::FSQRT, MVT::v2f64, 14 }, // Haswell from http://www.agner.org/ 2478 { ISD::FSQRT, MVT::v4f64, 28 }, // Haswell from http://www.agner.org/ 2479 }; 2480 static const CostTblEntry AVX1CostTbl[] = { 2481 { ISD::ABS, MVT::v4i64, 6 }, // VBLENDVPD(X,VPSUBQ(0,X),X) 2482 { ISD::ABS, MVT::v8i32, 3 }, 2483 { ISD::ABS, MVT::v16i16, 3 }, 2484 { ISD::ABS, MVT::v32i8, 3 }, 2485 { ISD::BITREVERSE, MVT::v4i64, 12 }, // 2 x 128-bit Op + extract/insert 2486 { ISD::BITREVERSE, MVT::v8i32, 12 }, // 2 x 128-bit Op + extract/insert 2487 { ISD::BITREVERSE, MVT::v16i16, 12 }, // 2 x 128-bit Op + extract/insert 2488 { ISD::BITREVERSE, MVT::v32i8, 12 }, // 2 x 128-bit Op + extract/insert 2489 { ISD::BSWAP, MVT::v4i64, 4 }, 2490 { ISD::BSWAP, MVT::v8i32, 4 }, 2491 { ISD::BSWAP, MVT::v16i16, 4 }, 2492 { ISD::CTLZ, MVT::v4i64, 48 }, // 2 x 128-bit Op + extract/insert 2493 { ISD::CTLZ, MVT::v8i32, 38 }, // 2 x 128-bit Op + extract/insert 2494 { ISD::CTLZ, MVT::v16i16, 30 }, // 2 x 128-bit Op + extract/insert 2495 { ISD::CTLZ, MVT::v32i8, 20 }, // 2 x 128-bit Op + extract/insert 2496 { ISD::CTPOP, MVT::v4i64, 16 }, // 2 x 128-bit Op + extract/insert 2497 { ISD::CTPOP, MVT::v8i32, 24 }, // 2 x 128-bit Op + extract/insert 2498 { ISD::CTPOP, MVT::v16i16, 20 }, // 2 x 128-bit Op + extract/insert 2499 { ISD::CTPOP, MVT::v32i8, 14 }, // 2 x 128-bit Op + extract/insert 2500 { ISD::CTTZ, MVT::v4i64, 22 }, // 2 x 128-bit Op + extract/insert 2501 { ISD::CTTZ, MVT::v8i32, 30 }, // 2 x 128-bit Op + extract/insert 2502 { ISD::CTTZ, MVT::v16i16, 26 }, // 2 x 128-bit Op + extract/insert 2503 { ISD::CTTZ, MVT::v32i8, 20 }, // 2 x 128-bit Op + extract/insert 2504 { ISD::SADDSAT, MVT::v16i16, 4 }, // 2 x 128-bit Op + extract/insert 2505 { ISD::SADDSAT, MVT::v32i8, 4 }, // 2 x 128-bit Op + extract/insert 2506 { ISD::SMAX, MVT::v8i32, 4 }, // 2 x 128-bit Op + extract/insert 2507 { ISD::SMAX, MVT::v16i16, 4 }, // 2 x 128-bit Op + extract/insert 2508 { ISD::SMAX, MVT::v32i8, 4 }, // 2 x 128-bit Op + extract/insert 2509 { ISD::SMIN, MVT::v8i32, 4 }, // 2 x 128-bit Op + extract/insert 2510 { ISD::SMIN, MVT::v16i16, 4 }, // 2 x 128-bit Op + extract/insert 2511 { ISD::SMIN, MVT::v32i8, 4 }, // 2 x 128-bit Op + extract/insert 2512 { ISD::SSUBSAT, MVT::v16i16, 4 }, // 2 x 128-bit Op + extract/insert 2513 { ISD::SSUBSAT, MVT::v32i8, 4 }, // 2 x 128-bit Op + extract/insert 2514 { ISD::UADDSAT, MVT::v16i16, 4 }, // 2 x 128-bit Op + extract/insert 2515 { ISD::UADDSAT, MVT::v32i8, 4 }, // 2 x 128-bit Op + extract/insert 2516 { ISD::UADDSAT, MVT::v8i32, 8 }, // 2 x 128-bit Op + extract/insert 2517 { ISD::UMAX, MVT::v8i32, 4 }, // 2 x 128-bit Op + extract/insert 2518 { ISD::UMAX, MVT::v16i16, 4 }, // 2 x 128-bit Op + extract/insert 2519 { ISD::UMAX, MVT::v32i8, 4 }, // 2 x 128-bit Op + extract/insert 2520 { ISD::UMIN, MVT::v8i32, 4 }, // 2 x 128-bit Op + extract/insert 2521 { ISD::UMIN, MVT::v16i16, 4 }, // 2 x 128-bit Op + extract/insert 2522 { ISD::UMIN, MVT::v32i8, 4 }, // 2 x 128-bit Op + extract/insert 2523 { ISD::USUBSAT, MVT::v16i16, 4 }, // 2 x 128-bit Op + extract/insert 2524 { ISD::USUBSAT, MVT::v32i8, 4 }, // 2 x 128-bit Op + extract/insert 2525 { ISD::USUBSAT, MVT::v8i32, 6 }, // 2 x 128-bit Op + extract/insert 2526 { ISD::FMAXNUM, MVT::f32, 3 }, 2527 { ISD::FMAXNUM, MVT::v4f32, 3 }, 2528 { ISD::FMAXNUM, MVT::v8f32, 5 }, 2529 { ISD::FMAXNUM, MVT::f64, 3 }, 2530 { ISD::FMAXNUM, MVT::v2f64, 3 }, 2531 { ISD::FMAXNUM, MVT::v4f64, 5 }, 2532 { ISD::FSQRT, MVT::f32, 14 }, // SNB from http://www.agner.org/ 2533 { ISD::FSQRT, MVT::v4f32, 14 }, // SNB from http://www.agner.org/ 2534 { ISD::FSQRT, MVT::v8f32, 28 }, // SNB from http://www.agner.org/ 2535 { ISD::FSQRT, MVT::f64, 21 }, // SNB from http://www.agner.org/ 2536 { ISD::FSQRT, MVT::v2f64, 21 }, // SNB from http://www.agner.org/ 2537 { ISD::FSQRT, MVT::v4f64, 43 }, // SNB from http://www.agner.org/ 2538 }; 2539 static const CostTblEntry GLMCostTbl[] = { 2540 { ISD::FSQRT, MVT::f32, 19 }, // sqrtss 2541 { ISD::FSQRT, MVT::v4f32, 37 }, // sqrtps 2542 { ISD::FSQRT, MVT::f64, 34 }, // sqrtsd 2543 { ISD::FSQRT, MVT::v2f64, 67 }, // sqrtpd 2544 }; 2545 static const CostTblEntry SLMCostTbl[] = { 2546 { ISD::FSQRT, MVT::f32, 20 }, // sqrtss 2547 { ISD::FSQRT, MVT::v4f32, 40 }, // sqrtps 2548 { ISD::FSQRT, MVT::f64, 35 }, // sqrtsd 2549 { ISD::FSQRT, MVT::v2f64, 70 }, // sqrtpd 2550 }; 2551 static const CostTblEntry SSE42CostTbl[] = { 2552 { ISD::ABS, MVT::v2i64, 3 }, // BLENDVPD(X,PSUBQ(0,X),X) 2553 { ISD::USUBSAT, MVT::v4i32, 2 }, // pmaxud + psubd 2554 { ISD::UADDSAT, MVT::v4i32, 3 }, // not + pminud + paddd 2555 { ISD::FSQRT, MVT::f32, 18 }, // Nehalem from http://www.agner.org/ 2556 { ISD::FSQRT, MVT::v4f32, 18 }, // Nehalem from http://www.agner.org/ 2557 }; 2558 static const CostTblEntry SSE41CostTbl[] = { 2559 { ISD::SMAX, MVT::v4i32, 1 }, 2560 { ISD::SMAX, MVT::v16i8, 1 }, 2561 { ISD::SMIN, MVT::v4i32, 1 }, 2562 { ISD::SMIN, MVT::v16i8, 1 }, 2563 { ISD::UMAX, MVT::v4i32, 1 }, 2564 { ISD::UMAX, MVT::v8i16, 1 }, 2565 { ISD::UMIN, MVT::v4i32, 1 }, 2566 { ISD::UMIN, MVT::v8i16, 1 }, 2567 }; 2568 static const CostTblEntry SSSE3CostTbl[] = { 2569 { ISD::ABS, MVT::v4i32, 1 }, 2570 { ISD::ABS, MVT::v8i16, 1 }, 2571 { ISD::ABS, MVT::v16i8, 1 }, 2572 { ISD::BITREVERSE, MVT::v2i64, 5 }, 2573 { ISD::BITREVERSE, MVT::v4i32, 5 }, 2574 { ISD::BITREVERSE, MVT::v8i16, 5 }, 2575 { ISD::BITREVERSE, MVT::v16i8, 5 }, 2576 { ISD::BSWAP, MVT::v2i64, 1 }, 2577 { ISD::BSWAP, MVT::v4i32, 1 }, 2578 { ISD::BSWAP, MVT::v8i16, 1 }, 2579 { ISD::CTLZ, MVT::v2i64, 23 }, 2580 { ISD::CTLZ, MVT::v4i32, 18 }, 2581 { ISD::CTLZ, MVT::v8i16, 14 }, 2582 { ISD::CTLZ, MVT::v16i8, 9 }, 2583 { ISD::CTPOP, MVT::v2i64, 7 }, 2584 { ISD::CTPOP, MVT::v4i32, 11 }, 2585 { ISD::CTPOP, MVT::v8i16, 9 }, 2586 { ISD::CTPOP, MVT::v16i8, 6 }, 2587 { ISD::CTTZ, MVT::v2i64, 10 }, 2588 { ISD::CTTZ, MVT::v4i32, 14 }, 2589 { ISD::CTTZ, MVT::v8i16, 12 }, 2590 { ISD::CTTZ, MVT::v16i8, 9 } 2591 }; 2592 static const CostTblEntry SSE2CostTbl[] = { 2593 { ISD::ABS, MVT::v2i64, 4 }, 2594 { ISD::ABS, MVT::v4i32, 3 }, 2595 { ISD::ABS, MVT::v8i16, 3 }, 2596 { ISD::ABS, MVT::v16i8, 3 }, 2597 { ISD::BITREVERSE, MVT::v2i64, 29 }, 2598 { ISD::BITREVERSE, MVT::v4i32, 27 }, 2599 { ISD::BITREVERSE, MVT::v8i16, 27 }, 2600 { ISD::BITREVERSE, MVT::v16i8, 20 }, 2601 { ISD::BSWAP, MVT::v2i64, 7 }, 2602 { ISD::BSWAP, MVT::v4i32, 7 }, 2603 { ISD::BSWAP, MVT::v8i16, 7 }, 2604 { ISD::CTLZ, MVT::v2i64, 25 }, 2605 { ISD::CTLZ, MVT::v4i32, 26 }, 2606 { ISD::CTLZ, MVT::v8i16, 20 }, 2607 { ISD::CTLZ, MVT::v16i8, 17 }, 2608 { ISD::CTPOP, MVT::v2i64, 12 }, 2609 { ISD::CTPOP, MVT::v4i32, 15 }, 2610 { ISD::CTPOP, MVT::v8i16, 13 }, 2611 { ISD::CTPOP, MVT::v16i8, 10 }, 2612 { ISD::CTTZ, MVT::v2i64, 14 }, 2613 { ISD::CTTZ, MVT::v4i32, 18 }, 2614 { ISD::CTTZ, MVT::v8i16, 16 }, 2615 { ISD::CTTZ, MVT::v16i8, 13 }, 2616 { ISD::SADDSAT, MVT::v8i16, 1 }, 2617 { ISD::SADDSAT, MVT::v16i8, 1 }, 2618 { ISD::SMAX, MVT::v8i16, 1 }, 2619 { ISD::SMIN, MVT::v8i16, 1 }, 2620 { ISD::SSUBSAT, MVT::v8i16, 1 }, 2621 { ISD::SSUBSAT, MVT::v16i8, 1 }, 2622 { ISD::UADDSAT, MVT::v8i16, 1 }, 2623 { ISD::UADDSAT, MVT::v16i8, 1 }, 2624 { ISD::UMAX, MVT::v8i16, 2 }, 2625 { ISD::UMAX, MVT::v16i8, 1 }, 2626 { ISD::UMIN, MVT::v8i16, 2 }, 2627 { ISD::UMIN, MVT::v16i8, 1 }, 2628 { ISD::USUBSAT, MVT::v8i16, 1 }, 2629 { ISD::USUBSAT, MVT::v16i8, 1 }, 2630 { ISD::FMAXNUM, MVT::f64, 4 }, 2631 { ISD::FMAXNUM, MVT::v2f64, 4 }, 2632 { ISD::FSQRT, MVT::f64, 32 }, // Nehalem from http://www.agner.org/ 2633 { ISD::FSQRT, MVT::v2f64, 32 }, // Nehalem from http://www.agner.org/ 2634 }; 2635 static const CostTblEntry SSE1CostTbl[] = { 2636 { ISD::FMAXNUM, MVT::f32, 4 }, 2637 { ISD::FMAXNUM, MVT::v4f32, 4 }, 2638 { ISD::FSQRT, MVT::f32, 28 }, // Pentium III from http://www.agner.org/ 2639 { ISD::FSQRT, MVT::v4f32, 56 }, // Pentium III from http://www.agner.org/ 2640 }; 2641 static const CostTblEntry BMI64CostTbl[] = { // 64-bit targets 2642 { ISD::CTTZ, MVT::i64, 1 }, 2643 }; 2644 static const CostTblEntry BMI32CostTbl[] = { // 32 or 64-bit targets 2645 { ISD::CTTZ, MVT::i32, 1 }, 2646 { ISD::CTTZ, MVT::i16, 1 }, 2647 { ISD::CTTZ, MVT::i8, 1 }, 2648 }; 2649 static const CostTblEntry LZCNT64CostTbl[] = { // 64-bit targets 2650 { ISD::CTLZ, MVT::i64, 1 }, 2651 }; 2652 static const CostTblEntry LZCNT32CostTbl[] = { // 32 or 64-bit targets 2653 { ISD::CTLZ, MVT::i32, 1 }, 2654 { ISD::CTLZ, MVT::i16, 1 }, 2655 { ISD::CTLZ, MVT::i8, 1 }, 2656 }; 2657 static const CostTblEntry POPCNT64CostTbl[] = { // 64-bit targets 2658 { ISD::CTPOP, MVT::i64, 1 }, 2659 }; 2660 static const CostTblEntry POPCNT32CostTbl[] = { // 32 or 64-bit targets 2661 { ISD::CTPOP, MVT::i32, 1 }, 2662 { ISD::CTPOP, MVT::i16, 1 }, 2663 { ISD::CTPOP, MVT::i8, 1 }, 2664 }; 2665 static const CostTblEntry X64CostTbl[] = { // 64-bit targets 2666 { ISD::BITREVERSE, MVT::i64, 14 }, 2667 { ISD::CTLZ, MVT::i64, 4 }, // BSR+XOR or BSR+XOR+CMOV 2668 { ISD::CTTZ, MVT::i64, 3 }, // TEST+BSF+CMOV/BRANCH 2669 { ISD::CTPOP, MVT::i64, 10 }, 2670 { ISD::SADDO, MVT::i64, 1 }, 2671 { ISD::UADDO, MVT::i64, 1 }, 2672 }; 2673 static const CostTblEntry X86CostTbl[] = { // 32 or 64-bit targets 2674 { ISD::BITREVERSE, MVT::i32, 14 }, 2675 { ISD::BITREVERSE, MVT::i16, 14 }, 2676 { ISD::BITREVERSE, MVT::i8, 11 }, 2677 { ISD::CTLZ, MVT::i32, 4 }, // BSR+XOR or BSR+XOR+CMOV 2678 { ISD::CTLZ, MVT::i16, 4 }, // BSR+XOR or BSR+XOR+CMOV 2679 { ISD::CTLZ, MVT::i8, 4 }, // BSR+XOR or BSR+XOR+CMOV 2680 { ISD::CTTZ, MVT::i32, 3 }, // TEST+BSF+CMOV/BRANCH 2681 { ISD::CTTZ, MVT::i16, 3 }, // TEST+BSF+CMOV/BRANCH 2682 { ISD::CTTZ, MVT::i8, 3 }, // TEST+BSF+CMOV/BRANCH 2683 { ISD::CTPOP, MVT::i32, 8 }, 2684 { ISD::CTPOP, MVT::i16, 9 }, 2685 { ISD::CTPOP, MVT::i8, 7 }, 2686 { ISD::SADDO, MVT::i32, 1 }, 2687 { ISD::SADDO, MVT::i16, 1 }, 2688 { ISD::SADDO, MVT::i8, 1 }, 2689 { ISD::UADDO, MVT::i32, 1 }, 2690 { ISD::UADDO, MVT::i16, 1 }, 2691 { ISD::UADDO, MVT::i8, 1 }, 2692 }; 2693 2694 Type *RetTy = ICA.getReturnType(); 2695 Type *OpTy = RetTy; 2696 Intrinsic::ID IID = ICA.getID(); 2697 unsigned ISD = ISD::DELETED_NODE; 2698 switch (IID) { 2699 default: 2700 break; 2701 case Intrinsic::abs: 2702 ISD = ISD::ABS; 2703 break; 2704 case Intrinsic::bitreverse: 2705 ISD = ISD::BITREVERSE; 2706 break; 2707 case Intrinsic::bswap: 2708 ISD = ISD::BSWAP; 2709 break; 2710 case Intrinsic::ctlz: 2711 ISD = ISD::CTLZ; 2712 break; 2713 case Intrinsic::ctpop: 2714 ISD = ISD::CTPOP; 2715 break; 2716 case Intrinsic::cttz: 2717 ISD = ISD::CTTZ; 2718 break; 2719 case Intrinsic::maxnum: 2720 case Intrinsic::minnum: 2721 // FMINNUM has same costs so don't duplicate. 2722 ISD = ISD::FMAXNUM; 2723 break; 2724 case Intrinsic::sadd_sat: 2725 ISD = ISD::SADDSAT; 2726 break; 2727 case Intrinsic::smax: 2728 ISD = ISD::SMAX; 2729 break; 2730 case Intrinsic::smin: 2731 ISD = ISD::SMIN; 2732 break; 2733 case Intrinsic::ssub_sat: 2734 ISD = ISD::SSUBSAT; 2735 break; 2736 case Intrinsic::uadd_sat: 2737 ISD = ISD::UADDSAT; 2738 break; 2739 case Intrinsic::umax: 2740 ISD = ISD::UMAX; 2741 break; 2742 case Intrinsic::umin: 2743 ISD = ISD::UMIN; 2744 break; 2745 case Intrinsic::usub_sat: 2746 ISD = ISD::USUBSAT; 2747 break; 2748 case Intrinsic::sqrt: 2749 ISD = ISD::FSQRT; 2750 break; 2751 case Intrinsic::sadd_with_overflow: 2752 case Intrinsic::ssub_with_overflow: 2753 // SSUBO has same costs so don't duplicate. 