1 //===-- X86TargetTransformInfo.cpp - X86 specific TTI pass ----------------===// 2 // 3 // The LLVM Compiler Infrastructure 4 // 5 // This file is distributed under the University of Illinois Open Source 6 // License. See LICENSE.TXT for details. 7 // 8 //===----------------------------------------------------------------------===// 9 /// \file 10 /// This file implements a TargetTransformInfo analysis pass specific to the 11 /// X86 target machine. It uses the target's detailed information to provide 12 /// more precise answers to certain TTI queries, while letting the target 13 /// independent and default TTI implementations handle the rest. 14 /// 15 //===----------------------------------------------------------------------===// 16 17 #include "X86TargetTransformInfo.h" 18 #include "llvm/Analysis/TargetTransformInfo.h" 19 #include "llvm/CodeGen/BasicTTIImpl.h" 20 #include "llvm/IR/IntrinsicInst.h" 21 #include "llvm/Support/Debug.h" 22 #include "llvm/Target/CostTable.h" 23 #include "llvm/Target/TargetLowering.h" 24 25 using namespace llvm; 26 27 #define DEBUG_TYPE "x86tti" 28 29 //===----------------------------------------------------------------------===// 30 // 31 // X86 cost model. 32 // 33 //===----------------------------------------------------------------------===// 34 35 TargetTransformInfo::PopcntSupportKind 36 X86TTIImpl::getPopcntSupport(unsigned TyWidth) { 37 assert(isPowerOf2_32(TyWidth) && "Ty width must be power of 2"); 38 // TODO: Currently the __builtin_popcount() implementation using SSE3 39 // instructions is inefficient. Once the problem is fixed, we should 40 // call ST->hasSSE3() instead of ST->hasPOPCNT(). 41 return ST->hasPOPCNT() ? TTI::PSK_FastHardware : TTI::PSK_Software; 42 } 43 44 unsigned X86TTIImpl::getNumberOfRegisters(bool Vector) { 45 if (Vector && !ST->hasSSE1()) 46 return 0; 47 48 if (ST->is64Bit()) { 49 if (Vector && ST->hasAVX512()) 50 return 32; 51 return 16; 52 } 53 return 8; 54 } 55 56 unsigned X86TTIImpl::getRegisterBitWidth(bool Vector) { 57 if (Vector) { 58 if (ST->hasAVX512()) return 512; 59 if (ST->hasAVX()) return 256; 60 if (ST->hasSSE1()) return 128; 61 return 0; 62 } 63 64 if (ST->is64Bit()) 65 return 64; 66 67 return 32; 68 } 69 70 unsigned X86TTIImpl::getMaxInterleaveFactor(unsigned VF) { 71 // If the loop will not be vectorized, don't interleave the loop. 72 // Let regular unroll to unroll the loop, which saves the overflow 73 // check and memory check cost. 74 if (VF == 1) 75 return 1; 76 77 if (ST->isAtom()) 78 return 1; 79 80 // Sandybridge and Haswell have multiple execution ports and pipelined 81 // vector units. 82 if (ST->hasAVX()) 83 return 4; 84 85 return 2; 86 } 87 88 int X86TTIImpl::getArithmeticInstrCost( 89 unsigned Opcode, Type *Ty, TTI::OperandValueKind Op1Info, 90 TTI::OperandValueKind Op2Info, TTI::OperandValueProperties Opd1PropInfo, 91 TTI::OperandValueProperties Opd2PropInfo) { 92 // Legalize the type. 93 std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, Ty); 94 95 int ISD = TLI->InstructionOpcodeToISD(Opcode); 96 assert(ISD && "Invalid opcode"); 97 98 if (ISD == ISD::SDIV && 99 Op2Info == TargetTransformInfo::OK_UniformConstantValue && 100 Opd2PropInfo == TargetTransformInfo::OP_PowerOf2) { 101 // On X86, vector signed division by constants power-of-two are 102 // normally expanded to the sequence SRA + SRL + ADD + SRA. 103 // The OperandValue properties many not be same as that of previous 104 // operation;conservatively assume OP_None. 105 int Cost = 2 * getArithmeticInstrCost(Instruction::AShr, Ty, Op1Info, 106 Op2Info, TargetTransformInfo::OP_None, 107 TargetTransformInfo::OP_None); 108 Cost += getArithmeticInstrCost(Instruction::LShr, Ty, Op1Info, Op2Info, 109 TargetTransformInfo::OP_None, 110 TargetTransformInfo::OP_None); 111 Cost += getArithmeticInstrCost(Instruction::Add, Ty, Op1Info, Op2Info, 112 TargetTransformInfo::OP_None, 113 TargetTransformInfo::OP_None); 114 115 return Cost; 116 } 117 118 static const CostTblEntry AVX2UniformConstCostTable[] = { 119 { ISD::SRA, MVT::v4i64, 4 }, // 2 x psrad + shuffle. 120 121 { ISD::SDIV, MVT::v16i16, 6 }, // vpmulhw sequence 122 { ISD::UDIV, MVT::v16i16, 6 }, // vpmulhuw sequence 123 { ISD::SDIV, MVT::v8i32, 15 }, // vpmuldq sequence 124 { ISD::UDIV, MVT::v8i32, 15 }, // vpmuludq sequence 125 }; 126 127 if (Op2Info == TargetTransformInfo::OK_UniformConstantValue && 128 ST->hasAVX2()) { 129 if (const auto *Entry = CostTableLookup(AVX2UniformConstCostTable, ISD, 130 LT.second)) 131 return LT.first * Entry->Cost; 132 } 133 134 static const CostTblEntry AVX512CostTable[] = { 135 { ISD::SHL, MVT::v16i32, 1 }, 136 { ISD::SRL, MVT::v16i32, 1 }, 137 { ISD::SRA, MVT::v16i32, 1 }, 138 { ISD::SHL, MVT::v8i64, 1 }, 139 { ISD::SRL, MVT::v8i64, 1 }, 140 { ISD::SRA, MVT::v8i64, 1 }, 141 }; 142 143 if (ST->hasAVX512()) { 144 if (const auto *Entry = CostTableLookup(AVX512CostTable, ISD, LT.second)) 145 return LT.first * Entry->Cost; 146 } 147 148 static const CostTblEntry AVX2CostTable[] = { 149 // Shifts on v4i64/v8i32 on AVX2 is legal even though we declare to 150 // customize them to detect the cases where shift amount is a scalar one. 151 { ISD::SHL, MVT::v4i32, 1 }, 152 { ISD::SRL, MVT::v4i32, 1 }, 153 { ISD::SRA, MVT::v4i32, 1 }, 154 { ISD::SHL, MVT::v8i32, 1 }, 155 { ISD::SRL, MVT::v8i32, 1 }, 156 { ISD::SRA, MVT::v8i32, 1 }, 157 { ISD::SHL, MVT::v2i64, 1 }, 158 { ISD::SRL, MVT::v2i64, 1 }, 159 { ISD::SHL, MVT::v4i64, 1 }, 160 { ISD::SRL, MVT::v4i64, 1 }, 161 }; 162 163 // Look for AVX2 lowering tricks. 164 if (ST->hasAVX2()) { 165 if (ISD == ISD::SHL && LT.second == MVT::v16i16 && 166 (Op2Info == TargetTransformInfo::OK_UniformConstantValue || 167 Op2Info == TargetTransformInfo::OK_NonUniformConstantValue)) 168 // On AVX2, a packed v16i16 shift left by a constant build_vector 169 // is lowered into a vector multiply (vpmullw). 170 return LT.first; 171 172 if (const auto *Entry = CostTableLookup(AVX2CostTable, ISD, LT.second)) 173 return LT.first * Entry->Cost; 174 } 175 176 static const CostTblEntry XOPCostTable[] = { 177 // 128bit shifts take 1cy, but right shifts require negation beforehand. 178 { ISD::SHL, MVT::v16i8, 1 }, 179 { ISD::SRL, MVT::v16i8, 2 }, 180 { ISD::SRA, MVT::v16i8, 2 }, 181 { ISD::SHL, MVT::v8i16, 1 }, 182 { ISD::SRL, MVT::v8i16, 2 }, 183 { ISD::SRA, MVT::v8i16, 2 }, 184 { ISD::SHL, MVT::v4i32, 1 }, 185 { ISD::SRL, MVT::v4i32, 2 }, 186 { ISD::SRA, MVT::v4i32, 2 }, 187 { ISD::SHL, MVT::v2i64, 1 }, 188 { ISD::SRL, MVT::v2i64, 2 }, 189 { ISD::SRA, MVT::v2i64, 2 }, 190 // 256bit shifts require splitting if AVX2 didn't catch them above. 191 { ISD::SHL, MVT::v32i8, 2 }, 192 { ISD::SRL, MVT::v32i8, 4 }, 193 { ISD::SRA, MVT::v32i8, 4 }, 194 { ISD::SHL, MVT::v16i16, 2 }, 195 { ISD::SRL, MVT::v16i16, 4 }, 196 { ISD::SRA, MVT::v16i16, 4 }, 197 { ISD::SHL, MVT::v8i32, 2 }, 198 { ISD::SRL, MVT::v8i32, 4 }, 199 { ISD::SRA, MVT::v8i32, 4 }, 200 { ISD::SHL, MVT::v4i64, 2 }, 201 { ISD::SRL, MVT::v4i64, 4 }, 202 { ISD::SRA, MVT::v4i64, 4 }, 203 }; 204 205 // Look for XOP lowering tricks. 206 if (ST->hasXOP()) { 207 if (const auto *Entry = CostTableLookup(XOPCostTable, ISD, LT.second)) 208 return LT.first * Entry->Cost; 209 } 210 211 static const CostTblEntry AVX2CustomCostTable[] = { 212 { ISD::SHL, MVT::v32i8, 11 }, // vpblendvb sequence. 213 { ISD::SHL, MVT::v16i16, 10 }, // extend/vpsrlvd/pack sequence. 