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::OperandValueKind Op1Info, 174 TTI::OperandValueKind Op2Info, 175 TTI::OperandValueProperties Opd1PropInfo, 176 TTI::OperandValueProperties Opd2PropInfo, 177 ArrayRef<const Value *> Args, 178 const Instruction *CxtI) { 179 // Legalize the type. 180 std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, Ty); 181 182 int ISD = TLI->InstructionOpcodeToISD(Opcode); 183 assert(ISD && "Invalid opcode"); 184 185 static const CostTblEntry GLMCostTable[] = { 186 { ISD::FDIV, MVT::f32, 18 }, // divss 187 { ISD::FDIV, MVT::v4f32, 35 }, // divps 188 { ISD::FDIV, MVT::f64, 33 }, // divsd 189 { ISD::FDIV, MVT::v2f64, 65 }, // divpd 190 }; 191 192 if (ST->useGLMDivSqrtCosts()) 193 if (const auto *Entry = CostTableLookup(GLMCostTable, ISD, 194 LT.second)) 195 return LT.first * Entry->Cost; 196 197 static const CostTblEntry SLMCostTable[] = { 198 { ISD::MUL, MVT::v4i32, 11 }, // pmulld 199 { ISD::MUL, MVT::v8i16, 2 }, // pmullw 200 { ISD::MUL, MVT::v16i8, 14 }, // extend/pmullw/trunc sequence. 201 { ISD::FMUL, MVT::f64, 2 }, // mulsd 202 { ISD::FMUL, MVT::v2f64, 4 }, // mulpd 203 { ISD::FMUL, MVT::v4f32, 2 }, // mulps 204 { ISD::FDIV, MVT::f32, 17 }, // divss 205 { ISD::FDIV, MVT::v4f32, 39 }, // divps 206 { ISD::FDIV, MVT::f64, 32 }, // divsd 207 { ISD::FDIV, MVT::v2f64, 69 }, // divpd 208 { ISD::FADD, MVT::v2f64, 2 }, // addpd 209 { ISD::FSUB, MVT::v2f64, 2 }, // subpd 210 // v2i64/v4i64 mul is custom lowered as a series of long: 211 // multiplies(3), shifts(3) and adds(2) 212 // slm muldq version throughput is 2 and addq throughput 4 213 // thus: 3X2 (muldq throughput) + 3X1 (shift throughput) + 214 // 3X4 (addq throughput) = 17 215 { ISD::MUL, MVT::v2i64, 17 }, 216 // slm addq\subq throughput is 4 217 { ISD::ADD, MVT::v2i64, 4 }, 218 { ISD::SUB, MVT::v2i64, 4 }, 219 }; 220 221 if (ST->isSLM()) { 222 if (Args.size() == 2 && ISD == ISD::MUL && LT.second == MVT::v4i32) { 223 // Check if the operands can be shrinked into a smaller datatype. 224 bool Op1Signed = false; 225 unsigned Op1MinSize = BaseT::minRequiredElementSize(Args[0], Op1Signed); 226 bool Op2Signed = false; 227 unsigned Op2MinSize = BaseT::minRequiredElementSize(Args[1], Op2Signed); 228 229 bool signedMode = Op1Signed | Op2Signed; 230 unsigned OpMinSize = std::max(Op1MinSize, Op2MinSize); 231 232 if (OpMinSize <= 7) 233 return LT.first * 3; // pmullw/sext 234 if (!signedMode && OpMinSize <= 8) 235 return LT.first * 3; // pmullw/zext 236 if (OpMinSize <= 15) 237 return LT.first * 5; // pmullw/pmulhw/pshuf 238 if (!signedMode && OpMinSize <= 16) 239 return LT.first * 5; // pmullw/pmulhw/pshuf 240 } 241 242 if (const auto *Entry = CostTableLookup(SLMCostTable, ISD, 243 LT.second)) { 244 return LT.first * Entry->Cost; 245 } 246 } 247 248 if ((ISD == ISD::SDIV || ISD == ISD::SREM || ISD == ISD::UDIV || 249 ISD == ISD::UREM) && 250 (Op2Info == TargetTransformInfo::OK_UniformConstantValue || 251 Op2Info == TargetTransformInfo::OK_NonUniformConstantValue) && 252 Opd2PropInfo == TargetTransformInfo::OP_PowerOf2) { 253 if (ISD == ISD::SDIV || ISD == ISD::SREM) { 254 // On X86, vector signed division by constants power-of-two are 255 // normally expanded to the sequence SRA + SRL + ADD + SRA. 256 // The OperandValue properties may not be the same as that of the previous 257 // operation; conservatively assume OP_None. 258 int Cost = 259 2 * getArithmeticInstrCost(Instruction::AShr, Ty, Op1Info, Op2Info, 260 TargetTransformInfo::OP_None, 261 TargetTransformInfo::OP_None); 262 Cost += getArithmeticInstrCost(Instruction::LShr, Ty, Op1Info, Op2Info, 263 TargetTransformInfo::OP_None, 264 TargetTransformInfo::OP_None); 265 Cost += getArithmeticInstrCost(Instruction::Add, Ty, Op1Info, Op2Info, 266 TargetTransformInfo::OP_None, 267 TargetTransformInfo::OP_None); 268 269 if (ISD == ISD::SREM) { 270 // For SREM: (X % C) is the equivalent of (X - (X/C)*C) 271 Cost += getArithmeticInstrCost(Instruction::Mul, Ty, Op1Info, Op2Info); 272 Cost += getArithmeticInstrCost(Instruction::Sub, Ty, Op1Info, Op2Info); 273 } 274 275 return Cost; 276 } 277 278 // Vector unsigned division/remainder will be simplified to shifts/masks. 279 if (ISD == ISD::UDIV) 280 return getArithmeticInstrCost(Instruction::LShr, Ty, Op1Info, Op2Info, 281 TargetTransformInfo::OP_None, 282 TargetTransformInfo::OP_None); 283 284 else // UREM 285 return getArithmeticInstrCost(Instruction::And, Ty, Op1Info, Op2Info, 286 TargetTransformInfo::OP_None, 287 TargetTransformInfo::OP_None); 288 } 289 290 static const CostTblEntry AVX512BWUniformConstCostTable[] = { 291 { ISD::SHL, MVT::v64i8, 2 }, // psllw + pand. 292 { ISD::SRL, MVT::v64i8, 2 }, // psrlw + pand. 293 { ISD::SRA, MVT::v64i8, 4 }, // psrlw, pand, pxor, psubb. 294 }; 295 296 if (Op2Info == TargetTransformInfo::OK_UniformConstantValue && 297 ST->hasBWI()) { 298 if (const auto *Entry = CostTableLookup(AVX512BWUniformConstCostTable, ISD, 299 LT.second)) 300 return LT.first * Entry->Cost; 301 } 302 303 static const CostTblEntry AVX512UniformConstCostTable[] = { 304 { ISD::SRA, MVT::v2i64, 1 }, 305 { ISD::SRA, MVT::v4i64, 1 }, 306 { ISD::SRA, MVT::v8i64, 1 }, 307 }; 308 309 if (Op2Info == TargetTransformInfo::OK_UniformConstantValue && 310 ST->hasAVX512()) { 311 if (const auto *Entry = CostTableLookup(AVX512UniformConstCostTable, ISD, 312 LT.second)) 313 return LT.first * Entry->Cost; 314 } 315 316 static const CostTblEntry AVX2UniformConstCostTable[] = { 317 { ISD::SHL, MVT::v32i8, 2 }, // psllw + pand. 318 { ISD::SRL, MVT::v32i8, 2 }, // psrlw + pand. 319 { ISD::SRA, MVT::v32i8, 4 }, // psrlw, pand, pxor, psubb. 320 321 { ISD::SRA, MVT::v4i64, 4 }, // 2 x psrad + shuffle. 322 }; 323 324 if (Op2Info == TargetTransformInfo::OK_UniformConstantValue && 325 ST->hasAVX2()) { 326 if (const auto *Entry = CostTableLookup(AVX2UniformConstCostTable, ISD, 327 LT.second)) 328 return LT.first * Entry->Cost; 329 } 330 331 static const CostTblEntry SSE2UniformConstCostTable[] = { 332 { ISD::SHL, MVT::v16i8, 2 }, // psllw + pand. 333 { ISD::SRL, MVT::v16i8, 2 }, // psrlw + pand. 334 { ISD::SRA, MVT::v16i8, 4 }, // psrlw, pand, pxor, psubb. 335 336 { ISD::SHL, MVT::v32i8, 4+2 }, // 2*(psllw + pand) + split. 337 { ISD::SRL, MVT::v32i8, 4+2 }, // 2*(psrlw + pand) + split. 338 { ISD::SRA, MVT::v32i8, 8+2 }, // 2*(psrlw, pand, pxor, psubb) + split. 339 }; 340 341 // XOP has faster vXi8 shifts. 342 if (Op2Info == TargetTransformInfo::OK_UniformConstantValue && 343 ST->hasSSE2() && !ST->hasXOP()) { 344 if (const auto *Entry = 345 CostTableLookup(SSE2UniformConstCostTable, ISD, LT.second)) 346 return LT.first * Entry->Cost; 347 } 348 349 static const CostTblEntry AVX512BWConstCostTable[] = { 350 { ISD::SDIV, MVT::v64i8, 14 }, // 2*ext+2*pmulhw sequence 351 { ISD::SREM, MVT::v64i8, 16 }, // 2*ext+2*pmulhw+mul+sub sequence 352 { ISD::UDIV, MVT::v64i8, 14 }, // 2*ext+2*pmulhw sequence 353 { ISD::UREM, MVT::v64i8, 16 }, // 2*ext+2*pmulhw+mul+sub sequence 354 { ISD::SDIV, MVT::v32i16, 6 }, // vpmulhw sequence 355 { ISD::SREM, MVT::v32i16, 8 }, // vpmulhw+mul+sub sequence 356 { ISD::UDIV, MVT::v32i16, 6 }, // vpmulhuw sequence 357 { ISD::UREM, MVT::v32i16, 8 }, // vpmulhuw+mul+sub sequence 358 }; 359 360 if ((Op2Info == TargetTransformInfo::OK_UniformConstantValue || 361 Op2Info == TargetTransformInfo::OK_NonUniformConstantValue) && 362 ST->hasBWI()) { 363 if (const auto *Entry = 364 CostTableLookup(AVX512BWConstCostTable, ISD, LT.second)) 365 return LT.first * Entry->Cost; 366 } 367 368 static const CostTblEntry AVX512ConstCostTable[] = { 369 { ISD::SDIV, MVT::v16i32, 15 }, // vpmuldq sequence 370 { ISD::SREM, MVT::v16i32, 17 }, // vpmuldq+mul+sub sequence 371 { ISD::UDIV, MVT::v16i32, 15 }, // vpmuludq sequence 372 { ISD::UREM, MVT::v16i32, 17 }, // vpmuludq+mul+sub sequence 373 }; 374 375 if ((Op2Info == TargetTransformInfo::OK_UniformConstantValue || 376 Op2Info == TargetTransformInfo::OK_NonUniformConstantValue) && 377 ST->hasAVX512()) { 378 if (const auto *Entry = 379 CostTableLookup(AVX512ConstCostTable, ISD, LT.second)) 380 return LT.first * Entry->Cost; 381 } 382 383 static const CostTblEntry AVX2ConstCostTable[] = { 384 { ISD::SDIV, MVT::v32i8, 14 }, // 2*ext+2*pmulhw sequence 385 { ISD::SREM, MVT::v32i8, 16 }, // 2*ext+2*pmulhw+mul+sub sequence 386 { ISD::UDIV, MVT::v32i8, 14 }, // 2*ext+2*pmulhw sequence 387 { ISD::UREM, MVT::v32i8, 16 }, // 2*ext+2*pmulhw+mul+sub sequence 388 { ISD::SDIV, MVT::v16i16, 6 }, // vpmulhw sequence 389 { ISD::SREM, MVT::v16i16, 8 }, // vpmulhw+mul+sub sequence 390 { ISD::UDIV, MVT::v16i16, 6 }, // vpmulhuw sequence 391 { ISD::UREM, MVT::v16i16, 8 }, // vpmulhuw+mul+sub sequence 392 { ISD::SDIV, MVT::v8i32, 15 }, // vpmuldq sequence 393 { ISD::SREM, MVT::v8i32, 19 }, // vpmuldq+mul+sub sequence 394 { ISD::UDIV, MVT::v8i32, 15 }, // vpmuludq sequence 395 { ISD::UREM, MVT::v8i32, 19 }, // vpmuludq+mul+sub sequence 396 }; 397 398 if ((Op2Info == TargetTransformInfo::OK_UniformConstantValue || 399 Op2Info == TargetTransformInfo::OK_NonUniformConstantValue) && 400 ST->hasAVX2()) { 401 if (const auto *Entry = CostTableLookup(AVX2ConstCostTable, ISD, LT.second)) 402 return LT.first * Entry->Cost; 403 } 404 405 static const CostTblEntry SSE2ConstCostTable[] = { 406 { ISD::SDIV, MVT::v32i8, 28+2 }, // 4*ext+4*pmulhw sequence + split. 407 { ISD::SREM, MVT::v32i8, 32+2 }, // 4*ext+4*pmulhw+mul+sub sequence + split. 408 { ISD::SDIV, MVT::v16i8, 14 }, // 2*ext+2*pmulhw sequence 409 { ISD::SREM, MVT::v16i8, 16 }, // 2*ext+2*pmulhw+mul+sub sequence 410 { ISD::UDIV, MVT::v32i8, 28+2 }, // 4*ext+4*pmulhw sequence + split. 411 { ISD::UREM, MVT::v32i8, 32+2 }, // 4*ext+4*pmulhw+mul+sub sequence + split. 412 { ISD::UDIV, MVT::v16i8, 14 }, // 2*ext+2*pmulhw sequence 413 { ISD::UREM, MVT::v16i8, 16 }, // 2*ext+2*pmulhw+mul+sub sequence 414 { ISD::SDIV, MVT::v16i16, 12+2 }, // 2*pmulhw sequence + split. 415 { ISD::SREM, MVT::v16i16, 16+2 }, // 2*pmulhw+mul+sub sequence + split. 416 { ISD::SDIV, MVT::v8i16, 6 }, // pmulhw sequence 417 { ISD::SREM, MVT::v8i16, 8 }, // pmulhw+mul+sub sequence 418 { ISD::UDIV, MVT::v16i16, 12+2 }, // 2*pmulhuw sequence + split. 419 { ISD::UREM, MVT::v16i16, 16+2 }, // 2*pmulhuw+mul+sub sequence + split. 420 { ISD::UDIV, MVT::v8i16, 6 }, // pmulhuw sequence 421 { ISD::UREM, MVT::v8i16, 8 }, // pmulhuw+mul+sub sequence 422 { ISD::SDIV, MVT::v8i32, 38+2 }, // 2*pmuludq sequence + split. 423 { ISD::SREM, MVT::v8i32, 48+2 }, // 2*pmuludq+mul+sub sequence + split. 424 { ISD::SDIV, MVT::v4i32, 19 }, // pmuludq sequence 425 { ISD::SREM, MVT::v4i32, 24 }, // pmuludq+mul+sub sequence 426 { ISD::UDIV, MVT::v8i32, 30+2 }, // 2*pmuludq sequence + split. 427 { ISD::UREM, MVT::v8i32, 40+2 }, // 2*pmuludq+mul+sub sequence + split. 428 { ISD::UDIV, MVT::v4i32, 15 }, // pmuludq sequence 429 { ISD::UREM, MVT::v4i32, 20 }, // pmuludq+mul+sub sequence 430 }; 431 432 if ((Op2Info == TargetTransformInfo::OK_UniformConstantValue || 433 Op2Info == TargetTransformInfo::OK_NonUniformConstantValue) && 434 ST->hasSSE2()) { 435 // pmuldq sequence. 436 if (ISD == ISD::SDIV && LT.second == MVT::v8i32 && ST->hasAVX()) 437 return LT.first * 32; 438 if (ISD == ISD::SREM && LT.second == MVT::v8i32 && ST->hasAVX()) 439 return LT.first * 38; 440 if (ISD == ISD::SDIV && LT.second == MVT::v4i32 && ST->hasSSE41()) 441 return LT.first * 15; 442 if (ISD == ISD::SREM && LT.second == MVT::v4i32 && ST->hasSSE41()) 443 return LT.first * 20; 444 445 if (const auto *Entry = CostTableLookup(SSE2ConstCostTable, ISD, LT.second)) 446 return LT.first * Entry->Cost; 447 } 448 449 static const CostTblEntry AVX2UniformCostTable[] = { 450 // Uniform splats are cheaper for the following instructions. 451 { ISD::SHL, MVT::v16i16, 1 }, // psllw. 452 { ISD::SRL, MVT::v16i16, 1 }, // psrlw. 453 { ISD::SRA, MVT::v16i16, 1 }, // psraw. 454 }; 455 456 if (ST->hasAVX2() && 457 ((Op2Info == TargetTransformInfo::OK_UniformConstantValue) || 458 (Op2Info == TargetTransformInfo::OK_UniformValue))) { 459 if (const auto *Entry = 460 CostTableLookup(AVX2UniformCostTable, ISD, LT.second)) 461 return LT.first * Entry->Cost; 462 } 463 464 static const CostTblEntry SSE2UniformCostTable[] = { 465 // Uniform splats are cheaper for the following instructions. 466 { ISD::SHL, MVT::v8i16, 1 }, // psllw. 467 { ISD::SHL, MVT::v4i32, 1 }, // pslld 468 { ISD::SHL, MVT::v2i64, 1 }, // psllq. 469 470 { ISD::SRL, MVT::v8i16, 1 }, // psrlw. 471 { ISD::SRL, MVT::v4i32, 1 }, // psrld. 472 { ISD::SRL, MVT::v2i64, 1 }, // psrlq. 473 474 { ISD::SRA, MVT::v8i16, 1 }, // psraw. 475 { ISD::SRA, MVT::v4i32, 1 }, // psrad. 476 }; 477 478 if (ST->hasSSE2() && 479 ((Op2Info == TargetTransformInfo::OK_UniformConstantValue) || 480 (Op2Info == TargetTransformInfo::OK_UniformValue))) { 481 if (const auto *Entry = 482 CostTableLookup(SSE2UniformCostTable, ISD, LT.second)) 483 return LT.first * Entry->Cost; 484 } 485 486 static const CostTblEntry AVX512DQCostTable[] = { 487 { ISD::MUL, MVT::v2i64, 1 }, 488 { ISD::MUL, MVT::v4i64, 1 }, 489 { ISD::MUL, MVT::v8i64, 1 } 490 }; 491 492 // Look for AVX512DQ lowering tricks for custom cases. 493 if (ST->hasDQI()) 494 if (const auto *Entry = CostTableLookup(AVX512DQCostTable, ISD, LT.second)) 495 return LT.first * Entry->Cost; 496 497 static const CostTblEntry AVX512BWCostTable[] = { 498 { ISD::SHL, MVT::v8i16, 1 }, // vpsllvw 499 { ISD::SRL, MVT::v8i16, 1 }, // vpsrlvw 500 { ISD::SRA, MVT::v8i16, 1 }, // vpsravw 501 502 { ISD::SHL, MVT::v16i16, 1 }, // vpsllvw 503 { ISD::SRL, MVT::v16i16, 1 }, // vpsrlvw 504 { ISD::SRA, MVT::v16i16, 1 }, // vpsravw 505 506 { ISD::SHL, MVT::v32i16, 1 }, // vpsllvw 507 { ISD::SRL, MVT::v32i16, 1 }, // vpsrlvw 508 { ISD::SRA, MVT::v32i16, 1 }, // vpsravw 509 510 { ISD::SHL, MVT::v64i8, 11 }, // vpblendvb sequence. 511 { ISD::SRL, MVT::v64i8, 11 }, // vpblendvb sequence. 512 { ISD::SRA, MVT::v64i8, 24 }, // vpblendvb sequence. 513 514 { ISD::MUL, MVT::v64i8, 11 }, // extend/pmullw/trunc sequence. 515 { ISD::MUL, MVT::v32i8, 4 }, // extend/pmullw/trunc sequence. 516 { ISD::MUL, MVT::v16i8, 4 }, // extend/pmullw/trunc sequence. 517 }; 518 519 // Look for AVX512BW lowering tricks for custom cases. 520 if (ST->hasBWI()) 521 if (const auto *Entry = CostTableLookup(AVX512BWCostTable, ISD, LT.second)) 522 return LT.first * Entry->Cost; 523 524 static const CostTblEntry AVX512CostTable[] = { 525 { ISD::SHL, MVT::v16i32, 1 }, 526 { ISD::SRL, MVT::v16i32, 1 }, 527 { ISD::SRA, MVT::v16i32, 1 }, 528 529 { ISD::SHL, MVT::v8i64, 1 }, 530 { ISD::SRL, MVT::v8i64, 1 }, 531 532 { ISD::SRA, MVT::v2i64, 1 }, 533 { ISD::SRA, MVT::v4i64, 1 }, 534 { ISD::SRA, MVT::v8i64, 1 }, 535 536 { ISD::MUL, MVT::v32i8, 13 }, // extend/pmullw/trunc sequence. 537 { ISD::MUL, MVT::v16i8, 5 }, // extend/pmullw/trunc sequence. 538 { ISD::MUL, MVT::v16i32, 1 }, // pmulld (Skylake from agner.org) 539 { ISD::MUL, MVT::v8i32, 1 }, // pmulld (Skylake from agner.org) 540 { ISD::MUL, MVT::v4i32, 1 }, // pmulld (Skylake from agner.org) 541 { ISD::MUL, MVT::v8i64, 8 }, // 3*pmuludq/3*shift/2*add 542 543 { ISD::FADD, MVT::v8f64, 1 }, // Skylake from http://www.agner.org/ 544 { ISD::FSUB, MVT::v8f64, 1 }, // Skylake from http://www.agner.org/ 545 { ISD::FMUL, MVT::v8f64, 1 }, // Skylake from http://www.agner.org/ 546 547 { ISD::FADD, MVT::v16f32, 1 }, // Skylake from http://www.agner.org/ 548 { ISD::FSUB, MVT::v16f32, 1 }, // Skylake from http://www.agner.org/ 549 { ISD::FMUL, MVT::v16f32, 1 }, // Skylake from http://www.agner.org/ 550 }; 551 552 if (ST->hasAVX512()) 553 if (const auto *Entry = CostTableLookup(AVX512CostTable, ISD, LT.second)) 554 return LT.first * Entry->Cost; 555 556 static const CostTblEntry AVX2ShiftCostTable[] = { 557 // Shifts on v4i64/v8i32 on AVX2 is legal even though we declare to 558 // customize them to detect the cases where shift amount is a scalar one. 559 { ISD::SHL, MVT::v4i32, 1 }, 560 { ISD::SRL, MVT::v4i32, 1 }, 561 { ISD::SRA, MVT::v4i32, 1 }, 562 { ISD::SHL, MVT::v8i32, 1 }, 563 { ISD::SRL, MVT::v8i32, 1 }, 564 { ISD::SRA, MVT::v8i32, 1 }, 565 { ISD::SHL, MVT::v2i64, 1 }, 566 { ISD::SRL, MVT::v2i64, 1 }, 567 { ISD::SHL, MVT::v4i64, 1 }, 568 { ISD::SRL, MVT::v4i64, 1 }, 569 }; 570 571 // Look for AVX2 lowering tricks. 572 if (ST->hasAVX2()) { 573 if (ISD == ISD::SHL && LT.second == MVT::v16i16 && 574 (Op2Info == TargetTransformInfo::OK_UniformConstantValue || 575 Op2Info == TargetTransformInfo::OK_NonUniformConstantValue)) 576 // On AVX2, a packed v16i16 shift left by a constant build_vector 577 // is lowered into a vector multiply (vpmullw). 578 return getArithmeticInstrCost(Instruction::Mul, Ty, Op1Info, Op2Info, 579 TargetTransformInfo::OP_None, 580 TargetTransformInfo::OP_None); 581 582 if (const auto *Entry = CostTableLookup(AVX2ShiftCostTable, ISD, LT.second)) 583 return LT.first * Entry->Cost; 584 } 585 586 static const CostTblEntry XOPShiftCostTable[] = { 587 // 128bit shifts take 1cy, but right shifts require negation beforehand. 588 { ISD::SHL, MVT::v16i8, 1 }, 589 { ISD::SRL, MVT::v16i8, 2 }, 590 { ISD::SRA, MVT::v16i8, 2 }, 591 { ISD::SHL, MVT::v8i16, 1 }, 592 { ISD::SRL, MVT::v8i16, 2 }, 593 { ISD::SRA, MVT::v8i16, 2 }, 594 { ISD::SHL, MVT::v4i32, 1 }, 595 { ISD::SRL, MVT::v4i32, 2 }, 596 { ISD::SRA, MVT::v4i32, 2 }, 597 { ISD::SHL, MVT::v2i64, 1 }, 598 { ISD::SRL, MVT::v2i64, 2 }, 599 { ISD::SRA, MVT::v2i64, 2 }, 600 // 256bit shifts require splitting if AVX2 didn't catch them above. 601 { ISD::SHL, MVT::v32i8, 2+2 }, 602 { ISD::SRL, MVT::v32i8, 4+2 }, 603 { ISD::SRA, MVT::v32i8, 4+2 }, 604 { ISD::SHL, MVT::v16i16, 2+2 }, 605 { ISD::SRL, MVT::v16i16, 4+2 }, 606 { ISD::SRA, MVT::v16i16, 4+2 }, 607 { ISD::SHL, MVT::v8i32, 2+2 }, 608 { ISD::SRL, MVT::v8i32, 4+2 }, 609 { ISD::SRA, MVT::v8i32, 4+2 }, 610 { ISD::SHL, MVT::v4i64, 2+2 }, 611 { ISD::SRL, MVT::v4i64, 4+2 }, 612 { ISD::SRA, MVT::v4i64, 4+2 }, 613 }; 614 615 // Look for XOP lowering tricks. 616 if (ST->hasXOP()) { 617 // If the right shift is constant then we'll fold the negation so 618 // it's as cheap as a left shift. 619 int ShiftISD = ISD; 620 if ((ShiftISD == ISD::SRL || ShiftISD == ISD::SRA) && 621 (Op2Info == TargetTransformInfo::OK_UniformConstantValue || 622 Op2Info == TargetTransformInfo::OK_NonUniformConstantValue)) 623 ShiftISD = ISD::SHL; 624 if (const auto *Entry = 625 CostTableLookup(XOPShiftCostTable, ShiftISD, LT.second)) 626 return LT.first * Entry->Cost; 627 } 628 629 static const CostTblEntry SSE2UniformShiftCostTable[] = { 630 // Uniform splats are cheaper for the following instructions. 631 { ISD::SHL, MVT::v16i16, 2+2 }, // 2*psllw + split. 632 { ISD::SHL, MVT::v8i32, 2+2 }, // 2*pslld + split. 633 { ISD::SHL, MVT::v4i64, 2+2 }, // 2*psllq + split. 634 635 { ISD::SRL, MVT::v16i16, 2+2 }, // 2*psrlw + split. 636 { ISD::SRL, MVT::v8i32, 2+2 }, // 2*psrld + split. 637 { ISD::SRL, MVT::v4i64, 2+2 }, // 2*psrlq + split. 638 639 { ISD::SRA, MVT::v16i16, 2+2 }, // 2*psraw + split. 640 { ISD::SRA, MVT::v8i32, 2+2 }, // 2*psrad + split. 641 { ISD::SRA, MVT::v2i64, 4 }, // 2*psrad + shuffle. 642 { ISD::SRA, MVT::v4i64, 8+2 }, // 2*(2*psrad + shuffle) + split. 643 }; 644 645 if (ST->hasSSE2() && 646 ((Op2Info == TargetTransformInfo::OK_UniformConstantValue) || 647 (Op2Info == TargetTransformInfo::OK_UniformValue))) { 648 649 // Handle AVX2 uniform v4i64 ISD::SRA, it's not worth a table. 650 if (ISD == ISD::SRA && LT.second == MVT::v4i64 && ST->hasAVX2()) 651 return LT.first * 4; // 2*psrad + shuffle. 