2754 ISD = ISD::SADDO; 2755 OpTy = RetTy->getContainedType(0); 2756 break; 2757 case Intrinsic::uadd_with_overflow: 2758 case Intrinsic::usub_with_overflow: 2759 // USUBO has same costs so don't duplicate. 2760 ISD = ISD::UADDO; 2761 OpTy = RetTy->getContainedType(0); 2762 break; 2763 } 2764 2765 if (ISD != ISD::DELETED_NODE) { 2766 // Legalize the type. 2767 std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, OpTy); 2768 MVT MTy = LT.second; 2769 2770 // Attempt to lookup cost. 2771 if (ST->useGLMDivSqrtCosts()) 2772 if (const auto *Entry = CostTableLookup(GLMCostTbl, ISD, MTy)) 2773 return LT.first * Entry->Cost; 2774 2775 if (ST->isSLM()) 2776 if (const auto *Entry = CostTableLookup(SLMCostTbl, ISD, MTy)) 2777 return LT.first * Entry->Cost; 2778 2779 if (ST->hasCDI()) 2780 if (const auto *Entry = CostTableLookup(AVX512CDCostTbl, ISD, MTy)) 2781 return LT.first * Entry->Cost; 2782 2783 if (ST->hasBWI()) 2784 if (const auto *Entry = CostTableLookup(AVX512BWCostTbl, ISD, MTy)) 2785 return LT.first * Entry->Cost; 2786 2787 if (ST->hasAVX512()) 2788 if (const auto *Entry = CostTableLookup(AVX512CostTbl, ISD, MTy)) 2789 return LT.first * Entry->Cost; 2790 2791 if (ST->hasXOP()) 2792 if (const auto *Entry = CostTableLookup(XOPCostTbl, ISD, MTy)) 2793 return LT.first * Entry->Cost; 2794 2795 if (ST->hasAVX2()) 2796 if (const auto *Entry = CostTableLookup(AVX2CostTbl, ISD, MTy)) 2797 return LT.first * Entry->Cost; 2798 2799 if (ST->hasAVX()) 2800 if (const auto *Entry = CostTableLookup(AVX1CostTbl, ISD, MTy)) 2801 return LT.first * Entry->Cost; 2802 2803 if (ST->hasSSE42()) 2804 if (const auto *Entry = CostTableLookup(SSE42CostTbl, ISD, MTy)) 2805 return LT.first * Entry->Cost; 2806 2807 if (ST->hasSSE41()) 2808 if (const auto *Entry = CostTableLookup(SSE41CostTbl, ISD, MTy)) 2809 return LT.first * Entry->Cost; 2810 2811 if (ST->hasSSSE3()) 2812 if (const auto *Entry = CostTableLookup(SSSE3CostTbl, ISD, MTy)) 2813 return LT.first * Entry->Cost; 2814 2815 if (ST->hasSSE2()) 2816 if (const auto *Entry = CostTableLookup(SSE2CostTbl, ISD, MTy)) 2817 return LT.first * Entry->Cost; 2818 2819 if (ST->hasSSE1()) 2820 if (const auto *Entry = CostTableLookup(SSE1CostTbl, ISD, MTy)) 2821 return LT.first * Entry->Cost; 2822 2823 if (ST->hasBMI()) { 2824 if (ST->is64Bit()) 2825 if (const auto *Entry = CostTableLookup(BMI64CostTbl, ISD, MTy)) 2826 return LT.first * Entry->Cost; 2827 2828 if (const auto *Entry = CostTableLookup(BMI32CostTbl, ISD, MTy)) 2829 return LT.first * Entry->Cost; 2830 } 2831 2832 if (ST->hasLZCNT()) { 2833 if (ST->is64Bit()) 2834 if (const auto *Entry = CostTableLookup(LZCNT64CostTbl, ISD, MTy)) 2835 return LT.first * Entry->Cost; 2836 2837 if (const auto *Entry = CostTableLookup(LZCNT32CostTbl, ISD, MTy)) 2838 return LT.first * Entry->Cost; 2839 } 2840 2841 if (ST->hasPOPCNT()) { 2842 if (ST->is64Bit()) 2843 if (const auto *Entry = CostTableLookup(POPCNT64CostTbl, ISD, MTy)) 2844 return LT.first * Entry->Cost; 2845 2846 if (const auto *Entry = CostTableLookup(POPCNT32CostTbl, ISD, MTy)) 2847 return LT.first * Entry->Cost; 2848 } 2849 2850 // TODO - add BMI (TZCNT) scalar handling 2851 2852 if (ST->is64Bit()) 2853 if (const auto *Entry = CostTableLookup(X64CostTbl, ISD, MTy)) 2854 return LT.first * Entry->Cost; 2855 2856 if (const auto *Entry = CostTableLookup(X86CostTbl, ISD, MTy)) 2857 return LT.first * Entry->Cost; 2858 } 2859 2860 return BaseT::getIntrinsicInstrCost(ICA, CostKind); 2861 } 2862 2863 int X86TTIImpl::getIntrinsicInstrCost(const IntrinsicCostAttributes &ICA, 2864 TTI::TargetCostKind CostKind) { 2865 if (ICA.isTypeBasedOnly()) 2866 return getTypeBasedIntrinsicInstrCost(ICA, CostKind); 2867 2868 static const CostTblEntry AVX512CostTbl[] = { 2869 { ISD::ROTL, MVT::v8i64, 1 }, 2870 { ISD::ROTL, MVT::v4i64, 1 }, 2871 { ISD::ROTL, MVT::v2i64, 1 }, 2872 { ISD::ROTL, MVT::v16i32, 1 }, 2873 { ISD::ROTL, MVT::v8i32, 1 }, 2874 { ISD::ROTL, MVT::v4i32, 1 }, 2875 { ISD::ROTR, MVT::v8i64, 1 }, 2876 { ISD::ROTR, MVT::v4i64, 1 }, 2877 { ISD::ROTR, MVT::v2i64, 1 }, 2878 { ISD::ROTR, MVT::v16i32, 1 }, 2879 { ISD::ROTR, MVT::v8i32, 1 }, 2880 { ISD::ROTR, MVT::v4i32, 1 } 2881 }; 2882 // XOP: ROTL = VPROT(X,Y), ROTR = VPROT(X,SUB(0,Y)) 2883 static const CostTblEntry XOPCostTbl[] = { 2884 { ISD::ROTL, MVT::v4i64, 4 }, 2885 { ISD::ROTL, MVT::v8i32, 4 }, 2886 { ISD::ROTL, MVT::v16i16, 4 }, 2887 { ISD::ROTL, MVT::v32i8, 4 }, 2888 { ISD::ROTL, MVT::v2i64, 1 }, 2889 { ISD::ROTL, MVT::v4i32, 1 }, 2890 { ISD::ROTL, MVT::v8i16, 1 }, 2891 { ISD::ROTL, MVT::v16i8, 1 }, 2892 { ISD::ROTR, MVT::v4i64, 6 }, 2893 { ISD::ROTR, MVT::v8i32, 6 }, 2894 { ISD::ROTR, MVT::v16i16, 6 }, 2895 { ISD::ROTR, MVT::v32i8, 6 }, 2896 { ISD::ROTR, MVT::v2i64, 2 }, 2897 { ISD::ROTR, MVT::v4i32, 2 }, 2898 { ISD::ROTR, MVT::v8i16, 2 }, 2899 { ISD::ROTR, MVT::v16i8, 2 } 2900 }; 2901 static const CostTblEntry X64CostTbl[] = { // 64-bit targets 2902 { ISD::ROTL, MVT::i64, 1 }, 2903 { ISD::ROTR, MVT::i64, 1 }, 2904 { ISD::FSHL, MVT::i64, 4 } 2905 }; 2906 static const CostTblEntry X86CostTbl[] = { // 32 or 64-bit targets 2907 { ISD::ROTL, MVT::i32, 1 }, 2908 { ISD::ROTL, MVT::i16, 1 }, 2909 { ISD::ROTL, MVT::i8, 1 }, 2910 { ISD::ROTR, MVT::i32, 1 }, 2911 { ISD::ROTR, MVT::i16, 1 }, 2912 { ISD::ROTR, MVT::i8, 1 }, 2913 { ISD::FSHL, MVT::i32, 4 }, 2914 { ISD::FSHL, MVT::i16, 4 }, 2915 { ISD::FSHL, MVT::i8, 4 } 2916 }; 2917 2918 Intrinsic::ID IID = ICA.getID(); 2919 Type *RetTy = ICA.getReturnType(); 2920 const SmallVectorImpl<const Value *> &Args = ICA.getArgs(); 2921 unsigned ISD = ISD::DELETED_NODE; 2922 switch (IID) { 2923 default: 2924 break; 2925 case Intrinsic::fshl: 2926 ISD = ISD::FSHL; 2927 if (Args[0] == Args[1]) 2928 ISD = ISD::ROTL; 2929 break; 2930 case Intrinsic::fshr: 2931 // FSHR has same costs so don't duplicate. 2932 ISD = ISD::FSHL; 2933 if (Args[0] == Args[1]) 2934 ISD = ISD::ROTR; 2935 break; 2936 } 2937 2938 if (ISD != ISD::DELETED_NODE) { 2939 // Legalize the type. 2940 std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, RetTy); 2941 MVT MTy = LT.second; 2942 2943 // Attempt to lookup cost. 2944 if (ST->hasAVX512()) 2945 if (const auto *Entry = CostTableLookup(AVX512CostTbl, ISD, MTy)) 2946 return LT.first * Entry->Cost; 2947 2948 if (ST->hasXOP()) 2949 if (const auto *Entry = CostTableLookup(XOPCostTbl, ISD, MTy)) 2950 return LT.first * Entry->Cost; 2951 2952 if (ST->is64Bit()) 2953 if (const auto *Entry = CostTableLookup(X64CostTbl, ISD, MTy)) 2954 return LT.first * Entry->Cost; 2955 2956 if (const auto *Entry = CostTableLookup(X86CostTbl, ISD, MTy)) 2957 return LT.first * Entry->Cost; 2958 } 2959 2960 return BaseT::getIntrinsicInstrCost(ICA, CostKind); 2961 } 2962 2963 int X86TTIImpl::getVectorInstrCost(unsigned Opcode, Type *Val, unsigned Index) { 2964 static const CostTblEntry SLMCostTbl[] = { 2965 { ISD::EXTRACT_VECTOR_ELT, MVT::i8, 4 }, 2966 { ISD::EXTRACT_VECTOR_ELT, MVT::i16, 4 }, 2967 { ISD::EXTRACT_VECTOR_ELT, MVT::i32, 4 }, 2968 { ISD::EXTRACT_VECTOR_ELT, MVT::i64, 7 } 2969 }; 2970 2971 assert(Val->isVectorTy() && "This must be a vector type"); 2972 Type *ScalarType = Val->getScalarType(); 2973 int RegisterFileMoveCost = 0; 2974 2975 if (Index != -1U && (Opcode == Instruction::ExtractElement || 2976 Opcode == Instruction::InsertElement)) { 2977 // Legalize the type. 2978 std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, Val); 2979 2980 // This type is legalized to a scalar type. 2981 if (!LT.second.isVector()) 2982 return 0; 2983 2984 // The type may be split. Normalize the index to the new type. 2985 unsigned NumElts = LT.second.getVectorNumElements(); 2986 unsigned SubNumElts = NumElts; 2987 Index = Index % NumElts; 2988 2989 // For >128-bit vectors, we need to extract higher 128-bit subvectors. 2990 // For inserts, we also need to insert the subvector back. 2991 if (LT.second.getSizeInBits() > 128) { 2992 assert((LT.second.getSizeInBits() % 128) == 0 && "Illegal vector"); 2993 unsigned NumSubVecs = LT.second.getSizeInBits() / 128; 2994 SubNumElts = NumElts / NumSubVecs; 2995 if (SubNumElts <= Index) { 2996 RegisterFileMoveCost += (Opcode == Instruction::InsertElement ? 2 : 1); 2997 Index %= SubNumElts; 2998 } 2999 } 3000 3001 if (Index == 0) { 3002 // Floating point scalars are already located in index #0. 3003 // Many insertions to #0 can fold away for scalar fp-ops, so let's assume 3004 // true for all. 3005 if (ScalarType->isFloatingPointTy()) 3006 return RegisterFileMoveCost; 3007 3008 // Assume movd/movq XMM -> GPR is relatively cheap on all targets. 3009 if (ScalarType->isIntegerTy() && Opcode == Instruction::ExtractElement) 3010 return 1 + RegisterFileMoveCost; 3011 } 3012 3013 int ISD = TLI->InstructionOpcodeToISD(Opcode); 3014 assert(ISD && "Unexpected vector opcode"); 3015 MVT MScalarTy = LT.second.getScalarType(); 3016 if (ST->isSLM()) 3017 if (auto *Entry = CostTableLookup(SLMCostTbl, ISD, MScalarTy)) 3018 return Entry->Cost + RegisterFileMoveCost; 3019 3020 // Assume pinsr/pextr XMM <-> GPR is relatively cheap on all targets. 3021 if ((MScalarTy == MVT::i16 && ST->hasSSE2()) || 3022 (MScalarTy.isInteger() && ST->hasSSE41())) 3023 return 1 + RegisterFileMoveCost; 3024 3025 // Assume insertps is relatively cheap on all targets. 3026 if (MScalarTy == MVT::f32 && ST->hasSSE41() && 3027 Opcode == Instruction::InsertElement) 3028 return 1 + RegisterFileMoveCost; 3029 3030 // For extractions we just need to shuffle the element to index 0, which 3031 // should be very cheap (assume cost = 1). For insertions we need to shuffle 3032 // the elements to its destination. In both cases we must handle the 3033 // subvector move(s). 3034 // If the vector type is already less than 128-bits then don't reduce it. 3035 // TODO: Under what circumstances should we shuffle using the full width? 3036 int ShuffleCost = 1; 3037 if (Opcode == Instruction::InsertElement) { 3038 auto *SubTy = cast<VectorType>(Val); 3039 EVT VT = TLI->getValueType(DL, Val); 3040 if (VT.getScalarType() != MScalarTy || VT.getSizeInBits() >= 128) 3041 SubTy = FixedVectorType::get(ScalarType, SubNumElts); 3042 ShuffleCost = getShuffleCost(TTI::SK_PermuteTwoSrc, SubTy, 0, SubTy); 3043 } 3044 int IntOrFpCost = ScalarType->isFloatingPointTy() ? 0 : 1; 3045 return ShuffleCost + IntOrFpCost + RegisterFileMoveCost; 3046 } 3047 3048 // Add to the base cost if we know that the extracted element of a vector is 3049 // destined to be moved to and used in the integer register file. 3050 if (Opcode == Instruction::ExtractElement && ScalarType->isPointerTy()) 3051 RegisterFileMoveCost += 1; 3052 3053 return BaseT::getVectorInstrCost(Opcode, Val, Index) + RegisterFileMoveCost; 3054 } 3055 3056 unsigned X86TTIImpl::getScalarizationOverhead(VectorType *Ty, 3057 const APInt &DemandedElts, 3058 bool Insert, bool Extract) { 3059 unsigned Cost = 0; 3060 3061 // For insertions, a ISD::BUILD_VECTOR style vector initialization can be much 3062 // cheaper than an accumulation of ISD::INSERT_VECTOR_ELT. 3063 if (Insert) { 3064 std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, Ty); 3065 MVT MScalarTy = LT.second.getScalarType(); 3066 3067 if ((MScalarTy == MVT::i16 && ST->hasSSE2()) || 3068 (MScalarTy.isInteger() && ST->hasSSE41()) || 3069 (MScalarTy == MVT::f32 && ST->hasSSE41())) { 3070 // For types we can insert directly, insertion into 128-bit sub vectors is 3071 // cheap, followed by a cheap chain of concatenations. 