| 1 | // Licensed to the .NET Foundation under one or more agreements. |
| 2 | // The .NET Foundation licenses this file to you under the MIT license. |
| 3 | // See the LICENSE file in the project root for more information. |
| 4 | |
| 5 | #include "jitpch.h" |
| 6 | #include "hwintrinsic.h" |
| 7 | |
| 8 | #ifdef FEATURE_HW_INTRINSICS |
| 9 | |
| 10 | namespace IsaFlag |
| 11 | { |
| 12 | enum Flag |
| 13 | { |
| 14 | #define HARDWARE_INTRINSIC_CLASS(flag, isa) isa = 1ULL << InstructionSet_##isa, |
| 15 | #include "hwintrinsiclistArm64.h" |
| 16 | None = 0, |
| 17 | Base = 1ULL << InstructionSet_Base, |
| 18 | EveryISA = ~0ULL |
| 19 | }; |
| 20 | |
| 21 | Flag operator|(Flag a, Flag b) |
| 22 | { |
| 23 | return Flag(uint64_t(a) | uint64_t(b)); |
| 24 | } |
| 25 | |
| 26 | Flag flag(InstructionSet isa) |
| 27 | { |
| 28 | return Flag(1ULL << isa); |
| 29 | } |
| 30 | } |
| 31 | |
| 32 | // clang-format off |
| 33 | static const HWIntrinsicInfo hwIntrinsicInfoArray[] = { |
| 34 | // Add lookupHWIntrinsic special cases see lookupHWIntrinsic() below |
| 35 | // NI_ARM64_IsSupported_True is used to expand get_IsSupported to const true |
| 36 | // NI_ARM64_IsSupported_False is used to expand get_IsSupported to const false |
| 37 | // NI_ARM64_PlatformNotSupported to throw PlatformNotSupported exception for every intrinsic not supported on the running platform |
| 38 | {NI_ARM64_IsSupported_True, "get_IsSupported" , IsaFlag::EveryISA, HWIntrinsicInfo::IsSupported, HWIntrinsicInfo::None, {}}, |
| 39 | {NI_ARM64_IsSupported_False, "::NI_ARM64_IsSupported_False" , IsaFlag::EveryISA, HWIntrinsicInfo::IsSupported, HWIntrinsicInfo::None, {}}, |
| 40 | {NI_ARM64_PlatformNotSupported, "::NI_ARM64_PlatformNotSupported" , IsaFlag::EveryISA, HWIntrinsicInfo::Unsupported, HWIntrinsicInfo::None, {}}, |
| 41 | #define HARDWARE_INTRINSIC(id, isa, name, form, i0, i1, i2, flags) \ |
| 42 | {id, #name, IsaFlag::isa, HWIntrinsicInfo::form, HWIntrinsicInfo::flags, { i0, i1, i2 }}, |
| 43 | #include "hwintrinsiclistArm64.h" |
| 44 | }; |
| 45 | // clang-format on |
| 46 | |
| 47 | //------------------------------------------------------------------------ |
| 48 | // lookup: Gets the HWIntrinsicInfo associated with a given NamedIntrinsic |
| 49 | // |
| 50 | // Arguments: |
| 51 | // id -- The NamedIntrinsic associated with the HWIntrinsic to lookup |
| 52 | // |
| 53 | // Return Value: |
| 54 | // The HWIntrinsicInfo associated with id |
| 55 | const HWIntrinsicInfo& HWIntrinsicInfo::lookup(NamedIntrinsic id) |
| 56 | { |
| 57 | assert(id != NI_Illegal); |
| 58 | |
| 59 | assert(id > NI_HW_INTRINSIC_START); |
| 60 | assert(id < NI_HW_INTRINSIC_END); |
| 61 | |
| 62 | return hwIntrinsicInfoArray[id - NI_HW_INTRINSIC_START - 1]; |
| 63 | } |
| 64 | |
| 65 | //------------------------------------------------------------------------ |
| 66 | // lookupHWIntrinsicISA: map class name to InstructionSet value |
| 67 | // |
| 68 | // Arguments: |
| 69 | // className -- class name in System.Runtime.Intrinsics.Arm.Arm64 |
| 70 | // |
| 71 | // Return Value: |
| 72 | // Id for the ISA class if enabled. |
| 73 | // |