652 653 if (const auto *Entry = 654 CostTableLookup(SSE2UniformShiftCostTable, ISD, LT.second)) 655 return LT.first * Entry->Cost; 656 } 657 658 if (ISD == ISD::SHL && 659 Op2Info == TargetTransformInfo::OK_NonUniformConstantValue) { 660 MVT VT = LT.second; 661 // Vector shift left by non uniform constant can be lowered 662 // into vector multiply. 663 if (((VT == MVT::v8i16 || VT == MVT::v4i32) && ST->hasSSE2()) || 664 ((VT == MVT::v16i16 || VT == MVT::v8i32) && ST->hasAVX())) 665 ISD = ISD::MUL; 666 } 667 668 static const CostTblEntry AVX2CostTable[] = { 669 { ISD::SHL, MVT::v32i8, 11 }, // vpblendvb sequence. 670 { ISD::SHL, MVT::v16i16, 10 }, // extend/vpsrlvd/pack sequence. 671 672 { ISD::SRL, MVT::v32i8, 11 }, // vpblendvb sequence. 673 { ISD::SRL, MVT::v16i16, 10 }, // extend/vpsrlvd/pack sequence. 674 675 { ISD::SRA, MVT::v32i8, 24 }, // vpblendvb sequence. 676 { ISD::SRA, MVT::v16i16, 10 }, // extend/vpsravd/pack sequence. 677 { ISD::SRA, MVT::v2i64, 4 }, // srl/xor/sub sequence. 678 { ISD::SRA, MVT::v4i64, 4 }, // srl/xor/sub sequence. 679 680 { ISD::SUB, MVT::v32i8, 1 }, // psubb 681 { ISD::ADD, MVT::v32i8, 1 }, // paddb 682 { ISD::SUB, MVT::v16i16, 1 }, // psubw 683 { ISD::ADD, MVT::v16i16, 1 }, // paddw 684 { ISD::SUB, MVT::v8i32, 1 }, // psubd 685 { ISD::ADD, MVT::v8i32, 1 }, // paddd 686 { ISD::SUB, MVT::v4i64, 1 }, // psubq 687 { ISD::ADD, MVT::v4i64, 1 }, // paddq 688 689 { ISD::MUL, MVT::v32i8, 17 }, // extend/pmullw/trunc sequence. 690 { ISD::MUL, MVT::v16i8, 7 }, // extend/pmullw/trunc sequence. 691 { ISD::MUL, MVT::v16i16, 1 }, // pmullw 692 { ISD::MUL, MVT::v8i32, 2 }, // pmulld (Haswell from agner.org) 693 { ISD::MUL, MVT::v4i64, 8 }, // 3*pmuludq/3*shift/2*add 694 695 { ISD::FADD, MVT::v4f64, 1 }, // Haswell from http://www.agner.org/ 696 { ISD::FADD, MVT::v8f32, 1 }, // Haswell from http://www.agner.org/ 697 { ISD::FSUB, MVT::v4f64, 1 }, // Haswell from http://www.agner.org/ 698 { ISD::FSUB, MVT::v8f32, 1 }, // Haswell from http://www.agner.org/ 699 { ISD::FMUL, MVT::v4f64, 1 }, // Haswell from http://www.agner.org/ 700 { ISD::FMUL, MVT::v8f32, 1 }, // Haswell from http://www.agner.org/ 701 702 { ISD::FDIV, MVT::f32, 7 }, // Haswell from http://www.agner.org/ 703 { ISD::FDIV, MVT::v4f32, 7 }, // Haswell from http://www.agner.org/ 704 { ISD::FDIV, MVT::v8f32, 14 }, // Haswell from http://www.agner.org/ 705 { ISD::FDIV, MVT::f64, 14 }, // Haswell from http://www.agner.org/ 706 { ISD::FDIV, MVT::v2f64, 14 }, // Haswell from http://www.agner.org/ 707 { ISD::FDIV, MVT::v4f64, 28 }, // Haswell from http://www.agner.org/ 708 }; 709 710 // Look for AVX2 lowering tricks for custom cases. 711 if (ST->hasAVX2()) 712 if (const auto *Entry = CostTableLookup(AVX2CostTable, ISD, LT.second)) 713 return LT.first * Entry->Cost; 714 715 static const CostTblEntry AVX1CostTable[] = { 716 // We don't have to scalarize unsupported ops. We can issue two half-sized 717 // operations and we only need to extract the upper YMM half. 718 // Two ops + 1 extract + 1 insert = 4. 719 { ISD::MUL, MVT::v16i16, 4 }, 720 { ISD::MUL, MVT::v8i32, 4 }, 721 { ISD::SUB, MVT::v32i8, 4 }, 722 { ISD::ADD, MVT::v32i8, 4 }, 723 { ISD::SUB, MVT::v16i16, 4 }, 724 { ISD::ADD, MVT::v16i16, 4 }, 725 { ISD::SUB, MVT::v8i32, 4 }, 726 { ISD::ADD, MVT::v8i32, 4 }, 727 { ISD::SUB, MVT::v4i64, 4 }, 728 { ISD::ADD, MVT::v4i64, 4 }, 729 730 // A v4i64 multiply is custom lowered as two split v2i64 vectors that then 731 // are lowered as a series of long multiplies(3), shifts(3) and adds(2) 732 // Because we believe v4i64 to be a legal type, we must also include the 733 // extract+insert in the cost table. Therefore, the cost here is 18 734 // instead of 8. 735 { ISD::MUL, MVT::v4i64, 18 }, 736 737 { ISD::MUL, MVT::v32i8, 26 }, // extend/pmullw/trunc sequence. 738 739 { ISD::FDIV, MVT::f32, 14 }, // SNB from http://www.agner.org/ 740 { ISD::FDIV, MVT::v4f32, 14 }, // SNB from http://www.agner.org/ 741 { ISD::FDIV, MVT::v8f32, 28 }, // SNB from http://www.agner.org/ 742 { ISD::FDIV, MVT::f64, 22 }, // SNB from http://www.agner.org/ 743 { ISD::FDIV, MVT::v2f64, 22 }, // SNB from http://www.agner.org/ 744 { ISD::FDIV, MVT::v4f64, 44 }, // SNB from http://www.agner.org/ 745 }; 746 747 if (ST->hasAVX()) 748 if (const auto *Entry = CostTableLookup(AVX1CostTable, ISD, LT.second)) 749 return LT.first * Entry->Cost; 750 751 static const CostTblEntry SSE42CostTable[] = { 752 { ISD::FADD, MVT::f64, 1 }, // Nehalem from http://www.agner.org/ 753 { ISD::FADD, MVT::f32, 1 }, // Nehalem from http://www.agner.org/ 754 { ISD::FADD, MVT::v2f64, 1 }, // Nehalem from http://www.agner.org/ 755 { ISD::FADD, MVT::v4f32, 1 }, // Nehalem from http://www.agner.org/ 756 757 { ISD::FSUB, MVT::f64, 1 }, // Nehalem from http://www.agner.org/ 758 { ISD::FSUB, MVT::f32 , 1 }, // Nehalem from http://www.agner.org/ 759 { ISD::FSUB, MVT::v2f64, 1 }, // Nehalem from http://www.agner.org/ 760 { ISD::FSUB, MVT::v4f32, 1 }, // Nehalem from http://www.agner.org/ 761 762 { ISD::FMUL, MVT::f64, 1 }, // Nehalem from http://www.agner.org/ 763 { ISD::FMUL, MVT::f32, 1 }, // Nehalem from http://www.agner.org/ 764 { ISD::FMUL, MVT::v2f64, 1 }, // Nehalem from http://www.agner.org/ 765 { ISD::FMUL, MVT::v4f32, 1 }, // Nehalem from http://www.agner.org/ 766 767 { ISD::FDIV, MVT::f32, 14 }, // Nehalem from http://www.agner.org/ 768 { ISD::FDIV, MVT::v4f32, 14 }, // Nehalem from http://www.agner.org/ 769 { ISD::FDIV, MVT::f64, 22 }, // Nehalem from http://www.agner.org/ 770 { ISD::FDIV, MVT::v2f64, 22 }, // Nehalem from http://www.agner.org/ 771 }; 772 773 if (ST->hasSSE42()) 774 if (const auto *Entry = CostTableLookup(SSE42CostTable, ISD, LT.second)) 775 return LT.first * Entry->Cost; 776 777 static const CostTblEntry SSE41CostTable[] = { 778 { ISD::SHL, MVT::v16i8, 11 }, // pblendvb sequence. 779 { ISD::SHL, MVT::v32i8, 2*11+2 }, // pblendvb sequence + split. 780 { ISD::SHL, MVT::v8i16, 14 }, // pblendvb sequence. 781 { ISD::SHL, MVT::v16i16, 2*14+2 }, // pblendvb sequence + split. 782 { ISD::SHL, MVT::v4i32, 4 }, // pslld/paddd/cvttps2dq/pmulld 783 { ISD::SHL, MVT::v8i32, 2*4+2 }, // pslld/paddd/cvttps2dq/pmulld + split 784 785 { ISD::SRL, MVT::v16i8, 12 }, // pblendvb sequence. 786 { ISD::SRL, MVT::v32i8, 2*12+2 }, // pblendvb sequence + split. 787 { ISD::SRL, MVT::v8i16, 14 }, // pblendvb sequence. 788 { ISD::SRL, MVT::v16i16, 2*14+2 }, // pblendvb sequence + split. 789 { ISD::SRL, MVT::v4i32, 11 }, // Shift each lane + blend. 790 { ISD::SRL, MVT::v8i32, 2*11+2 }, // Shift each lane + blend + split. 791 792 { ISD::SRA, MVT::v16i8, 24 }, // pblendvb sequence. 793 { ISD::SRA, MVT::v32i8, 2*24+2 }, // pblendvb sequence + split. 794 { ISD::SRA, MVT::v8i16, 14 }, // pblendvb sequence. 795 { ISD::SRA, MVT::v16i16, 2*14+2 }, // pblendvb sequence + split. 796 { ISD::SRA, MVT::v4i32, 12 }, // Shift each lane + blend. 797 { ISD::SRA, MVT::v8i32, 2*12+2 }, // Shift each lane + blend + split. 798 799 { ISD::MUL, MVT::v4i32, 2 } // pmulld (Nehalem from agner.org) 800 }; 801 802 if (ST->hasSSE41()) 803 if (const auto *Entry = CostTableLookup(SSE41CostTable, ISD, LT.second)) 804 return LT.first * Entry->Cost; 805 806 static const CostTblEntry SSE2CostTable[] = { 807 // We don't correctly identify costs of casts because they are marked as 808 // custom. 809 { ISD::SHL, MVT::v16i8, 26 }, // cmpgtb sequence. 810 { ISD::SHL, MVT::v8i16, 32 }, // cmpgtb sequence. 811 { ISD::SHL, MVT::v4i32, 2*5 }, // We optimized this using mul. 812 { ISD::SHL, MVT::v2i64, 4 }, // splat+shuffle sequence. 813 { ISD::SHL, MVT::v4i64, 2*4+2 }, // splat+shuffle sequence + split. 814 815 { ISD::SRL, MVT::v16i8, 26 }, // cmpgtb sequence. 816 { ISD::SRL, MVT::v8i16, 32 }, // cmpgtb sequence. 817 { ISD::SRL, MVT::v4i32, 16 }, // Shift each lane + blend. 818 { ISD::SRL, MVT::v2i64, 4 }, // splat+shuffle sequence. 819 { ISD::SRL, MVT::v4i64, 2*4+2 }, // splat+shuffle sequence + split. 820 821 { ISD::SRA, MVT::v16i8, 54 }, // unpacked cmpgtb sequence. 822 { ISD::SRA, MVT::v8i16, 32 }, // cmpgtb sequence. 823 { ISD::SRA, MVT::v4i32, 16 }, // Shift each lane + blend. 824 { ISD::SRA, MVT::v2i64, 12 }, // srl/xor/sub sequence. 825 { ISD::SRA, MVT::v4i64, 2*12+2 }, // srl/xor/sub sequence+split. 826 827 { ISD::MUL, MVT::v16i8, 12 }, // extend/pmullw/trunc sequence. 828 { ISD::MUL, MVT::v8i16, 1 }, // pmullw 829 { ISD::MUL, MVT::v4i32, 6 }, // 3*pmuludq/4*shuffle 830 { ISD::MUL, MVT::v2i64, 8 }, // 3*pmuludq/3*shift/2*add 831 832 { ISD::FDIV, MVT::f32, 23 }, // Pentium IV from http://www.agner.org/ 833 { ISD::FDIV, MVT::v4f32, 39 }, // Pentium IV from http://www.agner.org/ 834 { ISD::FDIV, MVT::f64, 38 }, // Pentium IV from http://www.agner.org/ 835 { ISD::FDIV, MVT::v2f64, 69 }, // Pentium IV from http://www.agner.org/ 836 837 { ISD::FADD, MVT::f32, 2 }, // Pentium IV from http://www.agner.org/ 838 { ISD::FADD, MVT::f64, 2 }, // Pentium IV from http://www.agner.org/ 839 840 { ISD::FSUB, MVT::f32, 2 }, // Pentium IV from http://www.agner.org/ 841 { ISD::FSUB, MVT::f64, 2 }, // Pentium IV from http://www.agner.org/ 842 }; 843 844 if (ST->hasSSE2()) 845 if (const auto *Entry = CostTableLookup(SSE2CostTable, ISD, LT.second)) 846 return LT.first * Entry->Cost; 847 848 static const CostTblEntry SSE1CostTable[] = { 849 { ISD::FDIV, MVT::f32, 17 }, // Pentium III from http://www.agner.org/ 850 { ISD::FDIV, MVT::v4f32, 34 }, // Pentium III from http://www.agner.org/ 851 852 { ISD::FADD, MVT::f32, 1 }, // Pentium III from http://www.agner.org/ 853 { ISD::FADD, MVT::v4f32, 2 }, // Pentium III from http://www.agner.org/ 854 855 { ISD::FSUB, MVT::f32, 1 }, // Pentium III from http://www.agner.org/ 856 { ISD::FSUB, MVT::v4f32, 2 }, // Pentium III from http://www.agner.org/ 857 858 { ISD::ADD, MVT::i8, 1 }, // Pentium III from http://www.agner.org/ 859 { ISD::ADD, MVT::i16, 1 }, // Pentium III from http://www.agner.org/ 860 { ISD::ADD, MVT::i32, 1 }, // Pentium III from http://www.agner.org/ 861 862 { ISD::SUB, MVT::i8, 1 }, // Pentium III from http://www.agner.org/ 863 { ISD::SUB, MVT::i16, 1 }, // Pentium III from http://www.agner.org/ 864 { ISD::SUB, MVT::i32, 1 }, // Pentium III from http://www.agner.org/ 865 }; 866 867 if (ST->hasSSE1()) 868 if (const auto *Entry = CostTableLookup(SSE1CostTable, ISD, LT.second)) 869 return LT.first * Entry->Cost; 870 871 // It is not a good idea to vectorize division. We have to scalarize it and 872 // in the process we will often end up having to spilling regular 873 // registers. The overhead of division is going to dominate most kernels 874 // anyways so try hard to prevent vectorization of division - it is 875 // generally a bad idea. Assume somewhat arbitrarily that we have to be able 876 // to hide "20 cycles" for each lane. 877 if (LT.second.isVector() && (ISD == ISD::SDIV || ISD == ISD::SREM || 878 ISD == ISD::UDIV || ISD == ISD::UREM)) { 879 int ScalarCost = getArithmeticInstrCost( 880 Opcode, Ty->getScalarType(), Op1Info, Op2Info, 881 TargetTransformInfo::OP_None, TargetTransformInfo::OP_None); 882 return 20 * LT.first * LT.second.getVectorNumElements() * ScalarCost; 883 } 884 885 // Fallback to the default implementation. 886 return BaseT::getArithmeticInstrCost(Opcode, Ty, Op1Info, Op2Info); 887 } 888 889 int X86TTIImpl::getShuffleCost(TTI::ShuffleKind Kind, Type *Tp, int Index, 890 Type *SubTp) { 891 // 64-bit packed float vectors (v2f32) are widened to type v4f32. 892 // 64-bit packed integer vectors (v2i32) are widened to type v4i32. 893 std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, Tp); 894 895 // Treat Transpose as 2-op shuffles - there's no difference in lowering. 896 if (Kind == TTI::SK_Transpose) 897 Kind = TTI::SK_PermuteTwoSrc; 898 899 // For Broadcasts we are splatting the first element from the first input 900 // register, so only need to reference that input and all the output 901 // registers are the same. 902 if (Kind == TTI::SK_Broadcast) 903 LT.first = 1; 904 905 // Subvector extractions are free if they start at the beginning of a 906 // vector and cheap if the subvectors are aligned. 907 if (Kind == TTI::SK_ExtractSubvector && LT.second.isVector()) { 908 int NumElts = LT.second.getVectorNumElements(); 909 if ((Index % NumElts) == 0) 910 return 0; 911 std::pair<int, MVT> SubLT = TLI->getTypeLegalizationCost(DL, SubTp); 912 if (SubLT.second.isVector()) { 913 int NumSubElts = SubLT.second.getVectorNumElements(); 914 if ((Index % NumSubElts) == 0 && (NumElts % NumSubElts) == 0) 915 return SubLT.first; 916 // Handle some cases for widening legalization. For now we only handle 917 // cases where the original subvector was naturally aligned and evenly 918 // fit in its legalized subvector type. 919 // FIXME: Remove some of the alignment restrictions. 920 // FIXME: We can use permq for 64-bit or larger extracts from 256-bit 921 // vectors. 922 int OrigSubElts = SubTp->getVectorNumElements(); 923 if (NumSubElts > OrigSubElts && 924 (Index % OrigSubElts) == 0 && (NumSubElts % OrigSubElts) == 0 && 925 LT.second.getVectorElementType() == 926 SubLT.second.getVectorElementType() && 927 LT.second.getVectorElementType().getSizeInBits() == 928 Tp->getVectorElementType()->getPrimitiveSizeInBits()) { 929 assert(NumElts >= NumSubElts && NumElts > OrigSubElts && 930 "Unexpected number of elements!"); 931 Type *VecTy = VectorType::get(Tp->getVectorElementType(), 932 LT.second.getVectorNumElements()); 933 Type *SubTy = VectorType::get(Tp->getVectorElementType(), 934 SubLT.second.getVectorNumElements()); 935 int ExtractIndex = alignDown((Index % NumElts), NumSubElts); 936 int ExtractCost = getShuffleCost(TTI::SK_ExtractSubvector, VecTy, 937 ExtractIndex, SubTy); 938 939 // If the original size is 32-bits or more, we can use pshufd. Otherwise 940 // if we have SSSE3 we can use pshufb. 941 if (SubTp->getPrimitiveSizeInBits() >= 32 || ST->hasSSSE3()) 942 return ExtractCost + 1; // pshufd or pshufb 943 944 assert(SubTp->getPrimitiveSizeInBits() == 16 && 945 "Unexpected vector size"); 946 947 return ExtractCost + 2; // worst case pshufhw + pshufd 948 } 949 } 950 } 951 952 // We are going to permute multiple sources and the result will be in multiple 953 // destinations. Providing an accurate cost only for splits where the element 954 // type remains the same. 955 if (Kind == TTI::SK_PermuteSingleSrc && LT.first != 1) { 956 MVT LegalVT = LT.second; 957 if (LegalVT.isVector() && 958 LegalVT.getVectorElementType().getSizeInBits() == 959 Tp->getVectorElementType()->getPrimitiveSizeInBits() && 960 LegalVT.getVectorNumElements() < Tp->getVectorNumElements()) { 961 962 unsigned VecTySize = DL.getTypeStoreSize(Tp); 963 unsigned LegalVTSize = LegalVT.getStoreSize(); 964 // Number of source vectors after legalization: 965 unsigned NumOfSrcs = (VecTySize + LegalVTSize - 1) / LegalVTSize; 966 // Number of destination vectors after legalization: 967 unsigned NumOfDests = LT.first; 968 969 Type *SingleOpTy = VectorType::get(Tp->getVectorElementType(), 970 LegalVT.getVectorNumElements()); 971 972 unsigned NumOfShuffles = (NumOfSrcs - 1) * NumOfDests; 973 return NumOfShuffles * 974 getShuffleCost(TTI::SK_PermuteTwoSrc, SingleOpTy, 0, nullptr); 975 } 976 977 return BaseT::getShuffleCost(Kind, Tp, Index, SubTp); 978 } 979 980 // For 2-input shuffles, we must account for splitting the 2 inputs into many. 981 if (Kind == TTI::SK_PermuteTwoSrc && LT.first != 1) { 982 // We assume that source and destination have the same vector type. 983 int NumOfDests = LT.first; 984 int NumOfShufflesPerDest = LT.first * 2 - 1; 985 LT.first = NumOfDests * NumOfShufflesPerDest; 986 } 987 988 static const CostTblEntry AVX512VBMIShuffleTbl[] = { 989 {TTI::SK_Reverse, MVT::v64i8, 1}, // vpermb 990 {TTI::SK_Reverse, MVT::v32i8, 1}, // vpermb 991 992 {TTI::SK_PermuteSingleSrc, MVT::v64i8, 1}, // vpermb 993 {TTI::SK_PermuteSingleSrc, MVT::v32i8, 1}, // vpermb 994 995 {TTI::SK_PermuteTwoSrc, MVT::v64i8, 1}, // vpermt2b 996 {TTI::SK_PermuteTwoSrc, MVT::v32i8, 1}, // vpermt2b 997 {TTI::SK_PermuteTwoSrc, MVT::v16i8, 1} // vpermt2b 998 }; 999 1000 if (ST->hasVBMI()) 1001 if (const auto *Entry = 1002 CostTableLookup(AVX512VBMIShuffleTbl, Kind, LT.second)) 1003 return LT.first * Entry->Cost; 1004 1005 static const CostTblEntry AVX512BWShuffleTbl[] = { 1006 {TTI::SK_Broadcast, MVT::v32i16, 1}, // vpbroadcastw 1007 {TTI::SK_Broadcast, MVT::v64i8, 1}, // vpbroadcastb 1008 1009 {TTI::SK_Reverse, MVT::v32i16, 1}, // vpermw 1010 {TTI::SK_Reverse, MVT::v16i16, 1}, // vpermw 1011 {TTI::SK_Reverse, MVT::v64i8, 2}, // pshufb + vshufi64x2 1012 1013 {TTI::SK_PermuteSingleSrc, MVT::v32i16, 1}, // vpermw 1014 {TTI::SK_PermuteSingleSrc, MVT::v16i16, 1}, // vpermw 1015 {TTI::SK_PermuteSingleSrc, MVT::v8i16, 1}, // vpermw 1016 {TTI::SK_PermuteSingleSrc, MVT::v64i8, 8}, // extend to v32i16 1017 {TTI::SK_PermuteSingleSrc, MVT::v32i8, 3}, // vpermw + zext/trunc 1018 1019 {TTI::SK_PermuteTwoSrc, MVT::v32i16, 1}, // vpermt2w 1020 {TTI::SK_PermuteTwoSrc, MVT::v16i16, 1}, // vpermt2w 1021 {TTI::SK_PermuteTwoSrc, MVT::v8i16, 1}, // vpermt2w 1022 {TTI::SK_PermuteTwoSrc, MVT::v32i8, 3}, // zext + vpermt2w + trunc 1023 {TTI::SK_PermuteTwoSrc, MVT::v64i8, 19}, // 6 * v32i8 + 1 1024 {TTI::SK_PermuteTwoSrc, MVT::v16i8, 3} // zext + vpermt2w + trunc 1025 }; 1026 1027 if (ST->hasBWI()) 1028 if (const auto *Entry = 1029 CostTableLookup(AVX512BWShuffleTbl, Kind, LT.second)) 1030 return LT.first * Entry->Cost; 1031 1032 static const CostTblEntry AVX512ShuffleTbl[] = { 1033 {TTI::SK_Broadcast, MVT::v8f64, 1}, // vbroadcastpd 1034 {TTI::SK_Broadcast, MVT::v16f32, 1}, // vbroadcastps 1035 {TTI::SK_Broadcast, MVT::v8i64, 1}, // vpbroadcastq 1036 {TTI::SK_Broadcast, MVT::v16i32, 1}, // vpbroadcastd 1037 1038 {TTI::SK_Reverse, MVT::v8f64, 1}, // vpermpd 1039 {TTI::SK_Reverse, MVT::v16f32, 1}, // vpermps 1040 {TTI::SK_Reverse, MVT::v8i64, 1}, // vpermq 1041 {TTI::SK_Reverse, MVT::v16i32, 1}, // vpermd 1042 1043 {TTI::SK_PermuteSingleSrc, MVT::v8f64, 1}, // vpermpd 1044 {TTI::SK_PermuteSingleSrc, MVT::v4f64, 1}, // vpermpd 1045 {TTI::SK_PermuteSingleSrc, MVT::v2f64, 1}, // vpermpd 1046 {TTI::SK_PermuteSingleSrc, MVT::v16f32, 1}, // vpermps 1047 {TTI::SK_PermuteSingleSrc, MVT::v8f32, 1}, // vpermps 1048 {TTI::SK_PermuteSingleSrc, MVT::v4f32, 1}, // vpermps 1049 {TTI::SK_PermuteSingleSrc, MVT::v8i64, 1}, // vpermq 1050 {TTI::SK_PermuteSingleSrc, MVT::v4i64, 1}, // vpermq 1051 {TTI::SK_PermuteSingleSrc, MVT::v2i64, 1}, // vpermq 1052 {TTI::SK_PermuteSingleSrc, MVT::v16i32, 1}, // vpermd 1053 {TTI::SK_PermuteSingleSrc, MVT::v8i32, 1}, // vpermd 1054 {TTI::SK_PermuteSingleSrc, MVT::v4i32, 1}, // vpermd 1055 {TTI::SK_PermuteSingleSrc, MVT::v16i8, 1}, // pshufb 1056 1057 {TTI::SK_PermuteTwoSrc, MVT::v8f64, 1}, // vpermt2pd 1058 {TTI::SK_PermuteTwoSrc, MVT::v16f32, 1}, // vpermt2ps 1059 {TTI::SK_PermuteTwoSrc, MVT::v8i64, 1}, // vpermt2q 1060 {TTI::SK_PermuteTwoSrc, MVT::v16i32, 1}, // vpermt2d 1061 {TTI::SK_PermuteTwoSrc, MVT::v4f64, 1}, // vpermt2pd 1062 {TTI::SK_PermuteTwoSrc, MVT::v8f32, 1}, // vpermt2ps 1063 {TTI::SK_PermuteTwoSrc, MVT::v4i64, 1}, // vpermt2q 1064 {TTI::SK_PermuteTwoSrc, MVT::v8i32, 1}, // vpermt2d 1065 {TTI::SK_PermuteTwoSrc, MVT::v2f64, 1}, // vpermt2pd 1066 {TTI::SK_PermuteTwoSrc, MVT::v4f32, 1}, // vpermt2ps 1067 {TTI::SK_PermuteTwoSrc, MVT::v2i64, 1}, // vpermt2q 1068 {TTI::SK_PermuteTwoSrc, MVT::v4i32, 1} // vpermt2d 1069 }; 1070 1071 if (ST->hasAVX512()) 1072 if (const auto *Entry = CostTableLookup(AVX512ShuffleTbl, Kind, LT.second)) 1073 return LT.first * Entry->Cost; 1074 1075 static const CostTblEntry