1//===-- X86.td - Target definition file for the Intel X86 --*- tablegen -*-===// 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// 9// This is a target description file for the Intel i386 architecture, referred 10// to here as the "X86" architecture. 11// 12//===----------------------------------------------------------------------===// 13 14// Get the target-independent interfaces which we are implementing... 15// 16include "llvm/Target/Target.td" 17 18//===----------------------------------------------------------------------===// 19// X86 Subtarget state 20// 21// disregarding specific ABI / programming model 22def Is64Bit : SubtargetFeature<"64bit-mode", "Is64Bit", "true", 23 "64-bit mode (x86_64)">; 24def Is32Bit : SubtargetFeature<"32bit-mode", "Is32Bit", "true", 25 "32-bit mode (80386)">; 26def Is16Bit : SubtargetFeature<"16bit-mode", "Is16Bit", "true", 27 "16-bit mode (i8086)">; 28 29//===----------------------------------------------------------------------===// 30// X86 Subtarget ISA features 31//===----------------------------------------------------------------------===// 32 33def FeatureX87 : SubtargetFeature<"x87","HasX87", "true", 34 "Enable X87 float instructions">; 35 36def FeatureNOPL : SubtargetFeature<"nopl", "HasNOPL", "true", 37 "Enable NOPL instruction (generally pentium pro+)">; 38 39def FeatureCMOV : SubtargetFeature<"cmov","HasCMOV", "true", 40 "Enable conditional move instructions">; 41 42def FeatureCX8 : SubtargetFeature<"cx8", "HasCX8", "true", 43 "Support CMPXCHG8B instructions">; 44 45def FeatureCRC32 : SubtargetFeature<"crc32", "HasCRC32", "true", 46 "Enable SSE 4.2 CRC32 instruction (used when SSE4.2 is supported but function is GPR only)">; 47 48def FeaturePOPCNT : SubtargetFeature<"popcnt", "HasPOPCNT", "true", 49 "Support POPCNT instruction">; 50 51def FeatureFXSR : SubtargetFeature<"fxsr", "HasFXSR", "true", 52 "Support fxsave/fxrestore instructions">; 53 54def FeatureXSAVE : SubtargetFeature<"xsave", "HasXSAVE", "true", 55 "Support xsave instructions">; 56 57def FeatureXSAVEOPT: SubtargetFeature<"xsaveopt", "HasXSAVEOPT", "true", 58 "Support xsaveopt instructions", 59 [FeatureXSAVE]>; 60 61def FeatureXSAVEC : SubtargetFeature<"xsavec", "HasXSAVEC", "true", 62 "Support xsavec instructions", 63 [FeatureXSAVE]>; 64 65def FeatureXSAVES : SubtargetFeature<"xsaves", "HasXSAVES", "true", 66 "Support xsaves instructions", 67 [FeatureXSAVE]>; 68 69def FeatureSSE1 : SubtargetFeature<"sse", "X86SSELevel", "SSE1", 70 "Enable SSE instructions">; 71def FeatureSSE2 : SubtargetFeature<"sse2", "X86SSELevel", "SSE2", 72 "Enable SSE2 instructions", 73 [FeatureSSE1]>; 74def FeatureSSE3 : SubtargetFeature<"sse3", "X86SSELevel", "SSE3", 75 "Enable SSE3 instructions", 76 [FeatureSSE2]>; 77def FeatureSSSE3 : SubtargetFeature<"ssse3", "X86SSELevel", "SSSE3", 78 "Enable SSSE3 instructions", 79 [FeatureSSE3]>; 80def FeatureSSE41 : SubtargetFeature<"sse4.1", "X86SSELevel", "SSE41", 81 "Enable SSE 4.1 instructions", 82 [FeatureSSSE3]>; 83def FeatureSSE42 : SubtargetFeature<"sse4.2", "X86SSELevel", "SSE42", 84 "Enable SSE 4.2 instructions", 85 [FeatureSSE41]>; 86// The MMX subtarget feature is separate from the rest of the SSE features 