| 74 | InstructionSet Compiler::lookupHWIntrinsicISA(const char* className) |
| 75 | { |
| 76 | if (className != nullptr) |
| 77 | { |
| 78 | if (strcmp(className, "Base" ) == 0) |
| 79 | return InstructionSet_Base; |
| 80 | #define HARDWARE_INTRINSIC_CLASS(flag, isa) \ |
| 81 | if (strcmp(className, #isa) == 0) \ |
| 82 | return InstructionSet_##isa; |
| 83 | #include "hwintrinsiclistArm64.h" |
| 84 | } |
| 85 | |
| 86 | return InstructionSet_NONE; |
| 87 | } |
| 88 | |
| 89 | //------------------------------------------------------------------------ |
| 90 | // lookupHWIntrinsic: map intrinsic name to named intrinsic value |
| 91 | // |
| 92 | // Arguments: |
| 93 | // methodName -- name of the intrinsic function. |
| 94 | // isa -- instruction set of the intrinsic. |
| 95 | // |
| 96 | // Return Value: |
| 97 | // Id for the hardware intrinsic. |
| 98 | // |
| 99 | // TODO-Throughput: replace sequential search by hash lookup |
| 100 | NamedIntrinsic Compiler::lookupHWIntrinsic(const char* className, const char* methodName) |
| 101 | { |
| 102 | InstructionSet isa = lookupHWIntrinsicISA(className); |
| 103 | NamedIntrinsic result = NI_Illegal; |
| 104 | if (isa != InstructionSet_NONE) |
| 105 | { |
| 106 | IsaFlag::Flag isaFlag = IsaFlag::flag(isa); |
| 107 | for (int i = 0; i < NI_HW_INTRINSIC_END - NI_HW_INTRINSIC_START; i++) |
| 108 | { |
| 109 | if ((isaFlag & hwIntrinsicInfoArray[i].isaflags) && strcmp(methodName, hwIntrinsicInfoArray[i].name) == 0) |
| 110 | { |
| 111 | if (compSupportsHWIntrinsic(isa)) |
| 112 | { |
| 113 | // Intrinsic is supported on platform |
| 114 | result = hwIntrinsicInfoArray[i].id; |
| 115 | } |
| 116 | else |
| 117 | { |
| 118 | // When the intrinsic class is not supported |
| 119 | // Return NI_ARM64_PlatformNotSupported for all intrinsics |
| 120 | // Return NI_ARM64_IsSupported_False for the IsSupported property |
| 121 | result = (hwIntrinsicInfoArray[i].id != NI_ARM64_IsSupported_True) ? NI_ARM64_PlatformNotSupported |
| 122 | : NI_ARM64_IsSupported_False; |
| 123 | } |
| 124 | break; |
| 125 | } |
| 126 | } |
| 127 | } |
| 128 | return result; |
| 129 | } |
| 130 | |
| 131 | //------------------------------------------------------------------------ |
| 132 | // impCheckImmediate: check if immediate is const and in range for inlining |
| 133 | // |
| 134 | bool Compiler::impCheckImmediate(GenTree* immediateOp, unsigned int max) |
| 135 | { |
| 136 | return immediateOp->IsCnsIntOrI() && (immediateOp->AsIntConCommon()->IconValue() < max); |
| 137 | } |
| 138 | |
| 139 | //------------------------------------------------------------------------ |
| 140 | // isFullyImplementedIsa: Gets a value that indicates whether the InstructionSet is fully implemented |
| 141 | // |
| 142 | // Arguments: |
| 143 | // isa - The InstructionSet to check |
| 144 | // |
| 145 | // Return Value: |
| 146 | // true if isa is supported; otherwise, false |
| 147 | // |
| 148 | // Notes: |
| 149 | // This currently returns true for all partially-implemented ISAs. |
| 150 | // TODO-Bug: Set this to return the correct values as GH 20427 is resolved. |
| 151 | // |
| 152 | bool HWIntrinsicInfo::isFullyImplementedIsa(InstructionSet isa) |