AVX2ShuffleTbl[] = { 1076 {TTI::SK_Broadcast, MVT::v4f64, 1}, // vbroadcastpd 1077 {TTI::SK_Broadcast, MVT::v8f32, 1}, // vbroadcastps 1078 {TTI::SK_Broadcast, MVT::v4i64, 1}, // vpbroadcastq 1079 {TTI::SK_Broadcast, MVT::v8i32, 1}, // vpbroadcastd 1080 {TTI::SK_Broadcast, MVT::v16i16, 1}, // vpbroadcastw 1081 {TTI::SK_Broadcast, MVT::v32i8, 1}, // vpbroadcastb 1082 1083 {TTI::SK_Reverse, MVT::v4f64, 1}, // vpermpd 1084 {TTI::SK_Reverse, MVT::v8f32, 1}, // vpermps 1085 {TTI::SK_Reverse, MVT::v4i64, 1}, // vpermq 1086 {TTI::SK_Reverse, MVT::v8i32, 1}, // vpermd 1087 {TTI::SK_Reverse, MVT::v16i16, 2}, // vperm2i128 + pshufb 1088 {TTI::SK_Reverse, MVT::v32i8, 2}, // vperm2i128 + pshufb 1089 1090 {TTI::SK_Select, MVT::v16i16, 1}, // vpblendvb 1091 {TTI::SK_Select, MVT::v32i8, 1}, // vpblendvb 1092 1093 {TTI::SK_PermuteSingleSrc, MVT::v4f64, 1}, // vpermpd 1094 {TTI::SK_PermuteSingleSrc, MVT::v8f32, 1}, // vpermps 1095 {TTI::SK_PermuteSingleSrc, MVT::v4i64, 1}, // vpermq 1096 {TTI::SK_PermuteSingleSrc, MVT::v8i32, 1}, // vpermd 1097 {TTI::SK_PermuteSingleSrc, MVT::v16i16, 4}, // vperm2i128 + 2*vpshufb 1098 // + vpblendvb 1099 {TTI::SK_PermuteSingleSrc, MVT::v32i8, 4}, // vperm2i128 + 2*vpshufb 1100 // + vpblendvb 1101 1102 {TTI::SK_PermuteTwoSrc, MVT::v4f64, 3}, // 2*vpermpd + vblendpd 1103 {TTI::SK_PermuteTwoSrc, MVT::v8f32, 3}, // 2*vpermps + vblendps 1104 {TTI::SK_PermuteTwoSrc, MVT::v4i64, 3}, // 2*vpermq + vpblendd 1105 {TTI::SK_PermuteTwoSrc, MVT::v8i32, 3}, // 2*vpermd + vpblendd 1106 {TTI::SK_PermuteTwoSrc, MVT::v16i16, 7}, // 2*vperm2i128 + 4*vpshufb 1107 // + vpblendvb 1108 {TTI::SK_PermuteTwoSrc, MVT::v32i8, 7}, // 2*vperm2i128 + 4*vpshufb 1109 // + vpblendvb 1110 }; 1111 1112 if (ST->hasAVX2()) 1113 if (const auto *Entry = CostTableLookup(AVX2ShuffleTbl, Kind, LT.second)) 1114 return LT.first * Entry->Cost; 1115 1116 static const CostTblEntry XOPShuffleTbl[] = { 1117 {TTI::SK_PermuteSingleSrc, MVT::v4f64, 2}, // vperm2f128 + vpermil2pd 1118 {TTI::SK_PermuteSingleSrc, MVT::v8f32, 2}, // vperm2f128 + vpermil2ps 1119 {TTI::SK_PermuteSingleSrc, MVT::v4i64, 2}, // vperm2f128 + vpermil2pd 1120 {TTI::SK_PermuteSingleSrc, MVT::v8i32, 2}, // vperm2f128 + vpermil2ps 1121 {TTI::SK_PermuteSingleSrc, MVT::v16i16, 4}, // vextractf128 + 2*vpperm 1122 // + vinsertf128 1123 {TTI::SK_PermuteSingleSrc, MVT::v32i8, 4}, // vextractf128 + 2*vpperm 1124 // + vinsertf128 1125 1126 {TTI::SK_PermuteTwoSrc, MVT::v16i16, 9}, // 2*vextractf128 + 6*vpperm 1127 // + vinsertf128 1128 {TTI::SK_PermuteTwoSrc, MVT::v8i16, 1}, // vpperm 1129 {TTI::SK_PermuteTwoSrc, MVT::v32i8, 9}, // 2*vextractf128 + 6*vpperm 1130 // + vinsertf128 1131 {TTI::SK_PermuteTwoSrc, MVT::v16i8, 1}, // vpperm 1132 }; 1133 1134 if (ST->hasXOP()) 1135 if (const auto *Entry = CostTableLookup(XOPShuffleTbl, Kind, LT.second)) 1136 return LT.first * Entry->Cost; 1137 1138 static const CostTblEntry AVX1ShuffleTbl[] = { 1139 {TTI::SK_Broadcast, MVT::v4f64, 2}, // vperm2f128 + vpermilpd 1140 {TTI::SK_Broadcast, MVT::v8f32, 2}, // vperm2f128 + vpermilps 1141 {TTI::SK_Broadcast, MVT::v4i64, 2}, // vperm2f128 + vpermilpd 1142 {TTI::SK_Broadcast, MVT::v8i32, 2}, // vperm2f128 + vpermilps 1143 {TTI::SK_Broadcast, MVT::v16i16, 3}, // vpshuflw + vpshufd + vinsertf128 1144 {TTI::SK_Broadcast, MVT::v32i8, 2}, // vpshufb + vinsertf128 1145 1146 {TTI::SK_Reverse, MVT::v4f64, 2}, // vperm2f128 + vpermilpd 1147 {TTI::SK_Reverse, MVT::v8f32, 2}, // vperm2f128 + vpermilps 1148 {TTI::SK_Reverse, MVT::v4i64, 2}, // vperm2f128 + vpermilpd 1149 {TTI::SK_Reverse, MVT::v8i32, 2}, // vperm2f128 + vpermilps 1150 {TTI::SK_Reverse, MVT::v16i16, 4}, // vextractf128 + 2*pshufb 1151 // + vinsertf128 1152 {TTI::SK_Reverse, MVT::v32i8, 4}, // vextractf128 + 2*pshufb 1153 // + vinsertf128 1154 1155 {TTI::SK_Select, MVT::v4i64, 1}, // vblendpd 1156 {TTI::SK_Select, MVT::v4f64, 1}, // vblendpd 1157 {TTI::SK_Select, MVT::v8i32, 1}, // vblendps 1158 {TTI::SK_Select, MVT::v8f32, 1}, // vblendps 1159 {TTI::SK_Select, MVT::v16i16, 3}, // vpand + vpandn + vpor 1160 {TTI::SK_Select, MVT::v32i8, 3}, // vpand + vpandn + vpor 1161 1162 {TTI::SK_PermuteSingleSrc, MVT::v4f64, 2}, // vperm2f128 + vshufpd 1163 {TTI::SK_PermuteSingleSrc, MVT::v4i64, 2}, // vperm2f128 + vshufpd 1164 {TTI::SK_PermuteSingleSrc, MVT::v8f32, 4}, // 2*vperm2f128 + 2*vshufps 1165 {TTI::SK_PermuteSingleSrc, MVT::v8i32, 4}, // 2*vperm2f128 + 2*vshufps 1166 {TTI::SK_PermuteSingleSrc, MVT::v16i16, 8}, // vextractf128 + 4*pshufb 1167 // + 2*por + vinsertf128 1168 {TTI::SK_PermuteSingleSrc, MVT::v32i8, 8}, // vextractf128 + 4*pshufb 1169 // + 2*por + vinsertf128 1170 1171 {TTI::SK_PermuteTwoSrc, MVT::v4f64, 3}, // 2*vperm2f128 + vshufpd 1172 {TTI::SK_PermuteTwoSrc, MVT::v4i64, 3}, // 2*vperm2f128 + vshufpd 1173 {TTI::SK_PermuteTwoSrc, MVT::v8f32, 4}, // 2*vperm2f128 + 2*vshufps 1174 {TTI::SK_PermuteTwoSrc, MVT::v8i32, 4}, // 2*vperm2f128 + 2*vshufps 1175 {TTI::SK_PermuteTwoSrc, MVT::v16i16, 15}, // 2*vextractf128 + 8*pshufb 1176 // + 4*por + vinsertf128 1177 {TTI::SK_PermuteTwoSrc, MVT::v32i8, 15}, // 2*vextractf128 + 8*pshufb 1178 // + 4*por + vinsertf128 1179 }; 1180 1181 if (ST->hasAVX()) 1182 if (const auto *Entry = CostTableLookup(AVX1ShuffleTbl, Kind, LT.second)) 1183 return LT.first * Entry->Cost; 1184 1185 static const CostTblEntry SSE41ShuffleTbl[] = { 1186 {TTI::SK_Select, MVT::v2i64, 1}, // pblendw 1187 {TTI::SK_Select, MVT::v2f64, 1}, // movsd 1188 {TTI::SK_Select, MVT::v4i32, 1}, // pblendw 1189 {TTI::SK_Select, MVT::v4f32, 1}, // blendps 1190 {TTI::SK_Select, MVT::v8i16, 1}, // pblendw 1191 {TTI::SK_Select, MVT::v16i8, 1} // pblendvb 1192 }; 1193 1194 if (ST->hasSSE41()) 1195 if (const auto *Entry = CostTableLookup(SSE41ShuffleTbl, Kind, LT.second)) 1196 return LT.first * Entry->Cost; 1197 1198 static const CostTblEntry SSSE3ShuffleTbl[] = { 1199 {TTI::SK_Broadcast, MVT::v8i16, 1}, // pshufb 1200 {TTI::SK_Broadcast, MVT::v16i8, 1}, // pshufb 1201 1202 {TTI::SK_Reverse, MVT::v8i16, 1}, // pshufb 1203 {TTI::SK_Reverse, MVT::v16i8, 1}, // pshufb 1204 1205 {TTI::SK_Select, MVT::v8i16, 3}, // 2*pshufb + por 1206 {TTI::SK_Select, MVT::v16i8, 3}, // 2*pshufb + por 1207 1208 {TTI::SK_PermuteSingleSrc, MVT::v8i16, 1}, // pshufb 1209 {TTI::SK_PermuteSingleSrc, MVT::v16i8, 1}, // pshufb 1210 1211 {TTI::SK_PermuteTwoSrc, MVT::v8i16, 3}, // 2*pshufb + por 1212 {TTI::SK_PermuteTwoSrc, MVT::v16i8, 3}, // 2*pshufb + por 1213 }; 1214 1215 if (ST->hasSSSE3()) 1216 if (const auto *Entry = CostTableLookup(SSSE3ShuffleTbl, Kind, LT.second)) 1217 return LT.first * Entry->Cost; 1218 1219 static const CostTblEntry SSE2ShuffleTbl[] = { 1220 {TTI::SK_Broadcast, MVT::v2f64, 1}, // shufpd 1221 {TTI::SK_Broadcast, MVT::v2i64, 1}, // pshufd 1222 {TTI::SK_Broadcast, MVT::v4i32, 1}, // pshufd 1223 {TTI::SK_Broadcast, MVT::v8i16, 2}, // pshuflw + pshufd 1224 {TTI::SK_Broadcast, MVT::v16i8, 3}, // unpck + pshuflw + pshufd 1225 1226 {TTI::SK_Reverse, MVT::v2f64, 1}, // shufpd 1227 {TTI::SK_Reverse, MVT::v2i64, 1}, // pshufd 1228 {TTI::SK_Reverse, MVT::v4i32, 1}, // pshufd 1229 {TTI::SK_Reverse, MVT::v8i16, 3}, // pshuflw + pshufhw + pshufd 1230 {TTI::SK_Reverse, MVT::v16i8, 9}, // 2*pshuflw + 2*pshufhw 1231 // + 2*pshufd + 2*unpck + packus 1232 1233 {TTI::SK_Select, MVT::v2i64, 1}, // movsd 1234 {TTI::SK_Select, MVT::v2f64, 1}, // movsd 1235 {TTI::SK_Select, MVT::v4i32, 2}, // 2*shufps 1236 {TTI::SK_Select, MVT::v8i16, 3}, // pand + pandn + por 1237 {TTI::SK_Select, MVT::v16i8, 3}, // pand + pandn + por 1238 1239 {TTI::SK_PermuteSingleSrc, MVT::v2f64, 1}, // shufpd 1240 {TTI::SK_PermuteSingleSrc, MVT::v2i64, 1}, // pshufd 1241 {TTI::SK_PermuteSingleSrc, MVT::v4i32, 1}, // pshufd 1242 {TTI::SK_PermuteSingleSrc, MVT::v8i16, 5}, // 2*pshuflw + 2*pshufhw 1243 // + pshufd/unpck 1244 { TTI::SK_PermuteSingleSrc, MVT::v16i8, 10 }, // 2*pshuflw + 2*pshufhw 1245 // + 2*pshufd + 2*unpck + 2*packus 1246 1247 { TTI::SK_PermuteTwoSrc, MVT::v2f64, 1 }, // shufpd 1248 { TTI::SK_PermuteTwoSrc, MVT::v2i64, 1 }, // shufpd 1249 { TTI::SK_PermuteTwoSrc, MVT::v4i32, 2 }, // 2*{unpck,movsd,pshufd} 1250 { TTI::SK_PermuteTwoSrc, MVT::v8i16, 8 }, // blend+permute 1251 { TTI::SK_PermuteTwoSrc, MVT::v16i8, 13 }, // blend+permute 1252 }; 1253 1254 if (ST->hasSSE2()) 1255 if (const auto *Entry = CostTableLookup(SSE2ShuffleTbl, Kind, LT.second)) 1256 return LT.first * Entry->Cost; 1257 1258 static const CostTblEntry SSE1ShuffleTbl[] = { 1259 { TTI::SK_Broadcast, MVT::v4f32, 1 }, // shufps 1260 { TTI::SK_Reverse, MVT::v4f32, 1 }, // shufps 1261 { TTI::SK_Select, MVT::v4f32, 2 }, // 2*shufps 1262 { TTI::SK_PermuteSingleSrc, MVT::v4f32, 1 }, // shufps 1263 { TTI::SK_PermuteTwoSrc, MVT::v4f32, 2 }, // 2*shufps 1264 }; 1265 1266 if (ST->hasSSE1()) 1267 if (const auto *Entry = CostTableLookup(SSE1ShuffleTbl, Kind, LT.second)) 1268 return LT.first * Entry->Cost; 1269 1270 return BaseT::getShuffleCost(Kind, Tp, Index, SubTp); 1271 } 1272 1273 int X86TTIImpl::getCastInstrCost(unsigned Opcode, Type *Dst, Type *Src, 1274 const Instruction *I) { 1275 int ISD = TLI->InstructionOpcodeToISD(Opcode); 1276 assert(ISD && "Invalid opcode"); 1277 1278 // FIXME: Need a better design of the cost table to handle non-simple types of 1279 // potential massive combinations (elem_num x src_type x dst_type). 1280 1281 static const TypeConversionCostTblEntry AVX512BWConversionTbl[] { 1282 { ISD::SIGN_EXTEND, MVT::v32i16, MVT::v32i8, 1 }, 1283 { ISD::ZERO_EXTEND, MVT::v32i16, MVT::v32i8, 1 }, 1284 1285 // Mask sign extend has an instruction. 1286 { ISD::SIGN_EXTEND, MVT::v8i16, MVT::v8i1, 1 }, 1287 { ISD::SIGN_EXTEND, MVT::v16i8, MVT::v16i1, 1 }, 1288 { ISD::SIGN_EXTEND, MVT::v16i16, MVT::v16i1, 1 }, 1289 { ISD::SIGN_EXTEND, MVT::v32i8, MVT::v32i1, 1 }, 1290 { ISD::SIGN_EXTEND, MVT::v32i16, MVT::v32i1, 1 }, 1291 { ISD::SIGN_EXTEND, MVT::v64i8, MVT::v64i1, 1 }, 1292 1293 // Mask zero extend is a load + broadcast. 1294 { ISD::ZERO_EXTEND, MVT::v8i16, MVT::v8i1, 2 }, 1295 { ISD::ZERO_EXTEND, MVT::v16i8, MVT::v16i1, 2 }, 1296 { ISD::ZERO_EXTEND, MVT::v16i16, MVT::v16i1, 2 }, 1297 { ISD::ZERO_EXTEND, MVT::v32i8, MVT::v32i1, 2 }, 1298 { ISD::ZERO_EXTEND, MVT::v32i16, MVT::v32i1, 2 }, 1299 { ISD::ZERO_EXTEND, MVT::v64i8, MVT::v64i1, 2 }, 1300 1301 { ISD::TRUNCATE, MVT::v32i8, MVT::v32i16, 1 }, 1302 }; 1303 1304 static const TypeConversionCostTblEntry AVX512DQConversionTbl[] = { 1305 { ISD::SINT_TO_FP, MVT::v2f32, MVT::v2i64, 1 }, 1306 { ISD::SINT_TO_FP, MVT::v2f64, MVT::v2i64, 1 }, 1307 { ISD::SINT_TO_FP, MVT::v4f32, MVT::v4i64, 1 }, 1308 { ISD::SINT_TO_FP, MVT::v4f64, MVT::v4i64, 1 }, 1309 { ISD::SINT_TO_FP, MVT::v8f32, MVT::v8i64, 1 }, 1310 { ISD::SINT_TO_FP, MVT::v8f64, MVT::v8i64, 1 }, 1311 1312 { ISD::UINT_TO_FP, MVT::v2f32, MVT::v2i64, 1 }, 1313 { ISD::UINT_TO_FP, MVT::v2f64, MVT::v2i64, 1 }, 1314 { ISD::UINT_TO_FP, MVT::v4f32, MVT::v4i64, 1 }, 1315 { ISD::UINT_TO_FP, MVT::v4f64, MVT::v4i64, 1 }, 1316 { ISD::UINT_TO_FP, MVT::v8f32, MVT::v8i64, 1 }, 1317 { ISD::UINT_TO_FP, MVT::v8f64, MVT::v8i64, 1 }, 1318 1319 { ISD::FP_TO_SINT, MVT::v2i64, MVT::v2f32, 1 }, 1320 { ISD::FP_TO_SINT, MVT::v4i64, MVT::v4f32, 1 }, 1321 { ISD::FP_TO_SINT, MVT::v8i64, MVT::v8f32, 1 }, 1322 { ISD::FP_TO_SINT, MVT::v2i64, MVT::v2f64, 1 }, 1323 { ISD::FP_TO_SINT, MVT::v4i64, MVT::v4f64, 1 }, 1324 { ISD::FP_TO_SINT, MVT::v8i64, MVT::v8f64, 1 }, 1325 1326 { ISD::FP_TO_UINT, MVT::v2i64, MVT::v2f32, 1 }, 1327 { ISD::FP_TO_UINT, MVT::v4i64, MVT::v4f32, 1 }, 1328 { ISD::FP_TO_UINT, MVT::v8i64, MVT::v8f32, 1 }, 1329 { ISD::FP_TO_UINT, MVT::v2i64, MVT::v2f64, 1 }, 1330 { ISD::FP_TO_UINT, MVT::v4i64, MVT::v4f64, 1 }, 1331 { ISD::FP_TO_UINT, MVT::v8i64, MVT::v8f64, 1 }, 1332 }; 1333 1334 // TODO: For AVX512DQ + AVX512VL, we also have cheap casts for 128-bit and 1335 // 256-bit wide vectors. 1336 1337 static const TypeConversionCostTblEntry AVX512FConversionTbl[] = { 1338 { ISD::FP_EXTEND, MVT::v8f64, MVT::v8f32, 1 }, 1339 { ISD::FP_EXTEND, MVT::v8f64, MVT::v16f32, 3 }, 1340 { ISD::FP_ROUND, MVT::v8f32, MVT::v8f64, 1 }, 1341 1342 { ISD::TRUNCATE, MVT::v16i8, MVT::v16i32, 1 }, 1343 { ISD::TRUNCATE, MVT::v16i16, MVT::v16i32, 1 }, 1344 { ISD::TRUNCATE, MVT::v8i16, MVT::v8i64, 1 }, 1345 { ISD::TRUNCATE, MVT::v8i32, MVT::v8i64, 1 }, 1346 1347 // v16i1 -> v16i32 - load + broadcast 1348 { ISD::SIGN_EXTEND, MVT::v16i32, MVT::v16i1, 2 }, 1349 { ISD::ZERO_EXTEND, MVT::v16i32, MVT::v16i1, 2 }, 1350 { ISD::SIGN_EXTEND, MVT::v16i32, MVT::v16i8, 1 }, 1351 { ISD::ZERO_EXTEND, MVT::v16i32, MVT::v16i8, 1 }, 1352 { ISD::SIGN_EXTEND, MVT::v16i32, MVT::v16i16, 1 }, 1353 { ISD::ZERO_EXTEND, MVT::v16i32, MVT::v16i16, 1 }, 1354 { ISD::SIGN_EXTEND, MVT::v8i64, MVT::v8i8, 1 }, 1355 { ISD::ZERO_EXTEND, MVT::v8i64, MVT::v8i8, 1 }, 1356 { ISD::SIGN_EXTEND, MVT::v8i64, MVT::v8i16, 1 }, 1357 { ISD::ZERO_EXTEND, MVT::v8i64, MVT::v8i16, 1 }, 1358 { ISD::SIGN_EXTEND, MVT::v8i64, MVT::v8i32, 1 }, 1359 { ISD::ZERO_EXTEND, MVT::v8i64, MVT::v8i32, 1 }, 1360 1361 { ISD::SINT_TO_FP, MVT::v8f64, MVT::v8i1, 4 }, 1362 { ISD::SINT_TO_FP, MVT::v16f32, MVT::v16i1, 3 }, 1363 { ISD::SINT_TO_FP, MVT::v8f64, MVT::v8i8, 2 }, 1364 { ISD::SINT_TO_FP, MVT::v16f32, MVT::v16i8, 2 }, 1365 { ISD::SINT_TO_FP, MVT::v8f64, MVT::v8i16, 2 }, 1366 { ISD::SINT_TO_FP, MVT::v16f32, MVT::v16i16, 2 }, 1367 { ISD::SINT_TO_FP, MVT::v16f32, MVT::v16i32, 1 }, 1368 { ISD::SINT_TO_FP, MVT::v8f64, MVT::v8i32, 1 }, 1369 1370 { ISD::UINT_TO_FP, MVT::v8f64, MVT::v8i1, 4 }, 1371 { ISD::UINT_TO_FP, MVT::v16f32, MVT::v16i1, 3 }, 1372 { ISD::UINT_TO_FP, MVT::v2f64, MVT::v2i8, 2 }, 1373 { ISD::UINT_TO_FP, MVT::v4f64, MVT::v4i8, 2 }, 1374 { ISD::UINT_TO_FP, MVT::v8f32, MVT::v8i8, 2 }, 1375 { ISD::UINT_TO_FP, MVT::v8f64, MVT::v8i8, 2 }, 1376 { ISD::UINT_TO_FP, MVT::v16f32, MVT::v16i8, 2 }, 1377 { ISD::UINT_TO_FP, MVT::v2f64, MVT::v2i16, 5 }, 1378 { ISD::UINT_TO_FP, MVT::v4f64, MVT::v4i16, 2 }, 1379 { ISD::UINT_TO_FP, MVT::v8f32, MVT::v8i16, 2 }, 1380 { ISD::UINT_TO_FP, MVT::v8f64, MVT::v8i16, 2 }, 1381 { ISD::UINT_TO_FP, MVT::v16f32, MVT::v16i16, 2 }, 1382 { ISD::UINT_TO_FP, MVT::v2f32, MVT::v2i32, 2 }, 1383 { ISD::UINT_TO_FP, MVT::v2f64, MVT::v2i32, 1 }, 1384 { ISD::UINT_TO_FP, MVT::v4f32, MVT::v4i32, 1 }, 1385 { ISD::UINT_TO_FP, MVT::v4f64, MVT::v4i32, 1 }, 1386 { ISD::UINT_TO_FP, MVT::v8f32, MVT::v8i32, 1 }, 1387 { ISD::UINT_TO_FP, MVT::v8f64, MVT::v8i32, 1 }, 1388 { ISD::UINT_TO_FP, MVT::v16f32, MVT::v16i32, 1 }, 1389 { ISD::UINT_TO_FP, MVT::v2f32, MVT::v2i64, 5 }, 1390 { ISD::UINT_TO_FP, MVT::v8f32, MVT::v8i64, 26 }, 1391 { ISD::UINT_TO_FP, MVT::v2f64, MVT::v2i64, 5 }, 1392 { ISD::UINT_TO_FP, MVT::v4f64, MVT::v4i64, 5 }, 1393 { ISD::UINT_TO_FP, MVT::v8f64, MVT::v8i64, 5 }, 1394 1395 { ISD::UINT_TO_FP, MVT::f32, MVT::i64, 1 }, 1396 { ISD::UINT_TO_FP, MVT::f64, MVT::i64, 1 }, 1397 { ISD::FP_TO_UINT, MVT::i64, MVT::f32, 1 }, 1398 { ISD::FP_TO_UINT, MVT::i64, MVT::f64, 1 }, 1399 1400 { ISD::FP_TO_UINT, MVT::v2i32, MVT::v2f32, 1 }, 1401 { ISD::FP_TO_UINT, MVT::v4i32, MVT::v4f32, 1 }, 1402 { ISD::FP_TO_UINT, MVT::v4i32, MVT::v4f64, 1 }, 1403 { ISD::FP_TO_UINT, MVT::v8i32, MVT::v8f32, 1 }, 1404 { ISD::FP_TO_UINT, MVT::v8i32, MVT::v8f64, 1 }, 1405 { ISD::FP_TO_UINT, MVT::v8i16, MVT::v8f64, 2 }, 1406 { ISD::FP_TO_UINT, MVT::v8i8, MVT::v8f64, 2 }, 1407 { ISD::FP_TO_UINT, MVT::v16i32, MVT::v16f32, 1 }, 1408 { ISD::FP_TO_UINT, MVT::v16i16, MVT::v16f32, 2 }, 1409 { ISD::FP_TO_UINT, MVT::v16i8, MVT::v16f32, 2 }, 1410 }; 1411 1412 static const TypeConversionCostTblEntry AVX2ConversionTbl[] = { 1413 { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v4i1, 3 }, 1414 { ISD::ZERO_EXTEND, MVT::v4i64, MVT::v4i1, 3 }, 1415 { ISD::SIGN_EXTEND, MVT::v8i32, MVT::v8i1, 3 }, 1416 { ISD::ZERO_EXTEND, MVT::v8i32, MVT::v8i1, 3 }, 1417 { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v4i8, 1 }, 1418 { ISD::ZERO_EXTEND, MVT::v4i64, MVT::v4i8, 1 }, 1419 { ISD::SIGN_EXTEND, MVT::v8i32, MVT::v8i8, 1 }, 1420 { ISD::ZERO_EXTEND, MVT::v8i32, MVT::v8i8, 1 }, 1421 { ISD::SIGN_EXTEND, MVT::v16i16, MVT::v16i8, 1 }, 1422 { ISD::ZERO_EXTEND, MVT::v16i16, MVT::v16i8, 1 }, 1423 { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v4i16, 1 }, 1424 { ISD::ZERO_EXTEND, MVT::v4i64, MVT::v4i16, 1 }, 1425 { ISD::SIGN_EXTEND, MVT::v8i32, MVT::v8i16, 1 }, 1426 { ISD::ZERO_EXTEND, MVT::v8i32, MVT::v8i16, 1 }, 1427 { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v4i32, 1 }, 1428 { ISD::ZERO_EXTEND, MVT::v4i64, MVT::v4i32, 1 }, 1429 { ISD::ZERO_EXTEND, MVT::v16i32, MVT::v16i16, 3 }, 1430 { ISD::SIGN_EXTEND, MVT::v16i32, MVT::v16i16, 3 }, 1431 1432 { ISD::TRUNCATE, MVT::v4i8, MVT::v4i64, 2 }, 1433 { ISD::TRUNCATE, MVT::v4i16, MVT::v4i64, 2 }, 1434 { ISD::TRUNCATE, MVT::v4i32, MVT::v4i64, 2 }, 1435 { ISD::TRUNCATE, MVT::v8i8, MVT::v8i32, 2 }, 1436 { ISD::TRUNCATE, MVT::v8i16, MVT::v8i32, 2 }, 1437 { ISD::TRUNCATE, MVT::v8i32, MVT::v8i64, 4 }, 1438 1439 { ISD::FP_EXTEND, MVT::v8f64, MVT::v8f32, 3 }, 1440 { ISD::FP_ROUND, MVT::v8f32, MVT::v8f64, 3 }, 1441 1442 { ISD::UINT_TO_FP, MVT::v8f32, MVT::v8i32, 8 }, 1443 }; 1444 1445 static const TypeConversionCostTblEntry AVXConversionTbl[] = { 1446 { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v4i1, 6 }, 1447 { ISD::ZERO_EXTEND, MVT::v4i64, MVT::v4i1, 4 }, 1448 { ISD::SIGN_EXTEND, MVT::v8i32, MVT::v8i1, 7 }, 1449 { ISD::ZERO_EXTEND, MVT::v8i32, MVT::v8i1, 4 }, 1450 { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v4i8, 4 }, 1451 { ISD::ZERO_EXTEND, MVT::v4i64, MVT::v4i8, 4 }, 1452 { ISD::SIGN_EXTEND, MVT::v8i32, MVT::v8i8, 4 }, 1453 { ISD::ZERO_EXTEND, MVT::v8i32, MVT::v8i8, 4 }, 1454 { ISD::SIGN_EXTEND, MVT::v16i16, MVT::v16i8, 4 }, 1455 { ISD::ZERO_EXTEND, MVT::v16i16, MVT::v16i8, 4 }, 1456 { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v4i16, 4 }, 1457 { ISD::ZERO_EXTEND, MVT::v4i64, MVT::v4i16, 3 }, 1458 { ISD::SIGN_EXTEND, MVT::v8i32, MVT::v8i16, 4 }, 1459 { ISD::ZERO_EXTEND, MVT::v8i32, MVT::v8i16, 4 }, 1460 { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v4i32, 4 }, 1461 { ISD::ZERO_EXTEND, MVT::v4i64, MVT::v4i32, 4 }, 1462 1463 { ISD::TRUNCATE, MVT::v16i8, MVT::v16i16, 4 }, 1464 { ISD::TRUNCATE, MVT::v8i8, MVT::v8i32, 4 }, 1465 { ISD::TRUNCATE, MVT::v8i16, MVT::v8i32, 5 }, 1466 { ISD::TRUNCATE, MVT::v4i8, MVT::v4i64, 4 }, 1467 { ISD::TRUNCATE, MVT::v4i16, MVT::v4i64, 4 }, 1468 { ISD::TRUNCATE, MVT::v4i32, MVT::v4i64, 4 }, 1469 { ISD::TRUNCATE, MVT::v8i8, MVT::v8i64, 11 }, 1470 { ISD::TRUNCATE, MVT::v8i16, MVT::v8i64, 9 }, 1471 { ISD::TRUNCATE, MVT::v8i32, MVT::v8i64, 9 }, 1472 { ISD::TRUNCATE, MVT::v16i8, MVT::v16i64, 11 }, 1473 1474 { ISD::SINT_TO_FP, MVT::v4f32, MVT::v4i1, 3 }, 1475 { ISD::SINT_TO_FP, MVT::v4f64, MVT::v4i1, 3 }, 1476 { ISD::SINT_TO_FP, MVT::v8f32, MVT::v8i1, 8 }, 1477 { ISD::SINT_TO_FP, MVT::v4f32, MVT::v4i8, 3 }, 1478 { ISD::SINT_TO_FP, MVT::v4f64, MVT::v4i8, 3 }, 1479 { ISD::SINT_TO_FP, MVT::v8f32, MVT::v8i8, 8 }, 1480 { ISD::SINT_TO_FP, MVT::v4f32, MVT::v4i16, 3 }, 1481 { ISD::SINT_TO_FP, MVT::v4f64, MVT::v4i16, 3 }, 1482 { ISD::SINT_TO_FP, MVT::v8f32, MVT::v8i16, 5 }, 1483 { ISD::SINT_TO_FP, MVT::v4f32, MVT::v4i32, 1 }, 1484 { ISD::SINT_TO_FP, MVT::v4f64, MVT::v4i32, 1 }, 1485 { ISD::SINT_TO_FP, MVT::v8f32, MVT::v8i32, 1 }, 1486 1487 { ISD::UINT_TO_FP, MVT::v4f32, MVT::v4i1, 7 }, 1488 { ISD::UINT_TO_FP, MVT::v4f64, MVT::v4i1, 7 }, 1489 { ISD::UINT_TO_FP, MVT::v8f32, MVT::v8i1, 6 }, 1490 { ISD::UINT_TO_FP, MVT::v4f32, MVT::v4i8, 2 }, 1491 { ISD::UINT_TO_FP, MVT::v4f64, MVT::v4i8, 2 }, 1492 { ISD::UINT_TO_FP, MVT::v8f32, MVT::v8i8, 5 }, 1493 { ISD::UINT_TO_FP, MVT::v4f32, MVT::v4i16, 2 }, 1494 { ISD::UINT_TO_FP, MVT::v4f64, MVT::v4i16, 2 }, 1495 { ISD::UINT_TO_FP, MVT::v8f32, MVT::v8i16, 5 }, 1496 { ISD::UINT_TO_FP, MVT::v2f64, MVT::v2i32, 6 }, 1497 { ISD::UINT_TO_FP, MVT::v4f32, MVT::v4i32, 6 }, 1498 { ISD::UINT_TO_FP, MVT::v4f64, MVT::v4i32, 6 }, 1499 { ISD::UINT_TO_FP, MVT::v8f32, MVT::v8i32, 9 }, 1500 { ISD::UINT_TO_FP, MVT::v2f64, MVT::v2i64, 5 }, 1501 { ISD::UINT_TO_FP, MVT::v4f64, MVT::v4i64, 6 }, 1502 // The generic code to compute the scalar overhead is currently broken. 