214 215 { ISD::SRL, MVT::v32i8, 11 }, // vpblendvb sequence. 216 { ISD::SRL, MVT::v16i16, 10 }, // extend/vpsrlvd/pack sequence. 217 218 { ISD::SRA, MVT::v32i8, 24 }, // vpblendvb sequence. 219 { ISD::SRA, MVT::v16i16, 10 }, // extend/vpsravd/pack sequence. 220 { ISD::SRA, MVT::v2i64, 4 }, // srl/xor/sub sequence. 221 { ISD::SRA, MVT::v4i64, 4 }, // srl/xor/sub sequence. 222 223 // Vectorizing division is a bad idea. See the SSE2 table for more comments. 224 { ISD::SDIV, MVT::v32i8, 32*20 }, 225 { ISD::SDIV, MVT::v16i16, 16*20 }, 226 { ISD::SDIV, MVT::v8i32, 8*20 }, 227 { ISD::SDIV, MVT::v4i64, 4*20 }, 228 { ISD::UDIV, MVT::v32i8, 32*20 }, 229 { ISD::UDIV, MVT::v16i16, 16*20 }, 230 { ISD::UDIV, MVT::v8i32, 8*20 }, 231 { ISD::UDIV, MVT::v4i64, 4*20 }, 232 }; 233 234 // Look for AVX2 lowering tricks for custom cases. 235 if (ST->hasAVX2()) { 236 if (const auto *Entry = CostTableLookup(AVX2CustomCostTable, ISD, 237 LT.second)) 238 return LT.first * Entry->Cost; 239 } 240 241 static const CostTblEntry 242 SSE2UniformConstCostTable[] = { 243 // We don't correctly identify costs of casts because they are marked as 244 // custom. 245 // Constant splats are cheaper for the following instructions. 246 { ISD::SHL, MVT::v16i8, 1 }, // psllw. 247 { ISD::SHL, MVT::v32i8, 2 }, // psllw. 248 { ISD::SHL, MVT::v8i16, 1 }, // psllw. 249 { ISD::SHL, MVT::v16i16, 2 }, // psllw. 250 { ISD::SHL, MVT::v4i32, 1 }, // pslld 251 { ISD::SHL, MVT::v8i32, 2 }, // pslld 252 { ISD::SHL, MVT::v2i64, 1 }, // psllq. 253 { ISD::SHL, MVT::v4i64, 2 }, // psllq. 254 255 { ISD::SRL, MVT::v16i8, 1 }, // psrlw. 256 { ISD::SRL, MVT::v32i8, 2 }, // psrlw. 257 { ISD::SRL, MVT::v8i16, 1 }, // psrlw. 258 { ISD::SRL, MVT::v16i16, 2 }, // psrlw. 259 { ISD::SRL, MVT::v4i32, 1 }, // psrld. 260 { ISD::SRL, MVT::v8i32, 2 }, // psrld. 261 { ISD::SRL, MVT::v2i64, 1 }, // psrlq. 262 { ISD::SRL, MVT::v4i64, 2 }, // psrlq. 263 264 { ISD::SRA, MVT::v16i8, 4 }, // psrlw, pand, pxor, psubb. 265 { ISD::SRA, MVT::v32i8, 8 }, // psrlw, pand, pxor, psubb. 266 { ISD::SRA, MVT::v8i16, 1 }, // psraw. 267 { ISD::SRA, MVT::v16i16, 2 }, // psraw. 268 { ISD::SRA, MVT::v4i32, 1 }, // psrad. 269 { ISD::SRA, MVT::v8i32, 2 }, // psrad. 270 { ISD::SRA, MVT::v2i64, 4 }, // 2 x psrad + shuffle. 271 { ISD::SRA, MVT::v4i64, 8 }, // 2 x psrad + shuffle. 272 273 { ISD::SDIV, MVT::v8i16, 6 }, // pmulhw sequence 274 { ISD::UDIV, MVT::v8i16, 6 }, // pmulhuw sequence 275 { ISD::SDIV, MVT::v4i32, 19 }, // pmuludq sequence 276 { ISD::UDIV, MVT::v4i32, 15 }, // pmuludq sequence 277 }; 278 279 if (Op2Info == TargetTransformInfo::OK_UniformConstantValue && 280 ST->hasSSE2()) { 281 // pmuldq sequence. 282 if (ISD == ISD::SDIV && LT.second == MVT::v4i32 && ST->hasSSE41()) 283 return LT.first * 15; 284 285 if (const auto *Entry = CostTableLookup(SSE2UniformConstCostTable, ISD, 286 LT.second)) 287 return LT.first * Entry->Cost; 288 } 289 290 if (ISD == ISD::SHL && 291 Op2Info == TargetTransformInfo::OK_NonUniformConstantValue) { 292 MVT VT = LT.second; 293 // Vector shift left by non uniform constant can be lowered 294 // into vector multiply (pmullw/pmulld). 295 if ((VT == MVT::v8i16 && ST->hasSSE2()) || 296 (VT == MVT::v4i32 && ST->hasSSE41())) 297 return LT.first; 298 299 // v16i16 and v8i32 shifts by non-uniform constants are lowered into a 300 // sequence of extract + two vector multiply + insert. 301 if ((VT == MVT::v8i32 || VT == MVT::v16i16) && 302 (ST->hasAVX() && !ST->hasAVX2())) 303 ISD = ISD::MUL; 304 305 // A vector shift left by non uniform constant is converted 306 // into a vector multiply; the new multiply is eventually 307 // lowered into a sequence of shuffles and 2 x pmuludq. 308 if (VT == MVT::v4i32 && ST->hasSSE2()) 309 ISD = ISD::MUL; 310 } 311 312 static const CostTblEntry SSE2CostTable[] = { 313 // We don't correctly identify costs of casts because they are marked as 314 // custom. 315 // For some cases, where the shift amount is a scalar we would be able 316 // to generate better code. Unfortunately, when this is the case the value 317 // (the splat) will get hoisted out of the loop, thereby making it invisible 318 // to ISel. The cost model must return worst case assumptions because it is 319 // used for vectorization and we don't want to make vectorized code worse 320 // than scalar code. 321 { ISD::SHL, MVT::v16i8, 26 }, // cmpgtb sequence. 322 { ISD::SHL, MVT::v32i8, 2*26 }, // cmpgtb sequence. 323 { ISD::SHL, MVT::v8i16, 32 }, // cmpgtb sequence. 324 { ISD::SHL, MVT::v16i16, 2*32 }, // cmpgtb sequence. 325 { ISD::SHL, MVT::v4i32, 2*5 }, // We optimized this using mul. 326 { ISD::SHL, MVT::v8i32, 2*2*5 }, // We optimized this using mul. 327 { ISD::SHL, MVT::v2i64, 4 }, // splat+shuffle sequence. 328 { ISD::SHL, MVT::v4i64, 2*4 }, // splat+shuffle sequence. 329 330 { ISD::SRL, MVT::v16i8, 26 }, // cmpgtb sequence. 331 { ISD::SRL, MVT::v32i8, 2*26 }, // cmpgtb sequence. 332 { ISD::SRL, MVT::v8i16, 32 }, // cmpgtb sequence. 333 { ISD::SRL, MVT::v16i16, 2*32 }, // cmpgtb sequence. 334 { ISD::SRL, MVT::v4i32, 16 }, // Shift each lane + blend. 335 { ISD::SRL, MVT::v8i32, 2*16 }, // Shift each lane + blend. 336 { ISD::SRL, MVT::v2i64, 4 }, // splat+shuffle sequence. 337 { ISD::SRL, MVT::v4i64, 2*4 }, // splat+shuffle sequence. 338 339 { ISD::SRA, MVT::v16i8, 54 }, // unpacked cmpgtb sequence. 340 { ISD::SRA, MVT::v32i8, 2*54 }, // unpacked cmpgtb sequence. 341 { ISD::SRA, MVT::v8i16, 32 }, // cmpgtb sequence. 342 { ISD::SRA, MVT::v16i16, 2*32 }, // cmpgtb sequence. 343 { ISD::SRA, MVT::v4i32, 16 }, // Shift each lane + blend. 344 { ISD::SRA, MVT::v8i32, 2*16 }, // Shift each lane + blend. 345 { ISD::SRA, MVT::v2i64, 12 }, // srl/xor/sub sequence. 346 { ISD::SRA, MVT::v4i64, 2*12 }, // srl/xor/sub sequence. 347 348 // It is not a good idea to vectorize division. We have to scalarize it and 349 // in the process we will often end up having to spilling regular 350 // registers. The overhead of division is going to dominate most kernels 351 // anyways so try hard to prevent vectorization of division - it is 352 // generally a bad idea. Assume somewhat arbitrarily that we have to be able 353 // to hide "20 cycles" for each lane. 354 { ISD::SDIV, MVT::v16i8, 16*20 }, 355 { ISD::SDIV, MVT::v8i16, 8*20 }, 356 { ISD::SDIV, MVT::v4i32, 4*20 }, 357 { ISD::SDIV, MVT::v2i64, 2*20 }, 358 { ISD::UDIV, MVT::v16i8, 16*20 }, 359 { ISD::UDIV, MVT::v8i16, 8*20 }, 360 { ISD::UDIV, MVT::v4i32, 4*20 }, 361 { ISD::UDIV, MVT::v2i64, 2*20 }, 362 }; 363 364 if (ST->hasSSE2()) { 365 if (const auto *Entry = CostTableLookup(SSE2CostTable, ISD, LT.second)) 366 return LT.first * Entry->Cost; 367 } 368 369 static const CostTblEntry AVX1CostTable[] = { 370 // We don't have to scalarize unsupported ops. We can issue two half-sized 371 // operations and we only need to extract the upper YMM half. 372 // Two ops + 1 extract + 1 insert = 4. 373 { ISD::MUL, MVT::v16i16, 4 }, 374 { ISD::MUL, MVT::v8i32, 4 }, 375 { ISD::SUB, MVT::v8i32, 4 }, 376 { ISD::ADD, MVT::v8i32, 4 }, 377 { ISD::SUB, MVT::v4i64, 4 }, 378 { ISD::ADD, MVT::v4i64, 4 }, 379 // A v4i64 multiply is custom lowered as two split v2i64 vectors that then 380 // are lowered as a series of long multiplies(3), shifts(4) and adds(2) 381 // Because we believe v4i64 to be a legal type, we must also include the 382 // split factor of two in the cost table. Therefore, the cost here is 18 383 // instead of 9. 