| 153 | { |
| 154 | switch (isa) |
| 155 | { |
| 156 | case InstructionSet_Base: |
| 157 | case InstructionSet_Crc32: |
| 158 | case InstructionSet_Aes: |
| 159 | case InstructionSet_Simd: |
| 160 | case InstructionSet_Sha1: |
| 161 | case InstructionSet_Sha256: |
| 162 | return true; |
| 163 | |
| 164 | default: |
| 165 | assert(!"Unexpected Arm64 HW intrinsics ISA" ); |
| 166 | return false; |
| 167 | } |
| 168 | } |
| 169 | |
| 170 | //------------------------------------------------------------------------ |
| 171 | // isScalarIsa: Gets a value that indicates whether the InstructionSet is scalar |
| 172 | // |
| 173 | // Arguments: |
| 174 | // isa - The InstructionSet to check |
| 175 | // |
| 176 | // Return Value: |
| 177 | // true if isa is scalar; otherwise, false |
| 178 | bool HWIntrinsicInfo::isScalarIsa(InstructionSet isa) |
| 179 | { |
| 180 | switch (isa) |
| 181 | { |
| 182 | case InstructionSet_Base: |
| 183 | case InstructionSet_Crc32: |
| 184 | return true; |
| 185 | |
| 186 | case InstructionSet_Aes: |
| 187 | case InstructionSet_Simd: |
| 188 | case InstructionSet_Sha1: |
| 189 | case InstructionSet_Sha256: |
| 190 | return false; |
| 191 | |
| 192 | default: |
| 193 | assert(!"Unexpected Arm64 HW intrinsics ISA" ); |
| 194 | return true; |
| 195 | } |
| 196 | } |
| 197 | |
| 198 | //------------------------------------------------------------------------ |
| 199 | // compSupportsHWIntrinsic: compiler support of hardware intrinsics |
| 200 | // |
| 201 | // Arguments: |
| 202 | // isa - Instruction set |
| 203 | // Return Value: |
| 204 | // true if |
| 205 | // - isa is a scalar ISA |
| 206 | // - isa is a SIMD ISA and featureSIMD=true |
| 207 | // - isa is fully implemented or EnableIncompleteISAClass=true |
| 208 | bool Compiler::compSupportsHWIntrinsic(InstructionSet isa) |
| 209 | { |
| 210 | return (featureSIMD || HWIntrinsicInfo::isScalarIsa(isa)) && ( |
| 211 | #ifdef DEBUG |
| 212 | JitConfig.EnableIncompleteISAClass() || |
| 213 | #endif |
| 214 | HWIntrinsicInfo::isFullyImplementedIsa(isa)); |
| 215 | } |
| 216 | |
| 217 | //------------------------------------------------------------------------ |
| 218 | // lookupNumArgs: gets the number of arguments for the hardware intrinsic. |
| 219 | // This attempts to do a table based lookup but will fallback to the number |
| 220 | // of operands in 'node' if the table entry is -1. |
| 221 | // |
| 222 | // Arguments: |
| 223 | // node -- GenTreeHWIntrinsic* node with nullptr default value |
| 224 | // |
| 225 | // Return Value: |
| 226 | // number of arguments |
| 227 | // |
| 228 | int HWIntrinsicInfo::lookupNumArgs(const GenTreeHWIntrinsic* node) |
| 229 | { |
| 230 | NamedIntrinsic intrinsic = node->gtHWIntrinsicId; |
| 231 | |
| 232 | assert(intrinsic != NI_Illegal); |
| 233 | assert(intrinsic > NI_HW_INTRINSIC_START && intrinsic < NI_HW_INTRINSIC_END); |
| 234 | |
| 235 | GenTree* op1 = node->gtGetOp1(); |
| 236 | GenTree* op2 = node->gtGetOp2(); |
| 237 | int numArgs = 0; |
| 238 | |
| 239 | if (op1 == nullptr) |
| 240 | { |
| 241 | return 0; |
| 242 | } |
| 243 | |
| 244 | if (op1->OperIsList()) |