1503 // Workaround this limitation by estimating the scalarization overhead 1504 // here. We have roughly 10 instructions per scalar element. 1505 // Multiply that by the vector width. 1506 // FIXME: remove that when PR19268 is fixed. 1507 { ISD::SINT_TO_FP, MVT::v4f64, MVT::v4i64, 13 }, 1508 { ISD::SINT_TO_FP, MVT::v4f64, MVT::v4i64, 13 }, 1509 1510 { ISD::FP_TO_SINT, MVT::v4i8, MVT::v4f32, 1 }, 1511 { ISD::FP_TO_SINT, MVT::v8i8, MVT::v8f32, 7 }, 1512 // This node is expanded into scalarized operations but BasicTTI is overly 1513 // optimistic estimating its cost. It computes 3 per element (one 1514 // vector-extract, one scalar conversion and one vector-insert). The 1515 // problem is that the inserts form a read-modify-write chain so latency 1516 // should be factored in too. Inflating the cost per element by 1. 1517 { ISD::FP_TO_UINT, MVT::v8i32, MVT::v8f32, 8*4 }, 1518 { ISD::FP_TO_UINT, MVT::v4i32, MVT::v4f64, 4*4 }, 1519 1520 { ISD::FP_EXTEND, MVT::v4f64, MVT::v4f32, 1 }, 1521 { ISD::FP_ROUND, MVT::v4f32, MVT::v4f64, 1 }, 1522 }; 1523 1524 static const TypeConversionCostTblEntry SSE41ConversionTbl[] = { 1525 { ISD::ZERO_EXTEND, MVT::v4i64, MVT::v4i8, 2 }, 1526 { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v4i8, 2 }, 1527 { ISD::ZERO_EXTEND, MVT::v4i64, MVT::v4i16, 2 }, 1528 { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v4i16, 2 }, 1529 { ISD::ZERO_EXTEND, MVT::v4i64, MVT::v4i32, 2 }, 1530 { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v4i32, 2 }, 1531 1532 { ISD::ZERO_EXTEND, MVT::v4i16, MVT::v4i8, 1 }, 1533 { ISD::SIGN_EXTEND, MVT::v4i16, MVT::v4i8, 2 }, 1534 { ISD::ZERO_EXTEND, MVT::v4i32, MVT::v4i8, 1 }, 1535 { ISD::SIGN_EXTEND, MVT::v4i32, MVT::v4i8, 1 }, 1536 { ISD::ZERO_EXTEND, MVT::v8i16, MVT::v8i8, 1 }, 1537 { ISD::SIGN_EXTEND, MVT::v8i16, MVT::v8i8, 1 }, 1538 { ISD::ZERO_EXTEND, MVT::v8i32, MVT::v8i8, 2 }, 1539 { ISD::SIGN_EXTEND, MVT::v8i32, MVT::v8i8, 2 }, 1540 { ISD::ZERO_EXTEND, MVT::v16i16, MVT::v16i8, 2 }, 1541 { ISD::SIGN_EXTEND, MVT::v16i16, MVT::v16i8, 2 }, 1542 { ISD::ZERO_EXTEND, MVT::v16i32, MVT::v16i8, 4 }, 1543 { ISD::SIGN_EXTEND, MVT::v16i32, MVT::v16i8, 4 }, 1544 { ISD::ZERO_EXTEND, MVT::v4i32, MVT::v4i16, 1 }, 1545 { ISD::SIGN_EXTEND, MVT::v4i32, MVT::v4i16, 1 }, 1546 { ISD::ZERO_EXTEND, MVT::v8i32, MVT::v8i16, 2 }, 1547 { ISD::SIGN_EXTEND, MVT::v8i32, MVT::v8i16, 2 }, 1548 { ISD::ZERO_EXTEND, MVT::v16i32, MVT::v16i16, 4 }, 1549 { ISD::SIGN_EXTEND, MVT::v16i32, MVT::v16i16, 4 }, 1550 1551 { ISD::TRUNCATE, MVT::v4i8, MVT::v4i16, 2 }, 1552 { ISD::TRUNCATE, MVT::v8i8, MVT::v8i16, 1 }, 1553 { ISD::TRUNCATE, MVT::v4i8, MVT::v4i32, 1 }, 1554 { ISD::TRUNCATE, MVT::v4i16, MVT::v4i32, 1 }, 1555 { ISD::TRUNCATE, MVT::v8i8, MVT::v8i32, 3 }, 1556 { ISD::TRUNCATE, MVT::v8i16, MVT::v8i32, 3 }, 1557 { ISD::TRUNCATE, MVT::v16i16, MVT::v16i32, 6 }, 1558 { ISD::TRUNCATE, MVT::v2i8, MVT::v2i64, 1 }, // PSHUFB 1559 1560 { ISD::UINT_TO_FP, MVT::f32, MVT::i64, 4 }, 1561 { ISD::UINT_TO_FP, MVT::f64, MVT::i64, 4 }, 1562 }; 1563 1564 static const TypeConversionCostTblEntry SSE2ConversionTbl[] = { 1565 // These are somewhat magic numbers justified by looking at the output of 1566 // Intel's IACA, running some kernels and making sure when we take 1567 // legalization into account the throughput will be overestimated. 1568 { ISD::SINT_TO_FP, MVT::v4f32, MVT::v16i8, 8 }, 1569 { ISD::SINT_TO_FP, MVT::v2f64, MVT::v16i8, 16*10 }, 1570 { ISD::SINT_TO_FP, MVT::v4f32, MVT::v8i16, 15 }, 1571 { ISD::SINT_TO_FP, MVT::v2f64, MVT::v8i16, 8*10 }, 1572 { ISD::SINT_TO_FP, MVT::v4f32, MVT::v4i32, 5 }, 1573 { ISD::SINT_TO_FP, MVT::v2f64, MVT::v4i32, 2*10 }, 1574 { ISD::SINT_TO_FP, MVT::v2f64, MVT::v2i32, 2*10 }, 1575 { ISD::SINT_TO_FP, MVT::v4f32, MVT::v2i64, 15 }, 1576 { ISD::SINT_TO_FP, MVT::v2f64, MVT::v2i64, 2*10 }, 1577 1578 { ISD::UINT_TO_FP, MVT::v2f64, MVT::v16i8, 16*10 }, 1579 { ISD::UINT_TO_FP, MVT::v4f32, MVT::v16i8, 8 }, 1580 { ISD::UINT_TO_FP, MVT::v4f32, MVT::v8i16, 15 }, 1581 { ISD::UINT_TO_FP, MVT::v2f64, MVT::v8i16, 8*10 }, 1582 { ISD::UINT_TO_FP, MVT::v2f64, MVT::v4i32, 4*10 }, 1583 { ISD::UINT_TO_FP, MVT::v4f32, MVT::v4i32, 8 }, 1584 { ISD::UINT_TO_FP, MVT::v2f64, MVT::v2i64, 6 }, 1585 { ISD::UINT_TO_FP, MVT::v4f32, MVT::v2i64, 15 }, 1586 1587 { ISD::FP_TO_SINT, MVT::v4i16, MVT::v4f32, 2 }, 1588 { ISD::FP_TO_SINT, MVT::v2i16, MVT::v2f64, 2 }, 1589 1590 { ISD::FP_TO_SINT, MVT::v2i32, MVT::v2f64, 3 }, 1591 1592 { ISD::UINT_TO_FP, MVT::f32, MVT::i64, 6 }, 1593 { ISD::UINT_TO_FP, MVT::f64, MVT::i64, 6 }, 1594 1595 { ISD::FP_TO_UINT, MVT::i64, MVT::f32, 4 }, 1596 { ISD::FP_TO_UINT, MVT::i64, MVT::f64, 4 }, 1597 1598 { ISD::ZERO_EXTEND, MVT::v4i16, MVT::v4i8, 1 }, 1599 { ISD::SIGN_EXTEND, MVT::v4i16, MVT::v4i8, 6 }, 1600 { ISD::ZERO_EXTEND, MVT::v4i32, MVT::v4i8, 2 }, 1601 { ISD::SIGN_EXTEND, MVT::v4i32, MVT::v4i8, 3 }, 1602 { ISD::ZERO_EXTEND, MVT::v4i64, MVT::v4i8, 4 }, 1603 { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v4i8, 8 }, 1604 { ISD::ZERO_EXTEND, MVT::v8i16, MVT::v8i8, 1 }, 1605 { ISD::SIGN_EXTEND, MVT::v8i16, MVT::v8i8, 2 }, 1606 { ISD::ZERO_EXTEND, MVT::v8i32, MVT::v8i8, 6 }, 1607 { ISD::SIGN_EXTEND, MVT::v8i32, MVT::v8i8, 6 }, 1608 { ISD::ZERO_EXTEND, MVT::v16i16, MVT::v16i8, 3 }, 1609 { ISD::SIGN_EXTEND, MVT::v16i16, MVT::v16i8, 4 }, 1610 { ISD::ZERO_EXTEND, MVT::v16i32, MVT::v16i8, 9 }, 1611 { ISD::SIGN_EXTEND, MVT::v16i32, MVT::v16i8, 12 }, 1612 { ISD::ZERO_EXTEND, MVT::v4i32, MVT::v4i16, 1 }, 1613 { ISD::SIGN_EXTEND, MVT::v4i32, MVT::v4i16, 2 }, 1614 { ISD::ZERO_EXTEND, MVT::v4i64, MVT::v4i16, 3 }, 1615 { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v4i16, 10 }, 1616 { ISD::ZERO_EXTEND, MVT::v8i32, MVT::v8i16, 3 }, 1617 { ISD::SIGN_EXTEND, MVT::v8i32, MVT::v8i16, 4 }, 1618 { ISD::ZERO_EXTEND, MVT::v16i32, MVT::v16i16, 6 }, 1619 { ISD::SIGN_EXTEND, MVT::v16i32, MVT::v16i16, 8 }, 1620 { ISD::ZERO_EXTEND, MVT::v4i64, MVT::v4i32, 3 }, 1621 { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v4i32, 5 }, 1622 1623 { ISD::TRUNCATE, MVT::v2i8, MVT::v2i16, 2 }, // PAND+PACKUSWB 1624 { ISD::TRUNCATE, MVT::v4i8, MVT::v4i16, 4 }, 1625 { ISD::TRUNCATE, MVT::v8i8, MVT::v8i16, 2 }, 1626 { ISD::TRUNCATE, MVT::v16i8, MVT::v16i16, 3 }, 1627 { ISD::TRUNCATE, MVT::v2i8, MVT::v2i32, 3 }, // PAND+3*PACKUSWB 1628 { ISD::TRUNCATE, MVT::v2i16, MVT::v2i32, 1 }, 1629 { ISD::TRUNCATE, MVT::v4i8, MVT::v4i32, 3 }, 1630 { ISD::TRUNCATE, MVT::v4i16, MVT::v4i32, 3 }, 1631 { ISD::TRUNCATE, MVT::v8i8, MVT::v8i32, 4 }, 1632 { ISD::TRUNCATE, MVT::v16i8, MVT::v16i32, 7 }, 1633 { ISD::TRUNCATE, MVT::v8i16, MVT::v8i32, 5 }, 1634 { ISD::TRUNCATE, MVT::v16i16, MVT::v16i32, 10 }, 1635 { ISD::TRUNCATE, MVT::v2i8, MVT::v2i64, 4 }, // PAND+3*PACKUSWB 1636 { ISD::TRUNCATE, MVT::v2i16, MVT::v2i64, 2 }, // PSHUFD+PSHUFLW 1637 { ISD::TRUNCATE, MVT::v2i32, MVT::v2i64, 1 }, // PSHUFD 1638 }; 1639 1640 std::pair<int, MVT> LTSrc = TLI->getTypeLegalizationCost(DL, Src); 1641 std::pair<int, MVT> LTDest = TLI->getTypeLegalizationCost(DL, Dst); 1642 1643 if (ST->hasSSE2() && !ST->hasAVX()) { 1644 if (const auto *Entry = ConvertCostTableLookup(SSE2ConversionTbl, ISD, 1645 LTDest.second, LTSrc.second)) 1646 return LTSrc.first * Entry->Cost; 1647 } 1648 1649 EVT SrcTy = TLI->getValueType(DL, Src); 1650 EVT DstTy = TLI->getValueType(DL, Dst); 1651 1652 // The function getSimpleVT only handles simple value types. 1653 if (!SrcTy.isSimple() || !DstTy.isSimple()) 1654 return BaseT::getCastInstrCost(Opcode, Dst, Src); 1655 1656 MVT SimpleSrcTy = SrcTy.getSimpleVT(); 1657 MVT SimpleDstTy = DstTy.getSimpleVT(); 1658 1659 // Make sure that neither type is going to be split before using the 1660 // AVX512 tables. This handles -mprefer-vector-width=256 1661 // with -min-legal-vector-width<=256 1662 if (TLI->getTypeAction(SimpleSrcTy) != TargetLowering::TypeSplitVector && 1663 TLI->getTypeAction(SimpleDstTy) != TargetLowering::TypeSplitVector) { 1664 if (ST->hasBWI()) 1665 if (const auto *Entry = ConvertCostTableLookup(AVX512BWConversionTbl, ISD, 1666 SimpleDstTy, SimpleSrcTy)) 1667 return Entry->Cost; 1668 1669 if (ST->hasDQI()) 1670 if (const auto *Entry = ConvertCostTableLookup(AVX512DQConversionTbl, ISD, 1671 SimpleDstTy, SimpleSrcTy)) 1672 return Entry->Cost; 1673 1674 if (ST->hasAVX512()) 1675 if (const auto *Entry = ConvertCostTableLookup(AVX512FConversionTbl, ISD, 1676 SimpleDstTy, SimpleSrcTy)) 1677 return Entry->Cost; 1678 } 1679 1680 if (ST->hasAVX2()) { 1681 if (const auto *Entry = ConvertCostTableLookup(AVX2ConversionTbl, ISD, 1682 SimpleDstTy, SimpleSrcTy)) 1683 return Entry->Cost; 1684 } 1685 1686 if (ST->hasAVX()) { 1687 if (const auto *Entry = ConvertCostTableLookup(AVXConversionTbl, ISD, 1688 SimpleDstTy, SimpleSrcTy)) 1689 return Entry->Cost; 1690 } 1691 1692 if (ST->hasSSE41()) { 1693 if (const auto *Entry = ConvertCostTableLookup(SSE41ConversionTbl, ISD, 1694 SimpleDstTy, SimpleSrcTy)) 1695 return Entry->Cost; 1696 } 1697 1698 if (ST->hasSSE2()) { 1699 if (const auto *Entry = ConvertCostTableLookup(SSE2ConversionTbl, ISD, 1700 SimpleDstTy, SimpleSrcTy)) 1701 return Entry->Cost; 1702 } 1703 1704 return BaseT::getCastInstrCost(Opcode, Dst, Src, I); 1705 } 1706 1707 int X86TTIImpl::getCmpSelInstrCost(unsigned Opcode, Type *ValTy, Type *CondTy, 1708 const Instruction *I) { 1709 // Legalize the type. 1710 std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, ValTy); 1711 1712 MVT MTy = LT.second; 1713 1714 int ISD = TLI->InstructionOpcodeToISD(Opcode); 1715 assert(ISD && "Invalid opcode"); 1716 1717 unsigned ExtraCost = 0; 1718 if (I && (Opcode == Instruction::ICmp || Opcode == Instruction::FCmp)) { 1719 // Some vector comparison predicates cost extra instructions. 1720 if (MTy.isVector() && 1721 !((ST->hasXOP() && (!ST->hasAVX2() || MTy.is128BitVector())) || 1722 (ST->hasAVX512() && 32 <= MTy.getScalarSizeInBits()) || 1723 ST->hasBWI())) { 1724 switch (cast<CmpInst>(I)->getPredicate()) { 1725 case CmpInst::Predicate::ICMP_NE: 1726 // xor(cmpeq(x,y),-1) 1727 ExtraCost = 1; 1728 break; 1729 case CmpInst::Predicate::ICMP_SGE: 1730 case CmpInst::Predicate::ICMP_SLE: 1731 // xor(cmpgt(x,y),-1) 1732 ExtraCost = 1; 1733 break; 1734 case CmpInst::Predicate::ICMP_ULT: 1735 case CmpInst::Predicate::ICMP_UGT: 1736 // cmpgt(xor(x,signbit),xor(y,signbit)) 1737 // xor(cmpeq(pmaxu(x,y),x),-1) 1738 ExtraCost = 2; 1739 break; 1740 case CmpInst::Predicate::ICMP_ULE: 1741 case CmpInst::Predicate::ICMP_UGE: 1742 if ((ST->hasSSE41() && MTy.getScalarSizeInBits() == 32) || 1743 (ST->hasSSE2() && MTy.getScalarSizeInBits() < 32)) { 1744 // cmpeq(psubus(x,y),0) 1745 // cmpeq(pminu(x,y),x) 1746 ExtraCost = 1; 1747 } else { 1748 // xor(cmpgt(xor(x,signbit),xor(y,signbit)),-1) 1749 ExtraCost = 3; 1750 } 1751 break; 1752 default: 1753 break; 1754 } 1755 } 1756 } 1757 1758 static const CostTblEntry SLMCostTbl[] = { 1759 // slm pcmpeq/pcmpgt throughput is 2 1760 { ISD::SETCC, MVT::v2i64, 2 }, 1761 }; 1762 1763 static const CostTblEntry AVX512BWCostTbl[] = { 1764 { ISD::SETCC, MVT::v32i16, 1 }, 1765 { ISD::SETCC, MVT::v64i8, 1 }, 1766 1767 { ISD::SELECT, MVT::v32i16, 1 }, 1768 { ISD::SELECT, MVT::v64i8, 1 }, 1769 }; 1770 1771 static const CostTblEntry AVX512CostTbl[] = { 1772 { ISD::SETCC, MVT::v8i64, 1 }, 1773 { ISD::SETCC, MVT::v16i32, 1 }, 1774 { ISD::SETCC, MVT::v8f64, 1 }, 1775 { ISD::SETCC, MVT::v16f32, 1 }, 1776 1777 { ISD::SELECT, MVT::v8i64, 1 }, 1778 { ISD::SELECT, MVT::v16i32, 1 }, 1779 { ISD::SELECT, MVT::v8f64, 1 }, 1780 { ISD::SELECT, MVT::v16f32, 1 }, 1781 }; 1782 1783 static const CostTblEntry AVX2CostTbl[] = { 1784 { ISD::SETCC, MVT::v4i64, 1 }, 1785 { ISD::SETCC, MVT::v8i32, 1 }, 1786 { ISD::SETCC, MVT::v16i16, 1 }, 1787 { ISD::SETCC, MVT::v32i8, 1 }, 1788 1789 { ISD::SELECT, MVT::v4i64, 1 }, // pblendvb 1790 { ISD::SELECT, MVT::v8i32, 1 }, // pblendvb 1791 { ISD::SELECT, MVT::v16i16, 1 }, // pblendvb 1792 { ISD::SELECT, MVT::v32i8, 1 }, // pblendvb 1793 }; 1794 1795 static const CostTblEntry AVX1CostTbl[] = { 1796 { ISD::SETCC, MVT::v4f64, 1 }, 1797 { ISD::SETCC, MVT::v8f32, 1 }, 1798 // AVX1 does not support 8-wide integer compare. 1799 { ISD::SETCC, MVT::v4i64, 4 }, 1800 { ISD::SETCC, MVT::v8i32, 4 }, 1801 { ISD::SETCC, MVT::v16i16, 4 }, 1802 { ISD::SETCC, MVT::v32i8, 4 }, 1803 1804 { ISD::SELECT, MVT::v4f64, 1 }, // vblendvpd 1805 { ISD::SELECT, MVT::v8f32, 1 }, // vblendvps 1806 { ISD::SELECT, MVT::v4i64, 1 }, // vblendvpd 1807 { ISD::SELECT, MVT::v8i32, 1 }, // vblendvps 1808 { ISD::SELECT, MVT::v16i16, 3 }, // vandps + vandnps + vorps 1809 { ISD::SELECT, MVT::v32i8, 3 }, // vandps + vandnps + vorps 1810 }; 1811 1812 static const CostTblEntry SSE42CostTbl[] = { 1813 { ISD::SETCC, MVT::v2f64, 1 }, 1814 { ISD::SETCC, MVT::v4f32, 1 }, 1815 { ISD::SETCC, MVT::v2i64, 1 }, 1816 }; 1817 1818 static const CostTblEntry SSE41CostTbl[] = { 1819 { ISD::SELECT, MVT::v2f64, 1 }, // blendvpd 1820 { ISD::SELECT, MVT::v4f32, 1 }, // blendvps 1821 { ISD::SELECT, MVT::v2i64, 1 }, // pblendvb 1822 { ISD::SELECT, MVT::v4i32, 1 }, // pblendvb 1823 { ISD::SELECT, MVT::v8i16, 1 }, // pblendvb 1824 { ISD::SELECT, MVT::v16i8, 1 }, // pblendvb 1825 }; 1826 1827 static const CostTblEntry SSE2CostTbl[] = { 1828 { ISD::SETCC, MVT::v2f64, 2 }, 1829 { ISD::SETCC, MVT::f64, 1 }, 1830 { ISD::SETCC, MVT::v2i64, 8 }, 1831 { ISD::SETCC, MVT::v4i32, 1 }, 1832 { ISD::SETCC, MVT::v8i16, 1 }, 1833 { ISD::SETCC, MVT::v16i8, 1 }, 1834 1835 { ISD::SELECT, MVT::v2f64, 3 }, // andpd + andnpd + orpd 1836 { ISD::SELECT, MVT::v2i64, 3 }, // pand + pandn + por 1837 { ISD::SELECT, MVT::v4i32, 3 }, // pand + pandn + por 1838 { ISD::SELECT, MVT::v8i16, 3 }, // pand + pandn + por 1839 { ISD::SELECT, MVT::v16i8, 3 }, // pand + pandn + por 1840 }; 1841 1842 static const CostTblEntry SSE1CostTbl[] = { 1843 { ISD::SETCC, MVT::v4f32, 2 }, 1844 { ISD::SETCC, MVT::f32, 1 }, 1845 1846 { ISD::SELECT, MVT::v4f32, 3 }, // andps + andnps + orps 1847 }; 1848 1849 if (ST->isSLM()) 1850 if (const auto *Entry = CostTableLookup(SLMCostTbl, ISD, MTy)) 1851 return LT.first * (ExtraCost + Entry->Cost); 1852 1853 if (ST->hasBWI()) 1854 if (const auto *Entry = CostTableLookup(AVX512BWCostTbl, ISD, MTy)) 1855 return LT.first * (ExtraCost + Entry->Cost); 1856 1857 if (ST->hasAVX512()) 1858 if (const auto *Entry = CostTableLookup(AVX512CostTbl, ISD, MTy)) 1859 return LT.first * (ExtraCost + Entry->Cost); 1860 1861 if (ST->hasAVX2()) 1862 if (const auto *Entry = CostTableLookup(AVX2CostTbl, ISD, MTy)) 1863 return LT.first * (ExtraCost + Entry->Cost); 1864 1865 if (ST->hasAVX()) 1866 if (const auto *Entry = CostTableLookup(AVX1CostTbl, ISD, MTy)) 1867 return LT.first * (ExtraCost + Entry->Cost); 1868 1869 if (ST->hasSSE42()) 1870 if (const auto *Entry = CostTableLookup(SSE42CostTbl, ISD, MTy)) 1871 return LT.first * (ExtraCost + Entry->Cost); 1872 1873 if (ST->hasSSE41()) 1874 if (const auto *Entry = CostTableLookup(SSE41CostTbl, ISD, MTy)) 1875 return LT.first * (ExtraCost + Entry->Cost); 1876 1877 if (ST->hasSSE2()) 1878 if (const auto *Entry = CostTableLookup(SSE2CostTbl, ISD, MTy)) 1879 return LT.first * (ExtraCost + Entry->Cost); 1880 1881 if (ST->hasSSE1()) 1882 if (const auto *Entry = CostTableLookup(SSE1CostTbl, ISD, MTy)) 1883 return LT.first * (ExtraCost + Entry->Cost); 1884 1885 return BaseT::getCmpSelInstrCost(Opcode, ValTy, CondTy, I); 1886 } 1887 1888 unsigned X86TTIImpl::getAtomicMemIntrinsicMaxElementSize() const { return 16; } 1889 1890 int X86TTIImpl::getIntrinsicInstrCost(Intrinsic::ID IID, Type *RetTy, 1891 ArrayRef<Type *> Tys, FastMathFlags FMF, 1892 unsigned ScalarizationCostPassed, 1893 const Instruction *I) { 1894 // Costs should match the codegen from: 1895 // BITREVERSE: llvm\test\CodeGen\X86\vector-bitreverse.ll 1896 // BSWAP: llvm\test\CodeGen\X86\bswap-vector.ll 1897 // CTLZ: llvm\test\CodeGen\X86\vector-lzcnt-*.ll 1898 // CTPOP: llvm\test\CodeGen\X86\vector-popcnt-*.ll 1899 // CTTZ: llvm\test\CodeGen\X86\vector-tzcnt-*.ll 1900 static const CostTblEntry AVX512CDCostTbl[] = { 1901 { ISD::CTLZ, MVT::v8i64, 1 }, 1902 { ISD::CTLZ, MVT::v16i32, 1 }, 1903 { ISD::CTLZ, MVT::v32i16, 8 }, 1904 { ISD::CTLZ, MVT::v64i8, 20 }, 1905 { ISD::CTLZ, MVT::v4i64, 1 }, 1906 { ISD::CTLZ, MVT::v8i32, 1 }, 1907 { ISD::CTLZ, MVT::v16i16, 4 }, 1908 { ISD::CTLZ, MVT::v32i8, 10 }, 1909 { ISD::CTLZ, MVT::v2i64, 1 }, 1910 { ISD::CTLZ, MVT::v4i32, 1 }, 1911 { ISD::CTLZ, MVT::v8i16, 4 }, 1912 { ISD::CTLZ, MVT::v16i8, 4 }, 1913 }; 1914 static const CostTblEntry AVX512BWCostTbl[] = { 1915 { ISD::BITREVERSE, MVT::v8i64, 5 }, 1916 { ISD::BITREVERSE, MVT::v16i32, 5 }, 1917 { ISD::BITREVERSE, MVT::v32i16, 5 }, 1918 { ISD::BITREVERSE, MVT::v64i8, 5 }, 1919 { ISD::CTLZ, MVT::v8i64, 23 }, 1920 { ISD::CTLZ, MVT::v16i32, 22 }, 1921 { ISD::CTLZ, MVT::v32i16, 18 }, 1922 { ISD::CTLZ, MVT::v64i8, 17 }, 1923 { ISD::CTPOP, MVT::v8i64, 7 }, 1924 { ISD::CTPOP, MVT::v16i32, 11 }, 1925 { ISD::CTPOP, MVT::v32i16, 9 }, 1926 { ISD::CTPOP, MVT::v64i8, 6 }, 1927 { ISD::CTTZ, MVT::v8i64, 10 }, 1928 { ISD::CTTZ, MVT::v16i32, 14 }, 1929 { ISD::CTTZ, MVT::v32i16, 12 }, 1930 { ISD::CTTZ, MVT::v64i8, 9 }, 1931 { ISD::SADDSAT, MVT::v32i16, 1 }, 1932 { ISD::SADDSAT, MVT::v64i8, 1 }, 1933 { ISD::SSUBSAT, MVT::v32i16, 1 }, 1934 { ISD::SSUBSAT, MVT::v64i8, 1 }, 1935 { ISD::UADDSAT, MVT::v32i16, 1 }, 1936 { ISD::UADDSAT, MVT::v64i8, 1 }, 1937 { ISD::USUBSAT, MVT::v32i16, 1 }, 1938 { ISD::USUBSAT, MVT::v64i8, 1 }, 1939 }; 1940 static const CostTblEntry AVX512CostTbl[] = { 1941 { ISD::BITREVERSE, MVT::v8i64, 36 }, 1942 { ISD::BITREVERSE, MVT::v16i32, 24 }, 1943 { ISD::CTLZ, MVT::v8i64, 29 }, 1944 { ISD::CTLZ, MVT::v16i32, 35 }, 1945 { ISD::CTPOP, MVT::v8i64, 16 }, 1946 { ISD::CTPOP, MVT::v16i32, 24 }, 1947 { ISD::CTTZ, MVT::v8i64, 20 }, 1948 { ISD::CTTZ, MVT::v16i32, 28 }, 1949 { ISD::USUBSAT, MVT::v16i32, 2 }, // pmaxud + psubd 1950 { ISD::USUBSAT, MVT::v2i64, 2 }, // pmaxuq + psubq 1951 { ISD::USUBSAT, MVT::v4i64, 2 }, // pmaxuq + psubq 1952 { ISD::USUBSAT, MVT::v8i64, 2 }, // pmaxuq + psubq 1953 { ISD::UADDSAT, MVT::v16i32, 3 }, // not + pminud + paddd 1954 { ISD::UADDSAT, MVT::v2i64, 3 }, // not + pminuq + paddq 1955 { ISD::UADDSAT, MVT::v4i64, 3 }, // not + pminuq + paddq 1956 { ISD::UADDSAT, MVT::v8i64, 3 }, // not + pminuq + paddq 1957 { ISD::FMAXNUM, MVT::f32, 2 }, 1958 { ISD::FMAXNUM, MVT::v4f32, 2 }, 1959 { ISD::FMAXNUM, MVT::v8f32, 2 }, 1960 { ISD::FMAXNUM, MVT::v16f32, 2 }, 1961 { ISD::FMAXNUM, MVT::f64, 2 }, 1962 { ISD::FMAXNUM, MVT::v2f64, 2 }, 1963 { ISD::FMAXNUM, MVT::v4f64, 2 }, 1964 { ISD::FMAXNUM, MVT::v8f64, 2 }, 1965 }; 1966 static const CostTblEntry XOPCostTbl[] = { 1967 { ISD::BITREVERSE, MVT::v4i64, 4 }, 1968 { ISD::BITREVERSE, MVT::v8i32, 4 }, 1969 { ISD::BITREVERSE, MVT::v16i16, 4 }, 1970 { ISD::BITREVERSE, MVT::v32i8, 4 }, 1971 { ISD::BITREVERSE, MVT::v2i64, 1 }, 1972 { ISD::BITREVERSE, MVT::v4i32, 1 }, 1973 { ISD::BITREVERSE, MVT::v8i16, 1 }, 1974 { ISD::BITREVERSE, MVT::v16i8, 1 }, 1975 { ISD::BITREVERSE, MVT::i64, 3 }, 1976 { ISD::BITREVERSE, MVT::i32, 3 }, 1977 { ISD::BITREVERSE, MVT::i16, 3 }, 1978 { ISD::BITREVERSE, MVT::i8, 3 } 1979 }; 1980 static const CostTblEntry AVX2CostTbl[] = { 1981 { ISD::BITREVERSE, MVT::v4i64, 5 }, 1982 { ISD::BITREVERSE, MVT::v8i32, 5 }, 1983 { ISD::BITREVERSE, MVT::v16i16, 5 }, 1984 { ISD::BITREVERSE, MVT::v32i8, 5 }, 1985 { ISD::BSWAP, MVT::v4i64, 1 }, 1986 { ISD::BSWAP, MVT::v8i32, 1 }, 1987 { ISD::BSWAP, MVT::v16i16, 1 }, 1988 { ISD::CTLZ, MVT::v4i64, 23 }, 1989 { ISD::CTLZ, MVT::v8i32, 18 }, 1990 { ISD::CTLZ, MVT::v16i16, 14 }, 1991 { ISD::CTLZ, MVT::v32i8, 9 }, 1992 { ISD::CTPOP, MVT::v4i64, 7 }, 1993 { ISD::CTPOP, MVT::v8i32, 11 }, 1994 { ISD::CTPOP, MVT::v16i16, 9 }, 1995 { ISD::CTPOP, MVT::v32i8, 6 }, 1996 { ISD::CTTZ, MVT::v4i64, 10 }, 1997 { ISD::CTTZ, MVT::v8i32, 14 }, 1998 { ISD::CTTZ, MVT::v16i16, 12 }, 1999 { ISD::CTTZ, MVT::v32i8, 9 }, 2000 { ISD::SADDSAT, MVT::v16i16, 1 }, 2001 { ISD::SADDSAT, MVT::v32i8, 1 }, 2002 { ISD::SSUBSAT, MVT::v16i16, 1 }, 2003 { ISD::SSUBSAT, MVT::v32i8, 1 }, 2004 { ISD::UADDSAT, MVT::v16i16, 1 }, 2005 { ISD::UADDSAT, MVT::v32i8, 1 }, 2006 { ISD::UADDSAT, MVT::v8i32, 3 }, // not + pminud + paddd 2007 { ISD::USUBSAT, MVT::v16i16, 1 }, 2008 { ISD::USUBSAT, MVT::v32i8, 1 }, 2009 { ISD::USUBSAT, MVT::v8i32, 2 }, // pmaxud + psubd 2010 { ISD::FSQRT, MVT::f32, 7 }, // Haswell from http://www.agner.org/ 2011 { ISD::FSQRT, MVT::v4f32, 7 }, // Haswell from http://www.agner.org/ 2012 { ISD::FSQRT, MVT::v8f32, 14 }, // Haswell from http://www.agner.org/ 2013 { ISD::FSQRT, MVT::f64, 14 }, // Haswell from http://www.agner.org/ 2014 { ISD::FSQRT, MVT::v2f64, 14 }, // Haswell from http://www.agner.org/ 2015 { ISD::FSQRT, MVT::v4f64, 28 }, // Haswell from http://www.agner.org/ 2016 }; 2017 static const CostTblEntry AVX1CostTbl[] = { 2018 { ISD::BITREVERSE, MVT::v4i64, 12 }, // 2 x 128-bit Op + extract/insert 2019 { ISD::BITREVERSE, MVT::v8i32, 12 }, // 2 x 128-bit Op + extract/insert 2020 { ISD::BITREVERSE, MVT::v16i16, 12 }, // 2 x 128-bit Op + extract/insert 2021 { ISD::BITREVERSE, MVT::v32i8, 12 }, // 2 x 128-bit Op + extract/insert 2022 { ISD::BSWAP, MVT::v4i64, 4 }, 2023 { ISD::BSWAP, MVT::v8i32, 4 }, 2024 { ISD::BSWAP, MVT::v16i16, 4 }, 2025 { ISD::CTLZ, MVT::v4i64, 48 }, // 2 x 128-bit Op + extract/insert 2026 { ISD::CTLZ, MVT::v8i32, 38 }, // 2 x 128-bit Op + extract/insert 2027 { ISD::CTLZ, MVT::v16i16, 30 }, // 2 x 128-bit Op + extract/insert 2028 { ISD::CTLZ, MVT::v32i8, 20 }, // 2 x 128-bit Op + extract/insert 2029 { ISD::CTPOP, MVT::v4i64, 16 }, // 2 x 128-bit Op + extract/insert 2030 { ISD::CTPOP, MVT::v8i32, 24 }, // 2 x 128-bit Op + extract/insert 2031 { ISD::CTPOP, MVT::v16i16, 20 }, // 2 x 128-bit Op + extract/insert 2032 { ISD::CTPOP, MVT::v32i8, 14 }, // 2 x 128-bit Op + extract/insert 2033 { ISD::CTTZ, MVT::v4i64, 22 }, // 2 x 128-bit Op + extract/insert 2034 { ISD::CTTZ, MVT::v8i32, 30 }, // 2 x 128-bit Op + extract/insert 2035 { ISD::CTTZ, MVT::v16i16, 26 }, // 2 x 128-bit Op + extract/insert 2036 { ISD::CTTZ, MVT::v32i8, 20 }, // 2 x 128-bit Op + extract/insert 2037 { ISD::SADDSAT, MVT::v16i16, 4 }, // 2 x 128-bit Op + extract/insert 2038 { ISD::SADDSAT, MVT::v32i8, 4 }, // 2 x 128-bit Op + extract/insert 2039 { ISD::SSUBSAT, MVT::v16i16, 4 }, // 2 x 128-bit Op + extract/insert 2040 { ISD::SSUBSAT, MVT::v32i8, 4 }, // 2 x 128-bit Op + extract/insert 2041 { ISD::UADDSAT, MVT::v16i16, 4 }, // 2 x 128-bit Op + extract/insert 2042 { ISD::UADDSAT, MVT::v32i8, 4 }, // 2 x 128-bit Op + extract/insert 2043 { ISD::UADDSAT, MVT::v8i32, 8 }, // 2 x 128-bit Op + extract/insert 2044 { ISD::USUBSAT, MVT::v16i16, 4 }, // 2 x 128-bit Op + extract/insert 2045 { ISD::USUBSAT, MVT::v32i8, 4 }, // 2 x 128-bit Op + extract/insert 2046 { ISD::USUBSAT, MVT::v8i32, 6 }, // 2 x 128-bit Op + extract/insert 2047 { ISD::FMAXNUM, MVT::f32, 3 }, 2048 { ISD::FMAXNUM, MVT::v4f32, 3 }, 2049 { ISD::FMAXNUM, MVT::v8f32, 5 }, 2050 { ISD::FMAXNUM, MVT::f64, 3 }, 2051 { ISD::FMAXNUM, MVT::v2f64, 3 }, 2052 { ISD::FMAXNUM, MVT::v4f64, 5 }, 2053 { ISD::FSQRT, MVT::f32, 14 }, // SNB from http://www.agner.org/ 2054 { ISD::FSQRT, MVT::v4f32, 14 }, // SNB from http://www.agner.org/ 2055 { ISD::FSQRT, MVT::v8f32, 28 }, // SNB from http://www.agner.org/ 2056 { ISD::FSQRT, MVT::f64, 21 }, // SNB from http://www.agner.org/ 2057 { ISD::FSQRT, MVT::v2f64, 21 }, // SNB from http://www.agner.org/ 2058 { ISD::FSQRT, MVT::v4f64, 43 }, // SNB from http://www.agner.org/ 2059 }; 2060 static const CostTblEntry GLMCostTbl[] = { 2061 { ISD::FSQRT, MVT::f32, 19 }, // sqrtss 2062 { ISD::FSQRT, MVT::v4f32, 37 }, // sqrtps 2063 { ISD::FSQRT, MVT::f64, 34 }, // sqrtsd 2064 { ISD::FSQRT, MVT::v2f64, 67 }, // sqrtpd 2065 }; 2066 static const CostTblEntry SLMCostTbl[] = { 2067 { ISD::FSQRT, MVT::f32, 20 }, // sqrtss 2068 { ISD::FSQRT, MVT::v4f32, 40 }, // sqrtps 2069 { ISD::FSQRT, MVT::f64, 35 }, // sqrtsd 2070 { ISD::FSQRT, MVT::v2f64, 70 }, // sqrtpd 2071 }; 2072 static const CostTblEntry SSE42CostTbl[] = { 2073 { ISD::USUBSAT, MVT::v4i32, 2 }, // pmaxud + psubd 2074 { ISD::UADDSAT, MVT::v4i32, 3 }, // not + pminud + paddd 2075 { ISD::FSQRT, MVT::f32, 18 }, // Nehalem from http://www.agner.org/ 2076 { ISD::FSQRT, MVT::v4f32, 18 }, // Nehalem from http://www.agner.org/ 2077 }; 2078 static const CostTblEntry SSSE3CostTbl[] = { 2079 { ISD::BITREVERSE, MVT::v2i64, 5 }, 2080 { ISD::BITREVERSE, MVT::v4i32, 5 }, 2081 { ISD::BITREVERSE, MVT::v8i16, 5 }, 2082 { ISD::BITREVERSE, MVT::v16i8, 5 }, 2083 { ISD::BSWAP, MVT::v2i64, 1 }, 2084 { ISD::BSWAP, MVT::v4i32, 1 }, 2085 { ISD::BSWAP, MVT::v8i16, 1 }, 2086 { ISD::CTLZ, MVT::v2i64, 23 }, 2087 { ISD::CTLZ, MVT::v4i32, 18 }, 2088 { ISD::CTLZ, MVT::v8i16, 14 }, 2089 { ISD::CTLZ, MVT::v16i8, 9 }, 2090 { ISD::CTPOP, MVT::v2i64, 7 }, 2091 { ISD::CTPOP, MVT::v4i32, 11 }, 2092 { ISD::CTPOP, MVT::v8i16, 9 }, 2093 { ISD::CTPOP, MVT::v16i8, 6 }, 2094 { ISD::CTTZ, MVT::v2i64, 10 }, 2095 { ISD::CTTZ, MVT::v4i32, 14 }, 2096 { ISD::CTTZ, MVT::v8i16, 12 }, 2097 { ISD::CTTZ, MVT::v16i8, 9 } 2098 }; 2099 static const CostTblEntry SSE2CostTbl[] = { 2100 { ISD::BITREVERSE, MVT::v2i64, 29 }, 2101 { ISD::BITREVERSE, MVT::v4i32, 27 }, 2102 { ISD::BITREVERSE, MVT::v8i16, 27 }, 2103 { ISD::BITREVERSE, MVT::v16i8, 20 }, 2104 { ISD::BSWAP, MVT::v2i64, 7 }, 2105 { ISD::BSWAP, MVT::v4i32, 7 }, 2106 { ISD::BSWAP, MVT::v8i16, 7 }, 2107 { ISD::CTLZ, MVT::v2i64, 25 }, 2108 { ISD::CTLZ, MVT::v4i32, 26 }, 2109 { ISD::CTLZ, MVT::v8i16, 20 }, 2110 { ISD::CTLZ, MVT::v16i8, 17 }, 2111 { ISD::CTPOP, MVT::v2i64, 12 }, 2112 { ISD::CTPOP, MVT::v4i32, 15 }, 2113 { ISD::CTPOP, MVT::v8i16, 13 }, 2114 { ISD::CTPOP, MVT::v16i8, 10 }, 2115 { ISD::CTTZ, MVT::v2i64, 14 }, 2116 { ISD::CTTZ, MVT::v4i32, 18 }, 2117 { ISD::CTTZ, MVT::v8i16, 16 }, 2118 { ISD::CTTZ, MVT::v16i8, 13 }, 2119 { ISD::SADDSAT, MVT::v8i16, 1 }, 2120 { ISD::SADDSAT, MVT::v16i8, 1 }, 2121 { ISD::SSUBSAT, MVT::v8i16, 1 }, 2122 { ISD::SSUBSAT, MVT::v16i8, 1 }, 2123 { ISD::UADDSAT, MVT::v8i16, 1 }, 2124 { ISD::UADDSAT, MVT::v16i8, 1 }, 2125 { ISD::USUBSAT, MVT::v8i16, 1 }, 2126 { ISD::USUBSAT, MVT::v16i8, 1 }, 2127 { ISD::FMAXNUM, MVT::f64, 4 }, 2128 { ISD::FMAXNUM, MVT::v2f64, 4 }, 2129 { ISD::FSQRT, MVT::f64, 32 }, // Nehalem from http://www.agner.org/ 2130 { ISD::FSQRT, MVT::v2f64, 32 }, // Nehalem from http://www.agner.org/ 2131 }; 2132 static const CostTblEntry SSE1CostTbl[] = { 2133 { ISD::FMAXNUM, MVT::f32, 4 }, 2134 { ISD::FMAXNUM, MVT::v4f32, 4 }, 2135 { ISD::FSQRT, MVT::f32, 28 }, // Pentium III from http://www.agner.org/ 2136 { ISD::FSQRT, MVT::v4f32, 56 }, // Pentium III from http://www.agner.org/ 2137 }; 2138 static const CostTblEntry BMI64CostTbl[] = { // 64-bit targets 2139 { ISD::CTTZ, MVT::i64, 1 }, 2140 }; 2141 static const CostTblEntry BMI32CostTbl[] = { // 32 or 64-bit targets 2142 { ISD::CTTZ, MVT::i32, 1 }, 2143 { ISD::CTTZ, MVT::i16, 1 }, 2144 { ISD::CTTZ, MVT::i8, 1 }, 2145 }; 2146 static const CostTblEntry LZCNT64CostTbl[] = { // 64-bit targets 2147 { ISD::CTLZ, MVT::i64, 1 }, 2148 }; 2149 static const CostTblEntry LZCNT32CostTbl[] = { // 32 or 64-bit targets 2150 { ISD::CTLZ, MVT::i32, 1 }, 2151 { ISD::CTLZ, MVT::i16, 1 }, 2152 { ISD::CTLZ, MVT::i8, 1 }, 2153 }; 2154 static const CostTblEntry POPCNT64CostTbl[] = { // 64-bit targets 2155 { ISD::CTPOP, MVT::i64, 1 }, 2156 }; 2157 static const CostTblEntry POPCNT32CostTbl[] = { // 32 or 64-bit targets 2158 { ISD::CTPOP, MVT::i32, 1 }, 2159 { ISD::CTPOP, MVT::i16, 1 }, 2160 { ISD::CTPOP, MVT::i8, 1 }, 2161 }; 2162 static const CostTblEntry X64CostTbl[] = { // 64-bit targets 2163 { ISD::BITREVERSE, MVT::i64, 14 }, 2164 { ISD::CTLZ, MVT::i64, 4 }, // BSR+XOR or BSR+XOR+CMOV 2165 { ISD::CTTZ, MVT::i64, 3 }, // TEST+BSF+CMOV/BRANCH 2166 { ISD::CTPOP, MVT::i64, 10 }, 2167 { ISD::SADDO, MVT::i64, 1 }, 2168 { ISD::UADDO, MVT::i64, 1 }, 2169 }; 2170 static const CostTblEntry X86CostTbl[] = { // 32 or 64-bit targets 2171 { ISD::BITREVERSE, MVT::i32, 14 }, 2172 { ISD::BITREVERSE, MVT::i16, 14 }, 2173 { ISD::BITREVERSE, MVT::i8, 11 }, 2174 { ISD::CTLZ, MVT::i32, 4 }, // BSR+XOR or BSR+XOR+CMOV 2175 { ISD::CTLZ, MVT::i16, 4 }, // BSR+XOR or BSR+XOR+CMOV 2176 { ISD::CTLZ, MVT::i8, 4 }, // BSR+XOR or BSR+XOR+CMOV 2177 { ISD::CTTZ, MVT::i32, 3 }, // TEST+BSF+CMOV/BRANCH 2178 { ISD::CTTZ, MVT::i16, 3 }, // TEST+BSF+CMOV/BRANCH 2179 { ISD::CTTZ, MVT::i8, 3 }, // TEST+BSF+CMOV/BRANCH 2180 { ISD::CTPOP, MVT::i32, 8 }, 2181 { ISD::CTPOP, MVT::i16, 9 }, 2182 { ISD::CTPOP, MVT::i8, 7 }, 2183 { ISD::SADDO, MVT::i32, 1 }, 2184 { ISD::SADDO, MVT::i16, 1 }, 2185 { ISD::SADDO, MVT::i8, 1 }, 2186 { ISD::UADDO, MVT::i32, 1 }, 2187 { ISD::UADDO, MVT::i16, 1 }, 2188 { ISD::UADDO, MVT::i8, 1 }, 2189 }; 2190 2191 Type *OpTy = RetTy; 2192 unsigned ISD = ISD::DELETED_NODE; 2193 switch (IID) { 2194 default: 2195 break; 2196 case Intrinsic::bitreverse: 2197 ISD = ISD::BITREVERSE; 2198 break; 2199 case Intrinsic::bswap: 2200 ISD = ISD::BSWAP; 2201 break; 2202 case Intrinsic::ctlz: 2203 ISD = ISD::CTLZ; 2204 break; 2205 case Intrinsic::ctpop: 2206 ISD = ISD::CTPOP; 2207 break; 2208 case Intrinsic::cttz: 2209 ISD = ISD::CTTZ; 2210 break; 2211 case Intrinsic::maxnum: 2212 case Intrinsic::minnum: 2213 // FMINNUM has same costs so don't duplicate. 2214 ISD = ISD::FMAXNUM; 2215 break; 2216 case Intrinsic::sadd_sat: 2217 ISD = ISD::SADDSAT; 2218 break; 2219 case Intrinsic::ssub_sat: 2220 ISD = ISD::SSUBSAT; 2221 break; 2222 case Intrinsic::uadd_sat: 2223 ISD = ISD::UADDSAT; 2224 break; 2225 case Intrinsic::usub_sat: 2226 ISD = ISD::USUBSAT; 2227 break; 2228 case Intrinsic::sqrt: 2229 ISD = ISD::FSQRT; 2230 break; 2231 case Intrinsic::sadd_with_overflow: 2232 case Intrinsic::ssub_with_overflow: 2233 // SSUBO has same costs so don't duplicate. 2234 ISD = ISD::SADDO; 2235 OpTy = RetTy->getContainedType(0); 2236 break; 2237 case Intrinsic::uadd_with_overflow: 2238 case Intrinsic::usub_with_overflow: 2239 // USUBO has same costs so don't duplicate. 2240 ISD = ISD::UADDO; 2241 OpTy = RetTy->getContainedType(0); 2242 break; 2243 } 2244 2245 if (ISD != ISD::DELETED_NODE) { 2246 // Legalize the type. 2247 std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, OpTy); 2248 MVT MTy = LT.second; 2249 2250 // Attempt to lookup cost. 2251 if (ST->useGLMDivSqrtCosts()) 2252 if (const auto *Entry = CostTableLookup(GLMCostTbl, ISD, MTy)) 2253 return LT.first * Entry->Cost; 2254 2255 if (ST->isSLM()) 2256 if (const auto *Entry = CostTableLookup(SLMCostTbl, ISD, MTy)) 2257 return LT.first * Entry->Cost; 2258 2259 if (ST->hasCDI()) 2260 if (const auto *Entry = CostTableLookup(AVX512CDCostTbl, ISD, MTy)) 2261 return LT.first * Entry->Cost; 2262 2263 if (ST->hasBWI()) 2264 if (const auto *Entry = CostTableLookup(AVX512BWCostTbl, ISD, MTy)) 2265 return LT.first * Entry->Cost; 2266 2267 if (ST->hasAVX512()) 2268 if (const auto *Entry = CostTableLookup(AVX512CostTbl, ISD, MTy)) 2269 return LT.first * Entry->Cost; 2270 2271 if (ST->hasXOP()) 2272 if (const auto *Entry = CostTableLookup(XOPCostTbl, ISD, MTy)) 2273 return LT.first * Entry->Cost; 2274 2275 if (ST->hasAVX2()) 2276 if (const auto *Entry = CostTableLookup(AVX2CostTbl, ISD, MTy)) 2277 return LT.first * Entry->Cost; 2278 2279 if (ST->hasAVX()) 2280 if (const auto *Entry = CostTableLookup(AVX1CostTbl, ISD, MTy)) 2281 return LT.first * Entry->Cost; 2282 2283 if (ST->hasSSE42()) 2284 if (const auto *Entry = CostTableLookup(SSE42CostTbl, ISD, MTy)) 2285 return LT.first * Entry->Cost; 2286 2287 if (ST->hasSSSE3()) 2288 if (const auto *Entry = CostTableLookup(SSSE3CostTbl, ISD, MTy)) 2289 return LT.first * Entry->Cost; 2290 2291 if (ST->hasSSE2()) 2292 if (const auto *Entry = CostTableLookup(SSE2CostTbl, ISD, MTy)) 2293 return LT.first * Entry->Cost; 2294 2295 if (ST->hasSSE1()) 2296 if (const auto *Entry = CostTableLookup(SSE1CostTbl, ISD, MTy)) 2297 return LT.first * Entry->Cost; 2298 2299 if (ST->hasBMI()) { 2300 if (ST->is64Bit()) 2301 if (const auto *Entry = CostTableLookup(BMI64CostTbl, ISD, MTy)) 2302 return LT.first * Entry->Cost; 2303 2304 if (const auto *Entry = CostTableLookup(BMI32CostTbl, ISD, MTy)) 2305 return LT.first * Entry->Cost; 2306 } 2307 2308 if (ST->hasLZCNT()) { 2309 if (ST->is64Bit()) 2310 if (const auto *Entry = CostTableLookup(LZCNT64CostTbl, ISD, MTy)) 2311 return LT.first * Entry->Cost; 2312 2313 if (const auto *Entry = CostTableLookup(LZCNT32CostTbl, ISD, MTy)) 2314 return LT.first * Entry->Cost; 2315 } 2316 2317 if (ST->hasPOPCNT()) { 2318 if (ST->is64Bit()) 2319 if (const auto *Entry = CostTableLookup(POPCNT64CostTbl, ISD, MTy)) 2320 return LT.first * Entry->Cost; 2321 2322 if (const auto *Entry = CostTableLookup(POPCNT32CostTbl, ISD, MTy)) 2323 return LT.first * Entry->Cost; 2324 } 2325 2326 // TODO - add BMI (TZCNT) scalar handling 2327 2328 if (ST->is64Bit()) 2329 if (const auto *Entry = CostTableLookup(X64CostTbl, ISD, MTy)) 2330 return LT.first * Entry->Cost; 2331 2332 if (const auto *Entry = CostTableLookup(X86CostTbl, ISD, MTy)) 2333 return LT.first * Entry->Cost; 2334 } 2335 2336 return BaseT::getIntrinsicInstrCost(IID, RetTy, Tys, FMF, 2337 ScalarizationCostPassed, I); 2338 } 2339 2340 int X86TTIImpl::getIntrinsicInstrCost(Intrinsic::ID IID, Type *RetTy, 2341 ArrayRef<Value *> Args, FastMathFlags FMF, 2342 unsigned VF, const Instruction *I) { 2343 static const CostTblEntry AVX512CostTbl[] = { 2344 { ISD::ROTL, MVT::v8i64, 1 }, 2345 { ISD::ROTL, MVT::v4i64, 1 }, 2346 { ISD::ROTL, MVT::v2i64, 1 }, 2347 { ISD::ROTL, MVT::v16i32, 1 }, 2348 { ISD::ROTL, MVT::v8i32, 1 }, 2349 { ISD::ROTL, MVT::v4i32, 1 }, 2350 { ISD::ROTR, MVT::v8i64, 1 }, 2351 { ISD::ROTR, MVT::v4i64, 1 }, 2352 { ISD::ROTR, MVT::v2i64, 1 }, 2353 { ISD::ROTR, MVT::v16i32, 1 }, 2354 { ISD::ROTR, MVT::v8i32, 1 }, 2355 { ISD::ROTR, MVT::v4i32, 1 } 2356 }; 2357 // XOP: ROTL = VPROT(X,Y), ROTR = VPROT(X,SUB(0,Y)) 2358 static const CostTblEntry XOPCostTbl[] = { 2359 { ISD::ROTL, MVT::v4i64, 4 }, 2360 { ISD::ROTL, MVT::v8i32, 4 }, 2361 { ISD::ROTL, MVT::v16i16, 4 }, 2362 { ISD::ROTL, MVT::v32i8, 4 }, 2363 { ISD::ROTL, MVT::v2i64, 1 }, 2364 { ISD::ROTL, MVT::v4i32, 1 }, 2365 { ISD::ROTL, MVT::v8i16, 1 }, 2366 { ISD::ROTL, MVT::v16i8, 1 }, 2367 { ISD::ROTR, MVT::v4i64, 6 }, 2368 { ISD::ROTR, MVT::v8i32, 6 }, 2369 { ISD::ROTR, MVT::v16i16, 6 }, 2370 { ISD::ROTR, MVT::v32i8, 6 }, 2371 { ISD::ROTR, MVT::v2i64, 2 }, 2372 { ISD::ROTR, MVT::v4i32, 2 }, 2373 { ISD::ROTR, MVT::v8i16, 2 }, 2374 { ISD::ROTR, MVT::v16i8, 2 } 2375 }; 2376 static const CostTblEntry X64CostTbl[] = { // 64-bit targets 2377 { ISD::ROTL, MVT::i64, 1 }, 2378 { ISD::ROTR, MVT::i64, 1 }, 2379 { ISD::FSHL, MVT::i64, 4 } 2380 }; 2381 static const CostTblEntry X86CostTbl[] = { // 32 or 64-bit targets 2382 { ISD::ROTL, MVT::i32, 1 }, 2383 { ISD::ROTL, MVT::i16, 1 }, 2384 { ISD::ROTL, MVT::i8, 1 }, 2385 { ISD::ROTR, MVT::i32, 1 }, 2386 { ISD::ROTR, MVT::i16, 1 }, 2387 { ISD::ROTR, MVT::i8, 1 }, 2388 { ISD::FSHL, MVT::i32, 4 }, 2389 { ISD::FSHL, MVT::i16, 4 }, 2390 { ISD::FSHL, MVT::i8, 4 } 2391 }; 2392 2393 unsigned ISD = ISD::DELETED_NODE; 2394 switch (IID) { 2395 default: 2396 break; 2397 case Intrinsic::fshl: 2398 ISD = ISD::FSHL; 2399 if (Args[0] == Args[1]) 2400 ISD = ISD::ROTL; 2401 break; 2402 case Intrinsic::fshr: 2403 // FSHR has same costs so don't duplicate. 2404 ISD = ISD::FSHL; 2405 if (Args[0] == Args[1]) 2406 ISD = ISD::ROTR; 2407 break; 2408 } 2409 2410 if (ISD != ISD::DELETED_NODE) { 2411 // Legalize the type. 2412 std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, RetTy); 2413 MVT MTy = LT.second; 2414 2415 // Attempt to lookup cost. 2416 if (ST->hasAVX512()) 2417 if (const auto *Entry = CostTableLookup(AVX512CostTbl, ISD, MTy)) 2418 return LT.first * Entry->Cost; 2419 2420 if (ST->hasXOP()) 2421 if (const auto *Entry = CostTableLookup(XOPCostTbl, ISD, MTy)) 2422 return LT.first * Entry->Cost; 2423 2424 if (ST->is64Bit()) 2425 if (const auto *Entry = CostTableLookup(X64CostTbl, ISD, MTy)) 2426 return LT.first * Entry->Cost; 2427 2428 if (const auto *Entry = CostTableLookup(X86CostTbl, ISD, MTy)) 2429 return LT.first * Entry->Cost; 2430 } 2431 2432 return BaseT::getIntrinsicInstrCost(IID, RetTy, Args, FMF, VF, I); 2433 } 2434 2435 int X86TTIImpl::getVectorInstrCost(unsigned Opcode, Type *Val, unsigned Index) { 2436 static const CostTblEntry SLMCostTbl[] = { 2437 { ISD::EXTRACT_VECTOR_ELT, MVT::i8, 4 }, 2438 { ISD::EXTRACT_VECTOR_ELT, MVT::i16, 4 }, 2439 { ISD::EXTRACT_VECTOR_ELT, MVT::i32, 4 }, 2440 { ISD::EXTRACT_VECTOR_ELT, MVT::i64, 7 } 2441 }; 2442 2443 assert(Val->isVectorTy() && "This must be a vector type"); 2444 Type *ScalarType = Val->getScalarType(); 2445 int RegisterFileMoveCost = 0; 2446 2447 if (Index != -1U && (Opcode == Instruction::ExtractElement || 2448 Opcode == Instruction::InsertElement)) { 2449 // Legalize the type. 2450 std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, Val); 2451 2452 // This type is legalized to a scalar type. 2453 if (!LT.second.isVector()) 2454 return 0; 2455 2456 // The type may be split. Normalize the index to the new type. 2457 unsigned NumElts = LT.second.getVectorNumElements(); 2458 unsigned SubNumElts = NumElts; 2459 Index = Index % NumElts; 2460 2461 // For >128-bit vectors, we need to extract higher 128-bit subvectors. 2462 // For inserts, we also need to insert the subvector back. 2463 if (LT.second.getSizeInBits() > 128) { 2464 assert((LT.second.getSizeInBits() % 128) == 0 && "Illegal vector"); 2465 unsigned NumSubVecs = LT.second.getSizeInBits() / 128; 2466 SubNumElts = NumElts / NumSubVecs; 2467 if (SubNumElts <= Index) { 2468 RegisterFileMoveCost += (Opcode == Instruction::InsertElement ? 2 : 1); 2469 Index %= SubNumElts; 2470 } 2471 } 2472 2473 if (Index == 0) { 2474 // Floating point scalars are already located in index #0. 2475 // Many insertions to #0 can fold away for scalar fp-ops, so let's assume 2476 // true for all. 2477 if (ScalarType->isFloatingPointTy()) 2478 return RegisterFileMoveCost; 2479 2480 // Assume movd/movq XMM -> GPR is relatively cheap on all targets. 2481 if (ScalarType->isIntegerTy() && Opcode == Instruction::ExtractElement) 2482 return 1 + RegisterFileMoveCost; 2483 } 2484 2485 int ISD = TLI->InstructionOpcodeToISD(Opcode); 2486 assert(ISD && "Unexpected vector opcode"); 2487 MVT MScalarTy = LT.second.getScalarType(); 2488 if (ST->isSLM()) 2489 if (auto *Entry = CostTableLookup(SLMCostTbl, ISD, MScalarTy)) 2490 return Entry->Cost + RegisterFileMoveCost; 2491 2492 // Assume pinsr/pextr XMM <-> GPR is relatively cheap on all targets. 