384 { ISD::MUL, MVT::v4i64, 18 }, 385 }; 386 387 // Look for AVX1 lowering tricks. 388 if (ST->hasAVX() && !ST->hasAVX2()) { 389 MVT VT = LT.second; 390 391 if (const auto *Entry = CostTableLookup(AVX1CostTable, ISD, VT)) 392 return LT.first * Entry->Cost; 393 } 394 395 // Custom lowering of vectors. 396 static const CostTblEntry CustomLowered[] = { 397 // A v2i64/v4i64 and multiply is custom lowered as a series of long 398 // multiplies(3), shifts(4) and adds(2). 399 { ISD::MUL, MVT::v2i64, 9 }, 400 { ISD::MUL, MVT::v4i64, 9 }, 401 }; 402 if (const auto *Entry = CostTableLookup(CustomLowered, ISD, LT.second)) 403 return LT.first * Entry->Cost; 404 405 // Special lowering of v4i32 mul on sse2, sse3: Lower v4i32 mul as 2x shuffle, 406 // 2x pmuludq, 2x shuffle. 407 if (ISD == ISD::MUL && LT.second == MVT::v4i32 && ST->hasSSE2() && 408 !ST->hasSSE41()) 409 return LT.first * 6; 410 411 // Fallback to the default implementation. 412 return BaseT::getArithmeticInstrCost(Opcode, Ty, Op1Info, Op2Info); 413 } 414 415 int X86TTIImpl::getShuffleCost(TTI::ShuffleKind Kind, Type *Tp, int Index, 416 Type *SubTp) { 417 // We only estimate the cost of reverse and alternate shuffles. 418 if (Kind != TTI::SK_Reverse && Kind != TTI::SK_Alternate) 419 return BaseT::getShuffleCost(Kind, Tp, Index, SubTp); 420 421 if (Kind == TTI::SK_Reverse) { 422 std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, Tp); 423 int Cost = 1; 424 if (LT.second.getSizeInBits() > 128) 425 Cost = 3; // Extract + insert + copy. 426 427 // Multiple by the number of parts. 428 return Cost * LT.first; 429 } 430 431 if (Kind == TTI::SK_Alternate) { 432 // 64-bit packed float vectors (v2f32) are widened to type v4f32. 433 // 64-bit packed integer vectors (v2i32) are promoted to type v2i64. 434 std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, Tp); 435 436 // The backend knows how to generate a single VEX.256 version of 437 // instruction VPBLENDW if the target supports AVX2. 438 if (ST->hasAVX2() && LT.second == MVT::v16i16) 439 return LT.first; 440 441 static const CostTblEntry AVXAltShuffleTbl[] = { 442 {ISD::VECTOR_SHUFFLE, MVT::v4i64, 1}, // vblendpd 443 {ISD::VECTOR_SHUFFLE, MVT::v4f64, 1}, // vblendpd 444 445 {ISD::VECTOR_SHUFFLE, MVT::v8i32, 1}, // vblendps 446 {ISD::VECTOR_SHUFFLE, MVT::v8f32, 1}, // vblendps 447 448 // This shuffle is custom lowered into a sequence of: 449 // 2x vextractf128 , 2x vpblendw , 1x vinsertf128 450 {ISD::VECTOR_SHUFFLE, MVT::v16i16, 5}, 451 452 // This shuffle is custom lowered into a long sequence of: 453 // 2x vextractf128 , 4x vpshufb , 2x vpor , 1x vinsertf128 454 {ISD::VECTOR_SHUFFLE, MVT::v32i8, 9} 455 }; 456 457 if (ST->hasAVX()) 458 if (const auto *Entry = CostTableLookup(AVXAltShuffleTbl, 459 ISD::VECTOR_SHUFFLE, LT.second)) 460 return LT.first * Entry->Cost; 461 462 static const CostTblEntry SSE41AltShuffleTbl[] = { 463 // These are lowered into movsd. 464 {ISD::VECTOR_SHUFFLE, MVT::v2i64, 1}, 465 {ISD::VECTOR_SHUFFLE, MVT::v2f64, 1}, 466 467 // packed float vectors with four elements are lowered into BLENDI dag 468 // nodes. A v4i32/v4f32 BLENDI generates a single 'blendps'/'blendpd'. 469 {ISD::VECTOR_SHUFFLE, MVT::v4i32, 1}, 470 {ISD::VECTOR_SHUFFLE, MVT::v4f32, 1}, 471 472 // This shuffle generates a single pshufw. 473 {ISD::VECTOR_SHUFFLE, MVT::v8i16, 1}, 474 475 // There is no instruction that matches a v16i8 alternate shuffle. 476 // The backend will expand it into the sequence 'pshufb + pshufb + or'. 477 {ISD::VECTOR_SHUFFLE, MVT::v16i8, 3} 478 }; 479 480 if (ST->hasSSE41()) 481 if (const auto *Entry = CostTableLookup(SSE41AltShuffleTbl, ISD::VECTOR_SHUFFLE, 482 LT.second)) 483 return LT.first * Entry->Cost; 484 485 static const CostTblEntry SSSE3AltShuffleTbl[] = { 486 {ISD::VECTOR_SHUFFLE, MVT::v2i64, 1}, // movsd 487 {ISD::VECTOR_SHUFFLE, MVT::v2f64, 1}, // movsd 488 489 // SSE3 doesn't have 'blendps'. The following shuffles are expanded into 490 // the sequence 'shufps + pshufd' 491 {ISD::VECTOR_SHUFFLE, MVT::v4i32, 2}, 492 {ISD::VECTOR_SHUFFLE, MVT::v4f32, 2}, 493 494 {ISD::VECTOR_SHUFFLE, MVT::v8i16, 3}, // pshufb + pshufb + or 495 {ISD::VECTOR_SHUFFLE, MVT::v16i8, 3} // pshufb + pshufb + or 496 }; 497 498 if (ST->hasSSSE3()) 499 if (const auto *Entry = CostTableLookup(SSSE3AltShuffleTbl, 500 ISD::VECTOR_SHUFFLE, LT.second)) 501 return LT.first * Entry->Cost; 502 503 static const CostTblEntry SSEAltShuffleTbl[] = { 504 {ISD::VECTOR_SHUFFLE, MVT::v2i64, 1}, // movsd 505 {ISD::VECTOR_SHUFFLE, MVT::v2f64, 1}, // movsd 506 507 {ISD::VECTOR_SHUFFLE, MVT::v4i32, 2}, // shufps + pshufd 508 {ISD::VECTOR_SHUFFLE, MVT::v4f32, 2}, // shufps + pshufd 509 510 // This is expanded into a long sequence of four extract + four insert. 511 {ISD::VECTOR_SHUFFLE, MVT::v8i16, 8}, // 4 x pextrw + 4 pinsrw. 512 513 // 8 x (pinsrw + pextrw + and + movb + movzb + or) 514 {ISD::VECTOR_SHUFFLE, MVT::v16i8, 48} 515 }; 516 517 // Fall-back (SSE3 and SSE2). 518 if (const auto *Entry = CostTableLookup(SSEAltShuffleTbl, 519 ISD::VECTOR_SHUFFLE, LT.second)) 520 return LT.first * Entry->Cost; 521 return BaseT::getShuffleCost(Kind, Tp, Index, SubTp); 522 } 523 524 return BaseT::getShuffleCost(Kind, Tp, Index, SubTp); 525 } 526 527 int X86TTIImpl::getCastInstrCost(unsigned Opcode, Type *Dst, Type *Src) { 528 int ISD = TLI->InstructionOpcodeToISD(Opcode); 529 assert(ISD && "Invalid opcode"); 530 531 // FIXME: Need a better design of the cost table to handle non-simple types of 532 // potential massive combinations (elem_num x src_type x dst_type). 533 534 static const TypeConversionCostTblEntry AVX512DQConversionTbl[] = { 535 { ISD::UINT_TO_FP, MVT::v2f64, MVT::v2i64, 1 }, 536 { ISD::UINT_TO_FP, MVT::v4f64, MVT::v4i64, 1 }, 537 { ISD::UINT_TO_FP, MVT::v8f64, MVT::v8i64, 1 }, 538 { ISD::UINT_TO_FP, MVT::v2f32, MVT::v2i64, 1 }, 539 { ISD::UINT_TO_FP, MVT::v4f32, MVT::v4i64, 1 }, 540 { ISD::UINT_TO_FP, MVT::v8f32, MVT::v8i64, 1 }, 541 542 { ISD::FP_TO_UINT, MVT::v2i64, MVT::v2f64, 1 }, 543 { ISD::FP_TO_UINT, MVT::v4i64, MVT::v4f64, 1 }, 544 { ISD::FP_TO_UINT, MVT::v8i64, MVT::v8f64, 1 }, 545 { ISD::FP_TO_UINT, MVT::v2i64, MVT::v2f32, 1 }, 546 { ISD::FP_TO_UINT, MVT::v4i64, MVT::v4f32, 1 }, 547 { ISD::FP_TO_UINT, MVT::v8i64, MVT::v8f32, 1 }, 548 }; 549 550 static const TypeConversionCostTblEntry AVX512FConversionTbl[] = { 551 { ISD::FP_EXTEND, MVT::v8f64, MVT::v8f32, 1 }, 552 { ISD::FP_EXTEND, MVT::v8f64, MVT::v16f32, 3 }, 553 { ISD::FP_ROUND, MVT::v8f32, MVT::v8f64, 1 }, 554 555 { ISD::TRUNCATE, MVT::v16i8, MVT::v16i32, 1 }, 556 { ISD::TRUNCATE, MVT::v16i16, MVT::v16i32, 1 }, 557 { ISD::TRUNCATE, MVT::v8i16, MVT::v8i64, 1 }, 558 { ISD::TRUNCATE, MVT::v8i32, MVT::v8i64, 1 }, 559 560 // v16i1 -> v16i32 - load + broadcast 561 { ISD::SIGN_EXTEND, MVT::v16i32, MVT::v16i1, 2 }, 562 { ISD::ZERO_EXTEND, MVT::v16i32, MVT::v16i1, 2 }, 563 564 { ISD::SIGN_EXTEND, MVT::v16i32, MVT::v16i8, 1 }, 565 { ISD::ZERO_EXTEND, MVT::v16i32, MVT::v16i8, 1 }, 566 { ISD::SIGN_EXTEND, MVT::v16i32, MVT::v16i16, 1 }, 567 { ISD::ZERO_EXTEND, MVT::v16i32, MVT::v16i16, 1 }, 568 { ISD::SIGN_EXTEND, MVT::v8i64, MVT::v8i32, 1 }, 569 { ISD::ZERO_EXTEND, MVT::v8i64, MVT::v8i32, 1 }, 570 { ISD::ZERO_EXTEND, MVT::v8i64, MVT::v8i16, 1 }, 571 { ISD::SIGN_EXTEND, MVT::v8i64, MVT::v8i16, 1 }, 572 573 { ISD::SINT_TO_FP, MVT::v16f32, MVT::v16i1, 3 }, 574 { ISD::SINT_TO_FP, MVT::v16f32, MVT::v16i8, 2 }, 575 { ISD::SINT_TO_FP, MVT::v16f32, MVT::v16i16, 