87// because it's important (for odd compatibility reasons) to be able to 88// turn it off explicitly while allowing SSE+ to be on. 89def FeatureMMX : SubtargetFeature<"mmx","X863DNowLevel", "MMX", 90 "Enable MMX instructions">; 91def Feature3DNow : SubtargetFeature<"3dnow", "X863DNowLevel", "ThreeDNow", 92 "Enable 3DNow! instructions", 93 [FeatureMMX]>; 94def Feature3DNowA : SubtargetFeature<"3dnowa", "X863DNowLevel", "ThreeDNowA", 95 "Enable 3DNow! Athlon instructions", 96 [Feature3DNow]>; 97// All x86-64 hardware has SSE2, but we don't mark SSE2 as an implied 98// feature, because SSE2 can be disabled (e.g. for compiling OS kernels) 99// without disabling 64-bit mode. Nothing should imply this feature bit. It 100// is used to enforce that only 64-bit capable CPUs are used in 64-bit mode. 101def FeatureX86_64 : SubtargetFeature<"64bit", "HasX86_64", "true", 102 "Support 64-bit instructions">; 103def FeatureCX16 : SubtargetFeature<"cx16", "HasCX16", "true", 104 "64-bit with cmpxchg16b (this is true for most x86-64 chips, but not the first AMD chips)", 105 [FeatureCX8]>; 106def FeatureSSE4A : SubtargetFeature<"sse4a", "HasSSE4A", "true", 107 "Support SSE 4a instructions", 108 [FeatureSSE3]>; 109 110def FeatureAVX : SubtargetFeature<"avx", "X86SSELevel", "AVX", 111 "Enable AVX instructions", 112 [FeatureSSE42]>; 113def FeatureAVX2 : SubtargetFeature<"avx2", "X86SSELevel", "AVX2", 114 "Enable AVX2 instructions", 115 [FeatureAVX]>; 116def FeatureFMA : SubtargetFeature<"fma", "HasFMA", "true", 117 "Enable three-operand fused multiple-add", 118 [FeatureAVX]>; 119def FeatureF16C : SubtargetFeature<"f16c", "HasF16C", "true", 120 "Support 16-bit floating point conversion instructions", 121 [FeatureAVX]>; 122def FeatureAVX512 : SubtargetFeature<"avx512f", "X86SSELevel", "AVX512", 123 "Enable AVX-512 instructions", 124 [FeatureAVX2, FeatureFMA, FeatureF16C]>; 125def FeatureERI : SubtargetFeature<"avx512er", "HasERI", "true", 126 "Enable AVX-512 Exponential and Reciprocal Instructions", 127 [FeatureAVX512]>; 128def FeatureCDI : SubtargetFeature<"avx512cd", "HasCDI", "true", 129 "Enable AVX-512 Conflict Detection Instructions", 130 [FeatureAVX512]>; 131def FeatureVPOPCNTDQ : SubtargetFeature<"avx512vpopcntdq", "HasVPOPCNTDQ", 132 "true", "Enable AVX-512 Population Count Instructions", 133 [FeatureAVX512]>; 134def FeaturePFI : SubtargetFeature<"avx512pf", "HasPFI", "true", 135 "Enable AVX-512 PreFetch Instructions", 136 [FeatureAVX512]>; 137def FeaturePREFETCHI : SubtargetFeature<"prefetchi", "HasPREFETCHI", 138 "true", 139 "Prefetch instruction with T0 or T1 Hint">; 140def FeaturePREFETCHWT1 : SubtargetFeature<"prefetchwt1", "HasPREFETCHWT1", 141 "true", 142 "Prefetch with Intent to Write and T1 Hint">; 143def FeatureDQI : SubtargetFeature<"avx512dq", "HasDQI", "true", 144 "Enable AVX-512 Doubleword and Quadword Instructions", 145 [FeatureAVX512]>; 146def FeatureBWI : SubtargetFeature<"avx512bw", "HasBWI", "true", 147 "Enable AVX-512 Byte and Word Instructions", 148 [FeatureAVX512]>; 149def FeatureVLX : SubtargetFeature<"avx512vl", "HasVLX", "true", 150 "Enable AVX-512 Vector Length eXtensions", 151 [FeatureAVX512]>; 152def FeatureVBMI : SubtargetFeature<"avx512vbmi", "HasVBMI", "true", 153 "Enable AVX-512 Vector Byte Manipulation Instructions", 154 [FeatureBWI]>; 155def FeatureVBMI2 : SubtargetFeature<"avx512vbmi2", "HasVBMI2", "true", 156 "Enable AVX-512 further Vector Byte Manipulation Instructions", 157 [FeatureBWI]>; 158def FeatureAVXIFMA : SubtargetFeature<"avxifma", "HasAVXIFMA", "true", 159 "Enable AVX-IFMA", 160 [FeatureAVX2]>; 161def FeatureIFMA : SubtargetFeature<"avx512ifma", "HasIFMA", "true", 162 "Enable AVX-512 Integer Fused Multiple-Add", 163 [FeatureAVX512]>; 164def FeaturePKU : SubtargetFeature<"pku", "HasPKU", "true", 165 "Enable protection keys">; 166def FeatureVNNI : SubtargetFeature<"avx512vnni", "HasVNNI", "true", 167 "Enable AVX-512 Vector Neural Network Instructions", 168 [FeatureAVX512]>; 169def FeatureAVXVNNI : SubtargetFeature<"avxvnni", "HasAVXVNNI", "true", 170 "Support AVX_VNNI encoding", 171 [FeatureAVX2]>; 172def FeatureBF16 : SubtargetFeature<"avx512bf16", "HasBF16", "true", 173 "Support bfloat16 floating point", 174 [FeatureBWI]>; 175def FeatureBITALG : SubtargetFeature<"avx512bitalg", "HasBITALG", "true", 176 "Enable AVX-512 Bit Algorithms", 177 [FeatureBWI]>; 178def FeatureVP2INTERSECT : SubtargetFeature<"avx512vp2intersect", 179 "HasVP2INTERSECT", "true", 180 "Enable AVX-512 vp2intersect", 181 [FeatureAVX512]>; 182// FIXME: FP16 scalar intrinsics use the type v8f16, which is supposed to be 183// guarded under condition hasVLX. So we imply it in FeatureFP16 currently. 184// FIXME: FP16 conversion between f16 and i64 customize type v8i64, which is 185// supposed to be guarded under condition hasDQI. So we imply it in FeatureFP16 186// currently. 187def FeatureFP16 : SubtargetFeature<"avx512fp16", "HasFP16", "true", 188 "Support 16-bit floating point", 189 [FeatureBWI, FeatureVLX, FeatureDQI]>; 190def FeatureAVXVNNIINT8 : SubtargetFeature<"avxvnniint8", 191 "HasAVXVNNIINT8", "true", 192 "Enable AVX-VNNI-INT8", 193 [FeatureAVX2]>; 194def FeaturePCLMUL : SubtargetFeature<"pclmul", "HasPCLMUL", "true", 195 "Enable packed carry-less multiplication instructions", 196 [FeatureSSE2]>; 197def FeatureGFNI : SubtargetFeature<"gfni", "HasGFNI", "true", 198 "Enable Galois Field Arithmetic Instructions", 199 [FeatureSSE2]>; 200def FeatureVPCLMULQDQ : SubtargetFeature<"vpclmulqdq", "HasVPCLMULQDQ", "true", 201 "Enable vpclmulqdq instructions", 202 [FeatureAVX, FeaturePCLMUL]>; 203def FeatureFMA4 : SubtargetFeature<"fma4", "HasFMA4", "true", 204 "Enable four-operand fused multiple-add", 205 [FeatureAVX, FeatureSSE4A]>; 206def FeatureXOP : SubtargetFeature<"xop", "HasXOP", "true", 207 "Enable XOP instructions", 208 [FeatureFMA4]>; 209def FeatureSSEUnalignedMem : SubtargetFeature<"sse-unaligned-mem", 210 "HasSSEUnalignedMem", "true", 211 "Allow unaligned memory operands with SSE instructions (this may require setting a configuration bit in the processor)">; 212def FeatureAES : SubtargetFeature<"aes", "HasAES", "true", 213 "Enable AES instructions", 214 [FeatureSSE2]>; 215def FeatureVAES : SubtargetFeature<"vaes", "HasVAES", "true", 216 "Promote selected AES instructions to AVX512/AVX registers", 217 [FeatureAVX, FeatureAES]>; 218def FeatureTBM : SubtargetFeature<"tbm", "HasTBM", "true", 