2493 if ((MScalarTy == MVT::i16 && ST->hasSSE2()) || 2494 (MScalarTy.isInteger() && ST->hasSSE41())) 2495 return 1 + RegisterFileMoveCost; 2496 2497 // Assume insertps is relatively cheap on all targets. 2498 if (MScalarTy == MVT::f32 && ST->hasSSE41() && 2499 Opcode == Instruction::InsertElement) 2500 return 1 + RegisterFileMoveCost; 2501 2502 // For extractions we just need to shuffle the element to index 0, which 2503 // should be very cheap (assume cost = 1). For insertions we need to shuffle 2504 // the elements to its destination. In both cases we must handle the 2505 // subvector move(s). 2506 // TODO: Under what circumstances should we shuffle using the full width? 2507 int ShuffleCost = 1; 2508 if (Opcode == Instruction::InsertElement) { 2509 Type *SubTy = VectorType::get(Val->getVectorElementType(), SubNumElts); 2510 ShuffleCost = getShuffleCost(TTI::SK_PermuteTwoSrc, SubTy, 0, SubTy); 2511 } 2512 int IntOrFpCost = ScalarType->isFloatingPointTy() ? 0 : 1; 2513 return ShuffleCost + IntOrFpCost + RegisterFileMoveCost; 2514 } 2515 2516 // Add to the base cost if we know that the extracted element of a vector is 2517 // destined to be moved to and used in the integer register file. 2518 if (Opcode == Instruction::ExtractElement && ScalarType->isPointerTy()) 2519 RegisterFileMoveCost += 1; 2520 2521 return BaseT::getVectorInstrCost(Opcode, Val, Index) + RegisterFileMoveCost; 2522 } 2523 2524 int X86TTIImpl::getMemoryOpCost(unsigned Opcode, Type *Src, 2525 MaybeAlign Alignment, unsigned AddressSpace, 2526 const Instruction *I) { 2527 // Handle non-power-of-two vectors such as <3 x float> 2528 if (VectorType *VTy = dyn_cast<VectorType>(Src)) { 2529 unsigned NumElem = VTy->getVectorNumElements(); 2530 2531 // Handle a few common cases: 2532 // <3 x float> 2533 if (NumElem == 3 && VTy->getScalarSizeInBits() == 32) 2534 // Cost = 64 bit store + extract + 32 bit store. 2535 return 3; 2536 2537 // <3 x double> 2538 if (NumElem == 3 && VTy->getScalarSizeInBits() == 64) 2539 // Cost = 128 bit store + unpack + 64 bit store. 2540 return 3; 2541 2542 // Assume that all other non-power-of-two numbers are scalarized. 2543 if (!isPowerOf2_32(NumElem)) { 2544 int Cost = BaseT::getMemoryOpCost(Opcode, VTy->getScalarType(), Alignment, 2545 AddressSpace); 2546 int SplitCost = getScalarizationOverhead(Src, Opcode == Instruction::Load, 2547 Opcode == Instruction::Store); 2548 return NumElem * Cost + SplitCost; 2549 } 2550 } 2551 2552 // Legalize the type. 2553 std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, Src); 2554 assert((Opcode == Instruction::Load || Opcode == Instruction::Store) && 2555 "Invalid Opcode"); 2556 2557 // Each load/store unit costs 1. 2558 int Cost = LT.first * 1; 2559 2560 // This isn't exactly right. We're using slow unaligned 32-byte accesses as a 2561 // proxy for a double-pumped AVX memory interface such as on Sandybridge. 2562 if (LT.second.getStoreSize() == 32 && ST->isUnalignedMem32Slow()) 2563 Cost *= 2; 2564 2565 return Cost; 2566 } 2567 2568 int X86TTIImpl::getMaskedMemoryOpCost(unsigned Opcode, Type *SrcTy, 2569 unsigned Alignment, 2570 unsigned AddressSpace) { 2571 bool IsLoad = (Instruction::Load == Opcode); 2572 bool IsStore = (Instruction::Store == Opcode); 2573 2574 VectorType *SrcVTy = dyn_cast<VectorType>(SrcTy); 2575 if (!SrcVTy) 2576 // To calculate scalar take the regular cost, without mask 2577 return getMemoryOpCost(Opcode, SrcTy, MaybeAlign(Alignment), AddressSpace); 2578 2579 unsigned NumElem = SrcVTy->getVectorNumElements(); 2580 VectorType *MaskTy = 2581 VectorType::get(Type::getInt8Ty(SrcVTy->getContext()), NumElem); 2582 if ((IsLoad && !isLegalMaskedLoad(SrcVTy, MaybeAlign(Alignment))) || 2583 (IsStore && !isLegalMaskedStore(SrcVTy, MaybeAlign(Alignment))) || 2584 !isPowerOf2_32(NumElem)) { 2585 // Scalarization 2586 int MaskSplitCost = getScalarizationOverhead(MaskTy, false, true); 2587 int ScalarCompareCost = getCmpSelInstrCost( 2588 Instruction::ICmp, Type::getInt8Ty(SrcVTy->getContext()), nullptr); 2589 int BranchCost = getCFInstrCost(Instruction::Br); 2590 int MaskCmpCost = NumElem * (BranchCost + ScalarCompareCost); 2591 2592 int ValueSplitCost = getScalarizationOverhead(SrcVTy, IsLoad, IsStore); 2593 int MemopCost = 2594 NumElem * BaseT::getMemoryOpCost(Opcode, SrcVTy->getScalarType(), 2595 MaybeAlign(Alignment), AddressSpace); 2596 return MemopCost + ValueSplitCost + MaskSplitCost + MaskCmpCost; 2597 } 2598 2599 // Legalize the type. 2600 std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, SrcVTy); 2601 auto VT = TLI->getValueType(DL, SrcVTy); 2602 int Cost = 0; 2603 if (VT.isSimple() && LT.second != VT.getSimpleVT() && 2604 LT.second.getVectorNumElements() == NumElem) 2605 // Promotion requires expand/truncate for data and a shuffle for mask. 2606 Cost += getShuffleCost(TTI::SK_PermuteTwoSrc, SrcVTy, 0, nullptr) + 2607 getShuffleCost(TTI::SK_PermuteTwoSrc, MaskTy, 0, nullptr); 2608 2609 else if (LT.second.getVectorNumElements() > NumElem) { 2610 VectorType *NewMaskTy = VectorType::get(MaskTy->getVectorElementType(), 2611 LT.second.getVectorNumElements()); 2612 // Expanding requires fill mask with zeroes 2613 Cost += getShuffleCost(TTI::SK_InsertSubvector, NewMaskTy, 0, MaskTy); 2614 } 2615 2616 // Pre-AVX512 - each maskmov load costs 2 + store costs ~8. 2617 if (!ST->hasAVX512()) 2618 return Cost + LT.first * (IsLoad ? 2 : 8); 2619 2620 // AVX-512 masked load/store is cheapper 2621 return Cost + LT.first; 2622 } 2623 2624 int X86TTIImpl::getAddressComputationCost(Type *Ty, ScalarEvolution *SE, 2625 const SCEV *Ptr) { 2626 // Address computations in vectorized code with non-consecutive addresses will 2627 // likely result in more instructions compared to scalar code where the 2628 // computation can more often be merged into the index mode. The resulting 2629 // extra micro-ops can significantly decrease throughput. 2630 const unsigned NumVectorInstToHideOverhead = 10; 2631 2632 // Cost modeling of Strided Access Computation is hidden by the indexing 2633 // modes of X86 regardless of the stride value. We dont believe that there 2634 // is a difference between constant strided access in gerenal and constant 2635 // strided value which is less than or equal to 64. 2636 // Even in the case of (loop invariant) stride whose value is not known at 2637 // compile time, the address computation will not incur more than one extra 2638 // ADD instruction. 2639 if (Ty->isVectorTy() && SE) { 2640 if (!BaseT::isStridedAccess(Ptr)) 2641 return NumVectorInstToHideOverhead; 2642 if (!BaseT::getConstantStrideStep(SE, Ptr)) 2643 return 1; 2644 } 2645 2646 return BaseT::getAddressComputationCost(Ty, SE, Ptr); 2647 } 2648 2649 int X86TTIImpl::getArithmeticReductionCost(unsigned Opcode, Type *ValTy, 2650 bool IsPairwise) { 2651 // Just use the default implementation for pair reductions. 2652 if (IsPairwise) 2653 return BaseT::getArithmeticReductionCost(Opcode, ValTy, IsPairwise); 2654 2655 // We use the Intel Architecture Code Analyzer(IACA) to measure the throughput 2656 // and make it as the cost. 2657 2658 static const CostTblEntry SLMCostTblNoPairWise[] = { 2659 { ISD::FADD, MVT::v2f64, 3 }, 2660 { ISD::ADD, MVT::v2i64, 5 }, 2661 }; 2662 2663 static const CostTblEntry SSE2CostTblNoPairWise[] = { 2664 { ISD::FADD, MVT::v2f64, 2 }, 2665 { ISD::FADD, MVT::v4f32, 4 }, 2666 { ISD::ADD, MVT::v2i64, 2 }, // The data reported by the IACA tool is "1.6". 2667 { ISD::ADD, MVT::v2i32, 2 }, // FIXME: chosen to be less than v4i32 2668 { ISD::ADD, MVT::v4i32, 3 }, // The data reported by the IACA tool is "3.3". 2669 { ISD::ADD, MVT::v2i16, 2 }, // The data reported by the IACA tool is "4.3". 2670 { ISD::ADD, MVT::v4i16, 3 }, // The data reported by the IACA tool is "4.3". 2671 { ISD::ADD, MVT::v8i16, 4 }, // The data reported by the IACA tool is "4.3". 2672 { ISD::ADD, MVT::v2i8, 2 }, 2673 { ISD::ADD, MVT::v4i8, 2 }, 2674 { ISD::ADD, MVT::v8i8, 2 }, 2675 { ISD::ADD, MVT::v16i8, 3 }, 2676 }; 2677 2678 static const CostTblEntry AVX1CostTblNoPairWise[] = { 2679 { ISD::FADD, MVT::v4f64, 3 }, 2680 { ISD::FADD, MVT::v4f32, 3 }, 2681 { ISD::FADD, MVT::v8f32, 4 }, 2682 { ISD::ADD, MVT::v2i64, 1 }, // The data reported by the IACA tool is "1.5". 2683 { ISD::ADD, MVT::v4i64, 3 }, 2684 { ISD::ADD, MVT::v8i32, 5 }, 2685 { ISD::ADD, MVT::v16i16, 5 }, 2686 { ISD::ADD, MVT::v32i8, 4 }, 2687 }; 2688 2689 int ISD = TLI->InstructionOpcodeToISD(Opcode); 2690 assert(ISD && "Invalid opcode"); 2691 2692 // Before legalizing the type, give a chance to look up illegal narrow types 2693 // in the table. 2694 // FIXME: Is there a better way to do this? 2695 EVT VT = TLI->getValueType(DL, ValTy); 2696 if (VT.isSimple()) { 2697 MVT MTy = VT.getSimpleVT(); 2698 if (ST->isSLM()) 2699 if (const auto *Entry = CostTableLookup(SLMCostTblNoPairWise, ISD, MTy)) 2700 return Entry->Cost; 2701 2702 if (ST->hasAVX()) 2703 if (const auto *Entry = CostTableLookup(AVX1CostTblNoPairWise, ISD, MTy)) 2704 return Entry->Cost; 2705 2706 if (ST->hasSSE2()) 2707 if (const auto *Entry = CostTableLookup(SSE2CostTblNoPairWise, ISD, MTy)) 2708 return Entry->Cost; 2709 } 2710 2711 std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, ValTy); 2712 2713 MVT MTy = LT.second; 2714 2715 unsigned ArithmeticCost = 0; 2716 if (LT.first != 1 && MTy.isVector() && 2717 MTy.getVectorNumElements() < ValTy->getVectorNumElements()) { 2718 // Type needs to be split. We need LT.first - 1 arithmetic ops. 2719 Type *SingleOpTy = VectorType::get(ValTy->getVectorElementType(), 2720 MTy.getVectorNumElements()); 2721 ArithmeticCost = getArithmeticInstrCost(Opcode, SingleOpTy); 2722 ArithmeticCost *= LT.first - 1; 2723 } 2724 2725 if (ST->isSLM()) 2726 if (const auto *Entry = CostTableLookup(SLMCostTblNoPairWise, ISD, MTy)) 2727 return ArithmeticCost + Entry->Cost; 2728 2729 if (ST->hasAVX()) 2730 if (const auto *Entry = CostTableLookup(AVX1CostTblNoPairWise, ISD, MTy)) 2731 return ArithmeticCost + Entry->Cost; 2732 2733 if (ST->hasSSE2()) 2734 if (const auto *Entry = CostTableLookup(SSE2CostTblNoPairWise, ISD, MTy)) 2735 return ArithmeticCost + Entry->Cost; 2736 2737 // FIXME: These assume a naive kshift+binop lowering, which is probably 2738 // conservative in most cases. 2739 static const CostTblEntry AVX512BoolReduction[] = { 2740 { ISD::AND, MVT::v2i1, 3 }, 2741 { ISD::AND, MVT::v4i1, 5 }, 2742 { ISD::AND, MVT::v8i1, 7 }, 2743 { ISD::AND, MVT::v16i1, 9 }, 2744 { ISD::AND, MVT::v32i1, 11 }, 2745 { ISD::AND, MVT::v64i1, 13 }, 2746 { ISD::OR, MVT::v2i1, 3 }, 2747 { ISD::OR, MVT::v4i1, 5 }, 2748 { ISD::OR, MVT::v8i1, 7 }, 2749 { ISD::OR, MVT::v16i1, 9 }, 2750 { ISD::OR, MVT::v32i1, 11 }, 2751 { ISD::OR, MVT::v64i1, 13 }, 2752 }; 2753 2754 static const CostTblEntry AVX2BoolReduction[] = { 2755 { ISD::AND, MVT::v16i16, 2 }, // vpmovmskb + cmp 2756 { ISD::AND, MVT::v32i8, 2 }, // vpmovmskb + cmp 2757 { ISD::OR, MVT::v16i16, 2 }, // vpmovmskb + cmp 2758 { ISD::OR, MVT::v32i8, 2 }, // vpmovmskb + cmp 2759 }; 2760 2761 static const CostTblEntry AVX1BoolReduction[] = { 2762 { ISD::AND, MVT::v4i64, 2 }, // vmovmskpd + cmp 2763 { ISD::AND, MVT::v8i32, 2 }, // vmovmskps + cmp 2764 { ISD::AND, MVT::v16i16, 4 }, // vextractf128 + vpand + vpmovmskb + cmp 2765 { ISD::AND, MVT::v32i8, 4 }, // vextractf128 + vpand + vpmovmskb + cmp 2766 { ISD::OR, MVT::v4i64, 2 }, // vmovmskpd + cmp 2767 { ISD::OR, MVT::v8i32, 2 }, // vmovmskps + cmp 2768 { ISD::OR, MVT::v16i16, 4 }, // vextractf128 + vpor + vpmovmskb + cmp 2769 { ISD::OR, MVT::v32i8, 4 }, // vextractf128 + vpor + vpmovmskb + cmp 2770 }; 2771 2772 static const CostTblEntry SSE2BoolReduction[] = { 2773 { ISD::AND, MVT::v2i64, 2 }, // movmskpd + cmp 2774 { ISD::AND, MVT::v4i32, 2 }, // movmskps + cmp 2775 { ISD::AND, MVT::v8i16, 2 }, // pmovmskb + cmp 2776 { ISD::AND, MVT::v16i8, 2 }, // pmovmskb + cmp 2777 { ISD::OR, MVT::v2i64, 2 }, // movmskpd + cmp 2778 { ISD::OR, MVT::v4i32, 2 }, // movmskps + cmp 2779 { ISD::OR, MVT::v8i16, 2 }, // pmovmskb + cmp 2780 { ISD::OR, MVT::v16i8, 2 }, // pmovmskb + cmp 2781 }; 2782 2783 // Handle bool allof/anyof patterns. 2784 if (ValTy->getVectorElementType()->isIntegerTy(1)) { 2785 unsigned ArithmeticCost = 0; 2786 if (LT.first != 1 && MTy.isVector() && 2787 MTy.getVectorNumElements() < ValTy->getVectorNumElements()) { 2788 // Type needs to be split. We need LT.first - 1 arithmetic ops. 2789 Type *SingleOpTy = VectorType::get(ValTy->getVectorElementType(), 2790 MTy.getVectorNumElements()); 2791 ArithmeticCost = getArithmeticInstrCost(Opcode, SingleOpTy); 2792 ArithmeticCost *= LT.first - 1; 2793 } 2794 2795 if (ST->hasAVX512()) 2796 if (const auto *Entry = CostTableLookup(AVX512BoolReduction, ISD, MTy)) 2797 return ArithmeticCost + Entry->Cost; 2798 if (ST->hasAVX2()) 2799 if (const auto *Entry = CostTableLookup(AVX2BoolReduction, ISD, MTy)) 2800 return ArithmeticCost + Entry->Cost; 2801 if (ST->hasAVX()) 2802 if (const auto *Entry = CostTableLookup(AVX1BoolReduction, ISD, MTy)) 2803 return ArithmeticCost + Entry->Cost; 2804 if (ST->hasSSE2()) 2805 if (const auto *Entry = CostTableLookup(SSE2BoolReduction, ISD, MTy)) 2806 return ArithmeticCost + Entry->Cost; 2807 2808 return BaseT::getArithmeticReductionCost(Opcode, ValTy, IsPairwise); 2809 } 2810 2811 unsigned NumVecElts = ValTy->getVectorNumElements(); 2812 unsigned ScalarSize = ValTy->getScalarSizeInBits(); 2813 2814 // Special case power of 2 reductions where the scalar type isn't changed 2815 // by type legalization. 2816 if (!isPowerOf2_32(NumVecElts) || ScalarSize != MTy.getScalarSizeInBits()) 2817 return BaseT::getArithmeticReductionCost(Opcode, ValTy, IsPairwise); 2818 2819 unsigned ReductionCost = 0; 2820 2821 Type *Ty = ValTy; 2822 if (LT.first != 1 && MTy.isVector() && 2823 MTy.getVectorNumElements() < ValTy->getVectorNumElements()) { 2824 // Type needs to be split. We need LT.first - 1 arithmetic ops. 2825 Ty = VectorType::get(ValTy->getVectorElementType(), 2826 MTy.getVectorNumElements()); 2827 ReductionCost = getArithmeticInstrCost(Opcode, Ty); 2828 ReductionCost *= LT.first - 1; 2829 NumVecElts = MTy.getVectorNumElements(); 2830 } 2831 2832 // Now handle reduction with the legal type, taking into account size changes 2833 // at each level. 2834 while (NumVecElts > 1) { 2835 // Determine the size of the remaining vector we need to reduce. 2836 unsigned Size = NumVecElts * ScalarSize; 2837 NumVecElts /= 2; 2838 // If we're reducing from 256/512 bits, use an extract_subvector. 2839 if (Size > 128) { 2840 Type *SubTy = VectorType::get(ValTy->getVectorElementType(), NumVecElts); 2841 ReductionCost += 2842 getShuffleCost(TTI::SK_ExtractSubvector, Ty, NumVecElts, SubTy); 2843 Ty = SubTy; 2844 } else if (Size == 128) { 2845 // Reducing from 128 bits is a permute of v2f64/v2i64. 2846 Type *ShufTy; 2847 if (ValTy->isFloatingPointTy()) 2848 ShufTy = VectorType::get(Type::getDoubleTy(ValTy->getContext()), 2); 2849 else 2850 ShufTy = VectorType::get(Type::getInt64Ty(ValTy->getContext()), 2); 2851 ReductionCost += 2852 getShuffleCost(TTI::SK_PermuteSingleSrc, ShufTy, 0, nullptr); 2853 } else if (Size == 64) { 2854 // Reducing from 64 bits is a shuffle of v4f32/v4i32. 2855 Type *ShufTy; 2856 if (ValTy->isFloatingPointTy()) 2857 ShufTy = VectorType::get(Type::getFloatTy(ValTy->getContext()), 4); 2858 else 2859 ShufTy = VectorType::get(Type::getInt32Ty(ValTy->getContext()), 4); 2860 ReductionCost += 2861 getShuffleCost(TTI::SK_PermuteSingleSrc, ShufTy, 0, nullptr); 2862 } else { 2863 // Reducing from smaller size is a shift by immediate. 2864 Type *ShiftTy = VectorType::get( 2865 Type::getIntNTy(ValTy->getContext(), Size), 128 / Size); 2866 ReductionCost += getArithmeticInstrCost( 2867 Instruction::LShr, ShiftTy, TargetTransformInfo::OK_AnyValue, 2868 TargetTransformInfo::OK_UniformConstantValue, 2869 TargetTransformInfo::OP_None, TargetTransformInfo::OP_None); 2870 } 2871 2872 // Add the arithmetic op for this level. 2873 ReductionCost += getArithmeticInstrCost(Opcode, Ty); 2874 } 2875 2876 // Add the final extract element to the cost. 2877 return ReductionCost + getVectorInstrCost(Instruction::ExtractElement, Ty, 0); 2878 } 2879 2880 int X86TTIImpl::getMinMaxReductionCost(Type *ValTy, Type *CondTy, 2881 bool IsPairwise, bool IsUnsigned) { 2882 std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, ValTy); 2883 2884 MVT MTy = LT.second; 2885 2886 int ISD; 2887 if (ValTy->isIntOrIntVectorTy()) { 2888 ISD = IsUnsigned ? ISD::UMIN : ISD::SMIN; 2889 } else { 2890 assert(ValTy->isFPOrFPVectorTy() && 2891 "Expected float point or integer vector type."); 2892 ISD = ISD::FMINNUM; 2893 } 2894 2895 // We use the Intel Architecture Code Analyzer(IACA) to measure the throughput 2896 // and make it as the cost. 2897 2898 static const CostTblEntry SSE1CostTblPairWise[] = { 2899 {ISD::FMINNUM, MVT::v4f32, 4}, 2900 }; 2901 2902 static const CostTblEntry SSE2CostTblPairWise[] = { 2903 {ISD::FMINNUM, MVT::v2f64, 3}, 2904 {ISD::SMIN, MVT::v2i64, 6}, 2905 {ISD::UMIN, MVT::v2i64, 8}, 2906 {ISD::SMIN, MVT::v4i32, 6}, 2907 {ISD::UMIN, MVT::v4i32, 8}, 2908 {ISD::SMIN, MVT::v8i16, 4}, 2909 {ISD::UMIN, MVT::v8i16, 6}, 2910 {ISD::SMIN, MVT::v16i8, 8}, 2911 {ISD::UMIN, MVT::v16i8, 6}, 2912 }; 2913 2914 static const CostTblEntry SSE41CostTblPairWise[] = { 2915 {ISD::FMINNUM, MVT::v4f32, 2}, 2916 {ISD::SMIN, MVT::v2i64, 9}, 2917 {ISD::UMIN, MVT::v2i64,10}, 2918 {ISD::SMIN, MVT::v4i32, 1}, // The data reported by the IACA is "1.5" 2919 {ISD::UMIN, MVT::v4i32, 2}, // The data reported by the IACA is "1.8" 2920 {ISD::SMIN, MVT::v8i16, 2}, 2921 {ISD::UMIN, MVT::v8i16, 2}, 2922 {ISD::SMIN, MVT::v16i8, 3}, 2923 {ISD::UMIN, MVT::v16i8, 3}, 2924 }; 2925 2926 static const CostTblEntry SSE42CostTblPairWise[] = { 2927 {ISD::SMIN, MVT::v2i64, 7}, // The data reported by the IACA is "6.8" 2928 {ISD::UMIN, MVT::v2i64, 8}, // The data reported by the IACA is "8.6" 2929 }; 2930 2931 static const CostTblEntry AVX1CostTblPairWise[] = { 2932 {ISD::FMINNUM, MVT::v4f32, 1}, 2933 {ISD::FMINNUM, MVT::v4f64, 1}, 2934 {ISD::FMINNUM, MVT::v8f32, 2}, 2935 {ISD::SMIN, MVT::v2i64, 3}, 2936 {ISD::UMIN, MVT::v2i64, 3}, 2937 {ISD::SMIN, MVT::v4i32, 1}, 2938 {ISD::UMIN, MVT::v4i32, 1}, 2939 {ISD::SMIN, MVT::v8i16, 1}, 2940 {ISD::UMIN, MVT::v8i16, 1}, 2941 {ISD::SMIN, MVT::v16i8, 2}, 2942 {ISD::UMIN, MVT::v16i8, 2}, 2943 {ISD::SMIN, MVT::v4i64, 7}, 2944 {ISD::UMIN, MVT::v4i64, 7}, 2945 {ISD::SMIN, MVT::v8i32, 3}, 2946 {ISD::UMIN, MVT::v8i32, 3}, 2947 {ISD::SMIN, MVT::v16i16, 3}, 2948 {ISD::UMIN, MVT::v16i16, 3}, 2949 {ISD::SMIN, MVT::v32i8, 3}, 2950 {ISD::UMIN, MVT::v32i8, 3}, 2951 }; 2952 2953 static const CostTblEntry AVX2CostTblPairWise[] = { 2954 {ISD::SMIN, MVT::v4i64, 2}, 2955 {ISD::UMIN, MVT::v4i64, 2}, 2956 {ISD::SMIN, MVT::v8i32, 1}, 2957 {ISD::UMIN, MVT::v8i32, 1}, 2958 {ISD::SMIN, MVT::v16i16, 1}, 2959 {ISD::UMIN, MVT::v16i16, 1}, 2960 {ISD::SMIN, MVT::v32i8, 2}, 2961 {ISD::UMIN, MVT::v32i8, 2}, 2962 }; 2963 2964 static const CostTblEntry AVX512CostTblPairWise[] = { 2965 {ISD::FMINNUM, MVT::v8f64, 1}, 2966 {ISD::FMINNUM, MVT::v16f32, 2}, 2967 {ISD::SMIN, MVT::v8i64, 2}, 2968 {ISD::UMIN, MVT::v8i64, 2}, 2969 {ISD::SMIN, MVT::v16i32, 1}, 2970 {ISD::UMIN, MVT::v16i32, 1}, 2971 }; 2972 2973 static const CostTblEntry SSE1CostTblNoPairWise[] = { 2974 {ISD::FMINNUM, MVT::v4f32, 4}, 2975 }; 2976 2977 static const CostTblEntry SSE2CostTblNoPairWise[] = { 2978 {ISD::FMINNUM, MVT::v2f64, 3}, 2979 {ISD::SMIN, MVT::v2i64, 6}, 2980 {ISD::UMIN, MVT::v2i64, 8}, 2981 {ISD::SMIN, MVT::v4i32, 6}, 2982 {ISD::UMIN, MVT::v4i32, 