2 }, 576 { ISD::SINT_TO_FP, MVT::v16f32, MVT::v16i32, 1 }, 577 { ISD::SINT_TO_FP, MVT::v8f64, MVT::v8i1, 4 }, 578 { ISD::SINT_TO_FP, MVT::v8f64, MVT::v8i8, 2 }, 579 { ISD::SINT_TO_FP, MVT::v8f64, MVT::v8i16, 2 }, 580 { ISD::SINT_TO_FP, MVT::v8f64, MVT::v8i32, 1 }, 581 582 { ISD::UINT_TO_FP, MVT::v16f32, MVT::v16i1, 3 }, 583 { ISD::UINT_TO_FP, MVT::v16f32, MVT::v16i8, 2 }, 584 { ISD::UINT_TO_FP, MVT::v16f32, MVT::v16i16, 2 }, 585 { ISD::UINT_TO_FP, MVT::v16f32, MVT::v16i32, 1 }, 586 { ISD::UINT_TO_FP, MVT::v8f32, MVT::v8i32, 1 }, 587 { ISD::UINT_TO_FP, MVT::v4f32, MVT::v4i32, 1 }, 588 { ISD::UINT_TO_FP, MVT::v8f64, MVT::v8i1, 4 }, 589 { ISD::UINT_TO_FP, MVT::v8f64, MVT::v8i16, 2 }, 590 { ISD::UINT_TO_FP, MVT::v8f64, MVT::v8i32, 1 }, 591 { ISD::UINT_TO_FP, MVT::v8f64, MVT::v8i8, 2 }, 592 { ISD::UINT_TO_FP, MVT::v8f32, MVT::v8i8, 2 }, 593 { ISD::UINT_TO_FP, MVT::v8f32, MVT::v8i16, 2 }, 594 { ISD::UINT_TO_FP, MVT::v4f64, MVT::v4i8, 2 }, 595 { ISD::UINT_TO_FP, MVT::v4f64, MVT::v4i16, 2 }, 596 { ISD::UINT_TO_FP, MVT::v4f64, MVT::v4i32, 1 }, 597 { ISD::UINT_TO_FP, MVT::v2f64, MVT::v2i8, 2 }, 598 { ISD::UINT_TO_FP, MVT::v2f64, MVT::v2i16, 5 }, 599 { ISD::UINT_TO_FP, MVT::v2f32, MVT::v2i32, 2 }, 600 { ISD::UINT_TO_FP, MVT::v2f64, MVT::v2i64, 5 }, 601 { ISD::UINT_TO_FP, MVT::v4f64, MVT::v4i64, 12 }, 602 { ISD::UINT_TO_FP, MVT::v8f64, MVT::v8i64, 26 }, 603 604 { ISD::FP_TO_UINT, MVT::v2i32, MVT::v2f32, 1 }, 605 { ISD::FP_TO_UINT, MVT::v4i32, MVT::v4f32, 1 }, 606 { ISD::FP_TO_UINT, MVT::v8i32, MVT::v8f32, 1 }, 607 { ISD::FP_TO_UINT, MVT::v16i32, MVT::v16f32, 1 }, 608 }; 609 610 static const TypeConversionCostTblEntry AVX2ConversionTbl[] = { 611 { ISD::SIGN_EXTEND, MVT::v16i16, MVT::v16i8, 1 }, 612 { ISD::ZERO_EXTEND, MVT::v16i16, MVT::v16i8, 1 }, 613 { ISD::SIGN_EXTEND, MVT::v8i32, MVT::v8i1, 3 }, 614 { ISD::ZERO_EXTEND, MVT::v8i32, MVT::v8i1, 3 }, 615 { ISD::SIGN_EXTEND, MVT::v8i32, MVT::v8i8, 3 }, 616 { ISD::ZERO_EXTEND, MVT::v8i32, MVT::v8i8, 3 }, 617 { ISD::SIGN_EXTEND, MVT::v8i32, MVT::v8i16, 1 }, 618 { ISD::ZERO_EXTEND, MVT::v8i32, MVT::v8i16, 1 }, 619 { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v4i1, 3 }, 620 { ISD::ZERO_EXTEND, MVT::v4i64, MVT::v4i1, 3 }, 621 { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v4i8, 3 }, 622 { ISD::ZERO_EXTEND, MVT::v4i64, MVT::v4i8, 3 }, 623 { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v4i16, 3 }, 624 { ISD::ZERO_EXTEND, MVT::v4i64, MVT::v4i16, 3 }, 625 { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v4i32, 1 }, 626 { ISD::ZERO_EXTEND, MVT::v4i64, MVT::v4i32, 1 }, 627 628 { ISD::TRUNCATE, MVT::v4i8, MVT::v4i64, 2 }, 629 { ISD::TRUNCATE, MVT::v4i16, MVT::v4i64, 2 }, 630 { ISD::TRUNCATE, MVT::v4i32, MVT::v4i64, 2 }, 631 { ISD::TRUNCATE, MVT::v8i8, MVT::v8i32, 2 }, 632 { ISD::TRUNCATE, MVT::v8i16, MVT::v8i32, 2 }, 633 { ISD::TRUNCATE, MVT::v8i32, MVT::v8i64, 4 }, 634 635 { ISD::FP_EXTEND, MVT::v8f64, MVT::v8f32, 3 }, 636 { ISD::FP_ROUND, MVT::v8f32, MVT::v8f64, 3 }, 637 638 { ISD::UINT_TO_FP, MVT::v8f32, MVT::v8i32, 8 }, 639 }; 640 641 static const TypeConversionCostTblEntry AVXConversionTbl[] = { 642 { ISD::SIGN_EXTEND, MVT::v16i16, MVT::v16i8, 4 }, 643 { ISD::ZERO_EXTEND, MVT::v16i16, MVT::v16i8, 4 }, 644 { ISD::SIGN_EXTEND, MVT::v8i32, MVT::v8i1, 7 }, 645 { ISD::ZERO_EXTEND, MVT::v8i32, MVT::v8i1, 4 }, 646 { ISD::SIGN_EXTEND, MVT::v8i32, MVT::v8i8, 7 }, 647 { ISD::ZERO_EXTEND, MVT::v8i32, MVT::v8i8, 4 }, 648 { ISD::SIGN_EXTEND, MVT::v8i32, MVT::v8i16, 4 }, 649 { ISD::ZERO_EXTEND, MVT::v8i32, MVT::v8i16, 4 }, 650 { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v4i1, 6 }, 651 { ISD::ZERO_EXTEND, MVT::v4i64, MVT::v4i1, 4 }, 652 { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v4i8, 6 }, 653 { ISD::ZERO_EXTEND, MVT::v4i64, MVT::v4i8, 4 }, 654 { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v4i16, 6 }, 655 { ISD::ZERO_EXTEND, MVT::v4i64, MVT::v4i16, 3 }, 656 { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v4i32, 4 }, 657 { ISD::ZERO_EXTEND, MVT::v4i64, MVT::v4i32, 4 }, 658 659 { ISD::TRUNCATE, MVT::v4i8, MVT::v4i64, 4 }, 660 { ISD::TRUNCATE, MVT::v4i16, MVT::v4i64, 4 }, 661 { ISD::TRUNCATE, MVT::v4i32, MVT::v4i64, 4 }, 662 { ISD::TRUNCATE, MVT::v8i8, MVT::v8i32, 4 }, 663 { ISD::TRUNCATE, MVT::v8i16, MVT::v8i32, 5 }, 664 { ISD::TRUNCATE, MVT::v16i8, MVT::v16i16, 4 }, 665 { ISD::TRUNCATE, MVT::v8i32, MVT::v8i64, 9 }, 666 667 { ISD::SINT_TO_FP, MVT::v8f32, MVT::v8i1, 8 }, 668 { ISD::SINT_TO_FP, MVT::v8f32, MVT::v8i8, 8 }, 669 { ISD::SINT_TO_FP, MVT::v8f32, MVT::v8i16, 5 }, 670 { ISD::SINT_TO_FP, MVT::v8f32, MVT::v8i32, 1 }, 671 { ISD::SINT_TO_FP, MVT::v4f32, MVT::v4i1, 3 }, 672 { ISD::SINT_TO_FP, MVT::v4f32, MVT::v4i8, 3 }, 673 { ISD::SINT_TO_FP, MVT::v4f32, MVT::v4i16, 3 }, 674 { ISD::SINT_TO_FP, MVT::v4f32, MVT::v4i32, 1 }, 675 { ISD::SINT_TO_FP, MVT::v4f64, MVT::v4i1, 3 }, 676 { ISD::SINT_TO_FP, MVT::v4f64, MVT::v4i8, 3 }, 677 { ISD::SINT_TO_FP, MVT::v4f64, MVT::v4i16, 3 }, 678 { ISD::SINT_TO_FP, MVT::v4f64, MVT::v4i32, 1 }, 679 680 { ISD::UINT_TO_FP, MVT::v8f32, MVT::v8i1, 6 }, 681 { ISD::UINT_TO_FP, MVT::v8f32, MVT::v8i8, 5 }, 682 { ISD::UINT_TO_FP, MVT::v8f32, MVT::v8i16, 5 }, 683 { ISD::UINT_TO_FP, MVT::v8f32, MVT::v8i32, 9 }, 684 { ISD::UINT_TO_FP, MVT::v4f32, MVT::v4i1, 7 }, 685 { ISD::UINT_TO_FP, MVT::v4f32, MVT::v4i8, 2 }, 686 { ISD::UINT_TO_FP, MVT::v4f32, MVT::v4i16, 2 }, 687 { ISD::UINT_TO_FP, MVT::v4f32, MVT::v4i32, 6 }, 688 { ISD::UINT_TO_FP, MVT::v4f64, MVT::v4i1, 7 }, 689 { ISD::UINT_TO_FP, MVT::v4f64, MVT::v4i8, 2 }, 690 { ISD::UINT_TO_FP, MVT::v4f64, MVT::v4i16, 2 }, 691 { ISD::UINT_TO_FP, MVT::v4f64, MVT::v4i32, 6 }, 692 // The generic code to compute the scalar overhead is currently broken. 693 // Workaround this limitation by estimating the scalarization overhead 694 // here. We have roughly 10 instructions per scalar element. 695 // Multiply that by the vector width. 696 // FIXME: remove that when PR19268 is fixed. 697 { ISD::UINT_TO_FP, MVT::v2f64, MVT::v2i64, 2*10 }, 698 { ISD::UINT_TO_FP, MVT::v4f64, MVT::v4i64, 4*10 }, 699 700 { ISD::FP_TO_SINT, MVT::v8i8, MVT::v8f32, 7 }, 701 { ISD::FP_TO_SINT, MVT::v4i8, MVT::v4f32, 1 }, 702 // This node is expanded into scalarized operations but BasicTTI is overly 703 // optimistic estimating its cost. It computes 3 per element (one 704 // vector-extract, one scalar conversion and one vector-insert). The 705 // problem is that the inserts form a read-modify-write chain so latency 706 // should be factored in too. Inflating the cost per element by 1. 