| 245 | { |
| 246 | numArgs = 0; |
| 247 | GenTreeArgList* list = op1->AsArgList(); |
| 248 | |
| 249 | while (list != nullptr) |
| 250 | { |
| 251 | numArgs++; |
| 252 | list = list->Rest(); |
| 253 | } |
| 254 | |
| 255 | // We should only use a list if we have 3 operands. |
| 256 | assert(numArgs >= 3); |
| 257 | return numArgs; |
| 258 | } |
| 259 | |
| 260 | if (op2 == nullptr) |
| 261 | { |
| 262 | return 1; |
| 263 | } |
| 264 | |
| 265 | return 2; |
| 266 | } |
| 267 | |
| 268 | //------------------------------------------------------------------------ |
| 269 | // impHWIntrinsic: dispatch hardware intrinsics to their own implementation |
| 270 | // function |
| 271 | // |
| 272 | // Arguments: |
| 273 | // intrinsic -- id of the intrinsic function. |
| 274 | // method -- method handle of the intrinsic function. |
| 275 | // sig -- signature of the intrinsic call |
| 276 | // |
| 277 | // Return Value: |
| 278 | // the expanded intrinsic. |
| 279 | // |
| 280 | GenTree* Compiler::impHWIntrinsic(NamedIntrinsic intrinsic, |
| 281 | CORINFO_METHOD_HANDLE method, |
| 282 | CORINFO_SIG_INFO* sig, |
| 283 | bool mustExpand) |
| 284 | { |
| 285 | GenTree* retNode = nullptr; |
| 286 | GenTree* op1 = nullptr; |
| 287 | GenTree* op2 = nullptr; |
| 288 | GenTree* op3 = nullptr; |
| 289 | CORINFO_CLASS_HANDLE simdClass = nullptr; |
| 290 | var_types simdType = TYP_UNKNOWN; |
| 291 | var_types simdBaseType = TYP_UNKNOWN; |
| 292 | unsigned simdSizeBytes = 0; |
| 293 | |
| 294 | switch (HWIntrinsicInfo::lookup(intrinsic).form) |
| 295 | { |
| 296 | case HWIntrinsicInfo::SimdBinaryOp: |
| 297 | case HWIntrinsicInfo::SimdInsertOp: |
| 298 | case HWIntrinsicInfo::SimdSelectOp: |
| 299 | case HWIntrinsicInfo::SimdSetAllOp: |
| 300 | case HWIntrinsicInfo::SimdUnaryOp: |
| 301 | case HWIntrinsicInfo::SimdBinaryRMWOp: |
| 302 | case HWIntrinsicInfo::SimdTernaryRMWOp: |
| 303 | case HWIntrinsicInfo::Sha1HashOp: |
| 304 | simdClass = sig->retTypeClass; |
| 305 | break; |
| 306 | case HWIntrinsicInfo::SimdExtractOp: |
| 307 | info.compCompHnd->getArgType(sig, sig->args, &simdClass); |
| 308 | break; |
| 309 | default: |
| 310 | break; |
| 311 | } |
| 312 | |
| 313 | // Simd instantiation type check |
| 314 | if (simdClass != nullptr) |
| 315 | { |
| 316 | if (featureSIMD) |
| 317 | { |
| 318 | compFloatingPointUsed = true; |
| 319 | |
| 320 | simdBaseType = getBaseTypeAndSizeOfSIMDType(simdClass, &simdSizeBytes); |
| 321 | } |
| 322 | |
| 323 | if (simdBaseType == TYP_UNKNOWN) |
| 324 | { |
| 325 | return impUnsupportedHWIntrinsic(CORINFO_HELP_THROW_PLATFORM_NOT_SUPPORTED, method, sig, mustExpand); |
| 326 | } |
| 327 | simdType = getSIMDTypeForSize(simdSizeBytes); |
| 328 | } |
| 329 | |
| 330 | switch (HWIntrinsicInfo::lookup(intrinsic).form) |
| 331 | { |
| 332 | case HWIntrinsicInfo::IsSupported: |
| 333 | return gtNewIconNode((intrinsic == NI_ARM64_IsSupported_True) ? 1 : 0); |
| 334 | |
| 335 | case HWIntrinsicInfo::Unsupported: |
| 336 | return impUnsupportedHWIntrinsic(CORINFO_HELP_THROW_PLATFORM_NOT_SUPPORTED, method, sig, mustExpand); |
| 337 | |
| 338 | case HWIntrinsicInfo::UnaryOp: |
| 339 | op1 = impPopStack().val; |