3072 if (LT.second.getSizeInBits() <= 128) { 3073 Cost += 3074 BaseT::getScalarizationOverhead(Ty, DemandedElts, Insert, false); 3075 } else { 3076 // In each 128-lane, if at least one index is demanded but not all 3077 // indices are demanded and this 128-lane is not the first 128-lane of 3078 // the legalized-vector, then this 128-lane needs a extracti128; If in 3079 // each 128-lane, there is at least one demanded index, this 128-lane 3080 // needs a inserti128. 3081 3082 // The following cases will help you build a better understanding: 3083 // Assume we insert several elements into a v8i32 vector in avx2, 3084 // Case#1: inserting into 1th index needs vpinsrd + inserti128. 3085 // Case#2: inserting into 5th index needs extracti128 + vpinsrd + 3086 // inserti128. 3087 // Case#3: inserting into 4,5,6,7 index needs 4*vpinsrd + inserti128. 3088 unsigned Num128Lanes = LT.second.getSizeInBits() / 128 * LT.first; 3089 unsigned NumElts = LT.second.getVectorNumElements() * LT.first; 3090 APInt WidenedDemandedElts = DemandedElts.zextOrSelf(NumElts); 3091 unsigned Scale = NumElts / Num128Lanes; 3092 // We iterate each 128-lane, and check if we need a 3093 // extracti128/inserti128 for this 128-lane. 3094 for (unsigned I = 0; I < NumElts; I += Scale) { 3095 APInt Mask = WidenedDemandedElts.getBitsSet(NumElts, I, I + Scale); 3096 APInt MaskedDE = Mask & WidenedDemandedElts; 3097 unsigned Population = MaskedDE.countPopulation(); 3098 Cost += (Population > 0 && Population != Scale && 3099 I % LT.second.getVectorNumElements() != 0); 3100 Cost += Population > 0; 3101 } 3102 Cost += DemandedElts.countPopulation(); 3103 3104 // For vXf32 cases, insertion into the 0'th index in each v4f32 3105 // 128-bit vector is free. 3106 // NOTE: This assumes legalization widens vXf32 vectors. 3107 if (MScalarTy == MVT::f32) 3108 for (unsigned i = 0, e = cast<FixedVectorType>(Ty)->getNumElements(); 3109 i < e; i += 4) 3110 if (DemandedElts[i]) 3111 Cost--; 3112 } 3113 } else if (LT.second.isVector()) { 3114 // Without fast insertion, we need to use MOVD/MOVQ to pass each demanded 3115 // integer element as a SCALAR_TO_VECTOR, then we build the vector as a 3116 // series of UNPCK followed by CONCAT_VECTORS - all of these can be 3117 // considered cheap. 3118 if (Ty->isIntOrIntVectorTy()) 3119 Cost += DemandedElts.countPopulation(); 3120 3121 // Get the smaller of the legalized or original pow2-extended number of 3122 // vector elements, which represents the number of unpacks we'll end up 3123 // performing. 3124 unsigned NumElts = LT.second.getVectorNumElements(); 3125 unsigned Pow2Elts = 3126 PowerOf2Ceil(cast<FixedVectorType>(Ty)->getNumElements()); 3127 Cost += (std::min<unsigned>(NumElts, Pow2Elts) - 1) * LT.first; 3128 } 3129 } 3130 3131 // TODO: Use default extraction for now, but we should investigate extending this 3132 // to handle repeated subvector extraction. 3133 if (Extract) 3134 Cost += BaseT::getScalarizationOverhead(Ty, DemandedElts, false, Extract); 3135 3136 return Cost; 3137 } 3138 3139 int X86TTIImpl::getMemoryOpCost(unsigned Opcode, Type *Src, 3140 MaybeAlign Alignment, unsigned AddressSpace, 3141 TTI::TargetCostKind CostKind, 3142 const Instruction *I) { 3143 // TODO: Handle other cost kinds. 3144 if (CostKind != TTI::TCK_RecipThroughput) { 3145 if (isa_and_nonnull<StoreInst>(I)) { 3146 Value *Ptr = I->getOperand(1); 3147 // Store instruction with index and scale costs 2 Uops. 3148 // Check the preceding GEP to identify non-const indices. 3149 if (auto *GEP = dyn_cast<GetElementPtrInst>(Ptr)) { 3150 if (!all_of(GEP->indices(), [](Value *V) { return isa<Constant>(V); })) 3151 return TTI::TCC_Basic * 2; 3152 } 3153 } 3154 return TTI::TCC_Basic; 3155 } 3156 3157 // Handle non-power-of-two vectors such as <3 x float> 3158 if (auto *VTy = dyn_cast<FixedVectorType>(Src)) { 3159 unsigned NumElem = VTy->getNumElements(); 3160 3161 // Handle a few common cases: 3162 // <3 x float> 3163 if (NumElem == 3 && VTy->getScalarSizeInBits() == 32) 3164 // Cost = 64 bit store + extract + 32 bit store. 3165 return 3; 3166 3167 // <3 x double> 3168 if (NumElem == 3 && VTy->getScalarSizeInBits() == 64) 3169 // Cost = 128 bit store + unpack + 64 bit store. 3170 return 3; 3171 3172 // Assume that all other non-power-of-two numbers are scalarized. 3173 if (!isPowerOf2_32(NumElem)) { 3174 APInt DemandedElts = APInt::getAllOnesValue(NumElem); 3175 int Cost = BaseT::getMemoryOpCost(Opcode, VTy->getScalarType(), Alignment, 3176 AddressSpace, CostKind); 3177 int SplitCost = getScalarizationOverhead(VTy, DemandedElts, 3178 Opcode == Instruction::Load, 3179 Opcode == Instruction::Store); 3180 return NumElem * Cost + SplitCost; 3181 } 3182 } 3183 3184 // Type legalization can't handle structs 3185 if (TLI->getValueType(DL, Src, true) == MVT::Other) 3186 return BaseT::getMemoryOpCost(Opcode, Src, Alignment, AddressSpace, 3187 CostKind); 3188 3189 // Legalize the type. 3190 std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, Src); 3191 assert((Opcode == Instruction::Load || Opcode == Instruction::Store) && 3192 "Invalid Opcode"); 3193 3194 // Each load/store unit costs 1. 3195 int Cost = LT.first * 1; 3196 3197 // This isn't exactly right. We're using slow unaligned 32-byte accesses as a 3198 // proxy for a double-pumped AVX memory interface such as on Sandybridge. 3199 if (LT.second.getStoreSize() == 32 && ST->isUnalignedMem32Slow()) 3200 Cost *= 2; 3201 3202 return Cost; 3203 } 3204 3205 int X86TTIImpl::getMaskedMemoryOpCost(unsigned Opcode, Type *SrcTy, 3206 Align Alignment, unsigned AddressSpace, 3207 TTI::TargetCostKind CostKind) { 3208 bool IsLoad = (Instruction::Load == Opcode); 3209 bool IsStore = (Instruction::Store == Opcode); 3210 3211 auto *SrcVTy = dyn_cast<FixedVectorType>(SrcTy); 3212 if (!SrcVTy) 3213 // To calculate scalar take the regular cost, without mask 3214 return getMemoryOpCost(Opcode, SrcTy, Alignment, AddressSpace, CostKind); 3215 3216 unsigned NumElem = SrcVTy->getNumElements(); 3217 auto *MaskTy = 3218 FixedVectorType::get(Type::getInt8Ty(SrcVTy->getContext()), NumElem); 3219 if ((IsLoad && !isLegalMaskedLoad(SrcVTy, Alignment)) || 3220 (IsStore && !isLegalMaskedStore(SrcVTy, Alignment)) || 3221 !isPowerOf2_32(NumElem)) { 3222 // Scalarization 3223 APInt DemandedElts = APInt::getAllOnesValue(NumElem); 3224 int MaskSplitCost = 3225 getScalarizationOverhead(MaskTy, DemandedElts, false, true); 3226 int ScalarCompareCost = getCmpSelInstrCost( 3227 Instruction::ICmp, Type::getInt8Ty(SrcVTy->getContext()), nullptr, 3228 CmpInst::BAD_ICMP_PREDICATE, CostKind); 3229 int BranchCost = getCFInstrCost(Instruction::Br, CostKind); 3230 int MaskCmpCost = NumElem * (BranchCost + ScalarCompareCost); 3231 int ValueSplitCost = 3232 getScalarizationOverhead(SrcVTy, DemandedElts, IsLoad, IsStore); 3233 int MemopCost = 3234 NumElem * BaseT::getMemoryOpCost(Opcode, SrcVTy->getScalarType(), 3235 Alignment, AddressSpace, CostKind); 3236 return MemopCost + ValueSplitCost + MaskSplitCost + MaskCmpCost; 3237 } 3238 3239 // Legalize the type. 3240 std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, SrcVTy); 3241 auto VT = TLI->getValueType(DL, SrcVTy); 3242 int Cost = 0; 3243 if (VT.isSimple() && LT.second != VT.getSimpleVT() && 3244 LT.second.getVectorNumElements() == NumElem) 3245 // Promotion requires expand/truncate for data and a shuffle for mask. 3246 Cost += getShuffleCost(TTI::SK_PermuteTwoSrc, SrcVTy, 0, nullptr) + 3247 getShuffleCost(TTI::SK_PermuteTwoSrc, MaskTy, 0, nullptr); 3248 3249 else if (LT.second.getVectorNumElements() > NumElem) { 3250 auto *NewMaskTy = FixedVectorType::get(MaskTy->getElementType(), 3251 LT.second.getVectorNumElements()); 3252 // Expanding requires fill mask with zeroes 3253 Cost += getShuffleCost(TTI::SK_InsertSubvector, NewMaskTy, 0, MaskTy); 3254 } 3255 3256 // Pre-AVX512 - each maskmov load costs 2 + store costs ~8. 3257 if (!ST->hasAVX512()) 3258 return Cost + LT.first * (IsLoad ? 2 : 8); 3259 3260 // AVX-512 masked load/store is cheapper 3261 return Cost + LT.first; 3262 } 3263 3264 int X86TTIImpl::getAddressComputationCost(Type *Ty, ScalarEvolution *SE, 3265 const SCEV *Ptr) { 3266 // Address computations in vectorized code with non-consecutive addresses will 3267 // likely result in more instructions compared to scalar code where the 3268 // computation can more often be merged into the index mode. The resulting 3269 // extra micro-ops can significantly decrease throughput. 3270 const unsigned NumVectorInstToHideOverhead = 10; 3271 3272 // Cost modeling of Strided Access Computation is hidden by the indexing 3273 // modes of X86 regardless of the stride value. We dont believe that there 3274 // is a difference between constant strided access in gerenal and constant 3275 // strided value which is less than or equal to 64. 3276 // Even in the case of (loop invariant) stride whose value is not known at 3277 // compile time, the address computation will not incur more than one extra 3278 // ADD instruction. 3279 if (Ty->isVectorTy() && SE) { 3280 if (!BaseT::isStridedAccess(Ptr)) 3281 return NumVectorInstToHideOverhead; 3282 if (!BaseT::getConstantStrideStep(SE, Ptr)) 3283 return 1; 3284 } 3285 3286 return BaseT::getAddressComputationCost(Ty, SE, Ptr); 3287 } 3288 3289 int X86TTIImpl::getArithmeticReductionCost(unsigned Opcode, VectorType *ValTy, 3290 bool IsPairwise, 3291 TTI::TargetCostKind CostKind) { 3292 // Just use the default implementation for pair reductions. 3293 if (IsPairwise) 3294 return BaseT::getArithmeticReductionCost(Opcode, ValTy, IsPairwise, CostKind); 3295 3296 // We use the Intel Architecture Code Analyzer(IACA) to measure the throughput 3297 // and make it as the cost. 3298 3299 static const CostTblEntry SLMCostTblNoPairWise[] = { 3300 { ISD::FADD, MVT::v2f64, 3 }, 3301 { ISD::ADD, MVT::v2i64, 5 }, 3302 }; 3303 3304 static const CostTblEntry SSE2CostTblNoPairWise[] = { 3305 { ISD::FADD, MVT::v2f64, 2 }, 3306 { ISD::FADD, MVT::v4f32, 4 }, 3307 { ISD::ADD, MVT::v2i64, 2 }, // The data reported by the IACA tool is "1.6". 3308 { ISD::ADD, MVT::v2i32, 2 }, // FIXME: chosen to be less than v4i32 3309 { ISD::ADD, MVT::v4i32, 3 }, // The data reported by the IACA tool is "3.3". 3310 { ISD::ADD, MVT::v2i16, 2 }, // The data reported by the IACA tool is "4.3". 3311 { ISD::ADD, MVT::v4i16, 3 }, // The data reported by the IACA tool is "4.3". 3312 { ISD::ADD, MVT::v8i16, 4 }, // The data reported by the IACA tool is "4.3". 