707 { ISD::FP_TO_UINT, MVT::v8i32, MVT::v8f32, 8*4 }, 708 { ISD::FP_TO_UINT, MVT::v4i32, MVT::v4f64, 4*4 }, 709 }; 710 711 static const TypeConversionCostTblEntry SSE41ConversionTbl[] = { 712 { ISD::ZERO_EXTEND, MVT::v16i32, MVT::v16i16, 4 }, 713 { ISD::SIGN_EXTEND, MVT::v16i32, MVT::v16i16, 4 }, 714 { ISD::ZERO_EXTEND, MVT::v8i32, MVT::v8i16, 2 }, 715 { ISD::SIGN_EXTEND, MVT::v8i32, MVT::v8i16, 2 }, 716 { ISD::ZERO_EXTEND, MVT::v4i32, MVT::v4i16, 1 }, 717 { ISD::SIGN_EXTEND, MVT::v4i32, MVT::v4i16, 1 }, 718 { ISD::ZERO_EXTEND, MVT::v16i32, MVT::v16i8, 4 }, 719 { ISD::SIGN_EXTEND, MVT::v16i32, MVT::v16i8, 4 }, 720 { ISD::ZERO_EXTEND, MVT::v8i32, MVT::v8i8, 2 }, 721 { ISD::SIGN_EXTEND, MVT::v8i32, MVT::v8i8, 2 }, 722 { ISD::ZERO_EXTEND, MVT::v4i32, MVT::v4i8, 1 }, 723 { ISD::SIGN_EXTEND, MVT::v4i32, MVT::v4i8, 1 }, 724 { ISD::ZERO_EXTEND, MVT::v16i16, MVT::v16i8, 2 }, 725 { ISD::SIGN_EXTEND, MVT::v16i16, MVT::v16i8, 2 }, 726 { ISD::ZERO_EXTEND, MVT::v8i16, MVT::v8i8, 1 }, 727 { ISD::SIGN_EXTEND, MVT::v8i16, MVT::v8i8, 1 }, 728 { ISD::ZERO_EXTEND, MVT::v4i16, MVT::v4i8, 1 }, 729 { ISD::SIGN_EXTEND, MVT::v4i16, MVT::v4i8, 2 }, 730 731 { ISD::TRUNCATE, MVT::v16i16, MVT::v16i32, 6 }, 732 { ISD::TRUNCATE, MVT::v8i16, MVT::v8i32, 3 }, 733 { ISD::TRUNCATE, MVT::v4i16, MVT::v4i32, 1 }, 734 { ISD::TRUNCATE, MVT::v8i8, MVT::v8i32, 3 }, 735 { ISD::TRUNCATE, MVT::v4i8, MVT::v4i32, 1 }, 736 { ISD::TRUNCATE, MVT::v8i8, MVT::v8i16, 1 }, 737 { ISD::TRUNCATE, MVT::v4i8, MVT::v4i16, 2 }, 738 }; 739 740 static const TypeConversionCostTblEntry SSE2ConversionTbl[] = { 741 // These are somewhat magic numbers justified by looking at the output of 742 // Intel's IACA, running some kernels and making sure when we take 743 // legalization into account the throughput will be overestimated. 744 { ISD::UINT_TO_FP, MVT::v2f64, MVT::v2i64, 2*10 }, 745 { ISD::UINT_TO_FP, MVT::v2f64, MVT::v4i32, 4*10 }, 746 { ISD::UINT_TO_FP, MVT::v2f64, MVT::v8i16, 8*10 }, 747 { ISD::UINT_TO_FP, MVT::v2f64, MVT::v16i8, 16*10 }, 748 { ISD::SINT_TO_FP, MVT::v2f64, MVT::v2i64, 2*10 }, 749 { ISD::SINT_TO_FP, MVT::v2f64, MVT::v4i32, 4*10 }, 750 { ISD::SINT_TO_FP, MVT::v2f64, MVT::v8i16, 8*10 }, 751 { ISD::SINT_TO_FP, MVT::v2f64, MVT::v16i8, 16*10 }, 752 // There are faster sequences for float conversions. 753 { ISD::UINT_TO_FP, MVT::v4f32, MVT::v2i64, 15 }, 754 { ISD::UINT_TO_FP, MVT::v4f32, MVT::v4i32, 8 }, 755 { ISD::UINT_TO_FP, MVT::v4f32, MVT::v8i16, 15 }, 756 { ISD::UINT_TO_FP, MVT::v4f32, MVT::v16i8, 8 }, 757 { ISD::SINT_TO_FP, MVT::v4f32, MVT::v2i64, 15 }, 758 { ISD::SINT_TO_FP, MVT::v4f32, MVT::v4i32, 15 }, 759 { ISD::SINT_TO_FP, MVT::v4f32, MVT::v8i16, 15 }, 760 { ISD::SINT_TO_FP, MVT::v4f32, MVT::v16i8, 8 }, 761 762 { ISD::ZERO_EXTEND, MVT::v16i32, MVT::v16i16, 6 }, 763 { ISD::SIGN_EXTEND, MVT::v16i32, MVT::v16i16, 8 }, 764 { ISD::ZERO_EXTEND, MVT::v8i32, MVT::v8i16, 3 }, 765 { ISD::SIGN_EXTEND, MVT::v8i32, MVT::v8i16, 4 }, 766 { ISD::ZERO_EXTEND, MVT::v4i32, MVT::v4i16, 1 }, 767 { ISD::SIGN_EXTEND, MVT::v4i32, MVT::v4i16, 2 }, 768 { ISD::ZERO_EXTEND, MVT::v16i32, MVT::v16i8, 9 }, 769 { ISD::SIGN_EXTEND, MVT::v16i32, MVT::v16i8, 12 }, 770 { ISD::ZERO_EXTEND, MVT::v8i32, MVT::v8i8, 6 }, 771 { ISD::SIGN_EXTEND, MVT::v8i32, MVT::v8i8, 6 }, 772 { ISD::ZERO_EXTEND, MVT::v4i32, MVT::v4i8, 2 }, 773 { ISD::SIGN_EXTEND, MVT::v4i32, MVT::v4i8, 3 }, 774 { ISD::ZERO_EXTEND, MVT::v16i16, MVT::v16i8, 3 }, 775 { ISD::SIGN_EXTEND, MVT::v16i16, MVT::v16i8, 4 }, 776 { ISD::ZERO_EXTEND, MVT::v8i16, MVT::v8i8, 1 }, 777 { ISD::SIGN_EXTEND, MVT::v8i16, MVT::v8i8, 2 }, 778 { ISD::ZERO_EXTEND, MVT::v4i16, MVT::v4i8, 1 }, 779 { ISD::SIGN_EXTEND, MVT::v4i16, MVT::v4i8, 6 }, 780 781 { ISD::TRUNCATE, MVT::v16i16, MVT::v16i32, 10 }, 782 { ISD::TRUNCATE, MVT::v8i16, MVT::v8i32, 5 }, 783 { ISD::TRUNCATE, MVT::v4i16, MVT::v4i32, 3 }, 784 { ISD::TRUNCATE, MVT::v16i8, MVT::v16i32, 7 }, 785 { ISD::TRUNCATE, MVT::v8i8, MVT::v8i32, 4 }, 786 { ISD::TRUNCATE, MVT::v4i8, MVT::v4i32, 3 }, 787 { ISD::TRUNCATE, MVT::v16i8, MVT::v16i16, 3 }, 788 { ISD::TRUNCATE, MVT::v8i8, MVT::v8i16, 2 }, 789 { ISD::TRUNCATE, MVT::v4i8, MVT::v4i16, 4 }, 790 }; 791 792 std::pair<int, MVT> LTSrc = TLI->getTypeLegalizationCost(DL, Src); 793 std::pair<int, MVT> LTDest = TLI->getTypeLegalizationCost(DL, Dst); 794 795 if (ST->hasSSE2() && !ST->hasAVX()) { 796 if (const auto *Entry = ConvertCostTableLookup(SSE2ConversionTbl, ISD, 797 LTDest.second, LTSrc.second)) 798 return LTSrc.first * Entry->Cost; 799 } 800 801 EVT SrcTy = TLI->getValueType(DL, Src); 802 EVT DstTy = TLI->getValueType(DL, Dst); 803 804 // The function getSimpleVT only handles simple value types. 805 if (!SrcTy.isSimple() || !DstTy.isSimple()) 806 return BaseT::getCastInstrCost(Opcode, Dst, Src); 807 808 if (ST->hasDQI()) 809 if (const auto *Entry = ConvertCostTableLookup(AVX512DQConversionTbl, ISD, 810 DstTy.getSimpleVT(), 811 SrcTy.getSimpleVT())) 812 return Entry->Cost; 813 814 if (ST->hasAVX512()) 815 if (const auto *Entry = ConvertCostTableLookup(AVX512FConversionTbl, ISD, 816 DstTy.getSimpleVT(), 817 SrcTy.getSimpleVT())) 818 return Entry->Cost; 819 820 if (ST->hasAVX2()) { 821 if (const auto *Entry = ConvertCostTableLookup(AVX2ConversionTbl, ISD, 822 DstTy.getSimpleVT(), 823 SrcTy.getSimpleVT())) 824 return Entry->Cost; 825 } 826 827 if (ST->hasAVX()) { 828 if (const auto *Entry = ConvertCostTableLookup(AVXConversionTbl, ISD, 829 DstTy.getSimpleVT(), 830 SrcTy.getSimpleVT())) 831 return Entry->Cost; 832 } 833 834 if (ST->hasSSE41()) { 835 if (const auto *Entry = ConvertCostTableLookup(SSE41ConversionTbl, ISD, 836 DstTy.getSimpleVT(), 837 SrcTy.getSimpleVT())) 838 return Entry->Cost; 839 } 840 841 if (ST->hasSSE2()) { 842 if (const auto *Entry = ConvertCostTableLookup(SSE2ConversionTbl, ISD, 843 DstTy.getSimpleVT(), 844 SrcTy.getSimpleVT())) 845 return Entry->Cost; 846 } 847 848 return BaseT::getCastInstrCost(Opcode, Dst, Src); 849 } 850 851 int X86TTIImpl::getCmpSelInstrCost(unsigned Opcode, Type *ValTy, Type *CondTy) { 852 // Legalize the type. 853 std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, ValTy); 854 855 MVT MTy = LT.second; 856 857 int ISD = TLI->InstructionOpcodeToISD(Opcode); 858 assert(ISD && "Invalid opcode"); 859 860 static const CostTblEntry SSE2CostTbl[] = { 861 { ISD::SETCC, MVT::v2i64, 8 }, 862 { ISD::SETCC, MVT::v4i32, 1 }, 863 { ISD::SETCC, MVT::v8i16, 1 }, 864 { ISD::SETCC, MVT::v16i8, 1 }, 865 }; 866 867 static const CostTblEntry SSE42CostTbl[] = { 868 { ISD::SETCC, MVT::v2f64, 1 }, 869 { ISD::SETCC, MVT::v4f32, 1 }, 870 { ISD::SETCC, MVT::v2i64, 1 }, 871 }; 872 873 static const CostTblEntry AVX1CostTbl[] = { 874 { ISD::SETCC, MVT::v4f64, 1 }, 875 { ISD::SETCC, MVT::v8f32, 1 }, 876 // AVX1 does not support 8-wide integer compare. 877 { ISD::SETCC, MVT::v4i64, 4 }, 878 { ISD::SETCC, MVT::v8i32, 4 }, 879 { ISD::SETCC, MVT::v16i16, 4 }, 880 { ISD::SETCC, MVT::v32i8, 4 }, 881 }; 882 883 static const CostTblEntry AVX2CostTbl[] = { 884 { ISD::SETCC, MVT::v4i64, 1 }, 885 { ISD::SETCC, MVT::v8i32, 1 }, 886 { ISD::SETCC, MVT::v16i16, 1 }, 887 { ISD::SETCC, MVT::v32i8, 1 }, 888 }; 889 890 static const CostTblEntry AVX512CostTbl[] = { 891 { ISD::SETCC, MVT::v8i64, 1 }, 892 { ISD::SETCC, MVT::v16i32, 1 }, 893 { ISD::SETCC, MVT::v8f64, 1 }, 894 { ISD::SETCC, MVT::v16f32, 1 }, 895 }; 896 897 if (ST->hasAVX512()) 898 if (const auto *Entry = CostTableLookup(AVX512CostTbl, ISD, MTy)) 899 return LT.first * Entry->Cost; 900 901 if (ST->hasAVX2()) 902 if (const auto *Entry = CostTableLookup(AVX2CostTbl, ISD, MTy)) 903 return LT.first * Entry->Cost; 904 905 if (ST->hasAVX()) 906 if (const auto *Entry = CostTableLookup(AVX1CostTbl, ISD, MTy)) 907 return LT.first * Entry->Cost; 908 909 if (ST->hasSSE42()) 910 if (const auto *Entry = CostTableLookup(SSE42CostTbl, ISD, MTy)) 911 return LT.first * Entry->Cost; 912 913 if (ST->hasSSE2()) 914 if (const auto *Entry = CostTableLookup(SSE2CostTbl, ISD, MTy)) 915 return LT.first * Entry->Cost; 916 917 return BaseT::getCmpSelInstrCost(Opcode, ValTy, CondTy); 918 } 919 920 int X86TTIImpl::getVectorInstrCost(unsigned Opcode, Type *Val, unsigned Index) { 921 assert(Val->isVectorTy() && "This must be a vector type"); 922 923 if (Index != -1U) { 924 // Legalize the type. 925 std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, Val); 926 927 // This type is legalized to a scalar type. 928 if (!LT.second.isVector()) 929 return 0; 930 931 // The type may be split. Normalize the index to the new type. 932 unsigned Width = LT.second.getVectorNumElements(); 933 Index = Index % Width; 934 935 // Floating point scalars are already located in index #0. 