219 "Enable TBM instructions">; 220def FeatureLWP : SubtargetFeature<"lwp", "HasLWP", "true", 221 "Enable LWP instructions">; 222def FeatureMOVBE : SubtargetFeature<"movbe", "HasMOVBE", "true", 223 "Support MOVBE instruction">; 224def FeatureRDRAND : SubtargetFeature<"rdrnd", "HasRDRAND", "true", 225 "Support RDRAND instruction">; 226def FeatureFSGSBase : SubtargetFeature<"fsgsbase", "HasFSGSBase", "true", 227 "Support FS/GS Base instructions">; 228def FeatureLZCNT : SubtargetFeature<"lzcnt", "HasLZCNT", "true", 229 "Support LZCNT instruction">; 230def FeatureBMI : SubtargetFeature<"bmi", "HasBMI", "true", 231 "Support BMI instructions">; 232def FeatureBMI2 : SubtargetFeature<"bmi2", "HasBMI2", "true", 233 "Support BMI2 instructions">; 234def FeatureRTM : SubtargetFeature<"rtm", "HasRTM", "true", 235 "Support RTM instructions">; 236def FeatureADX : SubtargetFeature<"adx", "HasADX", "true", 237 "Support ADX instructions">; 238def FeatureSHA : SubtargetFeature<"sha", "HasSHA", "true", 239 "Enable SHA instructions", 240 [FeatureSSE2]>; 241// Processor supports CET SHSTK - Control-Flow Enforcement Technology 242// using Shadow Stack 243def FeatureSHSTK : SubtargetFeature<"shstk", "HasSHSTK", "true", 244 "Support CET Shadow-Stack instructions">; 245def FeaturePRFCHW : SubtargetFeature<"prfchw", "HasPRFCHW", "true", 246 "Support PRFCHW instructions">; 247def FeatureRDSEED : SubtargetFeature<"rdseed", "HasRDSEED", "true", 248 "Support RDSEED instruction">; 249def FeatureLAHFSAHF64 : SubtargetFeature<"sahf", "HasLAHFSAHF64", "true", 250 "Support LAHF and SAHF instructions in 64-bit mode">; 251def FeatureMWAITX : SubtargetFeature<"mwaitx", "HasMWAITX", "true", 252 "Enable MONITORX/MWAITX timer functionality">; 253def FeatureCLZERO : SubtargetFeature<"clzero", "HasCLZERO", "true", 254 "Enable Cache Line Zero">; 255def FeatureCLDEMOTE : SubtargetFeature<"cldemote", "HasCLDEMOTE", "true", 256 "Enable Cache Line Demote">; 257def FeaturePTWRITE : SubtargetFeature<"ptwrite", "HasPTWRITE", "true", 258 "Support ptwrite instruction">; 259def FeatureAMXTILE : SubtargetFeature<"amx-tile", "HasAMXTILE", "true", 260 "Support AMX-TILE instructions">; 261def FeatureAMXINT8 : SubtargetFeature<"amx-int8", "HasAMXINT8", "true", 262 "Support AMX-INT8 instructions", 263 [FeatureAMXTILE]>; 264def FeatureAMXBF16 : SubtargetFeature<"amx-bf16", "HasAMXBF16", "true", 265 "Support AMX-BF16 instructions", 266 [FeatureAMXTILE]>; 267def FeatureAMXFP16 : SubtargetFeature<"amx-fp16", "HasAMXFP16", "true", 268 "Support AMX amx-fp16 instructions", 269 [FeatureAMXTILE]>; 270def FeatureCMPCCXADD : SubtargetFeature<"cmpccxadd", "HasCMPCCXADD", "true", 271 "Support CMPCCXADD instructions">; 272def FeatureRAOINT : SubtargetFeature<"raoint", "HasRAOINT", "true", 273 "Support RAO-INT instructions", 274 []>; 275def FeatureAVXNECONVERT : SubtargetFeature<"avxneconvert", "HasAVXNECONVERT", "true", 276 "Support AVX-NE-CONVERT instructions", 277 [FeatureAVX2]>; 278def FeatureINVPCID : SubtargetFeature<"invpcid", "HasINVPCID", "true", 279 "Invalidate Process-Context Identifier">; 280def FeatureSGX : SubtargetFeature<"sgx", "HasSGX", "true", 281 "Enable Software Guard Extensions">; 282def FeatureCLFLUSHOPT : SubtargetFeature<"clflushopt", "HasCLFLUSHOPT", "true", 283 "Flush A Cache Line Optimized">; 284def FeatureCLWB : SubtargetFeature<"clwb", "HasCLWB", "true", 285 "Cache Line Write Back">; 286def FeatureWBNOINVD : SubtargetFeature<"wbnoinvd", "HasWBNOINVD", "true", 287 "Write Back No Invalidate">; 288def FeatureRDPID : SubtargetFeature<"rdpid", "HasRDPID", "true", 289 "Support RDPID instructions">; 290def FeatureRDPRU : SubtargetFeature<"rdpru", "HasRDPRU", "true", 291 "Support RDPRU instructions">; 292def FeatureWAITPKG : SubtargetFeature<"waitpkg", "HasWAITPKG", "true", 293 "Wait and pause enhancements">; 294def FeatureENQCMD : SubtargetFeature<"enqcmd", "HasENQCMD", "true", 295 "Has ENQCMD instructions">; 296def FeatureKL : SubtargetFeature<"kl", "HasKL", "true", 297 "Support Key Locker kl Instructions", 298 [FeatureSSE2]>; 299def FeatureWIDEKL : SubtargetFeature<"widekl", "HasWIDEKL", "true", 300 "Support Key Locker wide Instructions", 301 [FeatureKL]>; 302def FeatureHRESET : SubtargetFeature<"hreset", "HasHRESET", "true", 303 "Has hreset instruction">; 304def FeatureSERIALIZE : SubtargetFeature<"serialize", "HasSERIALIZE", "true", 305 "Has serialize instruction">; 306def FeatureTSXLDTRK : SubtargetFeature<"tsxldtrk", "HasTSXLDTRK", "true", 307 "Support TSXLDTRK instructions">; 308def FeatureUINTR : SubtargetFeature<"uintr", "HasUINTR", "true", 309 "Has UINTR Instructions">; 310def FeaturePCONFIG : SubtargetFeature<"pconfig", "HasPCONFIG", "true", 311 "platform configuration instruction">; 312def FeatureMOVDIRI : SubtargetFeature<"movdiri", "HasMOVDIRI", "true", 313 "Support movdiri instruction (direct store integer)">; 314def FeatureMOVDIR64B : SubtargetFeature<"movdir64b", "HasMOVDIR64B", "true", 315 "Support movdir64b instruction (direct store 64 bytes)">; 316 317// Ivy Bridge and newer processors have enhanced REP MOVSB and STOSB (aka 318// "string operations"). See "REP String Enhancement" in the Intel Software 319// Development Manual. This feature essentially means that REP MOVSB will copy 320// using the largest available size instead of copying bytes one by one, making 321// it at least as fast as REPMOVS{W,D,Q}. 322def FeatureERMSB 323 : SubtargetFeature< 324 "ermsb", "HasERMSB", "true", 325 "REP MOVS/STOS are fast">; 326 327// Icelake and newer processors have Fast Short REP MOV. 328def FeatureFSRM 329 : SubtargetFeature< 330 "fsrm", "HasFSRM", "true", 331 "REP MOVSB of short lengths is faster">; 332 333def FeatureSoftFloat 334 : SubtargetFeature<"soft-float", "UseSoftFloat", "true", 335 "Use software floating point features">; 336 337//===----------------------------------------------------------------------===// 338// X86 Subtarget Security Mitigation features 339//===----------------------------------------------------------------------===// 340 341// Lower indirect calls using a special construct called a `retpoline` to 342// mitigate potential Spectre v2 attacks against them. 343def FeatureRetpolineIndirectCalls 344 : SubtargetFeature< 345 "retpoline-indirect-calls", "UseRetpolineIndirectCalls", "true", 346 "Remove speculation of indirect calls from the generated code">; 347 348// Lower indirect branches and switches either using