3313 { ISD::ADD, MVT::v2i8, 2 }, 3314 { ISD::ADD, MVT::v4i8, 2 }, 3315 { ISD::ADD, MVT::v8i8, 2 }, 3316 { ISD::ADD, MVT::v16i8, 3 }, 3317 }; 3318 3319 static const CostTblEntry AVX1CostTblNoPairWise[] = { 3320 { ISD::FADD, MVT::v4f64, 3 }, 3321 { ISD::FADD, MVT::v4f32, 3 }, 3322 { ISD::FADD, MVT::v8f32, 4 }, 3323 { ISD::ADD, MVT::v2i64, 1 }, // The data reported by the IACA tool is "1.5". 3324 { ISD::ADD, MVT::v4i64, 3 }, 3325 { ISD::ADD, MVT::v8i32, 5 }, 3326 { ISD::ADD, MVT::v16i16, 5 }, 3327 { ISD::ADD, MVT::v32i8, 4 }, 3328 }; 3329 3330 int ISD = TLI->InstructionOpcodeToISD(Opcode); 3331 assert(ISD && "Invalid opcode"); 3332 3333 // Before legalizing the type, give a chance to look up illegal narrow types 3334 // in the table. 3335 // FIXME: Is there a better way to do this? 3336 EVT VT = TLI->getValueType(DL, ValTy); 3337 if (VT.isSimple()) { 3338 MVT MTy = VT.getSimpleVT(); 3339 if (ST->isSLM()) 3340 if (const auto *Entry = CostTableLookup(SLMCostTblNoPairWise, ISD, MTy)) 3341 return Entry->Cost; 3342 3343 if (ST->hasAVX()) 3344 if (const auto *Entry = CostTableLookup(AVX1CostTblNoPairWise, ISD, MTy)) 3345 return Entry->Cost; 3346 3347 if (ST->hasSSE2()) 3348 if (const auto *Entry = CostTableLookup(SSE2CostTblNoPairWise, ISD, MTy)) 3349 return Entry->Cost; 3350 } 3351 3352 std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, ValTy); 3353 3354 MVT MTy = LT.second; 3355 3356 auto *ValVTy = cast<FixedVectorType>(ValTy); 3357 3358 unsigned ArithmeticCost = 0; 3359 if (LT.first != 1 && MTy.isVector() && 3360 MTy.getVectorNumElements() < ValVTy->getNumElements()) { 3361 // Type needs to be split. We need LT.first - 1 arithmetic ops. 3362 auto *SingleOpTy = FixedVectorType::get(ValVTy->getElementType(), 3363 MTy.getVectorNumElements()); 3364 ArithmeticCost = getArithmeticInstrCost(Opcode, SingleOpTy, CostKind); 3365 ArithmeticCost *= LT.first - 1; 3366 } 3367 3368 if (ST->isSLM()) 3369 if (const auto *Entry = CostTableLookup(SLMCostTblNoPairWise, ISD, MTy)) 3370 return ArithmeticCost + Entry->Cost; 3371 3372 if (ST->hasAVX()) 3373 if (const auto *Entry = CostTableLookup(AVX1CostTblNoPairWise, ISD, MTy)) 3374 return ArithmeticCost + Entry->Cost; 3375 3376 if (ST->hasSSE2()) 3377 if (const auto *Entry = CostTableLookup(SSE2CostTblNoPairWise, ISD, MTy)) 3378 return ArithmeticCost + Entry->Cost; 3379 3380 // FIXME: These assume a naive kshift+binop lowering, which is probably 3381 // conservative in most cases. 3382 static const CostTblEntry AVX512BoolReduction[] = { 3383 { ISD::AND, MVT::v2i1, 3 }, 3384 { ISD::AND, MVT::v4i1, 5 }, 3385 { ISD::AND, MVT::v8i1, 7 }, 3386 { ISD::AND, MVT::v16i1, 9 }, 3387 { ISD::AND, MVT::v32i1, 11 }, 3388 { ISD::AND, MVT::v64i1, 13 }, 3389 { ISD::OR, MVT::v2i1, 3 }, 3390 { ISD::OR, MVT::v4i1, 5 }, 3391 { ISD::OR, MVT::v8i1, 7 }, 3392 { ISD::OR, MVT::v16i1, 9 }, 3393 { ISD::OR, MVT::v32i1, 11 }, 3394 { ISD::OR, MVT::v64i1, 13 }, 3395 }; 3396 3397 static const CostTblEntry AVX2BoolReduction[] = { 3398 { ISD::AND, MVT::v16i16, 2 }, // vpmovmskb + cmp 3399 { ISD::AND, MVT::v32i8, 2 }, // vpmovmskb + cmp 3400 { ISD::OR, MVT::v16i16, 2 }, // vpmovmskb + cmp 3401 { ISD::OR, MVT::v32i8, 2 }, // vpmovmskb + cmp 3402 }; 3403 3404 static const CostTblEntry AVX1BoolReduction[] = { 3405 { ISD::AND, MVT::v4i64, 2 }, // vmovmskpd + cmp 3406 { ISD::AND, MVT::v8i32, 2 }, // vmovmskps + cmp 3407 { ISD::AND, MVT::v16i16, 4 }, // vextractf128 + vpand + vpmovmskb + cmp 3408 { ISD::AND, MVT::v32i8, 4 }, // vextractf128 + vpand + vpmovmskb + cmp 3409 { ISD::OR, MVT::v4i64, 2 }, // vmovmskpd + cmp 3410 { ISD::OR, MVT::v8i32, 2 }, // vmovmskps + cmp 3411 { ISD::OR, MVT::v16i16, 4 }, // vextractf128 + vpor + vpmovmskb + cmp 3412 { ISD::OR, MVT::v32i8, 4 }, // vextractf128 + vpor + vpmovmskb + cmp 3413 }; 3414 3415 static const CostTblEntry SSE2BoolReduction[] = { 3416 { ISD::AND, MVT::v2i64, 2 }, // movmskpd + cmp 3417 { ISD::AND, MVT::v4i32, 2 }, // movmskps + cmp 3418 { ISD::AND, MVT::v8i16, 2 }, // pmovmskb + cmp 3419 { ISD::AND, MVT::v16i8, 2 }, // pmovmskb + cmp 3420 { ISD::OR, MVT::v2i64, 2 }, // movmskpd + cmp 3421 { ISD::OR, MVT::v4i32, 2 }, // movmskps + cmp 3422 { ISD::OR, MVT::v8i16, 2 }, // pmovmskb + cmp 3423 { ISD::OR, MVT::v16i8, 2 }, // pmovmskb + cmp 3424 }; 3425 3426 // Handle bool allof/anyof patterns. 3427 if (ValVTy->getElementType()->isIntegerTy(1)) { 3428 unsigned ArithmeticCost = 0; 3429 if (LT.first != 1 && MTy.isVector() && 3430 MTy.getVectorNumElements() < ValVTy->getNumElements()) { 3431 // Type needs to be split. We need LT.first - 1 arithmetic ops. 3432 auto *SingleOpTy = FixedVectorType::get(ValVTy->getElementType(), 3433 MTy.getVectorNumElements()); 3434 ArithmeticCost = getArithmeticInstrCost(Opcode, SingleOpTy, CostKind); 3435 ArithmeticCost *= LT.first - 1; 3436 } 3437 3438 if (ST->hasAVX512()) 3439 if (const auto *Entry = CostTableLookup(AVX512BoolReduction, ISD, MTy)) 3440 return ArithmeticCost + Entry->Cost; 3441 if (ST->hasAVX2()) 3442 if (const auto *Entry = CostTableLookup(AVX2BoolReduction, ISD, MTy)) 3443 return ArithmeticCost + Entry->Cost; 3444 if (ST->hasAVX()) 3445 if (const auto *Entry = CostTableLookup(AVX1BoolReduction, ISD, MTy)) 3446 return ArithmeticCost + Entry->Cost; 3447 if (ST->hasSSE2()) 3448 if (const auto *Entry = CostTableLookup(SSE2BoolReduction, ISD, MTy)) 3449 return ArithmeticCost + Entry->Cost; 3450 3451 return BaseT::getArithmeticReductionCost(Opcode, ValVTy, IsPairwise, 3452 CostKind); 3453 } 3454 3455 unsigned NumVecElts = ValVTy->getNumElements(); 3456 unsigned ScalarSize = ValVTy->getScalarSizeInBits(); 3457 3458 // Special case power of 2 reductions where the scalar type isn't changed 3459 // by type legalization. 3460 if (!isPowerOf2_32(NumVecElts) || ScalarSize != MTy.getScalarSizeInBits()) 3461 return BaseT::getArithmeticReductionCost(Opcode, ValVTy, IsPairwise, 3462 CostKind); 3463 3464 unsigned ReductionCost = 0; 3465 3466 auto *Ty = ValVTy; 3467 if (LT.first != 1 && MTy.isVector() && 3468 MTy.getVectorNumElements() < ValVTy->getNumElements()) { 3469 // Type needs to be split. We need LT.first - 1 arithmetic ops. 3470 Ty = FixedVectorType::get(ValVTy->getElementType(), 3471 MTy.getVectorNumElements()); 3472 ReductionCost = getArithmeticInstrCost(Opcode, Ty, CostKind); 3473 ReductionCost *= LT.first - 1; 3474 NumVecElts = MTy.getVectorNumElements(); 3475 } 3476 3477 // Now handle reduction with the legal type, taking into account size changes 3478 // at each level. 3479 while (NumVecElts > 1) { 3480 // Determine the size of the remaining vector we need to reduce. 3481 unsigned Size = NumVecElts * ScalarSize; 3482 NumVecElts /= 2; 3483 // If we're reducing from 256/512 bits, use an extract_subvector. 3484 if (Size > 128) { 3485 auto *SubTy = FixedVectorType::get(ValVTy->getElementType(), NumVecElts); 3486 ReductionCost += 3487 getShuffleCost(TTI::SK_ExtractSubvector, Ty, NumVecElts, SubTy); 3488 Ty = SubTy; 3489 } else if (Size == 128) { 3490 // Reducing from 128 bits is a permute of v2f64/v2i64. 3491 FixedVectorType *ShufTy; 3492 if (ValVTy->isFloatingPointTy()) 3493 ShufTy = 3494 FixedVectorType::get(Type::getDoubleTy(ValVTy->getContext()), 2); 3495 else 3496 ShufTy = 3497 FixedVectorType::get(Type::getInt64Ty(ValVTy->getContext()), 2); 3498 ReductionCost += 3499 getShuffleCost(TTI::SK_PermuteSingleSrc, ShufTy, 0, nullptr); 3500 } else if (Size == 64) { 3501 // Reducing from 64 bits is a shuffle of v4f32/v4i32. 3502 FixedVectorType *ShufTy; 3503 if (ValVTy->isFloatingPointTy()) 3504 ShufTy = 3505 FixedVectorType::get(Type::getFloatTy(ValVTy->getContext()), 4); 3506 else 3507 ShufTy = 3508 FixedVectorType::get(Type::getInt32Ty(ValVTy->getContext()), 4); 3509 ReductionCost += 3510 getShuffleCost(TTI::SK_PermuteSingleSrc, ShufTy, 0, nullptr); 3511 } else { 3512 // Reducing from smaller size is a shift by immediate. 3513 auto *ShiftTy = FixedVectorType::get( 3514 Type::getIntNTy(ValVTy->getContext(), Size), 128 / Size); 3515 ReductionCost += getArithmeticInstrCost( 3516 Instruction::LShr, ShiftTy, CostKind, 3517 TargetTransformInfo::OK_AnyValue, 3518 TargetTransformInfo::OK_UniformConstantValue, 3519 TargetTransformInfo::OP_None, TargetTransformInfo::OP_None); 3520 } 3521 3522 // Add the arithmetic op for this level. 3523 ReductionCost += getArithmeticInstrCost(Opcode, Ty, CostKind); 3524 } 3525 3526 // Add the final extract element to the cost. 3527 return ReductionCost + getVectorInstrCost(Instruction::ExtractElement, Ty, 0); 3528 } 3529 3530 int X86TTIImpl::getMinMaxCost(Type *Ty, Type *CondTy, bool IsUnsigned) { 3531 std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, Ty); 3532 3533 MVT MTy = LT.second; 3534 3535 int ISD; 3536 if (Ty->isIntOrIntVectorTy()) { 3537 ISD = IsUnsigned ? ISD::UMIN : ISD::SMIN; 3538 } else { 3539 assert(Ty->isFPOrFPVectorTy() && 3540 "Expected float point or integer vector type."); 3541 ISD = ISD::FMINNUM; 3542 } 3543 3544 static const CostTblEntry SSE1CostTbl[] = { 3545 {ISD::FMINNUM, MVT::v4f32, 1}, 3546 }; 3547 3548 static const CostTblEntry SSE2CostTbl[] = { 3549 {ISD::FMINNUM, MVT::v2f64, 1}, 3550 {ISD::SMIN, MVT::v8i16, 1}, 3551 {ISD::UMIN, MVT::v16i8, 1}, 3552 }; 3553 3554 static const CostTblEntry SSE41CostTbl[] = { 3555 {ISD::SMIN, MVT::v4i32, 1}, 3556 {ISD::UMIN, MVT::v4i32, 1}, 3557 {ISD::UMIN, MVT::v8i16, 1}, 3558 {ISD::SMIN, MVT::v16i8, 1}, 3559 }; 3560 3561 static const CostTblEntry SSE42CostTbl[] = { 3562 {ISD::UMIN, MVT::v2i64, 3}, // xor+pcmpgtq+blendvpd 3563 }; 3564 3565 static const CostTblEntry AVX1CostTbl[] = { 3566 {ISD::FMINNUM, MVT::v8f32, 1}, 3567 {ISD::FMINNUM, MVT::v4f64, 1}, 3568 {ISD::SMIN, MVT::v8i32, 3}, 3569 {ISD::UMIN, MVT::v8i32, 3}, 3570 {ISD::SMIN, MVT::v16i16, 3}, 3571 {ISD::UMIN, MVT::v16i16, 3}, 3572 {ISD::SMIN, MVT::v32i8, 3}, 3573 {ISD::UMIN, MVT::v32i8, 3}, 3574 }; 3575 3576 static const CostTblEntry AVX2CostTbl[] = { 3577 {ISD::SMIN, MVT::v8i32, 1}, 3578 {ISD::UMIN, MVT::v8i32, 1}, 3579 {ISD::SMIN, MVT::v16i16, 1}, 3580 {ISD::UMIN, MVT::v16i16, 1}, 3581 {ISD::SMIN, MVT::v32i8, 1}, 3582 {ISD::UMIN, MVT::v32i8, 1}, 3583 }; 3584 3585 static const CostTblEntry AVX512CostTbl[] = { 3586 {ISD::FMINNUM, MVT::v16f32, 1}, 3587 {ISD::FMINNUM, MVT::v8f64, 1}, 3588 {ISD::SMIN, MVT::v2i64, 1}, 3589 {ISD::UMIN, MVT::v2i64, 1}, 3590 {ISD::SMIN, MVT::v4i64, 1}, 3591 {ISD::UMIN, MVT::v4i64, 1}, 3592 {ISD::SMIN, MVT::v8i64, 1}, 3593 {ISD::UMIN, MVT::v8i64, 1}, 3594 {ISD::SMIN, MVT::v16i32, 1}, 3595 {ISD::UMIN, MVT::v16i32, 1}, 3596 }; 3597 3598 static const CostTblEntry AVX512BWCostTbl[] = { 3599 {ISD::SMIN, MVT::v32i16, 1}, 3600 {ISD::UMIN, MVT::v32i16, 1}, 3601 {ISD::SMIN, MVT::v64i8, 1}, 3602 {ISD::UMIN, MVT::v64i8, 1}, 3603 }; 3604 3605 // If we have a native MIN/MAX instruction for this type, use it. 3606 if (ST->hasBWI()) 3607 if (const auto *Entry = CostTableLookup(AVX512BWCostTbl, ISD, MTy)) 3608 return LT.first * Entry->Cost; 3609 3610 if (ST->hasAVX512()) 3611 if (const auto *Entry = CostTableLookup(AVX512CostTbl, ISD, MTy)) 3612 return LT.first * Entry->Cost; 3613 3614 if (ST->hasAVX2()) 3615 if (const auto *Entry = CostTableLookup(AVX2CostTbl, ISD, MTy)) 3616 return LT.first * Entry->Cost; 3617 3618 if (ST->hasAVX()) 3619 if (const auto *Entry = CostTableLookup(AVX1CostTbl, ISD, MTy)) 3620 return LT.first * Entry->Cost; 3621 3622 if (ST->hasSSE42()) 3623 if (const auto *Entry = CostTableLookup(SSE42CostTbl, ISD, MTy)) 3624 return LT.first * Entry->Cost; 3625 3626 if (ST->hasSSE41()) 3627 if (const auto *Entry = CostTableLookup(SSE41CostTbl, ISD, MTy)) 3628 return LT.first * Entry->Cost; 3629 3630 if (ST->hasSSE2()) 3631 if (const auto *Entry = CostTableLookup(SSE2CostTbl, ISD, MTy)) 3632 return LT.first * Entry->Cost; 3633 3634 if (ST->hasSSE1()) 3635 if (const auto *Entry = CostTableLookup(SSE1CostTbl, ISD, MTy)) 3636 return LT.first * Entry->Cost; 3637 3638 unsigned CmpOpcode; 3639 if (Ty->isFPOrFPVectorTy()) { 3640 CmpOpcode = Instruction::FCmp; 3641 } else { 3642 assert(Ty->isIntOrIntVectorTy() && 3643 "expecting floating point or integer type for min/max reduction"); 3644 CmpOpcode = Instruction::ICmp; 3645 } 3646 3647 TTI::TargetCostKind CostKind = TTI::TCK_RecipThroughput; 3648 // Otherwise fall back to cmp+select. 