936 if (Val->getScalarType()->isFloatingPointTy() && Index == 0) 937 return 0; 938 } 939 940 return BaseT::getVectorInstrCost(Opcode, Val, Index); 941 } 942 943 int X86TTIImpl::getScalarizationOverhead(Type *Ty, bool Insert, bool Extract) { 944 assert (Ty->isVectorTy() && "Can only scalarize vectors"); 945 int Cost = 0; 946 947 for (int i = 0, e = Ty->getVectorNumElements(); i < e; ++i) { 948 if (Insert) 949 Cost += getVectorInstrCost(Instruction::InsertElement, Ty, i); 950 if (Extract) 951 Cost += getVectorInstrCost(Instruction::ExtractElement, Ty, i); 952 } 953 954 return Cost; 955 } 956 957 int X86TTIImpl::getMemoryOpCost(unsigned Opcode, Type *Src, unsigned Alignment, 958 unsigned AddressSpace) { 959 // Handle non-power-of-two vectors such as <3 x float> 960 if (VectorType *VTy = dyn_cast<VectorType>(Src)) { 961 unsigned NumElem = VTy->getVectorNumElements(); 962 963 // Handle a few common cases: 964 // <3 x float> 965 if (NumElem == 3 && VTy->getScalarSizeInBits() == 32) 966 // Cost = 64 bit store + extract + 32 bit store. 967 return 3; 968 969 // <3 x double> 970 if (NumElem == 3 && VTy->getScalarSizeInBits() == 64) 971 // Cost = 128 bit store + unpack + 64 bit store. 972 return 3; 973 974 // Assume that all other non-power-of-two numbers are scalarized. 975 if (!isPowerOf2_32(NumElem)) { 976 int Cost = BaseT::getMemoryOpCost(Opcode, VTy->getScalarType(), Alignment, 977 AddressSpace); 978 int SplitCost = getScalarizationOverhead(Src, Opcode == Instruction::Load, 979 Opcode == Instruction::Store); 980 return NumElem * Cost + SplitCost; 981 } 982 } 983 984 // Legalize the type. 985 std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, Src); 986 assert((Opcode == Instruction::Load || Opcode == Instruction::Store) && 987 "Invalid Opcode"); 988 989 // Each load/store unit costs 1. 990 int Cost = LT.first * 1; 991 992 // This isn't exactly right. We're using slow unaligned 32-byte accesses as a 993 // proxy for a double-pumped AVX memory interface such as on Sandybridge. 994 if (LT.second.getStoreSize() == 32 && ST->isUnalignedMem32Slow()) 995 Cost *= 2; 996 997 return Cost; 998 } 999 1000 int X86TTIImpl::getMaskedMemoryOpCost(unsigned Opcode, Type *SrcTy, 1001 unsigned Alignment, 1002 unsigned AddressSpace) { 1003 VectorType *SrcVTy = dyn_cast<VectorType>(SrcTy); 1004 if (!SrcVTy) 1005 // To calculate scalar take the regular cost, without mask 1006 return getMemoryOpCost(Opcode, SrcTy, Alignment, AddressSpace); 1007 1008 unsigned NumElem = SrcVTy->getVectorNumElements(); 1009 VectorType *MaskTy = 1010 VectorType::get(Type::getInt8Ty(SrcVTy->getContext()), NumElem); 1011 if ((Opcode == Instruction::Load && !isLegalMaskedLoad(SrcVTy)) || 1012 (Opcode == Instruction::Store && !isLegalMaskedStore(SrcVTy)) || 1013 !isPowerOf2_32(NumElem)) { 1014 // Scalarization 1015 int MaskSplitCost = getScalarizationOverhead(MaskTy, false, true); 1016 int ScalarCompareCost = getCmpSelInstrCost( 1017 Instruction::ICmp, Type::getInt8Ty(SrcVTy->getContext()), nullptr); 1018 int BranchCost = getCFInstrCost(Instruction::Br); 1019 int MaskCmpCost = NumElem * (BranchCost + ScalarCompareCost); 1020 1021 int ValueSplitCost = getScalarizationOverhead( 1022 SrcVTy, Opcode == Instruction::Load, Opcode == Instruction::Store); 1023 int MemopCost = 1024 NumElem * BaseT::getMemoryOpCost(Opcode, SrcVTy->getScalarType(), 1025 Alignment, AddressSpace); 1026 return MemopCost + ValueSplitCost + MaskSplitCost + MaskCmpCost; 1027 } 1028 1029 // Legalize the type. 1030 std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, SrcVTy); 1031 auto VT = TLI->getValueType(DL, SrcVTy); 1032 int Cost = 0; 1033 if (VT.isSimple() && LT.second != VT.getSimpleVT() && 1034 LT.second.getVectorNumElements() == NumElem) 1035 // Promotion requires expand/truncate for data and a shuffle for mask. 1036 Cost += getShuffleCost(TTI::SK_Alternate, SrcVTy, 0, nullptr) + 1037 getShuffleCost(TTI::SK_Alternate, MaskTy, 0, nullptr); 1038 1039 else if (LT.second.getVectorNumElements() > NumElem) { 1040 VectorType *NewMaskTy = VectorType::get(MaskTy->getVectorElementType(), 1041 LT.second.getVectorNumElements()); 1042 // Expanding requires fill mask with zeroes 1043 Cost += getShuffleCost(TTI::SK_InsertSubvector, NewMaskTy, 0, MaskTy); 1044 } 1045 if (!ST->hasAVX512()) 1046 return Cost + LT.first*4; // Each maskmov costs 4 1047 1048 // AVX-512 masked load/store is cheapper 1049 return Cost+LT.first; 1050 } 1051 1052 int X86TTIImpl::getAddressComputationCost(Type *Ty, bool IsComplex) { 1053 // Address computations in vectorized code with non-consecutive addresses will 1054 // likely result in more instructions compared to scalar code where the 1055 // computation can more often be merged into the index mode. The resulting 1056 // extra micro-ops can significantly decrease throughput. 1057 unsigned NumVectorInstToHideOverhead = 10; 1058 1059 if (Ty->isVectorTy() && IsComplex) 1060 return NumVectorInstToHideOverhead; 1061 1062 return BaseT::getAddressComputationCost(Ty, IsComplex); 1063 } 1064 1065 int X86TTIImpl::getReductionCost(unsigned Opcode, Type *ValTy, 1066 bool IsPairwise) { 1067 1068 std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, ValTy); 1069 1070 MVT MTy = LT.second; 1071 1072 int ISD = TLI->InstructionOpcodeToISD(Opcode); 1073 assert(ISD && "Invalid opcode"); 1074 1075 // We use the Intel Architecture Code Analyzer(IACA) to measure the throughput 1076 // and make it as the cost. 1077 1078 static const CostTblEntry SSE42CostTblPairWise[] = { 1079 { ISD::FADD, MVT::v2f64, 2 }, 1080 { ISD::FADD, MVT::v4f32, 4 }, 1081 { ISD::ADD, MVT::v2i64, 2 }, // The data reported by the IACA tool is "1.6". 1082 { ISD::ADD, MVT::v4i32, 3 }, // The data reported by the IACA tool is "3.5". 1083 { ISD::ADD, MVT::v8i16, 5 }, 1084 }; 1085 1086 static const CostTblEntry AVX1CostTblPairWise[] = { 1087 { ISD::FADD, MVT::v4f32, 4 }, 1088 { ISD::FADD, MVT::v4f64, 5 }, 1089 { ISD::FADD, MVT::v8f32, 7 }, 1090 { ISD::ADD, MVT::v2i64, 1 }, // The data reported by the IACA tool is "1.5". 1091 { ISD::ADD, MVT::v4i32, 3 }, // The data reported by the IACA tool is "3.5". 1092 { ISD::ADD, MVT::v4i64, 5 }, // The data reported by the IACA tool is "4.8". 1093 { ISD::ADD, MVT::v8i16, 5 }, 1094 { ISD::ADD, MVT::v8i32, 5 }, 1095 }; 1096 1097 static const CostTblEntry SSE42CostTblNoPairWise[] = { 1098 { ISD::FADD, MVT::v2f64, 2 }, 1099 { ISD::FADD, MVT::v4f32, 4 }, 1100 { ISD::ADD, MVT::v2i64, 2 }, // The data reported by the IACA tool is "1.6". 1101 { ISD::ADD, MVT::v4i32, 3 }, // The data reported by the IACA tool is "3.3". 1102 { ISD::ADD, MVT::v8i16, 4 }, // The data reported by the IACA tool is "4.3". 1103 }; 1104 1105 static const CostTblEntry AVX1CostTblNoPairWise[] = { 1106 { ISD::FADD, MVT::v4f32, 3 }, 1107 { ISD::FADD, MVT::v4f64, 3 }, 1108 { ISD::FADD, MVT::v8f32, 4 }, 1109 { ISD::ADD, MVT::v2i64, 1 }, // The data reported by the IACA tool is "1.5". 1110 { ISD::ADD, MVT::v4i32, 3 }, // The data reported by the IACA tool is "2.8". 