8}, 2983 {ISD::SMIN, MVT::v8i16, 4}, 2984 {ISD::UMIN, MVT::v8i16, 6}, 2985 {ISD::SMIN, MVT::v16i8, 8}, 2986 {ISD::UMIN, MVT::v16i8, 6}, 2987 }; 2988 2989 static const CostTblEntry SSE41CostTblNoPairWise[] = { 2990 {ISD::FMINNUM, MVT::v4f32, 3}, 2991 {ISD::SMIN, MVT::v2i64, 9}, 2992 {ISD::UMIN, MVT::v2i64,11}, 2993 {ISD::SMIN, MVT::v4i32, 1}, // The data reported by the IACA is "1.5" 2994 {ISD::UMIN, MVT::v4i32, 2}, // The data reported by the IACA is "1.8" 2995 {ISD::SMIN, MVT::v8i16, 1}, // The data reported by the IACA is "1.5" 2996 {ISD::UMIN, MVT::v8i16, 2}, // The data reported by the IACA is "1.8" 2997 {ISD::SMIN, MVT::v16i8, 3}, 2998 {ISD::UMIN, MVT::v16i8, 3}, 2999 }; 3000 3001 static const CostTblEntry SSE42CostTblNoPairWise[] = { 3002 {ISD::SMIN, MVT::v2i64, 7}, // The data reported by the IACA is "6.8" 3003 {ISD::UMIN, MVT::v2i64, 9}, // The data reported by the IACA is "8.6" 3004 }; 3005 3006 static const CostTblEntry AVX1CostTblNoPairWise[] = { 3007 {ISD::FMINNUM, MVT::v4f32, 1}, 3008 {ISD::FMINNUM, MVT::v4f64, 1}, 3009 {ISD::FMINNUM, MVT::v8f32, 1}, 3010 {ISD::SMIN, MVT::v2i64, 3}, 3011 {ISD::UMIN, MVT::v2i64, 3}, 3012 {ISD::SMIN, MVT::v4i32, 1}, 3013 {ISD::UMIN, MVT::v4i32, 1}, 3014 {ISD::SMIN, MVT::v8i16, 1}, 3015 {ISD::UMIN, MVT::v8i16, 1}, 3016 {ISD::SMIN, MVT::v16i8, 2}, 3017 {ISD::UMIN, MVT::v16i8, 2}, 3018 {ISD::SMIN, MVT::v4i64, 7}, 3019 {ISD::UMIN, MVT::v4i64, 7}, 3020 {ISD::SMIN, MVT::v8i32, 2}, 3021 {ISD::UMIN, MVT::v8i32, 2}, 3022 {ISD::SMIN, MVT::v16i16, 2}, 3023 {ISD::UMIN, MVT::v16i16, 2}, 3024 {ISD::SMIN, MVT::v32i8, 2}, 3025 {ISD::UMIN, MVT::v32i8, 2}, 3026 }; 3027 3028 static const CostTblEntry AVX2CostTblNoPairWise[] = { 3029 {ISD::SMIN, MVT::v4i64, 1}, 3030 {ISD::UMIN, MVT::v4i64, 1}, 3031 {ISD::SMIN, MVT::v8i32, 1}, 3032 {ISD::UMIN, MVT::v8i32, 1}, 3033 {ISD::SMIN, MVT::v16i16, 1}, 3034 {ISD::UMIN, MVT::v16i16, 1}, 3035 {ISD::SMIN, MVT::v32i8, 1}, 3036 {ISD::UMIN, MVT::v32i8, 1}, 3037 }; 3038 3039 static const CostTblEntry AVX512CostTblNoPairWise[] = { 3040 {ISD::FMINNUM, MVT::v8f64, 1}, 3041 {ISD::FMINNUM, MVT::v16f32, 2}, 3042 {ISD::SMIN, MVT::v8i64, 1}, 3043 {ISD::UMIN, MVT::v8i64, 1}, 3044 {ISD::SMIN, MVT::v16i32, 1}, 3045 {ISD::UMIN, MVT::v16i32, 1}, 3046 }; 3047 3048 if (IsPairwise) { 3049 if (ST->hasAVX512()) 3050 if (const auto *Entry = CostTableLookup(AVX512CostTblPairWise, ISD, MTy)) 3051 return LT.first * Entry->Cost; 3052 3053 if (ST->hasAVX2()) 3054 if (const auto *Entry = CostTableLookup(AVX2CostTblPairWise, ISD, MTy)) 3055 return LT.first * Entry->Cost; 3056 3057 if (ST->hasAVX()) 3058 if (const auto *Entry = CostTableLookup(AVX1CostTblPairWise, ISD, MTy)) 3059 return LT.first * Entry->Cost; 3060 3061 if (ST->hasSSE42()) 3062 if (const auto *Entry = CostTableLookup(SSE42CostTblPairWise, ISD, MTy)) 3063 return LT.first * Entry->Cost; 3064 3065 if (ST->hasSSE41()) 3066 if (const auto *Entry = CostTableLookup(SSE41CostTblPairWise, ISD, MTy)) 3067 return LT.first * Entry->Cost; 3068 3069 if (ST->hasSSE2()) 3070 if (const auto *Entry = CostTableLookup(SSE2CostTblPairWise, ISD, MTy)) 3071 return LT.first * Entry->Cost; 3072 3073 if (ST->hasSSE1()) 3074 if (const auto *Entry = CostTableLookup(SSE1CostTblPairWise, ISD, MTy)) 3075 return LT.first * Entry->Cost; 3076 } else { 3077 if (ST->hasAVX512()) 3078 if (const auto *Entry = 3079 CostTableLookup(AVX512CostTblNoPairWise, ISD, MTy)) 3080 return LT.first * Entry->Cost; 3081 3082 if (ST->hasAVX2()) 3083 if (const auto *Entry = CostTableLookup(AVX2CostTblNoPairWise, ISD, MTy)) 3084 return LT.first * Entry->Cost; 3085 3086 if (ST->hasAVX()) 3087 if (const auto *Entry = CostTableLookup(AVX1CostTblNoPairWise, ISD, MTy)) 3088 return LT.first * Entry->Cost; 3089 3090 if (ST->hasSSE42()) 3091 if (const auto *Entry = CostTableLookup(SSE42CostTblNoPairWise, ISD, MTy)) 3092 return LT.first * Entry->Cost; 3093 3094 if (ST->hasSSE41()) 3095 if (const auto *Entry = CostTableLookup(SSE41CostTblNoPairWise, ISD, MTy)) 3096 return LT.first * Entry->Cost; 3097 3098 if (ST->hasSSE2()) 3099 if (const auto *Entry = CostTableLookup(SSE2CostTblNoPairWise, ISD, MTy)) 3100 return LT.first * Entry->Cost; 3101 3102 if (ST->hasSSE1()) 3103 if (const auto *Entry = CostTableLookup(SSE1CostTblNoPairWise, ISD, MTy)) 3104 return LT.first * Entry->Cost; 3105 } 3106 3107 return BaseT::getMinMaxReductionCost(ValTy, CondTy, IsPairwise, IsUnsigned); 3108 } 3109 3110 /// Calculate the cost of materializing a 64-bit value. This helper 3111 /// method might only calculate a fraction of a larger immediate. Therefore it 3112 /// is valid to return a cost of ZERO. 3113 int X86TTIImpl::getIntImmCost(int64_t Val) { 3114 if (Val == 0) 3115 return TTI::TCC_Free; 3116 3117 if (isInt<32>(Val)) 3118 return TTI::TCC_Basic; 3119 3120 return 2 * TTI::TCC_Basic; 3121 } 3122 3123 int X86TTIImpl::getIntImmCost(const APInt &Imm, Type *Ty) { 3124 assert(Ty->isIntegerTy()); 3125 3126 unsigned BitSize = Ty->getPrimitiveSizeInBits(); 3127 if (BitSize == 0) 3128 return ~0U; 3129 3130 // Never hoist constants larger than 128bit, because this might lead to 3131 // incorrect code generation or assertions in codegen. 3132 // Fixme: Create a cost model for types larger than i128 once the codegen 3133 // issues have been fixed. 3134 if (BitSize > 128) 3135 return TTI::TCC_Free; 3136 3137 if (Imm == 0) 3138 return TTI::TCC_Free; 3139 3140 // Sign-extend all constants to a multiple of 64-bit. 3141 APInt ImmVal = Imm; 3142 if (BitSize % 64 != 0) 3143 ImmVal = Imm.sext(alignTo(BitSize, 64)); 3144 3145 // Split the constant into 64-bit chunks and calculate the cost for each 3146 // chunk. 3147 int Cost = 0; 3148 for (unsigned ShiftVal = 0; ShiftVal < BitSize; ShiftVal += 64) { 3149 APInt Tmp = ImmVal.ashr(ShiftVal).sextOrTrunc(64); 3150 int64_t Val = Tmp.getSExtValue(); 3151 Cost += getIntImmCost(Val); 3152 } 3153 // We need at least one instruction to materialize the constant. 3154 return std::max(1, Cost); 3155 } 3156 3157 int X86TTIImpl::getIntImmCostInst(unsigned Opcode, unsigned Idx, const APInt &Imm, 3158 Type *Ty) { 3159 assert(Ty->isIntegerTy()); 3160 3161 unsigned BitSize = Ty->getPrimitiveSizeInBits(); 3162 // There is no cost model for constants with a bit size of 0. Return TCC_Free 3163 // here, so that constant hoisting will ignore this constant. 3164 if (BitSize == 0) 3165 return TTI::TCC_Free; 3166 3167 unsigned ImmIdx = ~0U; 3168 switch (Opcode) { 3169 default: 3170 return TTI::TCC_Free; 3171 case Instruction::GetElementPtr: 3172 // Always hoist the base address of a GetElementPtr. This prevents the 3173 // creation of new constants for every base constant that gets constant 3174 // folded with the offset. 3175 if (Idx == 0) 3176 return 2 * TTI::TCC_Basic; 3177 return TTI::TCC_Free; 3178 case Instruction::Store: 3179 ImmIdx = 0; 3180 break; 3181 case Instruction::ICmp: 3182 // This is an imperfect hack to prevent constant hoisting of 3183 // compares that might be trying to check if a 64-bit value fits in 3184 // 32-bits. The backend can optimize these cases using a right shift by 32. 3185 // Ideally we would check the compare predicate here. There also other 3186 // similar immediates the backend can use shifts for. 3187 if (Idx == 1 && Imm.getBitWidth() == 64) { 3188 uint64_t ImmVal = Imm.getZExtValue(); 3189 if (ImmVal == 0x100000000ULL || ImmVal == 0xffffffff) 3190 return TTI::TCC_Free; 3191 } 3192 ImmIdx = 1; 3193 break; 3194 case Instruction::And: 3195 // We support 64-bit ANDs with immediates with 32-bits of leading zeroes 3196 // by using a 32-bit operation with implicit zero extension. Detect such 3197 // immediates here as the normal path expects bit 31 to be sign extended. 3198 if (Idx == 1 && Imm.getBitWidth() == 64 && isUInt<32>(Imm.getZExtValue())) 3199 return TTI::TCC_Free; 3200 ImmIdx = 1; 3201 break; 3202 case Instruction::Add: 3203 case Instruction::Sub: 3204 // For add/sub, we can use the opposite instruction for INT32_MIN. 3205 if (Idx == 1 && Imm.getBitWidth() == 64 && Imm.getZExtValue() == 0x80000000) 3206 return TTI::TCC_Free; 3207 ImmIdx = 1; 3208 break; 3209 case Instruction::UDiv: 3210 case Instruction::SDiv: 3211 case Instruction::URem: 3212 case Instruction::SRem: 3213 // Division by constant is typically expanded later into a different 3214 // instruction sequence. This completely changes the constants. 3215 // Report them as "free" to stop ConstantHoist from marking them as opaque. 3216 return TTI::TCC_Free; 3217 case Instruction::Mul: 3218 case Instruction::Or: 3219 case Instruction::Xor: 3220 ImmIdx = 1; 3221 break; 3222 // Always return TCC_Free for the shift value of a shift instruction. 3223 case Instruction::Shl: 3224 case Instruction::LShr: 3225 case Instruction::AShr: 3226 if (Idx == 1) 3227 return TTI::TCC_Free; 3228 break; 3229 case Instruction::Trunc: 3230 case Instruction::ZExt: 3231 case Instruction::SExt: 3232 case Instruction::IntToPtr: 3233 case Instruction::PtrToInt: 3234 case Instruction::BitCast: 3235 case Instruction::PHI: 3236 case Instruction::Call: 3237 case Instruction::Select: 3238 case Instruction::Ret: 3239 case Instruction::Load: 3240 break; 3241 } 3242 3243 if (Idx == ImmIdx) { 3244 int NumConstants = divideCeil(BitSize, 64); 3245 int Cost = X86TTIImpl::getIntImmCost(Imm, Ty); 3246 return (Cost <= NumConstants * TTI::TCC_Basic) 3247 ? static_cast<int>(TTI::TCC_Free) 3248 : Cost; 3249 } 3250 3251 return X86TTIImpl::getIntImmCost(Imm, Ty); 3252 } 3253 3254 int X86TTIImpl::getIntImmCostIntrin(Intrinsic::ID IID, unsigned Idx, 3255 const APInt &Imm, Type *Ty) { 3256 assert(Ty->isIntegerTy()); 3257 3258 unsigned BitSize = Ty->getPrimitiveSizeInBits(); 3259 // There is no cost model for constants with a bit size of 0. Return TCC_Free 3260 // here, so that constant hoisting will ignore this constant. 3261 if (BitSize == 0) 3262 return TTI::TCC_Free; 3263 3264 switch (IID) { 3265 default: 3266 return TTI::TCC_Free; 3267 case Intrinsic::sadd_with_overflow: 3268 case Intrinsic::uadd_with_overflow: 3269 case Intrinsic::ssub_with_overflow: 3270 case Intrinsic::usub_with_overflow: 3271 case Intrinsic::smul_with_overflow: 3272 case Intrinsic::umul_with_overflow: 3273 if ((Idx == 1) && Imm.getBitWidth() <= 64 && isInt<32>(Imm.getSExtValue())) 3274 return TTI::TCC_Free; 3275 break; 3276 case Intrinsic::experimental_stackmap: 3277 if ((Idx < 2) || (Imm.getBitWidth() <= 64 && isInt<64>(Imm.getSExtValue()))) 3278 return TTI::TCC_Free; 3279 break; 3280 case Intrinsic::experimental_patchpoint_void: 3281 case Intrinsic::experimental_patchpoint_i64: 3282 if ((Idx < 4) || (Imm.getBitWidth() <= 64 && isInt<64>(Imm.getSExtValue()))) 3283 return TTI::TCC_Free; 3284 break; 3285 } 3286 return X86TTIImpl::getIntImmCost(Imm, Ty); 3287 } 3288 3289 unsigned X86TTIImpl::getUserCost(const User *U, 3290 ArrayRef<const Value *> Operands) { 3291 if (isa<StoreInst>(U)) { 3292 Value *Ptr = U->getOperand(1); 3293 // Store instruction with index and scale costs 2 Uops. 3294 // Check the preceding GEP to identify non-const indices. 3295 if (auto GEP = dyn_cast<GetElementPtrInst>(Ptr)) { 3296 if (!all_of(GEP->indices(), [](Value *V) { return isa<Constant>(V); })) 3297 return TTI::TCC_Basic * 2; 3298 } 3299 return TTI::TCC_Basic; 3300 } 3301 return BaseT::getUserCost(U, Operands); 3302 } 3303 3304 // Return an average cost of Gather / Scatter instruction, maybe improved later 3305 int X86TTIImpl::getGSVectorCost(unsigned Opcode, Type *SrcVTy, Value *Ptr, 3306 unsigned Alignment, unsigned AddressSpace) { 3307 3308 assert(isa<VectorType>(SrcVTy) && "Unexpected type in getGSVectorCost"); 3309 unsigned VF = SrcVTy->getVectorNumElements(); 3310 3311 // Try to reduce index size from 64 bit (default for GEP) 3312 // to 32. It is essential for VF 16. If the index can't be reduced to 32, the 3313 // operation will use 16 x 64 indices which do not fit in a zmm and needs 3314 // to split. Also check that the base pointer is the same for all lanes, 3315 // and that there's at most one variable index. 3316 auto getIndexSizeInBits = [](Value *Ptr, const DataLayout& DL) { 3317 unsigned IndexSize = DL.getPointerSizeInBits(); 3318 GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr); 3319 if (IndexSize < 64 || !GEP) 3320 return IndexSize; 3321 3322 unsigned NumOfVarIndices = 0; 3323 Value *Ptrs = GEP->getPointerOperand(); 3324 if (Ptrs->getType()->isVectorTy() && !getSplatValue(Ptrs)) 3325 return IndexSize; 3326 for (unsigned i = 1; i < GEP->getNumOperands(); ++i) { 3327 if (isa<Constant>(GEP->getOperand(i))) 3328 continue; 3329 Type *IndxTy = GEP->getOperand(i)->getType(); 3330 if (IndxTy->isVectorTy()) 3331 IndxTy = IndxTy->getVectorElementType(); 3332 if ((IndxTy->getPrimitiveSizeInBits() == 64 && 3333 !isa<SExtInst>(GEP->getOperand(i))) || 3334 ++NumOfVarIndices > 1) 3335 return IndexSize; // 64 3336 } 3337 return (unsigned)32; 3338 }; 3339 3340 3341 // Trying to reduce IndexSize to 32 bits for vector 16. 3342 // By default the IndexSize is equal to pointer size. 3343 unsigned IndexSize = (ST->hasAVX512() && VF >= 16) 3344 ? getIndexSizeInBits(Ptr, DL) 3345 : DL.getPointerSizeInBits(); 3346 3347 Type *IndexVTy = VectorType::get(IntegerType::get(SrcVTy->getContext(), 3348 IndexSize), VF); 3349 std::pair<int, MVT> IdxsLT = TLI->getTypeLegalizationCost(DL, IndexVTy); 3350 std::pair<int, MVT> SrcLT = TLI->getTypeLegalizationCost(DL, SrcVTy); 3351 int SplitFactor = std::max(IdxsLT.first, SrcLT.first); 3352 if (SplitFactor > 1) { 3353 // Handle splitting of vector of pointers 3354 Type *SplitSrcTy = VectorType::get(SrcVTy->getScalarType(), VF / SplitFactor); 3355 return SplitFactor * getGSVectorCost(Opcode, SplitSrcTy, Ptr, Alignment, 3356 AddressSpace); 3357 } 3358 3359 // The gather / scatter cost is given by Intel architects. It is a rough 3360 // number since we are looking at one instruction in a time. 3361 const int GSOverhead = (Opcode == Instruction::Load) 3362 ? ST->getGatherOverhead() 3363 : ST->getScatterOverhead(); 3364 return GSOverhead + VF * getMemoryOpCost(Opcode, SrcVTy->getScalarType(), 3365 MaybeAlign(Alignment), AddressSpace); 3366 } 3367 3368 /// Return the cost of full scalarization of gather / scatter operation. 3369 /// 3370 /// Opcode - Load or Store instruction. 3371 /// SrcVTy - The type of the data vector that should be gathered or scattered. 3372 /// VariableMask - The mask is non-constant at compile time. 3373 /// Alignment - Alignment for one element. 3374 /// AddressSpace - pointer[s] address space. 3375 /// 3376 int X86TTIImpl::getGSScalarCost(unsigned Opcode, Type *SrcVTy, 3377 bool VariableMask, unsigned Alignment, 3378 unsigned AddressSpace) { 3379 unsigned VF = SrcVTy->getVectorNumElements(); 3380 3381 int MaskUnpackCost = 0; 3382 if (VariableMask) { 3383 VectorType *MaskTy = 3384 VectorType::get(Type::getInt1Ty(SrcVTy->getContext()), VF); 3385 MaskUnpackCost = getScalarizationOverhead(MaskTy, false, true); 3386 int ScalarCompareCost = 3387 getCmpSelInstrCost(Instruction::ICmp, Type::getInt1Ty(SrcVTy->getContext()), 3388 nullptr); 3389 int BranchCost = getCFInstrCost(Instruction::Br); 3390 MaskUnpackCost += VF * (BranchCost + ScalarCompareCost); 3391 } 3392 3393 // The cost of the scalar loads/stores. 3394 int MemoryOpCost = VF * getMemoryOpCost(Opcode, SrcVTy->getScalarType(), 3395 MaybeAlign(Alignment), AddressSpace); 3396 3397 int InsertExtractCost = 0; 3398 if (Opcode == Instruction::Load) 3399 for (unsigned i = 0; i < VF; ++i) 3400 // Add the cost of inserting each scalar load into the vector 3401 InsertExtractCost += 3402 getVectorInstrCost(Instruction::InsertElement, SrcVTy, i); 3403 else 3404 for (unsigned i = 0; i < VF; ++i) 3405 // Add the cost of extracting each element out of the data vector 3406 InsertExtractCost += 3407 getVectorInstrCost(Instruction::ExtractElement, SrcVTy, i); 3408 3409 return MemoryOpCost + MaskUnpackCost + InsertExtractCost; 3410 } 3411 3412 /// Calculate the cost of Gather / Scatter operation 3413 int X86TTIImpl::getGatherScatterOpCost(unsigned Opcode, Type *SrcVTy, 3414 Value *Ptr, bool VariableMask, 3415 unsigned Alignment, 3416 const Instruction *I = nullptr) { 3417 assert(SrcVTy->isVectorTy() && "Unexpected data type for Gather/Scatter"); 3418 unsigned VF = SrcVTy->getVectorNumElements(); 3419 PointerType *PtrTy = dyn_cast<PointerType>(Ptr->getType()); 3420 if (!PtrTy && Ptr->getType()->isVectorTy()) 3421 PtrTy = dyn_cast<PointerType>(Ptr->getType()->getVectorElementType()); 3422 assert(PtrTy && "Unexpected type for Ptr argument"); 3423 unsigned AddressSpace = PtrTy->getAddressSpace(); 3424 3425 bool Scalarize = false; 3426 if ((Opcode == Instruction::Load && 3427 !isLegalMaskedGather(SrcVTy, MaybeAlign(Alignment))) || 3428 (Opcode == Instruction::Store && 3429 !isLegalMaskedScatter(SrcVTy, MaybeAlign(Alignment)))) 3430 Scalarize = true; 3431 // Gather / Scatter for vector 2 is not profitable on KNL / SKX 3432 // Vector-4 of gather/scatter instruction does not exist on KNL. 3433 // We can extend it to 8 elements, but zeroing upper bits of 3434 // the mask vector will add more instructions. Right now we give the scalar 3435 // cost of vector-4 for KNL. TODO: Check, maybe the gather/scatter instruction 3436 // is better in the VariableMask case. 3437 if (ST->hasAVX512() && (VF == 2 || (VF == 4 && !ST->hasVLX()))) 3438 Scalarize = true; 3439 3440 if (Scalarize) 3441 return getGSScalarCost(Opcode, SrcVTy, VariableMask, Alignment, 3442 AddressSpace); 3443 3444 return getGSVectorCost(Opcode, SrcVTy, Ptr, Alignment, AddressSpace); 3445 } 3446 3447 bool X86TTIImpl::isLSRCostLess(TargetTransformInfo::LSRCost &C1, 3448 TargetTransformInfo::LSRCost &C2) { 3449 // X86 specific here are "instruction number 1st priority". 3450 return std::tie(C1.Insns, C1.NumRegs, C1.AddRecCost, 3451 C1.NumIVMuls, C1.NumBaseAdds, 3452 C1.ScaleCost, C1.ImmCost, C1.SetupCost) < 3453 std::tie(C2.Insns, C2.NumRegs, C2.AddRecCost, 3454 C2.NumIVMuls, C2.NumBaseAdds, 3455 C2.ScaleCost, C2.ImmCost, C2.SetupCost); 3456 } 3457 3458 bool X86TTIImpl::canMacroFuseCmp() { 3459 return ST->hasMacroFusion() || ST->hasBranchFusion(); 3460 } 3461 3462 bool X86TTIImpl::isLegalMaskedLoad(Type *DataTy, MaybeAlign Alignment) { 3463 if (!ST->hasAVX()) 3464 return false; 3465 3466 // The backend can't handle a single element vector. 3467 if (isa<VectorType>(DataTy) && DataTy->getVectorNumElements() == 1) 3468 return false; 3469 Type *ScalarTy = DataTy->getScalarType(); 3470 3471 if (ScalarTy->isPointerTy()) 3472 return true; 3473 3474 if (ScalarTy->isFloatTy() || ScalarTy->isDoubleTy()) 3475 return true; 3476 3477 if (!ScalarTy->isIntegerTy()) 3478 return false; 3479 3480 unsigned IntWidth = ScalarTy->getIntegerBitWidth(); 3481 return IntWidth == 32 || IntWidth == 64 || 3482 ((IntWidth == 8 || IntWidth == 16) && ST->hasBWI()); 3483 } 3484 3485 bool X86TTIImpl::isLegalMaskedStore(Type *DataType, MaybeAlign Alignment) { 3486 return isLegalMaskedLoad(DataType, Alignment); 3487 } 3488 3489 bool X86TTIImpl::isLegalNTLoad(Type *DataType, Align Alignment) { 3490 unsigned DataSize = DL.getTypeStoreSize(DataType); 3491 // The only supported nontemporal loads are for aligned vectors of 16 or 32 3492 // bytes. Note that 32-byte nontemporal vector loads are supported by AVX2 3493 // (the equivalent stores only require AVX). 3494 if (Alignment >= DataSize && (DataSize == 16 || DataSize == 32)) 3495 return DataSize == 16 ? ST->hasSSE1() : ST->hasAVX2(); 3496 3497 return false; 3498 } 3499 3500 bool X86TTIImpl::isLegalNTStore(Type *DataType, Align Alignment) { 3501 unsigned DataSize = DL.getTypeStoreSize(DataType); 3502 3503 // SSE4A supports nontemporal stores of float and double at arbitrary 3504 // alignment. 