3649 return getCmpSelInstrCost(CmpOpcode, Ty, CondTy, CmpInst::BAD_ICMP_PREDICATE, 3650 CostKind) + 3651 getCmpSelInstrCost(Instruction::Select, Ty, CondTy, 3652 CmpInst::BAD_ICMP_PREDICATE, CostKind); 3653 } 3654 3655 int X86TTIImpl::getMinMaxReductionCost(VectorType *ValTy, VectorType *CondTy, 3656 bool IsPairwise, bool IsUnsigned, 3657 TTI::TargetCostKind CostKind) { 3658 // Just use the default implementation for pair reductions. 3659 if (IsPairwise) 3660 return BaseT::getMinMaxReductionCost(ValTy, CondTy, IsPairwise, IsUnsigned, 3661 CostKind); 3662 3663 std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, ValTy); 3664 3665 MVT MTy = LT.second; 3666 3667 int ISD; 3668 if (ValTy->isIntOrIntVectorTy()) { 3669 ISD = IsUnsigned ? ISD::UMIN : ISD::SMIN; 3670 } else { 3671 assert(ValTy->isFPOrFPVectorTy() && 3672 "Expected float point or integer vector type."); 3673 ISD = ISD::FMINNUM; 3674 } 3675 3676 // We use the Intel Architecture Code Analyzer(IACA) to measure the throughput 3677 // and make it as the cost. 3678 3679 static const CostTblEntry SSE2CostTblNoPairWise[] = { 3680 {ISD::UMIN, MVT::v2i16, 5}, // need pxors to use pminsw/pmaxsw 3681 {ISD::UMIN, MVT::v4i16, 7}, // need pxors to use pminsw/pmaxsw 3682 {ISD::UMIN, MVT::v8i16, 9}, // need pxors to use pminsw/pmaxsw 3683 }; 3684 3685 static const CostTblEntry SSE41CostTblNoPairWise[] = { 3686 {ISD::SMIN, MVT::v2i16, 3}, // same as sse2 3687 {ISD::SMIN, MVT::v4i16, 5}, // same as sse2 3688 {ISD::UMIN, MVT::v2i16, 5}, // same as sse2 3689 {ISD::UMIN, MVT::v4i16, 7}, // same as sse2 3690 {ISD::SMIN, MVT::v8i16, 4}, // phminposuw+xor 3691 {ISD::UMIN, MVT::v8i16, 4}, // FIXME: umin is cheaper than umax 3692 {ISD::SMIN, MVT::v2i8, 3}, // pminsb 3693 {ISD::SMIN, MVT::v4i8, 5}, // pminsb 3694 {ISD::SMIN, MVT::v8i8, 7}, // pminsb 3695 {ISD::SMIN, MVT::v16i8, 6}, 3696 {ISD::UMIN, MVT::v2i8, 3}, // same as sse2 3697 {ISD::UMIN, MVT::v4i8, 5}, // same as sse2 3698 {ISD::UMIN, MVT::v8i8, 7}, // same as sse2 3699 {ISD::UMIN, MVT::v16i8, 6}, // FIXME: umin is cheaper than umax 3700 }; 3701 3702 static const CostTblEntry AVX1CostTblNoPairWise[] = { 3703 {ISD::SMIN, MVT::v16i16, 6}, 3704 {ISD::UMIN, MVT::v16i16, 6}, // FIXME: umin is cheaper than umax 3705 {ISD::SMIN, MVT::v32i8, 8}, 3706 {ISD::UMIN, MVT::v32i8, 8}, 3707 }; 3708 3709 static const CostTblEntry AVX512BWCostTblNoPairWise[] = { 3710 {ISD::SMIN, MVT::v32i16, 8}, 3711 {ISD::UMIN, MVT::v32i16, 8}, // FIXME: umin is cheaper than umax 3712 {ISD::SMIN, MVT::v64i8, 10}, 3713 {ISD::UMIN, MVT::v64i8, 10}, 3714 }; 3715 3716 // Before legalizing the type, give a chance to look up illegal narrow types 3717 // in the table. 3718 // FIXME: Is there a better way to do this? 3719 EVT VT = TLI->getValueType(DL, ValTy); 3720 if (VT.isSimple()) { 3721 MVT MTy = VT.getSimpleVT(); 3722 if (ST->hasBWI()) 3723 if (const auto *Entry = CostTableLookup(AVX512BWCostTblNoPairWise, ISD, MTy)) 3724 return Entry->Cost; 3725 3726 if (ST->hasAVX()) 3727 if (const auto *Entry = CostTableLookup(AVX1CostTblNoPairWise, ISD, MTy)) 3728 return Entry->Cost; 3729 3730 if (ST->hasSSE41()) 3731 if (const auto *Entry = CostTableLookup(SSE41CostTblNoPairWise, ISD, MTy)) 3732 return Entry->Cost; 3733 3734 if (ST->hasSSE2()) 3735 if (const auto *Entry = CostTableLookup(SSE2CostTblNoPairWise, ISD, MTy)) 3736 return Entry->Cost; 3737 } 3738 3739 auto *ValVTy = cast<FixedVectorType>(ValTy); 3740 unsigned NumVecElts = ValVTy->getNumElements(); 3741 3742 auto *Ty = ValVTy; 3743 unsigned MinMaxCost = 0; 3744 if (LT.first != 1 && MTy.isVector() && 3745 MTy.getVectorNumElements() < ValVTy->getNumElements()) { 3746 // Type needs to be split. We need LT.first - 1 operations ops. 3747 Ty = FixedVectorType::get(ValVTy->getElementType(), 3748 MTy.getVectorNumElements()); 3749 auto *SubCondTy = FixedVectorType::get(CondTy->getElementType(), 3750 MTy.getVectorNumElements()); 3751 MinMaxCost = getMinMaxCost(Ty, SubCondTy, IsUnsigned); 3752 MinMaxCost *= LT.first - 1; 3753 NumVecElts = MTy.getVectorNumElements(); 3754 } 3755 3756 if (ST->hasBWI()) 3757 if (const auto *Entry = CostTableLookup(AVX512BWCostTblNoPairWise, ISD, MTy)) 3758 return MinMaxCost + Entry->Cost; 3759 3760 if (ST->hasAVX()) 3761 if (const auto *Entry = CostTableLookup(AVX1CostTblNoPairWise, ISD, MTy)) 3762 return MinMaxCost + Entry->Cost; 3763 3764 if (ST->hasSSE41()) 3765 if (const auto *Entry = CostTableLookup(SSE41CostTblNoPairWise, ISD, MTy)) 3766 return MinMaxCost + Entry->Cost; 3767 3768 if (ST->hasSSE2()) 3769 if (const auto *Entry = CostTableLookup(SSE2CostTblNoPairWise, ISD, MTy)) 3770 return MinMaxCost + Entry->Cost; 3771 3772 unsigned ScalarSize = ValTy->getScalarSizeInBits(); 3773 3774 // Special case power of 2 reductions where the scalar type isn't changed 3775 // by type legalization. 3776 if (!isPowerOf2_32(ValVTy->getNumElements()) || 3777 ScalarSize != MTy.getScalarSizeInBits()) 3778 return BaseT::getMinMaxReductionCost(ValTy, CondTy, IsPairwise, IsUnsigned, 3779 CostKind); 3780 3781 // Now handle reduction with the legal type, taking into account size changes 3782 // at each level. 3783 while (NumVecElts > 1) { 3784 // Determine the size of the remaining vector we need to reduce. 3785 unsigned Size = NumVecElts * ScalarSize; 3786 NumVecElts /= 2; 3787 // If we're reducing from 256/512 bits, use an extract_subvector. 3788 if (Size > 128) { 3789 auto *SubTy = FixedVectorType::get(ValVTy->getElementType(), NumVecElts); 3790 MinMaxCost += 3791 getShuffleCost(TTI::SK_ExtractSubvector, Ty, NumVecElts, SubTy); 3792 Ty = SubTy; 3793 } else if (Size == 128) { 3794 // Reducing from 128 bits is a permute of v2f64/v2i64. 3795 VectorType *ShufTy; 3796 if (ValTy->isFloatingPointTy()) 3797 ShufTy = 3798 FixedVectorType::get(Type::getDoubleTy(ValTy->getContext()), 2); 3799 else 3800 ShufTy = FixedVectorType::get(Type::getInt64Ty(ValTy->getContext()), 2); 3801 MinMaxCost += 3802 getShuffleCost(TTI::SK_PermuteSingleSrc, ShufTy, 0, nullptr); 3803 } else if (Size == 64) { 3804 // Reducing from 64 bits is a shuffle of v4f32/v4i32. 3805 FixedVectorType *ShufTy; 3806 if (ValTy->isFloatingPointTy()) 3807 ShufTy = FixedVectorType::get(Type::getFloatTy(ValTy->getContext()), 4); 3808 else 3809 ShufTy = FixedVectorType::get(Type::getInt32Ty(ValTy->getContext()), 4); 3810 MinMaxCost += 3811 getShuffleCost(TTI::SK_PermuteSingleSrc, ShufTy, 0, nullptr); 3812 } else { 3813 // Reducing from smaller size is a shift by immediate. 3814 auto *ShiftTy = FixedVectorType::get( 3815 Type::getIntNTy(ValTy->getContext(), Size), 128 / Size); 3816 MinMaxCost += getArithmeticInstrCost( 3817 Instruction::LShr, ShiftTy, TTI::TCK_RecipThroughput, 3818 TargetTransformInfo::OK_AnyValue, 3819 TargetTransformInfo::OK_UniformConstantValue, 3820 TargetTransformInfo::OP_None, TargetTransformInfo::OP_None); 3821 } 3822 3823 // Add the arithmetic op for this level. 3824 auto *SubCondTy = 3825 FixedVectorType::get(CondTy->getElementType(), Ty->getNumElements()); 3826 MinMaxCost += getMinMaxCost(Ty, SubCondTy, IsUnsigned); 3827 } 3828 3829 // Add the final extract element to the cost. 3830 return MinMaxCost + getVectorInstrCost(Instruction::ExtractElement, Ty, 0); 3831 } 3832 3833 /// Calculate the cost of materializing a 64-bit value. This helper 3834 /// method might only calculate a fraction of a larger immediate. Therefore it 3835 /// is valid to return a cost of ZERO. 3836 int X86TTIImpl::getIntImmCost(int64_t Val) { 3837 if (Val == 0) 3838 return TTI::TCC_Free; 3839 3840 if (isInt<32>(Val)) 3841 return TTI::TCC_Basic; 3842 3843 return 2 * TTI::TCC_Basic; 3844 } 3845 3846 int X86TTIImpl::getIntImmCost(const APInt &Imm, Type *Ty, 3847 TTI::TargetCostKind CostKind) { 3848 assert(Ty->isIntegerTy()); 3849 3850 unsigned BitSize = Ty->getPrimitiveSizeInBits(); 3851 if (BitSize == 0) 3852 return ~0U; 3853 3854 // Never hoist constants larger than 128bit, because this might lead to 3855 // incorrect code generation or assertions in codegen. 3856 // Fixme: Create a cost model for types larger than i128 once the codegen 3857 // issues have been fixed. 3858 if (BitSize > 128) 3859 return TTI::TCC_Free; 3860 3861 if (Imm == 0) 3862 return TTI::TCC_Free; 3863 3864 // Sign-extend all constants to a multiple of 64-bit. 3865 APInt ImmVal = Imm; 3866 if (BitSize % 64 != 0) 3867 ImmVal = Imm.sext(alignTo(BitSize, 64)); 3868 3869 // Split the constant into 64-bit chunks and calculate the cost for each 3870 // chunk. 3871 int Cost = 0; 3872 for (unsigned ShiftVal = 0; ShiftVal < BitSize; ShiftVal += 64) { 3873 APInt Tmp = ImmVal.ashr(ShiftVal).sextOrTrunc(64); 3874 int64_t Val = Tmp.getSExtValue(); 3875 Cost += getIntImmCost(Val); 3876 } 3877 // We need at least one instruction to materialize the constant. 3878 return std::max(1, Cost); 3879 } 3880 3881 int X86TTIImpl::getIntImmCostInst(unsigned Opcode, unsigned Idx, 3882 const APInt &Imm, Type *Ty, 3883 TTI::TargetCostKind CostKind, 3884 Instruction *Inst) { 3885 assert(Ty->isIntegerTy()); 3886 3887 unsigned BitSize = Ty->getPrimitiveSizeInBits(); 3888 // There is no cost model for constants with a bit size of 0. Return TCC_Free 3889 // here, so that constant hoisting will ignore this constant. 3890 if (BitSize == 0) 3891 return TTI::TCC_Free; 3892 3893 unsigned ImmIdx = ~0U; 3894 switch (Opcode) { 3895 default: 3896 return TTI::TCC_Free; 3897 case Instruction::GetElementPtr: 3898 // Always hoist the base address of a GetElementPtr. This prevents the 3899 // creation of new constants for every base constant that gets constant 3900 // folded with the offset. 3901 if (Idx == 0) 3902 return 2 * TTI::TCC_Basic; 3903 return TTI::TCC_Free; 3904 case Instruction::Store: 3905 ImmIdx = 0; 3906 break; 3907 case Instruction::ICmp: 3908 // This is an imperfect hack to prevent constant hoisting of 3909 // compares that might be trying to check if a 64-bit value fits in 3910 // 32-bits. The backend can optimize these cases using a right shift by 32. 