1111 { ISD::ADD, MVT::v4i64, 3 }, 1112 { ISD::ADD, MVT::v8i16, 4 }, 1113 { ISD::ADD, MVT::v8i32, 5 }, 1114 }; 1115 1116 if (IsPairwise) { 1117 if (ST->hasAVX()) 1118 if (const auto *Entry = CostTableLookup(AVX1CostTblPairWise, ISD, MTy)) 1119 return LT.first * Entry->Cost; 1120 1121 if (ST->hasSSE42()) 1122 if (const auto *Entry = CostTableLookup(SSE42CostTblPairWise, ISD, MTy)) 1123 return LT.first * Entry->Cost; 1124 } else { 1125 if (ST->hasAVX()) 1126 if (const auto *Entry = CostTableLookup(AVX1CostTblNoPairWise, ISD, MTy)) 1127 return LT.first * Entry->Cost; 1128 1129 if (ST->hasSSE42()) 1130 if (const auto *Entry = CostTableLookup(SSE42CostTblNoPairWise, ISD, MTy)) 1131 return LT.first * Entry->Cost; 1132 } 1133 1134 return BaseT::getReductionCost(Opcode, ValTy, IsPairwise); 1135 } 1136 1137 /// \brief Calculate the cost of materializing a 64-bit value. This helper 1138 /// method might only calculate a fraction of a larger immediate. Therefore it 1139 /// is valid to return a cost of ZERO. 1140 int X86TTIImpl::getIntImmCost(int64_t Val) { 1141 if (Val == 0) 1142 return TTI::TCC_Free; 1143 1144 if (isInt<32>(Val)) 1145 return TTI::TCC_Basic; 1146 1147 return 2 * TTI::TCC_Basic; 1148 } 1149 1150 int X86TTIImpl::getIntImmCost(const APInt &Imm, Type *Ty) { 1151 assert(Ty->isIntegerTy()); 1152 1153 unsigned BitSize = Ty->getPrimitiveSizeInBits(); 1154 if (BitSize == 0) 1155 return ~0U; 1156 1157 // Never hoist constants larger than 128bit, because this might lead to 1158 // incorrect code generation or assertions in codegen. 1159 // Fixme: Create a cost model for types larger than i128 once the codegen 1160 // issues have been fixed. 1161 if (BitSize > 128) 1162 return TTI::TCC_Free; 1163 1164 if (Imm == 0) 1165 return TTI::TCC_Free; 1166 1167 // Sign-extend all constants to a multiple of 64-bit. 1168 APInt ImmVal = Imm; 1169 if (BitSize & 0x3f) 1170 ImmVal = Imm.sext((BitSize + 63) & ~0x3fU); 1171 1172 // Split the constant into 64-bit chunks and calculate the cost for each 1173 // chunk. 1174 int Cost = 0; 1175 for (unsigned ShiftVal = 0; ShiftVal < BitSize; ShiftVal += 64) { 1176 APInt Tmp = ImmVal.ashr(ShiftVal).sextOrTrunc(64); 1177 int64_t Val = Tmp.getSExtValue(); 1178 Cost += getIntImmCost(Val); 1179 } 1180 // We need at least one instruction to materialize the constant. 1181 return std::max(1, Cost); 1182 } 1183 1184 int X86TTIImpl::getIntImmCost(unsigned Opcode, unsigned Idx, const APInt &Imm, 1185 Type *Ty) { 1186 assert(Ty->isIntegerTy()); 1187 1188 unsigned BitSize = Ty->getPrimitiveSizeInBits(); 1189 // There is no cost model for constants with a bit size of 0. Return TCC_Free 1190 // here, so that constant hoisting will ignore this constant. 1191 if (BitSize == 0) 1192 return TTI::TCC_Free; 1193 1194 unsigned ImmIdx = ~0U; 1195 switch (Opcode) { 1196 default: 1197 return TTI::TCC_Free; 1198 case Instruction::GetElementPtr: 1199 // Always hoist the base address of a GetElementPtr. This prevents the 1200 // creation of new constants for every base constant that gets constant 1201 // folded with the offset. 1202 if (Idx == 0) 1203 return 2 * TTI::TCC_Basic; 1204 return TTI::TCC_Free; 1205 case Instruction::Store: 1206 ImmIdx = 0; 1207 break; 1208 case Instruction::ICmp: 1209 // This is an imperfect hack to prevent constant hoisting of 1210 // compares that might be trying to check if a 64-bit value fits in 1211 // 32-bits. The backend can optimize these cases using a right shift by 32. 1212 // Ideally we would check the compare predicate here. There also other 1213 // similar immediates the backend can use shifts for. 1214 if (Idx == 1 && Imm.getBitWidth() == 64) { 1215 uint64_t ImmVal = Imm.getZExtValue(); 1216 if (ImmVal == 0x100000000ULL || ImmVal == 0xffffffff) 1217 return TTI::TCC_Free; 1218 } 1219 ImmIdx = 1; 1220 break; 1221 case Instruction::And: 1222 // We support 64-bit ANDs with immediates with 32-bits of leading zeroes 1223 // by using a 32-bit operation with implicit zero extension. Detect such 1224 // immediates here as the normal path expects bit 31 to be sign extended. 1225 if (Idx == 1 && Imm.getBitWidth() == 64 && isUInt<32>(Imm.getZExtValue())) 1226 return TTI::TCC_Free; 1227 // Fallthrough 1228 case Instruction::Add: 1229 case Instruction::Sub: 1230 case Instruction::Mul: 1231 case Instruction::UDiv: 1232 case Instruction::SDiv: 1233 case Instruction::URem: 1234 case Instruction::SRem: 1235 case Instruction::Or: 1236 case Instruction::Xor: 1237 ImmIdx = 1; 1238 break; 1239 // Always return TCC_Free for the shift value of a shift instruction. 1240 case Instruction::Shl: 1241 case Instruction::LShr: 1242 case Instruction::AShr: 1243 if (Idx == 1) 1244 return TTI::TCC_Free; 1245 break; 1246 case Instruction::Trunc: 1247 case Instruction::ZExt: 1248 case Instruction::SExt: 1249 case Instruction::IntToPtr: 1250 case Instruction::PtrToInt: 1251 case Instruction::BitCast: 1252 case Instruction::PHI: 1253 case Instruction::Call: 1254 case Instruction::Select: 1255 case Instruction::Ret: 1256 case Instruction::Load: 1257 break; 1258 } 1259 1260 if (Idx == ImmIdx) { 1261 int NumConstants = (BitSize + 63) / 64; 1262 int Cost = X86TTIImpl::getIntImmCost(Imm, Ty); 1263 return (Cost <= NumConstants * TTI::TCC_Basic) 1264 ? static_cast<int>(TTI::TCC_Free) 1265 : Cost; 1266 } 1267 1268 return X86TTIImpl::getIntImmCost(Imm, Ty); 1269 } 1270 1271 int X86TTIImpl::getIntImmCost(Intrinsic::ID IID, unsigned Idx, const APInt &Imm, 1272 Type *Ty) { 1273 assert(Ty->isIntegerTy()); 1274 1275 unsigned BitSize = Ty->getPrimitiveSizeInBits(); 1276 // There is no cost model for constants with a bit size of 0. Return TCC_Free 1277 // here, so that constant hoisting will ignore this constant. 1278 if (BitSize == 0) 1279 return TTI::TCC_Free; 1280 1281 switch (IID) { 1282 default: 1283 return TTI::TCC_Free; 1284 case Intrinsic::sadd_with_overflow: 1285 case Intrinsic::uadd_with_overflow: 1286 case Intrinsic::ssub_with_overflow: 1287 case Intrinsic::usub_with_overflow: 1288 case Intrinsic::smul_with_overflow: 1289 case Intrinsic::umul_with_overflow: 1290 if ((Idx == 1) && Imm.getBitWidth() <= 64 && isInt<32>(Imm.getSExtValue())) 1291 return TTI::TCC_Free; 1292 break; 1293 case Intrinsic::experimental_stackmap: 1294 if ((Idx < 2) || (Imm.getBitWidth() <= 64 && isInt<64>(Imm.getSExtValue()))) 1295 return TTI::TCC_Free; 1296 break; 1297 case Intrinsic::experimental_patchpoint_void: 1298 case Intrinsic::experimental_patchpoint_i64: 1299 if ((Idx < 4) || (Imm.getBitWidth() <= 64 && isInt<64>(Imm.getSExtValue()))) 1300 return TTI::TCC_Free; 1301 break; 1302 } 1303 return X86TTIImpl::getIntImmCost(Imm, Ty); 1304 } 1305 1306 // Return an average cost of Gather / Scatter instruction, maybe improved later 1307 int X86TTIImpl::getGSVectorCost(unsigned Opcode, Type *SrcVTy, Value *Ptr, 1308 unsigned Alignment, unsigned AddressSpace) { 1309 1310 assert(isa<VectorType>(SrcVTy) && "Unexpected type in getGSVectorCost"); 1311 unsigned VF = SrcVTy->getVectorNumElements(); 1312 1313 // Try to reduce index size from 64 bit (default for GEP) 1314 // to 32. It is essential for VF 16. If the index can't be reduced to 32, the 1315 // operation will use 16 x 64 indices which do not fit in a zmm and needs 1316 // to split. Also check that the base pointer is the same for all lanes, 1317 // and that there's at most one variable index. 