conditional branch trees 349// or using a special construct called a `retpoline` to mitigate potential 350// Spectre v2 attacks against them. 351def FeatureRetpolineIndirectBranches 352 : SubtargetFeature< 353 "retpoline-indirect-branches", "UseRetpolineIndirectBranches", "true", 354 "Remove speculation of indirect branches from the generated code">; 355 356// Deprecated umbrella feature for enabling both `retpoline-indirect-calls` and 357// `retpoline-indirect-branches` above. 358def FeatureRetpoline 359 : SubtargetFeature<"retpoline", "DeprecatedUseRetpoline", "true", 360 "Remove speculation of indirect branches from the " 361 "generated code, either by avoiding them entirely or " 362 "lowering them with a speculation blocking construct", 363 [FeatureRetpolineIndirectCalls, 364 FeatureRetpolineIndirectBranches]>; 365 366// Rely on external thunks for the emitted retpoline calls. This allows users 367// to provide their own custom thunk definitions in highly specialized 368// environments such as a kernel that does boot-time hot patching. 369def FeatureRetpolineExternalThunk 370 : SubtargetFeature< 371 "retpoline-external-thunk", "UseRetpolineExternalThunk", "true", 372 "When lowering an indirect call or branch using a `retpoline`, rely " 373 "on the specified user provided thunk rather than emitting one " 374 "ourselves. Only has effect when combined with some other retpoline " 375 "feature", [FeatureRetpolineIndirectCalls]>; 376 377// Mitigate LVI attacks against indirect calls/branches and call returns 378def FeatureLVIControlFlowIntegrity 379 : SubtargetFeature< 380 "lvi-cfi", "UseLVIControlFlowIntegrity", "true", 381 "Prevent indirect calls/branches from using a memory operand, and " 382 "precede all indirect calls/branches from a register with an " 383 "LFENCE instruction to serialize control flow. Also decompose RET " 384 "instructions into a POP+LFENCE+JMP sequence.">; 385 386// Enable SESES to mitigate speculative execution attacks 387def FeatureSpeculativeExecutionSideEffectSuppression 388 : SubtargetFeature< 389 "seses", "UseSpeculativeExecutionSideEffectSuppression", "true", 390 "Prevent speculative execution side channel timing attacks by " 391 "inserting a speculation barrier before memory reads, memory writes, " 392 "and conditional branches. Implies LVI Control Flow integrity.", 393 [FeatureLVIControlFlowIntegrity]>; 394 395// Mitigate LVI attacks against data loads 396def FeatureLVILoadHardening 397 : SubtargetFeature< 398 "lvi-load-hardening", "UseLVILoadHardening", "true", 399 "Insert LFENCE instructions to prevent data speculatively injected " 400 "into loads from being used maliciously.">; 401 402def FeatureTaggedGlobals 403 : SubtargetFeature< 404 "tagged-globals", "AllowTaggedGlobals", "true", 405 "Use an instruction sequence for taking the address of a global " 406 "that allows a memory tag in the upper address bits.">; 407 408// Control codegen mitigation against Straight Line Speculation vulnerability. 