3505 if (ST->hasSSE4A() && (DataType->isFloatTy() || DataType->isDoubleTy())) 3506 return true; 3507 3508 // Besides the SSE4A subtarget exception above, only aligned stores are 3509 // available nontemporaly on any other subtarget. And only stores with a size 3510 // of 4..32 bytes (powers of 2, only) are permitted. 3511 if (Alignment < DataSize || DataSize < 4 || DataSize > 32 || 3512 !isPowerOf2_32(DataSize)) 3513 return false; 3514 3515 // 32-byte vector nontemporal stores are supported by AVX (the equivalent 3516 // loads require AVX2). 3517 if (DataSize == 32) 3518 return ST->hasAVX(); 3519 else if (DataSize == 16) 3520 return ST->hasSSE1(); 3521 return true; 3522 } 3523 3524 bool X86TTIImpl::isLegalMaskedExpandLoad(Type *DataTy) { 3525 if (!isa<VectorType>(DataTy)) 3526 return false; 3527 3528 if (!ST->hasAVX512()) 3529 return false; 3530 3531 // The backend can't handle a single element vector. 3532 if (DataTy->getVectorNumElements() == 1) 3533 return false; 3534 3535 Type *ScalarTy = DataTy->getVectorElementType(); 3536 3537 if (ScalarTy->isFloatTy() || ScalarTy->isDoubleTy()) 3538 return true; 3539 3540 if (!ScalarTy->isIntegerTy()) 3541 return false; 3542 3543 unsigned IntWidth = ScalarTy->getIntegerBitWidth(); 3544 return IntWidth == 32 || IntWidth == 64 || 3545 ((IntWidth == 8 || IntWidth == 16) && ST->hasVBMI2()); 3546 } 3547 3548 bool X86TTIImpl::isLegalMaskedCompressStore(Type *DataTy) { 3549 return isLegalMaskedExpandLoad(DataTy); 3550 } 3551 3552 bool X86TTIImpl::isLegalMaskedGather(Type *DataTy, MaybeAlign Alignment) { 3553 // Some CPUs have better gather performance than others. 3554 // TODO: Remove the explicit ST->hasAVX512()?, That would mean we would only 3555 // enable gather with a -march. 3556 if (!(ST->hasAVX512() || (ST->hasFastGather() && ST->hasAVX2()))) 3557 return false; 3558 3559 // This function is called now in two cases: from the Loop Vectorizer 3560 // and from the Scalarizer. 3561 // When the Loop Vectorizer asks about legality of the feature, 3562 // the vectorization factor is not calculated yet. The Loop Vectorizer 3563 // sends a scalar type and the decision is based on the width of the 3564 // scalar element. 3565 // Later on, the cost model will estimate usage this intrinsic based on 3566 // the vector type. 3567 // The Scalarizer asks again about legality. It sends a vector type. 3568 // In this case we can reject non-power-of-2 vectors. 3569 // We also reject single element vectors as the type legalizer can't 3570 // scalarize it. 3571 if (isa<VectorType>(DataTy)) { 3572 unsigned NumElts = DataTy->getVectorNumElements(); 3573 if (NumElts == 1 || !isPowerOf2_32(NumElts)) 3574 return false; 3575 } 3576 Type *ScalarTy = DataTy->getScalarType(); 3577 if (ScalarTy->isPointerTy()) 3578 return true; 3579 3580 if (ScalarTy->isFloatTy() || ScalarTy->isDoubleTy()) 3581 return true; 3582 3583 if (!ScalarTy->isIntegerTy()) 3584 return false; 3585 3586 unsigned IntWidth = ScalarTy->getIntegerBitWidth(); 3587 return IntWidth == 32 || IntWidth == 64; 3588 } 3589 3590 bool X86TTIImpl::isLegalMaskedScatter(Type *DataType, MaybeAlign Alignment) { 3591 // AVX2 doesn't support scatter 3592 if (!ST->hasAVX512()) 3593 return false; 3594 return isLegalMaskedGather(DataType, Alignment); 3595 } 3596 3597 bool X86TTIImpl::hasDivRemOp(Type *DataType, bool IsSigned) { 3598 EVT VT = TLI->getValueType(DL, DataType); 3599 return TLI->isOperationLegal(IsSigned ? ISD::SDIVREM : ISD::UDIVREM, VT); 3600 } 3601 3602 bool X86TTIImpl::isFCmpOrdCheaperThanFCmpZero(Type *Ty) { 3603 return false; 3604 } 3605 3606 bool X86TTIImpl::areInlineCompatible(const Function *Caller, 3607 const Function *Callee) const { 3608 const TargetMachine &TM = getTLI()->getTargetMachine(); 3609 3610 // Work this as a subsetting of subtarget features. 3611 const FeatureBitset &CallerBits = 3612 TM.getSubtargetImpl(*Caller)->getFeatureBits(); 3613 const FeatureBitset &CalleeBits = 3614 TM.getSubtargetImpl(*Callee)->getFeatureBits(); 3615 3616 FeatureBitset RealCallerBits = CallerBits & ~InlineFeatureIgnoreList; 3617 FeatureBitset RealCalleeBits = CalleeBits & ~InlineFeatureIgnoreList; 3618 return (RealCallerBits & RealCalleeBits) == RealCalleeBits; 3619 } 3620 3621 bool X86TTIImpl::areFunctionArgsABICompatible( 3622 const Function *Caller, const Function *Callee, 3623 SmallPtrSetImpl<Argument *> &Args) const { 3624 if (!BaseT::areFunctionArgsABICompatible(Caller, Callee, Args)) 3625 return false; 3626 3627 // If we get here, we know the target features match. If one function 3628 // considers 512-bit vectors legal and the other does not, consider them 3629 // incompatible. 3630 // FIXME Look at the arguments and only consider 512 bit or larger vectors? 3631 const TargetMachine &TM = getTLI()->getTargetMachine(); 3632 3633 return TM.getSubtarget<X86Subtarget>(*Caller).useAVX512Regs() == 3634 TM.getSubtarget<X86Subtarget>(*Callee).useAVX512Regs(); 3635 } 3636 3637 X86TTIImpl::TTI::MemCmpExpansionOptions 3638 X86TTIImpl::enableMemCmpExpansion(bool OptSize, bool IsZeroCmp) const { 3639 TTI::MemCmpExpansionOptions Options; 3640 Options.MaxNumLoads = TLI->getMaxExpandSizeMemcmp(OptSize); 3641 Options.NumLoadsPerBlock = 2; 3642 if (IsZeroCmp) { 3643 // Only enable vector loads for equality comparison. Right now the vector 3644 // version is not as fast for three way compare (see #33329). 3645 const unsigned PreferredWidth = ST->getPreferVectorWidth(); 3646 if (PreferredWidth >= 512 && ST->hasAVX512()) Options.LoadSizes.push_back(64); 3647 if (PreferredWidth >= 256 && ST->hasAVX()) Options.LoadSizes.push_back(32); 3648 if (PreferredWidth >= 128 && ST->hasSSE2()) Options.LoadSizes.push_back(16); 3649 // All GPR and vector loads can be unaligned. 3650 Options.AllowOverlappingLoads = true; 3651 } 3652 if (ST->is64Bit()) { 3653 Options.LoadSizes.push_back(8); 3654 } 3655 Options.LoadSizes.push_back(4); 3656 Options.LoadSizes.push_back(2); 3657 Options.LoadSizes.push_back(1); 3658 return Options; 3659 } 3660 3661 bool X86TTIImpl::enableInterleavedAccessVectorization() { 3662 // TODO: We expect this to be beneficial regardless of arch, 3663 // but there are currently some unexplained performance artifacts on Atom. 3664 // As a temporary solution, disable on Atom. 3665 return !(ST->isAtom()); 3666 } 3667 3668 // Get estimation for interleaved load/store operations for AVX2. 3669 // \p Factor is the interleaved-access factor (stride) - number of 3670 // (interleaved) elements in the group. 3671 // \p Indices contains the indices for a strided load: when the 3672 // interleaved load has gaps they indicate which elements are used. 3673 // If Indices is empty (or if the number of indices is equal to the size 3674 // of the interleaved-access as given in \p Factor) the access has no gaps. 3675 // 3676 // As opposed to AVX-512, AVX2 does not have generic shuffles that allow 3677 // computing the cost using a generic formula as a function of generic 3678 // shuffles. We therefore use a lookup table instead, filled according to 3679 // the instruction sequences that codegen currently generates. 3680 int X86TTIImpl::getInterleavedMemoryOpCostAVX2(unsigned Opcode, Type *VecTy, 3681 unsigned Factor, 3682 ArrayRef<unsigned> Indices, 3683 unsigned Alignment, 3684 unsigned AddressSpace, 3685 bool UseMaskForCond, 3686 bool UseMaskForGaps) { 3687 3688 if (UseMaskForCond || UseMaskForGaps) 3689 return BaseT::getInterleavedMemoryOpCost(Opcode, VecTy, Factor, Indices, 3690 Alignment, AddressSpace, 3691 UseMaskForCond, UseMaskForGaps); 3692 3693 // We currently Support only fully-interleaved groups, with no gaps. 3694 // TODO: Support also strided loads (interleaved-groups with gaps). 3695 if (Indices.size() && Indices.size() != Factor) 3696 return BaseT::getInterleavedMemoryOpCost(Opcode, VecTy, Factor, Indices, 3697 Alignment, AddressSpace); 3698 3699 // VecTy for interleave memop is <VF*Factor x Elt>. 3700 // So, for VF=4, Interleave Factor = 3, Element type = i32 we have 3701 // VecTy = <12 x i32>. 3702 MVT LegalVT = getTLI()->getTypeLegalizationCost(DL, VecTy).second; 3703 3704 // This function can be called with VecTy=<6xi128>, Factor=3, in which case 3705 // the VF=2, while v2i128 is an unsupported MVT vector type 3706 // (see MachineValueType.h::getVectorVT()). 3707 if (!LegalVT.isVector()) 3708 return BaseT::getInterleavedMemoryOpCost(Opcode, VecTy, Factor, Indices, 3709 Alignment, AddressSpace); 3710 3711 unsigned VF = VecTy->getVectorNumElements() / Factor; 3712 Type *ScalarTy = VecTy->getVectorElementType(); 3713 3714 // Calculate the number of memory operations (NumOfMemOps), required 3715 // for load/store the VecTy. 3716 unsigned VecTySize = DL.getTypeStoreSize(VecTy); 3717 unsigned LegalVTSize = LegalVT.getStoreSize(); 3718 unsigned NumOfMemOps = (VecTySize + LegalVTSize - 1) / LegalVTSize; 3719 3720 // Get the cost of one memory operation. 3721 Type *SingleMemOpTy = VectorType::get(VecTy->getVectorElementType(), 3722 LegalVT.getVectorNumElements()); 3723 unsigned MemOpCost = getMemoryOpCost(Opcode, SingleMemOpTy, 3724 MaybeAlign(Alignment), AddressSpace); 3725 3726 VectorType *VT = VectorType::get(ScalarTy, VF); 3727 EVT ETy = TLI->getValueType(DL, VT); 3728 if (!ETy.isSimple()) 3729 return BaseT::getInterleavedMemoryOpCost(Opcode, VecTy, Factor, Indices, 3730 Alignment, AddressSpace); 3731 3732 // TODO: Complete for other data-types and strides. 3733 // Each combination of Stride, ElementTy and VF results in a different 3734 // sequence; The cost tables are therefore accessed with: 3735 // Factor (stride) and VectorType=VFxElemType. 3736 // The Cost accounts only for the shuffle sequence; 3737 // The cost of the loads/stores is accounted for separately. 3738 // 3739 static const CostTblEntry AVX2InterleavedLoadTbl[] = { 3740 { 2, MVT::v4i64, 6 }, //(load 8i64 and) deinterleave into 2 x 4i64 3741 { 2, MVT::v4f64, 6 }, //(load 8f64 and) deinterleave into 2 x 4f64 3742 3743 { 3, MVT::v2i8, 10 }, //(load 6i8 and) deinterleave into 3 x 2i8 3744 { 3, MVT::v4i8, 4 }, //(load 12i8 and) deinterleave into 3 x 4i8 3745 { 3, MVT::v8i8, 9 }, //(load 24i8 and) deinterleave into 3 x 8i8 3746 { 3, MVT::v16i8, 11}, //(load 48i8 and) deinterleave into 3 x 16i8 3747 { 3, MVT::v32i8, 13}, //(load 96i8 and) deinterleave into 3 x 32i8 3748 { 3, MVT::v8f32, 17 }, //(load 24f32 and)deinterleave into 3 x 8f32 3749 3750 { 4, MVT::v2i8, 12 }, //(load 8i8 and) deinterleave into 4 x 2i8 3751 { 4, MVT::v4i8, 4 }, //(load 16i8 and) deinterleave into 4 x 4i8 3752 { 4, MVT::v8i8, 20 }, //(load 32i8 and) deinterleave into 4 x 8i8 3753 { 4, MVT::v16i8, 39 }, //(load 64i8 and) deinterleave into 4 x 16i8 3754 { 4, MVT::v32i8, 80 }, //(load 128i8 and) deinterleave into 4 x 32i8 3755 3756 { 8, MVT::v8f32, 40 } //(load 64f32 and)deinterleave into 8 x 8f32 3757 }; 3758 3759 static const CostTblEntry AVX2InterleavedStoreTbl[] = { 3760 { 2, MVT::v4i64, 6 }, //interleave into 2 x 4i64 into 8i64 (and store) 3761 { 2, MVT::v4f64, 6 }, //interleave into 2 x 4f64 into 8f64 (and store) 3762 3763 { 3, MVT::v2i8, 7 }, //interleave 3 x 2i8 into 6i8 (and store) 3764 { 3, MVT::v4i8, 8 }, //interleave 3 x 4i8 into 12i8 (and store) 3765 { 3, MVT::v8i8, 11 }, //interleave 3 x 8i8 into 24i8 (and store) 3766 { 3, MVT::v16i8, 11 }, //interleave 3 x 16i8 into 48i8 (and store) 3767 { 3, MVT::v32i8, 13 }, //interleave 3 x 32i8 into 96i8 (and store) 3768 3769 { 4, MVT::v2i8, 12 }, //interleave 4 x 2i8 into 8i8 (and store) 3770 { 4, MVT::v4i8, 9 }, //interleave 4 x 4i8 into 16i8 (and store) 3771 { 4, MVT::v8i8, 10 }, //interleave 4 x 8i8 into 32i8 (and store) 3772 { 4, MVT::v16i8, 10 }, //interleave 4 x 16i8 into 64i8 (and store) 3773 { 4, MVT::v32i8, 12 } //interleave 4 x 32i8 into 128i8 (and store) 3774 }; 3775 3776 if (Opcode == Instruction::Load) { 3777 if (const auto *Entry = 3778 CostTableLookup(AVX2InterleavedLoadTbl, Factor, ETy.getSimpleVT())) 3779 return NumOfMemOps * MemOpCost + Entry->Cost; 3780 } else { 3781 assert(Opcode == Instruction::Store && 3782 "Expected Store Instruction at this point"); 3783 if (const auto *Entry = 3784 CostTableLookup(AVX2InterleavedStoreTbl, Factor, ETy.getSimpleVT())) 3785 return NumOfMemOps * MemOpCost + Entry->Cost; 3786 } 3787 3788 return BaseT::getInterleavedMemoryOpCost(Opcode, VecTy, Factor, Indices, 3789 Alignment, AddressSpace); 3790 } 3791 3792 // Get estimation for interleaved load/store operations and strided load. 3793 // \p Indices contains indices for strided load. 3794 // \p Factor - the factor of interleaving. 3795 // AVX-512 provides 3-src shuffles that significantly reduces the cost. 3796 int X86TTIImpl::getInterleavedMemoryOpCostAVX512(unsigned Opcode, Type *VecTy, 3797 unsigned Factor, 3798 ArrayRef<unsigned> Indices, 3799 unsigned Alignment, 3800 unsigned AddressSpace, 3801 bool UseMaskForCond, 3802 bool UseMaskForGaps) { 3803 3804 if (UseMaskForCond || UseMaskForGaps) 3805 return BaseT::getInterleavedMemoryOpCost(Opcode, VecTy, Factor, Indices, 3806 Alignment, AddressSpace, 3807 UseMaskForCond, UseMaskForGaps); 3808 3809 // VecTy for interleave memop is <VF*Factor x Elt>. 3810 // So, for VF=4, Interleave Factor = 3, Element type = i32 we have 3811 // VecTy = <12 x i32>. 3812 3813 // Calculate the number of memory operations (NumOfMemOps), required 3814 // for load/store the VecTy. 3815 MVT LegalVT = getTLI()->getTypeLegalizationCost(DL, VecTy).second; 3816 unsigned VecTySize = DL.getTypeStoreSize(VecTy); 3817 unsigned LegalVTSize = LegalVT.getStoreSize(); 3818 unsigned NumOfMemOps = (VecTySize + LegalVTSize - 1) / LegalVTSize; 3819 3820 // Get the cost of one memory operation. 3821 Type *SingleMemOpTy = VectorType::get(VecTy->getVectorElementType(), 3822 LegalVT.getVectorNumElements()); 3823 unsigned MemOpCost = getMemoryOpCost(Opcode, SingleMemOpTy, 3824 MaybeAlign(Alignment), AddressSpace); 3825 3826 unsigned VF = VecTy->getVectorNumElements() / Factor; 3827 MVT VT = MVT::getVectorVT(MVT::getVT(VecTy->getScalarType()), VF); 3828 3829 if (Opcode == Instruction::Load) { 3830 // The tables (AVX512InterleavedLoadTbl and AVX512InterleavedStoreTbl) 3831 // contain the cost of the optimized shuffle sequence that the 3832 // X86InterleavedAccess pass will generate. 3833 // The cost of loads and stores are computed separately from the table. 3834 3835 // X86InterleavedAccess support only the following interleaved-access group. 3836 static const CostTblEntry AVX512InterleavedLoadTbl[] = { 3837 {3, MVT::v16i8, 12}, //(load 48i8 and) deinterleave into 3 x 16i8 3838 {3, MVT::v32i8, 14}, //(load 96i8 and) deinterleave into 3 x 32i8 3839 {3, MVT::v64i8, 22}, //(load 96i8 and) deinterleave into 3 x 32i8 3840 }; 3841 3842 if (const auto *Entry = 3843 CostTableLookup(AVX512InterleavedLoadTbl, Factor, VT)) 3844 return NumOfMemOps * MemOpCost + Entry->Cost; 3845 //If an entry does not exist, fallback to the default implementation. 3846 3847 // Kind of shuffle depends on number of loaded values. 3848 // If we load the entire data in one register, we can use a 1-src shuffle. 3849 // Otherwise, we'll merge 2 sources in each operation. 3850 TTI::ShuffleKind ShuffleKind = 3851 (NumOfMemOps > 1) ? TTI::SK_PermuteTwoSrc : TTI::SK_PermuteSingleSrc; 3852 3853 unsigned ShuffleCost = 3854 getShuffleCost(ShuffleKind, SingleMemOpTy, 0, nullptr); 3855 3856 unsigned NumOfLoadsInInterleaveGrp = 3857 Indices.size() ? Indices.size() : Factor; 3858 Type *ResultTy = VectorType::get(VecTy->getVectorElementType(), 3859 VecTy->getVectorNumElements() / Factor); 3860 unsigned NumOfResults = 3861 getTLI()->getTypeLegalizationCost(DL, ResultTy).first * 3862 NumOfLoadsInInterleaveGrp; 3863 3864 // About a half of the loads may be folded in shuffles when we have only 3865 // one result. If we have more than one result, we do not fold loads at all. 3866 unsigned NumOfUnfoldedLoads = 3867 NumOfResults > 1 ? NumOfMemOps : NumOfMemOps / 2; 3868 3869 // Get a number of shuffle operations per result. 3870 unsigned NumOfShufflesPerResult = 3871 std::max((unsigned)1, (unsigned)(NumOfMemOps - 1)); 3872 3873 // The SK_MergeTwoSrc shuffle clobbers one of src operands. 3874 // When we have more than one destination, we need additional instructions 3875 // to keep sources. 3876 unsigned NumOfMoves = 0; 3877 if (NumOfResults > 1 && ShuffleKind == TTI::SK_PermuteTwoSrc) 3878 NumOfMoves = NumOfResults * NumOfShufflesPerResult / 2; 3879 3880 int Cost = NumOfResults * NumOfShufflesPerResult * ShuffleCost + 3881 NumOfUnfoldedLoads * MemOpCost + NumOfMoves; 3882 3883 return Cost; 3884 } 3885 3886 // Store. 3887 assert(Opcode == Instruction::Store && 3888 "Expected Store Instruction at this point"); 3889 // X86InterleavedAccess support only the following interleaved-access group. 3890 static const CostTblEntry AVX512InterleavedStoreTbl[] = { 3891 {3, MVT::v16i8, 12}, // interleave 3 x 16i8 into 48i8 (and store) 3892 {3, MVT::v32i8, 14}, // interleave 3 x 32i8 into 96i8 (and store) 3893 {3, MVT::v64i8, 26}, // interleave 3 x 64i8 into 96i8 (and store) 3894 3895 {4, MVT::v8i8, 10}, // interleave 4 x 8i8 into 32i8 (and store) 3896 {4, MVT::v16i8, 11}, // interleave 4 x 16i8 into 64i8 (and store) 3897 {4, MVT::v32i8, 14}, // interleave 4 x 32i8 into 128i8 (and store) 3898 {4, MVT::v64i8, 24} // interleave 4 x 32i8 into 256i8 (and store) 3899 }; 3900 3901 if (const auto *Entry = 3902 CostTableLookup(AVX512InterleavedStoreTbl, Factor, VT)) 3903 return NumOfMemOps * MemOpCost + Entry->Cost; 3904 //If an entry does not exist, fallback to the default implementation. 3905 3906 // There is no strided stores meanwhile. And store can't be folded in 3907 // shuffle. 3908 unsigned NumOfSources = Factor; // The number of values to be merged. 3909 unsigned ShuffleCost = 3910 getShuffleCost(TTI::SK_PermuteTwoSrc, SingleMemOpTy, 0, nullptr); 3911 unsigned NumOfShufflesPerStore = NumOfSources - 1; 3912 3913 // The SK_MergeTwoSrc shuffle clobbers one of src operands. 3914 // We need additional instructions to keep sources. 3915 unsigned NumOfMoves = NumOfMemOps * NumOfShufflesPerStore / 2; 3916 int Cost = NumOfMemOps * (MemOpCost + NumOfShufflesPerStore * ShuffleCost) + 3917 NumOfMoves; 3918 return Cost; 3919 } 3920 3921 int X86TTIImpl::getInterleavedMemoryOpCost(unsigned Opcode, Type *VecTy, 3922 unsigned Factor, 3923 ArrayRef<unsigned> Indices, 3924 unsigned Alignment, 3925 unsigned AddressSpace, 3926 bool UseMaskForCond, 3927 bool UseMaskForGaps) { 3928 auto isSupportedOnAVX512 = [](Type *VecTy, bool HasBW) { 3929 Type *EltTy = VecTy->getVectorElementType(); 3930 if (EltTy->isFloatTy() || EltTy->isDoubleTy() || EltTy->isIntegerTy(64) || 3931 EltTy->isIntegerTy(32) || EltTy->isPointerTy()) 3932 return true; 3933 if (EltTy->isIntegerTy(16) || EltTy->isIntegerTy(8)) 3934 return HasBW; 3935 return false; 3936 }; 3937 if (ST->hasAVX512() && isSupportedOnAVX512(VecTy, ST->hasBWI())) 3938 return getInterleavedMemoryOpCostAVX512(Opcode, VecTy, Factor, Indices, 3939 Alignment, AddressSpace, 3940 UseMaskForCond, UseMaskForGaps); 3941 if (ST->hasAVX2()) 3942 return getInterleavedMemoryOpCostAVX2(Opcode, VecTy, Factor, Indices, 3943 Alignment, AddressSpace, 3944 UseMaskForCond, UseMaskForGaps); 3945 3946 return BaseT::getInterleavedMemoryOpCost(Opcode, VecTy, Factor, Indices, 3947 Alignment, AddressSpace, 3948 UseMaskForCond, UseMaskForGaps); 3949 } 3950