3911 // Ideally we would check the compare predicate here. There also other 3912 // similar immediates the backend can use shifts for. 3913 if (Idx == 1 && Imm.getBitWidth() == 64) { 3914 uint64_t ImmVal = Imm.getZExtValue(); 3915 if (ImmVal == 0x100000000ULL || ImmVal == 0xffffffff) 3916 return TTI::TCC_Free; 3917 } 3918 ImmIdx = 1; 3919 break; 3920 case Instruction::And: 3921 // We support 64-bit ANDs with immediates with 32-bits of leading zeroes 3922 // by using a 32-bit operation with implicit zero extension. Detect such 3923 // immediates here as the normal path expects bit 31 to be sign extended. 3924 if (Idx == 1 && Imm.getBitWidth() == 64 && isUInt<32>(Imm.getZExtValue())) 3925 return TTI::TCC_Free; 3926 ImmIdx = 1; 3927 break; 3928 case Instruction::Add: 3929 case Instruction::Sub: 3930 // For add/sub, we can use the opposite instruction for INT32_MIN. 3931 if (Idx == 1 && Imm.getBitWidth() == 64 && Imm.getZExtValue() == 0x80000000) 3932 return TTI::TCC_Free; 3933 ImmIdx = 1; 3934 break; 3935 case Instruction::UDiv: 3936 case Instruction::SDiv: 3937 case Instruction::URem: 3938 case Instruction::SRem: 3939 // Division by constant is typically expanded later into a different 3940 // instruction sequence. This completely changes the constants. 3941 // Report them as "free" to stop ConstantHoist from marking them as opaque. 3942 return TTI::TCC_Free; 3943 case Instruction::Mul: 3944 case Instruction::Or: 3945 case Instruction::Xor: 3946 ImmIdx = 1; 3947 break; 3948 // Always return TCC_Free for the shift value of a shift instruction. 3949 case Instruction::Shl: 3950 case Instruction::LShr: 3951 case Instruction::AShr: 3952 if (Idx == 1) 3953 return TTI::TCC_Free; 3954 break; 3955 case Instruction::Trunc: 3956 case Instruction::ZExt: 3957 case Instruction::SExt: 3958 case Instruction::IntToPtr: 3959 case Instruction::PtrToInt: 3960 case Instruction::BitCast: 3961 case Instruction::PHI: 3962 case Instruction::Call: 3963 case Instruction::Select: 3964 case Instruction::Ret: 3965 case Instruction::Load: 3966 break; 3967 } 3968 3969 if (Idx == ImmIdx) { 3970 int NumConstants = divideCeil(BitSize, 64); 3971 int Cost = X86TTIImpl::getIntImmCost(Imm, Ty, CostKind); 3972 return (Cost <= NumConstants * TTI::TCC_Basic) 3973 ? static_cast<int>(TTI::TCC_Free) 3974 : Cost; 3975 } 3976 3977 return X86TTIImpl::getIntImmCost(Imm, Ty, CostKind); 3978 } 3979 3980 int X86TTIImpl::getIntImmCostIntrin(Intrinsic::ID IID, unsigned Idx, 3981 const APInt &Imm, Type *Ty, 3982 TTI::TargetCostKind CostKind) { 3983 assert(Ty->isIntegerTy()); 3984 3985 unsigned BitSize = Ty->getPrimitiveSizeInBits(); 3986 // There is no cost model for constants with a bit size of 0. Return TCC_Free 3987 // here, so that constant hoisting will ignore this constant. 3988 if (BitSize == 0) 3989 return TTI::TCC_Free; 3990 3991 switch (IID) { 3992 default: 3993 return TTI::TCC_Free; 3994 case Intrinsic::sadd_with_overflow: 3995 case Intrinsic::uadd_with_overflow: 3996 case Intrinsic::ssub_with_overflow: 3997 case Intrinsic::usub_with_overflow: 3998 case Intrinsic::smul_with_overflow: 3999 case Intrinsic::umul_with_overflow: 4000 if ((Idx == 1) && Imm.getBitWidth() <= 64 && isInt<32>(Imm.getSExtValue())) 4001 return TTI::TCC_Free; 4002 break; 4003 case Intrinsic::experimental_stackmap: 4004 if ((Idx < 2) || (Imm.getBitWidth() <= 64 && isInt<64>(Imm.getSExtValue()))) 4005 return TTI::TCC_Free; 4006 break; 4007 case Intrinsic::experimental_patchpoint_void: 4008 case Intrinsic::experimental_patchpoint_i64: 4009 if ((Idx < 4) || (Imm.getBitWidth() <= 64 && isInt<64>(Imm.getSExtValue()))) 4010 return TTI::TCC_Free; 4011 break; 4012 } 4013 return X86TTIImpl::getIntImmCost(Imm, Ty, CostKind); 4014 } 4015 4016 unsigned 4017 X86TTIImpl::getCFInstrCost(unsigned Opcode, TTI::TargetCostKind CostKind) { 4018 if (CostKind != TTI::TCK_RecipThroughput) 4019 return Opcode == Instruction::PHI ? 0 : 1; 4020 // Branches are assumed to be predicted. 4021 return CostKind == TTI::TCK_RecipThroughput ? 0 : 1; 4022 } 4023 4024 int X86TTIImpl::getGatherOverhead() const { 4025 // Some CPUs have more overhead for gather. The specified overhead is relative 4026 // to the Load operation. "2" is the number provided by Intel architects. This 4027 // parameter is used for cost estimation of Gather Op and comparison with 4028 // other alternatives. 4029 // TODO: Remove the explicit hasAVX512()?, That would mean we would only 4030 // enable gather with a -march. 4031 if (ST->hasAVX512() || (ST->hasAVX2() && ST->hasFastGather())) 4032 return 2; 4033 4034 return 1024; 4035 } 4036 4037 int X86TTIImpl::getScatterOverhead() const { 4038 if (ST->hasAVX512()) 4039 return 2; 4040 4041 return 1024; 4042 } 4043 4044 // Return an average cost of Gather / Scatter instruction, maybe improved later 4045 int X86TTIImpl::getGSVectorCost(unsigned Opcode, Type *SrcVTy, const Value *Ptr, 4046 Align Alignment, unsigned AddressSpace) { 4047 4048 assert(isa<VectorType>(SrcVTy) && "Unexpected type in getGSVectorCost"); 4049 unsigned VF = cast<FixedVectorType>(SrcVTy)->getNumElements(); 4050 4051 // Try to reduce index size from 64 bit (default for GEP) 4052 // to 32. It is essential for VF 16. If the index can't be reduced to 32, the 4053 // operation will use 16 x 64 indices which do not fit in a zmm and needs 4054 // to split. Also check that the base pointer is the same for all lanes, 4055 // and that there's at most one variable index. 4056 auto getIndexSizeInBits = [](const Value *Ptr, const DataLayout &DL) { 4057 unsigned IndexSize = DL.getPointerSizeInBits(); 4058 const GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr); 4059 if (IndexSize < 64 || !GEP) 4060 return IndexSize; 4061 4062 unsigned NumOfVarIndices = 0; 4063 const Value *Ptrs = GEP->getPointerOperand(); 4064 if (Ptrs->getType()->isVectorTy() && !getSplatValue(Ptrs)) 4065 return IndexSize; 4066 for (unsigned i = 1; i < GEP->getNumOperands(); ++i) { 4067 if (isa<Constant>(GEP->getOperand(i))) 4068 continue; 4069 Type *IndxTy = GEP->getOperand(i)->getType(); 4070 if (auto *IndexVTy = dyn_cast<VectorType>(IndxTy)) 4071 IndxTy = IndexVTy->getElementType(); 4072 if ((IndxTy->getPrimitiveSizeInBits() == 64 && 4073 !isa<SExtInst>(GEP->getOperand(i))) || 4074 ++NumOfVarIndices > 1) 4075 return IndexSize; // 64 4076 } 4077 return (unsigned)32; 4078 }; 4079 4080 // Trying to reduce IndexSize to 32 bits for vector 16. 4081 // By default the IndexSize is equal to pointer size. 4082 unsigned IndexSize = (ST->hasAVX512() && VF >= 16) 4083 ? getIndexSizeInBits(Ptr, DL) 4084 : DL.getPointerSizeInBits(); 4085 4086 auto *IndexVTy = FixedVectorType::get( 4087 IntegerType::get(SrcVTy->getContext(), IndexSize), VF); 4088 std::pair<int, MVT> IdxsLT = TLI->getTypeLegalizationCost(DL, IndexVTy); 4089 std::pair<int, MVT> SrcLT = TLI->getTypeLegalizationCost(DL, SrcVTy); 4090 int SplitFactor = std::max(IdxsLT.first, SrcLT.first); 4091 if (SplitFactor > 1) { 4092 // Handle splitting of vector of pointers 4093 auto *SplitSrcTy = 4094 FixedVectorType::get(SrcVTy->getScalarType(), VF / SplitFactor); 4095 return SplitFactor * getGSVectorCost(Opcode, SplitSrcTy, Ptr, Alignment, 4096 AddressSpace); 4097 } 4098 4099 // The gather / scatter cost is given by Intel architects. It is a rough 4100 // number since we are looking at one instruction in a time. 4101 const int GSOverhead = (Opcode == Instruction::Load) 4102 ? getGatherOverhead() 4103 : getScatterOverhead(); 4104 return GSOverhead + VF * getMemoryOpCost(Opcode, SrcVTy->getScalarType(), 4105 MaybeAlign(Alignment), AddressSpace, 4106 TTI::TCK_RecipThroughput); 4107 } 4108 4109 /// Return the cost of full scalarization of gather / scatter operation. 4110 /// 4111 /// Opcode - Load or Store instruction. 4112 /// SrcVTy - The type of the data vector that should be gathered or scattered. 4113 /// VariableMask - The mask is non-constant at compile time. 4114 /// Alignment - Alignment for one element. 4115 /// AddressSpace - pointer[s] address space. 4116 /// 4117 int X86TTIImpl::getGSScalarCost(unsigned Opcode, Type *SrcVTy, 4118 bool VariableMask, Align Alignment, 4119 unsigned AddressSpace) { 4120 unsigned VF = cast<FixedVectorType>(SrcVTy)->getNumElements(); 4121 APInt DemandedElts = APInt::getAllOnesValue(VF); 4122 TTI::TargetCostKind CostKind = TTI::TCK_RecipThroughput; 4123 4124 int MaskUnpackCost = 0; 4125 if (VariableMask) { 4126 auto *MaskTy = 4127 FixedVectorType::get(Type::getInt1Ty(SrcVTy->getContext()), VF); 4128 MaskUnpackCost = 4129 getScalarizationOverhead(MaskTy, DemandedElts, false, true); 4130 int ScalarCompareCost = getCmpSelInstrCost( 4131 Instruction::ICmp, Type::getInt1Ty(SrcVTy->getContext()), nullptr, 4132 CmpInst::BAD_ICMP_PREDICATE, CostKind); 4133 int BranchCost = getCFInstrCost(Instruction::Br, CostKind); 4134 MaskUnpackCost += VF * (BranchCost + ScalarCompareCost); 4135 } 4136 4137 // The cost of the scalar loads/stores. 4138 int MemoryOpCost = VF * getMemoryOpCost(Opcode, SrcVTy->getScalarType(), 4139 MaybeAlign(Alignment), AddressSpace, 4140 CostKind); 4141 4142 int InsertExtractCost = 0; 4143 if (Opcode == Instruction::Load) 4144 for (unsigned i = 0; i < VF; ++i) 4145 // Add the cost of inserting each scalar load into the vector 4146 InsertExtractCost += 4147 getVectorInstrCost(Instruction::InsertElement, SrcVTy, i); 4148 else 4149 for (unsigned i = 0; i < VF; ++i) 4150 // Add the cost of extracting each element out of the data vector 4151 InsertExtractCost += 4152 getVectorInstrCost(Instruction::ExtractElement, SrcVTy, i); 4153 4154 return MemoryOpCost + MaskUnpackCost + InsertExtractCost; 4155 } 4156 4157 /// Calculate the cost of Gather / Scatter operation 4158 int X86TTIImpl::getGatherScatterOpCost(unsigned Opcode, Type *SrcVTy, 4159 const Value *Ptr, bool VariableMask, 4160 Align Alignment, 4161 TTI::TargetCostKind CostKind, 4162 const Instruction *I = nullptr) { 4163 4164 if (CostKind != TTI::TCK_RecipThroughput) 4165 return 1; 4166 4167 assert(SrcVTy->isVectorTy() && "Unexpected data type for Gather/Scatter"); 4168 unsigned VF = cast<FixedVectorType>(SrcVTy)->getNumElements(); 4169 PointerType *PtrTy = dyn_cast<PointerType>(Ptr->getType()); 4170 if (!PtrTy && Ptr->getType()->isVectorTy()) 4171 PtrTy = dyn_cast<PointerType>( 4172 cast<VectorType>(Ptr->getType())->getElementType()); 4173 assert(PtrTy && "Unexpected type for Ptr argument"); 4174 unsigned AddressSpace = PtrTy->getAddressSpace(); 4175 4176 bool Scalarize = false; 4177 if ((Opcode == Instruction::Load && 4178 !isLegalMaskedGather(SrcVTy, Align(Alignment))) || 4179 (Opcode == Instruction::Store && 4180 !isLegalMaskedScatter(SrcVTy, Align(Alignment)))) 4181 Scalarize = true; 4182 // Gather / Scatter for vector 2 is not profitable on KNL / SKX 4183 // Vector-4 of gather/scatter instruction does not exist on KNL. 4184 // We can extend it to 8 elements, but zeroing upper bits of 4185 // the mask vector will add more instructions. Right now we give the scalar 4186 // cost of vector-4 for KNL. TODO: Check, maybe the gather/scatter instruction 4187 // is better in the VariableMask case. 4188 if (ST->hasAVX512() && (VF == 2 || (VF == 4 && !ST->hasVLX()))) 4189 Scalarize = true; 4190 4191 if (Scalarize) 4192 return getGSScalarCost(Opcode, SrcVTy, VariableMask, Alignment, 4193 AddressSpace); 4194 4195 return getGSVectorCost(Opcode, SrcVTy, Ptr, Alignment, AddressSpace); 4196 } 4197 4198 bool X86TTIImpl::isLSRCostLess(TargetTransformInfo::LSRCost &C1, 4199 TargetTransformInfo::LSRCost &C2) { 4200 // X86 specific here are "instruction number 1st priority". 