1318 auto getIndexSizeInBits = [](Value *Ptr, const DataLayout& DL) { 1319 unsigned IndexSize = DL.getPointerSizeInBits(); 1320 GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr); 1321 if (IndexSize < 64 || !GEP) 1322 return IndexSize; 1323 1324 unsigned NumOfVarIndices = 0; 1325 Value *Ptrs = GEP->getPointerOperand(); 1326 if (Ptrs->getType()->isVectorTy() && !getSplatValue(Ptrs)) 1327 return IndexSize; 1328 for (unsigned i = 1; i < GEP->getNumOperands(); ++i) { 1329 if (isa<Constant>(GEP->getOperand(i))) 1330 continue; 1331 Type *IndxTy = GEP->getOperand(i)->getType(); 1332 if (IndxTy->isVectorTy()) 1333 IndxTy = IndxTy->getVectorElementType(); 1334 if ((IndxTy->getPrimitiveSizeInBits() == 64 && 1335 !isa<SExtInst>(GEP->getOperand(i))) || 1336 ++NumOfVarIndices > 1) 1337 return IndexSize; // 64 1338 } 1339 return (unsigned)32; 1340 }; 1341 1342 1343 // Trying to reduce IndexSize to 32 bits for vector 16. 1344 // By default the IndexSize is equal to pointer size. 1345 unsigned IndexSize = (VF >= 16) ? getIndexSizeInBits(Ptr, DL) : 1346 DL.getPointerSizeInBits(); 1347 1348 Type *IndexVTy = VectorType::get(IntegerType::get(SrcVTy->getContext(), 1349 IndexSize), VF); 1350 std::pair<int, MVT> IdxsLT = TLI->getTypeLegalizationCost(DL, IndexVTy); 1351 std::pair<int, MVT> SrcLT = TLI->getTypeLegalizationCost(DL, SrcVTy); 1352 int SplitFactor = std::max(IdxsLT.first, SrcLT.first); 1353 if (SplitFactor > 1) { 1354 // Handle splitting of vector of pointers 1355 Type *SplitSrcTy = VectorType::get(SrcVTy->getScalarType(), VF / SplitFactor); 1356 return SplitFactor * getGSVectorCost(Opcode, SplitSrcTy, Ptr, Alignment, 1357 AddressSpace); 1358 } 1359 1360 // The gather / scatter cost is given by Intel architects. It is a rough 1361 // number since we are looking at one instruction in a time. 1362 const int GSOverhead = 2; 1363 return GSOverhead + VF * getMemoryOpCost(Opcode, SrcVTy->getScalarType(), 1364 Alignment, AddressSpace); 1365 } 1366 1367 /// Return the cost of full scalarization of gather / scatter operation. 1368 /// 1369 /// Opcode - Load or Store instruction. 1370 /// SrcVTy - The type of the data vector that should be gathered or scattered. 1371 /// VariableMask - The mask is non-constant at compile time. 1372 /// Alignment - Alignment for one element. 1373 /// AddressSpace - pointer[s] address space. 1374 /// 1375 int X86TTIImpl::getGSScalarCost(unsigned Opcode, Type *SrcVTy, 1376 bool VariableMask, unsigned Alignment, 1377 unsigned AddressSpace) { 1378 unsigned VF = SrcVTy->getVectorNumElements(); 1379 1380 int MaskUnpackCost = 0; 1381 if (VariableMask) { 1382 VectorType *MaskTy = 1383 VectorType::get(Type::getInt1Ty(SrcVTy->getContext()), VF); 1384 MaskUnpackCost = getScalarizationOverhead(MaskTy, false, true); 1385 int ScalarCompareCost = 1386 getCmpSelInstrCost(Instruction::ICmp, Type::getInt1Ty(SrcVTy->getContext()), 1387 nullptr); 1388 int BranchCost = getCFInstrCost(Instruction::Br); 1389 MaskUnpackCost += VF * (BranchCost + ScalarCompareCost); 1390 } 1391 1392 // The cost of the scalar loads/stores. 1393 int MemoryOpCost = VF * getMemoryOpCost(Opcode, SrcVTy->getScalarType(), 1394 Alignment, AddressSpace); 1395 1396 int InsertExtractCost = 0; 1397 if (Opcode == Instruction::Load) 1398 for (unsigned i = 0; i < VF; ++i) 1399 // Add the cost of inserting each scalar load into the vector 1400 InsertExtractCost += 1401 getVectorInstrCost(Instruction::InsertElement, SrcVTy, i); 1402 else 1403 for (unsigned i = 0; i < VF; ++i) 1404 // Add the cost of extracting each element out of the data vector 1405 InsertExtractCost += 1406 getVectorInstrCost(Instruction::ExtractElement, SrcVTy, i); 1407 1408 return MemoryOpCost + MaskUnpackCost + InsertExtractCost; 1409 } 1410 1411 /// Calculate the cost of Gather / Scatter operation 1412 int X86TTIImpl::getGatherScatterOpCost(unsigned Opcode, Type *SrcVTy, 1413 Value *Ptr, bool VariableMask, 1414 unsigned Alignment) { 1415 assert(SrcVTy->isVectorTy() && "Unexpected data type for Gather/Scatter"); 1416 unsigned VF = SrcVTy->getVectorNumElements(); 1417 PointerType *PtrTy = dyn_cast<PointerType>(Ptr->getType()); 1418 if (!PtrTy && Ptr->getType()->isVectorTy()) 1419 PtrTy = dyn_cast<PointerType>(Ptr->getType()->getVectorElementType()); 1420 assert(PtrTy && "Unexpected type for Ptr argument"); 1421 unsigned AddressSpace = PtrTy->getAddressSpace(); 1422 1423 bool Scalarize = false; 1424 if ((Opcode == Instruction::Load && !isLegalMaskedGather(SrcVTy)) || 1425 (Opcode == Instruction::Store && !isLegalMaskedScatter(SrcVTy))) 1426 Scalarize = true; 1427 // Gather / Scatter for vector 2 is not profitable on KNL / SKX 1428 // Vector-4 of gather/scatter instruction does not exist on KNL. 1429 // We can extend it to 8 elements, but zeroing upper bits of 1430 // the mask vector will add more instructions. Right now we give the scalar 1431 // cost of vector-4 for KNL. TODO: Check, maybe the gather/scatter instruction is 1432 // better in the VariableMask case. 1433 if (VF == 2 || (VF == 4 && !ST->hasVLX())) 1434 Scalarize = true; 1435 1436 if (Scalarize) 1437 return getGSScalarCost(Opcode, SrcVTy, VariableMask, Alignment, AddressSpace); 1438 1439 return getGSVectorCost(Opcode, SrcVTy, Ptr, Alignment, AddressSpace); 1440 } 1441 1442 bool X86TTIImpl::isLegalMaskedLoad(Type *DataTy) { 1443 Type *ScalarTy = DataTy->getScalarType(); 1444 int DataWidth = isa<PointerType>(ScalarTy) ? 1445 DL.getPointerSizeInBits() : ScalarTy->getPrimitiveSizeInBits(); 1446 1447 return (DataWidth >= 32 && ST->hasAVX()) || 1448 (DataWidth >= 8 && ST->hasBWI()); 1449 } 1450 1451 bool X86TTIImpl::isLegalMaskedStore(Type *DataType) { 1452 return isLegalMaskedLoad(DataType); 1453 } 1454 1455 bool X86TTIImpl::isLegalMaskedGather(Type *DataTy) { 1456 // This function is called now in two cases: from the Loop Vectorizer 1457 // and from the Scalarizer. 1458 // When the Loop Vectorizer asks about legality of the feature, 1459 // the vectorization factor is not calculated yet. The Loop Vectorizer 1460 // sends a scalar type and the decision is based on the width of the 1461 // scalar element. 1462 // Later on, the cost model will estimate usage this intrinsic based on 1463 // the vector type. 1464 // The Scalarizer asks again about legality. It sends a vector type. 1465 // In this case we can reject non-power-of-2 vectors. 1466 if (isa<VectorType>(DataTy) && !isPowerOf2_32(DataTy->getVectorNumElements())) 1467 return false; 1468 Type *ScalarTy = DataTy->getScalarType(); 1469 int DataWidth = isa<PointerType>(ScalarTy) ? 1470 DL.getPointerSizeInBits() : ScalarTy->getPrimitiveSizeInBits(); 1471 1472 // AVX-512 allows gather and scatter 1473 return DataWidth >= 32 && ST->hasAVX512(); 1474 } 1475 1476 bool X86TTIImpl::isLegalMaskedScatter(Type *DataType) { 1477 return isLegalMaskedGather(DataType); 1478 } 1479 1480 bool X86TTIImpl::areInlineCompatible(const Function *Caller, 1481 const Function *Callee) const { 1482 const TargetMachine &TM = getTLI()->getTargetMachine(); 1483 1484 // Work this as a subsetting of subtarget features. 1485 const FeatureBitset &CallerBits = 1486 TM.getSubtargetImpl(*Caller)->getFeatureBits(); 1487 const FeatureBitset &CalleeBits = 1488 TM.getSubtargetImpl(*Callee)->getFeatureBits(); 1489 1490 // FIXME: This is likely too limiting as it will include subtarget features 1491 // that we might not care about for inlining, but it is conservatively 1492 // correct. 1493 return (CallerBits & CalleeBits) == CalleeBits; 1494 } 1495