| 340 | |
| 341 | return gtNewScalarHWIntrinsicNode(JITtype2varType(sig->retType), op1, intrinsic); |
| 342 | |
| 343 | case HWIntrinsicInfo::SimdBinaryOp: |
| 344 | case HWIntrinsicInfo::SimdBinaryRMWOp: |
| 345 | // op1 is the first operand |
| 346 | // op2 is the second operand |
| 347 | op2 = impSIMDPopStack(simdType); |
| 348 | op1 = impSIMDPopStack(simdType); |
| 349 | |
| 350 | return gtNewSimdHWIntrinsicNode(simdType, op1, op2, intrinsic, simdBaseType, simdSizeBytes); |
| 351 | |
| 352 | case HWIntrinsicInfo::SimdTernaryRMWOp: |
| 353 | case HWIntrinsicInfo::SimdSelectOp: |
| 354 | // op1 is the first operand |
| 355 | // op2 is the second operand |
| 356 | // op3 is the third operand |
| 357 | op3 = impSIMDPopStack(simdType); |
| 358 | op2 = impSIMDPopStack(simdType); |
| 359 | op1 = impSIMDPopStack(simdType); |
| 360 | |
| 361 | return gtNewSimdHWIntrinsicNode(simdType, op1, op2, op3, intrinsic, simdBaseType, simdSizeBytes); |
| 362 | |
| 363 | case HWIntrinsicInfo::SimdSetAllOp: |
| 364 | op1 = impPopStack().val; |
| 365 | |
| 366 | return gtNewSimdHWIntrinsicNode(simdType, op1, intrinsic, simdBaseType, simdSizeBytes); |
| 367 | |
| 368 | case HWIntrinsicInfo::SimdUnaryOp: |
| 369 | op1 = impSIMDPopStack(simdType); |
| 370 | |
| 371 | return gtNewSimdHWIntrinsicNode(simdType, op1, intrinsic, simdBaseType, simdSizeBytes); |
| 372 | |
| 373 | case HWIntrinsicInfo::SimdExtractOp: |
| 374 | if (!mustExpand && !impCheckImmediate(impStackTop(0).val, getSIMDVectorLength(simdSizeBytes, simdBaseType))) |
| 375 | { |
| 376 | // Immediate lane not constant or out of range |
| 377 | return nullptr; |
| 378 | } |
| 379 | op2 = impPopStack().val; |
| 380 | op1 = impSIMDPopStack(simdType); |
| 381 | |
| 382 | return gtNewScalarHWIntrinsicNode(JITtype2varType(sig->retType), op1, op2, intrinsic); |
| 383 | |
| 384 | case HWIntrinsicInfo::SimdInsertOp: |
| 385 | if (!mustExpand && !impCheckImmediate(impStackTop(1).val, getSIMDVectorLength(simdSizeBytes, simdBaseType))) |
| 386 | { |
| 387 | // Immediate lane not constant or out of range |
| 388 | return nullptr; |
| 389 | } |
| 390 | op3 = impPopStack().val; |
| 391 | op2 = impPopStack().val; |
| 392 | op1 = impSIMDPopStack(simdType); |
| 393 | |
| 394 | return gtNewSimdHWIntrinsicNode(simdType, op1, op2, op3, intrinsic, simdBaseType, simdSizeBytes); |
| 395 | |
| 396 | case HWIntrinsicInfo::Sha1HashOp: |
| 397 | op3 = impSIMDPopStack(simdType); |
| 398 | op2 = impPopStack().val; |
| 399 | op1 = impSIMDPopStack(simdType); |
| 400 | |
| 401 | return gtNewSimdHWIntrinsicNode(simdType, op1, op2, op3, intrinsic, simdBaseType, simdSizeBytes); |
| 402 | |
| 403 | case HWIntrinsicInfo::Sha1RotateOp: |
| 404 | assert(sig->numArgs == 1); |
| 405 | compFloatingPointUsed = true; |
| 406 | return gtNewScalarHWIntrinsicNode(TYP_UINT, impPopStack().val, NI_ARM64_Sha1FixedRotate); |
| 407 | |
| 408 | default: |
| 409 | JITDUMP("Not implemented hardware intrinsic form" ); |
| 410 | assert(!"Unimplemented SIMD Intrinsic form" ); |
| 411 | |
| 412 | break; |
| 413 | } |
| 414 | return retNode; |
| 415 | } |
| 416 | |
| 417 | #endif // FEATURE_HW_INTRINSICS |
| 418 | |