4201 return std::tie(C1.Insns, C1.NumRegs, C1.AddRecCost, 4202 C1.NumIVMuls, C1.NumBaseAdds, 4203 C1.ScaleCost, C1.ImmCost, C1.SetupCost) < 4204 std::tie(C2.Insns, C2.NumRegs, C2.AddRecCost, 4205 C2.NumIVMuls, C2.NumBaseAdds, 4206 C2.ScaleCost, C2.ImmCost, C2.SetupCost); 4207 } 4208 4209 bool X86TTIImpl::canMacroFuseCmp() { 4210 return ST->hasMacroFusion() || ST->hasBranchFusion(); 4211 } 4212 4213 bool X86TTIImpl::isLegalMaskedLoad(Type *DataTy, Align Alignment) { 4214 if (!ST->hasAVX()) 4215 return false; 4216 4217 // The backend can't handle a single element vector. 4218 if (isa<VectorType>(DataTy) && 4219 cast<FixedVectorType>(DataTy)->getNumElements() == 1) 4220 return false; 4221 Type *ScalarTy = DataTy->getScalarType(); 4222 4223 if (ScalarTy->isPointerTy()) 4224 return true; 4225 4226 if (ScalarTy->isFloatTy() || ScalarTy->isDoubleTy()) 4227 return true; 4228 4229 if (!ScalarTy->isIntegerTy()) 4230 return false; 4231 4232 unsigned IntWidth = ScalarTy->getIntegerBitWidth(); 4233 return IntWidth == 32 || IntWidth == 64 || 4234 ((IntWidth == 8 || IntWidth == 16) && ST->hasBWI()); 4235 } 4236 4237 bool X86TTIImpl::isLegalMaskedStore(Type *DataType, Align Alignment) { 4238 return isLegalMaskedLoad(DataType, Alignment); 4239 } 4240 4241 bool X86TTIImpl::isLegalNTLoad(Type *DataType, Align Alignment) { 4242 unsigned DataSize = DL.getTypeStoreSize(DataType); 4243 // The only supported nontemporal loads are for aligned vectors of 16 or 32 4244 // bytes. Note that 32-byte nontemporal vector loads are supported by AVX2 4245 // (the equivalent stores only require AVX). 4246 if (Alignment >= DataSize && (DataSize == 16 || DataSize == 32)) 4247 return DataSize == 16 ? ST->hasSSE1() : ST->hasAVX2(); 4248 4249 return false; 4250 } 4251 4252 bool X86TTIImpl::isLegalNTStore(Type *DataType, Align Alignment) { 4253 unsigned DataSize = DL.getTypeStoreSize(DataType); 4254 4255 // SSE4A supports nontemporal stores of float and double at arbitrary 4256 // alignment. 4257 if (ST->hasSSE4A() && (DataType->isFloatTy() || DataType->isDoubleTy())) 4258 return true; 4259 4260 // Besides the SSE4A subtarget exception above, only aligned stores are 4261 // available nontemporaly on any other subtarget. And only stores with a size 4262 // of 4..32 bytes (powers of 2, only) are permitted. 4263 if (Alignment < DataSize || DataSize < 4 || DataSize > 32 || 4264 !isPowerOf2_32(DataSize)) 4265 return false; 4266 4267 // 32-byte vector nontemporal stores are supported by AVX (the equivalent 4268 // loads require AVX2). 4269 if (DataSize == 32) 4270 return ST->hasAVX(); 4271 else if (DataSize == 16) 4272 return ST->hasSSE1(); 4273 return true; 4274 } 4275 4276 bool X86TTIImpl::isLegalMaskedExpandLoad(Type *DataTy) { 4277 if (!isa<VectorType>(DataTy)) 4278 return false; 4279 4280 if (!ST->hasAVX512()) 4281 return false; 4282 4283 // The backend can't handle a single element vector. 4284 if (cast<FixedVectorType>(DataTy)->getNumElements() == 1) 4285 return false; 4286 4287 Type *ScalarTy = cast<VectorType>(DataTy)->getElementType(); 4288 4289 if (ScalarTy->isFloatTy() || ScalarTy->isDoubleTy()) 4290 return true; 4291 4292 if (!ScalarTy->isIntegerTy()) 4293 return false; 4294 4295 unsigned IntWidth = ScalarTy->getIntegerBitWidth(); 4296 return IntWidth == 32 || IntWidth == 64 || 4297 ((IntWidth == 8 || IntWidth == 16) && ST->hasVBMI2()); 4298 } 4299 4300 bool X86TTIImpl::isLegalMaskedCompressStore(Type *DataTy) { 4301 return isLegalMaskedExpandLoad(DataTy); 4302 } 4303 4304 bool X86TTIImpl::isLegalMaskedGather(Type *DataTy, Align Alignment) { 4305 // Some CPUs have better gather performance than others. 4306 // TODO: Remove the explicit ST->hasAVX512()?, That would mean we would only 4307 // enable gather with a -march. 4308 if (!(ST->hasAVX512() || (ST->hasFastGather() && ST->hasAVX2()))) 4309 return false; 4310 4311 // This function is called now in two cases: from the Loop Vectorizer 4312 // and from the Scalarizer. 4313 // When the Loop Vectorizer asks about legality of the feature, 4314 // the vectorization factor is not calculated yet. The Loop Vectorizer 4315 // sends a scalar type and the decision is based on the width of the 4316 // scalar element. 4317 // Later on, the cost model will estimate usage this intrinsic based on 4318 // the vector type. 4319 // The Scalarizer asks again about legality. It sends a vector type. 4320 // In this case we can reject non-power-of-2 vectors. 4321 // We also reject single element vectors as the type legalizer can't 4322 // scalarize it. 4323 if (auto *DataVTy = dyn_cast<FixedVectorType>(DataTy)) { 4324 unsigned NumElts = DataVTy->getNumElements(); 4325 if (NumElts == 1) 4326 return false; 4327 } 4328 Type *ScalarTy = DataTy->getScalarType(); 4329 if (ScalarTy->isPointerTy()) 4330 return true; 4331 4332 if (ScalarTy->isFloatTy() || ScalarTy->isDoubleTy()) 4333 return true; 4334 4335 if (!ScalarTy->isIntegerTy()) 4336 return false; 4337 4338 unsigned IntWidth = ScalarTy->getIntegerBitWidth(); 4339 return IntWidth == 32 || IntWidth == 64; 4340 } 4341 4342 bool X86TTIImpl::isLegalMaskedScatter(Type *DataType, Align Alignment) { 4343 // AVX2 doesn't support scatter 4344 if (!ST->hasAVX512()) 4345 return false; 4346 return isLegalMaskedGather(DataType, Alignment); 4347 } 4348 4349 bool X86TTIImpl::hasDivRemOp(Type *DataType, bool IsSigned) { 4350 EVT VT = TLI->getValueType(DL, DataType); 4351 return TLI->isOperationLegal(IsSigned ? ISD::SDIVREM : ISD::UDIVREM, VT); 4352 } 4353 4354 bool X86TTIImpl::isFCmpOrdCheaperThanFCmpZero(Type *Ty) { 4355 return false; 4356 } 4357 4358 bool X86TTIImpl::areInlineCompatible(const Function *Caller, 4359 const Function *Callee) const { 4360 const TargetMachine &TM = getTLI()->getTargetMachine(); 4361 4362 // Work this as a subsetting of subtarget features. 4363 const FeatureBitset &CallerBits = 4364 TM.getSubtargetImpl(*Caller)->getFeatureBits(); 4365 const FeatureBitset &CalleeBits = 4366 TM.getSubtargetImpl(*Callee)->getFeatureBits(); 4367 4368 FeatureBitset RealCallerBits = CallerBits & ~InlineFeatureIgnoreList; 4369 FeatureBitset RealCalleeBits = CalleeBits & ~InlineFeatureIgnoreList; 4370 return (RealCallerBits & RealCalleeBits) == RealCalleeBits; 4371 } 4372 4373 bool X86TTIImpl::areFunctionArgsABICompatible( 4374 const Function *Caller, const Function *Callee, 4375 SmallPtrSetImpl<Argument *> &Args) const { 4376 if (!BaseT::areFunctionArgsABICompatible(Caller, Callee, Args)) 4377 return false; 4378 4379 // If we get here, we know the target features match. If one function 4380 // considers 512-bit vectors legal and the other does not, consider them 4381 // incompatible. 4382 const TargetMachine &TM = getTLI()->getTargetMachine(); 4383 4384 if (TM.getSubtarget<X86Subtarget>(*Caller).useAVX512Regs() == 4385 TM.getSubtarget<X86Subtarget>(*Callee).useAVX512Regs()) 4386 return true; 4387 4388 // Consider the arguments compatible if they aren't vectors or aggregates. 4389 // FIXME: Look at the size of vectors. 4390 // FIXME: Look at the element types of aggregates to see if there are vectors. 4391 // FIXME: The API of this function seems intended to allow arguments 4392 // to be removed from the set, but the caller doesn't check if the set 4393 // becomes empty so that may not work in practice. 4394 return llvm::none_of(Args, [](Argument *A) { 4395 auto *EltTy = cast<PointerType>(A->getType())->getElementType(); 4396 return EltTy->isVectorTy() || EltTy->isAggregateType(); 4397 }); 4398 } 4399 4400 X86TTIImpl::TTI::MemCmpExpansionOptions 4401 X86TTIImpl::enableMemCmpExpansion(bool OptSize, bool IsZeroCmp) const { 4402 TTI::MemCmpExpansionOptions Options; 4403 Options.MaxNumLoads = TLI->getMaxExpandSizeMemcmp(OptSize); 4404 Options.NumLoadsPerBlock = 2; 4405 // All GPR and vector loads can be unaligned. 4406 Options.AllowOverlappingLoads = true; 4407 if (IsZeroCmp) { 4408 // Only enable vector loads for equality comparison. Right now the vector 4409 // version is not as fast for three way compare (see #33329). 4410 const unsigned PreferredWidth = ST->getPreferVectorWidth(); 4411 if (PreferredWidth >= 512 && ST->hasAVX512()) Options.LoadSizes.push_back(64); 4412 if (PreferredWidth >= 256 && ST->hasAVX()) Options.LoadSizes.push_back(32); 4413 if (PreferredWidth >= 128 && ST->hasSSE2()) Options.LoadSizes.push_back(16); 4414 } 4415 if (ST->is64Bit()) { 4416 Options.LoadSizes.push_back(8); 4417 } 4418 Options.LoadSizes.push_back(4); 4419 Options.LoadSizes.push_back(2); 4420 Options.LoadSizes.push_back(1); 4421 return Options; 4422 } 4423 4424 bool X86TTIImpl::enableInterleavedAccessVectorization() { 4425 // TODO: We expect this to be beneficial regardless of arch, 4426 // but there are currently some unexplained performance artifacts on Atom. 4427 // As a temporary solution, disable on Atom. 4428 return !(ST->isAtom()); 4429 } 4430 4431 // Get estimation for interleaved load/store operations for AVX2. 4432 // \p Factor is the interleaved-access factor (stride) - number of 4433 // (interleaved) elements in the group. 4434 // \p Indices contains the indices for a strided load: when the 4435 // interleaved load has gaps they indicate which elements are used. 4436 // If Indices is empty (or if the number of indices is equal to the size 4437 // of the interleaved-access as given in \p Factor) the access has no gaps. 4438 // 4439 // As opposed to AVX-512, AVX2 does not have generic shuffles that allow 4440 // computing the cost using a generic formula as a function of generic 4441 // shuffles. We therefore use a lookup table instead, filled according to 4442 // the instruction sequences that codegen currently generates. 4443 int X86TTIImpl::getInterleavedMemoryOpCostAVX2( 4444 unsigned Opcode, FixedVectorType *VecTy, unsigned Factor, 4445 ArrayRef<unsigned> Indices, Align Alignment, unsigned AddressSpace, 4446 TTI::TargetCostKind CostKind, bool UseMaskForCond, bool UseMaskForGaps) { 4447 4448 if (UseMaskForCond || UseMaskForGaps) 4449 return BaseT::getInterleavedMemoryOpCost(Opcode, VecTy, Factor, Indices, 4450 Alignment, AddressSpace, CostKind, 4451 UseMaskForCond, UseMaskForGaps); 4452 4453 // We currently Support only fully-interleaved groups, with no gaps. 4454 // TODO: Support also strided loads (interleaved-groups with gaps). 4455 if (Indices.size() && Indices.size() != Factor) 4456 return BaseT::getInterleavedMemoryOpCost(Opcode, VecTy, Factor, Indices, 4457 Alignment, AddressSpace, 4458 CostKind); 4459 4460 // VecTy for interleave memop is <VF*Factor x Elt>. 4461 // So, for VF=4, Interleave Factor = 3, Element type = i32 we have 4462 // VecTy = <12 x i32>. 4463 MVT LegalVT = getTLI()->getTypeLegalizationCost(DL, VecTy).second; 4464 4465 // This function can be called with VecTy=<6xi128>, Factor=3, in which case 4466 // the VF=2, while v2i128 is an unsupported MVT vector type 4467 // (see MachineValueType.h::getVectorVT()). 4468 if (!LegalVT.isVector()) 4469 return BaseT::getInterleavedMemoryOpCost(Opcode, VecTy, Factor, Indices, 4470 Alignment, AddressSpace, 4471 CostKind); 4472 4473 unsigned VF = VecTy->getNumElements() / Factor; 4474 Type *ScalarTy = VecTy->getElementType(); 4475 4476 // Calculate the number of memory operations (NumOfMemOps), required 4477 // for load/store the VecTy. 4478 unsigned VecTySize = DL.getTypeStoreSize(VecTy); 4479 unsigned LegalVTSize = LegalVT.getStoreSize(); 4480 unsigned NumOfMemOps = (VecTySize + LegalVTSize - 1) / LegalVTSize; 4481 4482 // Get the cost of one memory operation. 4483 auto *SingleMemOpTy = FixedVectorType::get(VecTy->getElementType(), 4484 LegalVT.getVectorNumElements()); 4485 unsigned MemOpCost = getMemoryOpCost(Opcode, SingleMemOpTy, 4486 MaybeAlign(Alignment), AddressSpace, 4487 CostKind); 4488 4489 auto *VT = FixedVectorType::get(ScalarTy, VF); 4490 EVT ETy = TLI->getValueType(DL, VT); 4491 if (!ETy.isSimple()) 4492 return BaseT::getInterleavedMemoryOpCost(Opcode, VecTy, Factor, Indices, 4493 Alignment, AddressSpace, 4494 CostKind); 4495 4496 // TODO: Complete for other data-types and strides. 4497 // Each combination of Stride, ElementTy and VF results in a different 4498 // sequence; The cost tables are therefore accessed with: 4499 // Factor (stride) and VectorType=VFxElemType. 4500 // The Cost accounts only for the shuffle sequence; 4501 // The cost of the loads/stores is accounted for separately. 4502 // 4503 static const CostTblEntry AVX2InterleavedLoadTbl[] = { 4504 { 2, MVT::v4i64, 6 }, //(load 8i64 and) deinterleave into 2 x 4i64 4505 { 2, MVT::v4f64, 6 }, //(load 8f64 and) deinterleave into 2 x 4f64 4506 4507 { 3, MVT::v2i8, 10 }, //(load 6i8 and) deinterleave into 3 x 2i8 4508 { 3, MVT::v4i8, 4 }, //(load 12i8 and) deinterleave into 3 x 4i8 4509 { 3, MVT::v8i8, 9 }, //(load 24i8 and) deinterleave into 3 x 8i8 4510 { 3, MVT::v16i8, 11}, //(load 48i8 and) deinterleave into 3 x 16i8 4511 { 3, MVT::v32i8, 13}, //(load 96i8 and) deinterleave into 3 x 32i8 4512 { 3, MVT::v8f32, 17 }, //(load 24f32 and)deinterleave into 3 x 8f32 4513 4514 { 4, MVT::v2i8, 12 }, //(load 8i8 and) deinterleave into 4 x 2i8 4515 { 4, MVT::v4i8, 4 }, //(load 16i8 and) deinterleave into 4 x 4i8 4516 { 4, MVT::v8i8, 20 }, //(load 32i8 and) deinterleave into 4 x 8i8 4517 { 4, MVT::v16i8, 39 }, //(load 64i8 and) deinterleave into 4 x 16i8 4518 { 4, MVT::v32i8, 80 }, //(load 128i8 and) deinterleave into 4 x 32i8 4519 4520 { 8, MVT::v8f32, 40 } //(load 64f32 and)deinterleave into 8 x 8f32 4521 }; 4522 4523 static const CostTblEntry AVX2InterleavedStoreTbl[] = { 4524 { 2, MVT::v4i64, 6 }, //interleave into 2 x 4i64 into 8i64 (and store) 4525 { 2, MVT::v4f64, 6 }, //interleave into 2 x 4f64 into 8f64 (and store) 4526 4527 { 3, MVT::v2i8, 7 }, //interleave 3 x 2i8 into 6i8 (and store) 4528 { 3, MVT::v4i8, 8 }, //interleave 3 x 4i8 into 12i8 (and store) 4529 { 3, MVT::v8i8, 11 }, //interleave 3 x 8i8 into 24i8 (and store) 4530 { 3, MVT::v16i8, 11 }, //interleave 3 x 16i8 into 48i8 (and store) 4531 { 3, MVT::v32i8, 13 }, //interleave 3 x 32i8 into 96i8 (and store) 4532 4533 { 4, MVT::v2i8, 12 }, //interleave 4 x 2i8 into 8i8 (and store) 4534 { 4, MVT::v4i8, 9 }, //interleave 4 x 4i8 into 16i8 (and store) 4535 { 4, MVT::v8i8, 10 }, //interleave 4 x 8i8 into 32i8 (and store) 4536 { 4, MVT::v16i8, 10 }, //interleave 4 x 16i8 into 64i8 (and store) 4537 { 4, MVT::v32i8, 12 } //interleave 4 x 32i8 into 128i8 (and store) 4538 }; 4539 4540 if (Opcode == Instruction::Load) { 4541 if (const auto *Entry = 4542 CostTableLookup(AVX2InterleavedLoadTbl, Factor, ETy.getSimpleVT())) 4543 return NumOfMemOps * MemOpCost + Entry->Cost; 4544 } else { 4545 assert(Opcode == Instruction::Store && 4546 "Expected Store Instruction at this point"); 4547 if (const auto *Entry = 4548 CostTableLookup(AVX2InterleavedStoreTbl, Factor, ETy.getSimpleVT())) 4549 return NumOfMemOps * MemOpCost + Entry->Cost; 4550 } 4551 4552 return BaseT::getInterleavedMemoryOpCost(Opcode, VecTy, Factor, Indices, 4553 Alignment, AddressSpace, CostKind); 4554 } 4555 4556 // Get estimation for interleaved load/store operations and strided load. 4557 // \p Indices contains indices for strided load. 4558 // \p Factor - the factor of interleaving. 4559 // AVX-512 provides 3-src shuffles that significantly reduces the cost. 4560 int X86TTIImpl::getInterleavedMemoryOpCostAVX512( 4561 unsigned Opcode, FixedVectorType *VecTy, unsigned Factor, 4562 ArrayRef<unsigned> Indices, Align Alignment, unsigned AddressSpace, 4563 TTI::TargetCostKind CostKind, bool UseMaskForCond, bool UseMaskForGaps) { 4564 4565 if (UseMaskForCond || UseMaskForGaps) 4566 return BaseT::getInterleavedMemoryOpCost(Opcode, VecTy, Factor, Indices, 4567 Alignment, AddressSpace, CostKind, 4568 UseMaskForCond, UseMaskForGaps); 4569 4570 // VecTy for interleave memop is <VF*Factor x Elt>. 4571 // So, for VF=4, Interleave Factor = 3, Element type = i32 we have 4572 // VecTy = <12 x i32>. 4573 4574 // Calculate the number of memory operations (NumOfMemOps), required 4575 // for load/store the VecTy. 4576 MVT LegalVT = getTLI()->getTypeLegalizationCost(DL, VecTy).second; 4577 unsigned VecTySize = DL.getTypeStoreSize(VecTy); 4578 unsigned LegalVTSize = LegalVT.getStoreSize(); 4579 unsigned NumOfMemOps = (VecTySize + LegalVTSize - 1) / LegalVTSize; 4580 4581 // Get the cost of one memory operation. 4582 auto *SingleMemOpTy = FixedVectorType::get(VecTy->getElementType(), 4583 LegalVT.getVectorNumElements()); 4584 unsigned MemOpCost = getMemoryOpCost(Opcode, SingleMemOpTy, 4585 MaybeAlign(Alignment), AddressSpace, 4586 CostKind); 4587 4588 unsigned VF = VecTy->getNumElements() / Factor; 4589 MVT VT = MVT::getVectorVT(MVT::getVT(VecTy->getScalarType()), VF); 4590 4591 if (Opcode == Instruction::Load) { 4592 // The tables (AVX512InterleavedLoadTbl and AVX512InterleavedStoreTbl) 4593 // contain the cost of the optimized shuffle sequence that the 4594 // X86InterleavedAccess pass will generate. 4595 // The cost of loads and stores are computed separately from the table. 4596 4597 // X86InterleavedAccess support only the following interleaved-access group. 4598 static const CostTblEntry AVX512InterleavedLoadTbl[] = { 4599 {3, MVT::v16i8, 12}, //(load 48i8 and) deinterleave into 3 x 16i8 4600 {3, MVT::v32i8, 14}, //(load 96i8 and) deinterleave into 3 x 32i8 4601 {3, MVT::v64i8, 22}, //(load 96i8 and) deinterleave into 3 x 32i8 4602 }; 4603 4604 if (const auto *Entry = 4605 CostTableLookup(AVX512InterleavedLoadTbl, Factor, VT)) 4606 return NumOfMemOps * MemOpCost + Entry->Cost; 4607 //If an entry does not exist, fallback to the default implementation. 4608 4609 // Kind of shuffle depends on number of loaded values. 4610 // If we load the entire data in one register, we can use a 1-src shuffle. 4611 // Otherwise, we'll merge 2 sources in each operation. 4612 TTI::ShuffleKind ShuffleKind = 4613 (NumOfMemOps > 1) ? TTI::SK_PermuteTwoSrc : TTI::SK_PermuteSingleSrc; 4614 4615 unsigned ShuffleCost = 4616 getShuffleCost(ShuffleKind, SingleMemOpTy, 0, nullptr); 4617 4618 unsigned NumOfLoadsInInterleaveGrp = 4619 Indices.size() ? Indices.size() : Factor; 4620 auto *ResultTy = FixedVectorType::get(VecTy->getElementType(), 4621 VecTy->getNumElements() / Factor); 4622 unsigned NumOfResults = 4623 getTLI()->getTypeLegalizationCost(DL, ResultTy).first * 4624 NumOfLoadsInInterleaveGrp; 4625 4626 // About a half of the loads may be folded in shuffles when we have only 4627 // one result. If we have more than one result, we do not fold loads at all. 4628 unsigned NumOfUnfoldedLoads = 4629 NumOfResults > 1 ? NumOfMemOps : NumOfMemOps / 2; 4630 4631 // Get a number of shuffle operations per result. 4632 unsigned NumOfShufflesPerResult = 4633 std::max((unsigned)1, (unsigned)(NumOfMemOps - 1)); 4634 4635 // The SK_MergeTwoSrc shuffle clobbers one of src operands. 4636 // When we have more than one destination, we need additional instructions 4637 // to keep sources. 4638 unsigned NumOfMoves = 0; 4639 if (NumOfResults > 1 && ShuffleKind == TTI::SK_PermuteTwoSrc) 4640 NumOfMoves = NumOfResults * NumOfShufflesPerResult / 2; 4641 4642 int Cost = NumOfResults * NumOfShufflesPerResult * ShuffleCost + 4643 NumOfUnfoldedLoads * MemOpCost + NumOfMoves; 4644 4645 return Cost; 4646 } 4647 4648 // Store. 4649 assert(Opcode == Instruction::Store && 4650 "Expected Store Instruction at this point"); 4651 // X86InterleavedAccess support only the following interleaved-access group. 4652 static const CostTblEntry AVX512InterleavedStoreTbl[] = { 4653 {3, MVT::v16i8, 12}, // interleave 3 x 16i8 into 48i8 (and store) 4654 {3, MVT::v32i8, 14}, // interleave 3 x 32i8 into 96i8 (and store) 4655 {3, MVT::v64i8, 26}, // interleave 3 x 64i8 into 96i8 (and store) 4656 4657 {4, MVT::v8i8, 10}, // interleave 4 x 8i8 into 32i8 (and store) 4658 {4, MVT::v16i8, 11}, // interleave 4 x 16i8 into 64i8 (and store) 4659 {4, MVT::v32i8, 14}, // interleave 4 x 32i8 into 128i8 (and store) 4660 {4, MVT::v64i8, 24} // interleave 4 x 32i8 into 256i8 (and store) 4661 }; 4662 4663 if (const auto *Entry = 4664 CostTableLookup(AVX512InterleavedStoreTbl, Factor, VT)) 4665 return NumOfMemOps * MemOpCost + Entry->Cost; 4666 //If an entry does not exist, fallback to the default implementation. 4667 4668 // There is no strided stores meanwhile. And store can't be folded in 4669 // shuffle. 4670 unsigned NumOfSources = Factor; // The number of values to be merged. 4671 unsigned ShuffleCost = 4672 getShuffleCost(TTI::SK_PermuteTwoSrc, SingleMemOpTy, 0, nullptr); 4673 unsigned NumOfShufflesPerStore = NumOfSources - 1; 4674 4675 // The SK_MergeTwoSrc shuffle clobbers one of src operands. 4676 // We need additional instructions to keep sources. 4677 unsigned NumOfMoves = NumOfMemOps * NumOfShufflesPerStore / 2; 4678 int Cost = NumOfMemOps * (MemOpCost + NumOfShufflesPerStore * ShuffleCost) + 4679 NumOfMoves; 4680 return Cost; 4681 } 4682 4683 int X86TTIImpl::getInterleavedMemoryOpCost( 4684 unsigned Opcode, Type *VecTy, unsigned Factor, ArrayRef<unsigned> Indices, 4685 Align Alignment, unsigned AddressSpace, TTI::TargetCostKind CostKind, 4686 bool UseMaskForCond, bool UseMaskForGaps) { 4687 auto isSupportedOnAVX512 = [](Type *VecTy, bool HasBW) { 4688 Type *EltTy = cast<VectorType>(VecTy)->getElementType(); 4689 if (EltTy->isFloatTy() || EltTy->isDoubleTy() || EltTy->isIntegerTy(64) || 4690 EltTy->isIntegerTy(32) || EltTy->isPointerTy()) 4691 return true; 4692 if (EltTy->isIntegerTy(16) || EltTy->isIntegerTy(8)) 4693 return HasBW; 4694 return false; 4695 }; 4696 if (ST->hasAVX512() && isSupportedOnAVX512(VecTy, ST->hasBWI())) 4697 return getInterleavedMemoryOpCostAVX512( 4698 Opcode, cast<FixedVectorType>(VecTy), Factor, Indices, Alignment, 4699 AddressSpace, CostKind, UseMaskForCond, UseMaskForGaps); 4700 if (ST->hasAVX2()) 4701 return getInterleavedMemoryOpCostAVX2( 4702 Opcode, cast<FixedVectorType>(VecTy), Factor, Indices, Alignment, 4703 AddressSpace, CostKind, UseMaskForCond, UseMaskForGaps); 4704 4705 return BaseT::getInterleavedMemoryOpCost(Opcode, VecTy, Factor, Indices, 4706 Alignment, AddressSpace, CostKind, 4707 UseMaskForCond, UseMaskForGaps); 4708 } 4709