409def FeatureHardenSlsRet 410 : SubtargetFeature< 411 "harden-sls-ret", "HardenSlsRet", "true", 412 "Harden against straight line speculation across RET instructions.">; 413 414def FeatureHardenSlsIJmp 415 : SubtargetFeature< 416 "harden-sls-ijmp", "HardenSlsIJmp", "true", 417 "Harden against straight line speculation across indirect JMP instructions.">; 418 419//===----------------------------------------------------------------------===// 420// X86 Subtarget Tuning features 421//===----------------------------------------------------------------------===// 422 423def TuningSlowSHLD : SubtargetFeature<"slow-shld", "IsSHLDSlow", "true", 424 "SHLD instruction is slow">; 425 426def TuningSlowPMULLD : SubtargetFeature<"slow-pmulld", "IsPMULLDSlow", "true", 427 "PMULLD instruction is slow (compared to PMULLW/PMULHW and PMULUDQ)">; 428 429def TuningSlowPMADDWD : SubtargetFeature<"slow-pmaddwd", "IsPMADDWDSlow", 430 "true", 431 "PMADDWD is slower than PMULLD">; 432 433// FIXME: This should not apply to CPUs that do not have SSE. 434def TuningSlowUAMem16 : SubtargetFeature<"slow-unaligned-mem-16", 435 "IsUnalignedMem16Slow", "true", 436 "Slow unaligned 16-byte memory access">; 437 438def TuningSlowUAMem32 : SubtargetFeature<"slow-unaligned-mem-32", 439 "IsUnalignedMem32Slow", "true", 440 "Slow unaligned 32-byte memory access">; 441 442def TuningLEAForSP : SubtargetFeature<"lea-sp", "UseLeaForSP", "true", 443 "Use LEA for adjusting the stack pointer (this is an optimization for Intel Atom processors)">; 444 445// True if 8-bit divisions are significantly faster than 446// 32-bit divisions and should be used when possible. 447def TuningSlowDivide32 : SubtargetFeature<"idivl-to-divb", 448 "HasSlowDivide32", "true", 449 "Use 8-bit divide for positive values less than 256">; 450 451// True if 32-bit divides are significantly faster than 452// 64-bit divisions and should be used when possible. 453def TuningSlowDivide64 : SubtargetFeature<"idivq-to-divl", 454 "HasSlowDivide64", "true", 455 "Use 32-bit divide for positive values less than 2^32">; 456 457def TuningPadShortFunctions : SubtargetFeature<"pad-short-functions", 458 "PadShortFunctions", "true", 459 "Pad short functions (to prevent a stall when returning too early)">; 460 461// On some processors, instructions that implicitly take two memory operands are 462// slow. In practice, this means that CALL, PUSH, and POP with memory operands 463// should be avoided in favor of a MOV + register CALL/PUSH/POP. 464def TuningSlowTwoMemOps : SubtargetFeature<"slow-two-mem-ops", 465 "SlowTwoMemOps", "true", 466 "Two memory operand instructions are slow">; 467 468// True if the LEA instruction inputs have to be ready at address generation 469// (AG) time. 470def TuningLEAUsesAG : SubtargetFeature<"lea-uses-ag", "LeaUsesAG", "true", 471 "LEA instruction needs inputs at AG stage">; 472 473def TuningSlowLEA : SubtargetFeature<"slow-lea", "SlowLEA", "true", 474 "LEA instruction with certain arguments is slow">; 475 476// True if the LEA instruction has all three source operands: base, index, 477// and offset or if the LEA instruction uses base and index registers where 478// the base is EBP, RBP,or R13 479def TuningSlow3OpsLEA : SubtargetFeature<"slow-3ops-lea", "Slow3OpsLEA", "true", 480 "LEA instruction with 3 ops or certain registers is slow">; 481 482// True if INC and DEC instructions are slow when writing to flags 483def TuningSlowIncDec : SubtargetFeature<"slow-incdec", "SlowIncDec", "true", 484 "INC and DEC instructions are slower than ADD and SUB">; 485 486def TuningPOPCNTFalseDeps : SubtargetFeature<"false-deps-popcnt", 487 "HasPOPCNTFalseDeps", "true", 488 "POPCNT has a false dependency on dest register">; 489 490def TuningLZCNTFalseDeps : SubtargetFeature<"false-deps-lzcnt-tzcnt", 491 "HasLZCNTFalseDeps", "true", 492 "LZCNT/TZCNT have a false dependency on dest register">; 493 494def TuningMULCFalseDeps : SubtargetFeature<"false-deps-mulc", 495 "HasMULCFalseDeps", "true", 496 "VF[C]MULCPH/SH has a false dependency on dest register">; 497 498def TuningPERMFalseDeps : SubtargetFeature<"false-deps-perm", 499 "HasPERMFalseDeps", "true", 500 "VPERMD/Q/PS/PD has a false dependency on dest register">; 501 502def TuningRANGEFalseDeps : SubtargetFeature<"false-deps-range", 503 "HasRANGEFalseDeps", "true", 504 "VRANGEPD/PS/SD/SS has a false dependency on dest register">; 505 506def TuningGETMANTFalseDeps : SubtargetFeature<"false-deps-getmant", 507 "HasGETMANTFalseDeps", "true", 508 "VGETMANTSS/SD/SH and VGETMANDPS/PD(memory version) has a" 509 " false dependency on dest register">; 510 511def TuningMULLQFalseDeps : SubtargetFeature<"false-deps-mullq", 512 "HasMULLQFalseDeps", "true", 513 "VPMULLQ has a false dependency on dest register">; 514 515def TuningSBBDepBreaking : SubtargetFeature<"sbb-dep-breaking", 516 "HasSBBDepBreaking", "true", 517 "SBB with same register has no source dependency">; 518 519// On recent X86 (port bound) processors, its preferable to combine to a single shuffle 520// using a variable mask over multiple fixed shuffles. 521def TuningFastVariableCrossLaneShuffle 522 : SubtargetFeature<"fast-variable-crosslane-shuffle", 523 "HasFastVariableCrossLaneShuffle", 524 "true", "Cross-lane shuffles with variable masks are fast">; 525def TuningFastVariablePerLaneShuffle 526 : SubtargetFeature<"fast-variable-perlane-shuffle", 527 "HasFastVariablePerLaneShuffle", 528 "true", "Per-lane shuffles with variable masks are fast">; 529 530// On some X86 processors, a vzeroupper instruction should be inserted after 531// using ymm/zmm registers before executing code that may use SSE instructions. 532def TuningInsertVZEROUPPER 533 : SubtargetFeature<"vzeroupper", 534 "InsertVZEROUPPER", 535 "true", "Should insert vzeroupper instructions">; 536 537// TuningFastScalarFSQRT should be enabled if scalar FSQRT has shorter latency 538// than the corresponding NR code. TuningFastVectorFSQRT should be enabled if 539// vector FSQRT has higher throughput than the corresponding NR code. 540// The idea is that throughput bound code is likely to be vectorized, so for 541// vectorized code we should care about the throughput of SQRT operations. 542// But if the code is scalar that probably means that the code has some kind of 543// dependency and we should care more about reducing the latency. 544 545// True if hardware SQRTSS instruction is at least as fast (latency) as 546// RSQRTSS followed by a Newton-Raphson iteration. 547def TuningFastScalarFSQRT 548 : SubtargetFeature<"fast-scalar-fsqrt", "HasFastScalarFSQRT", 549 "true", "Scalar SQRT is fast (disable Newton-Raphson)">; 550// True if hardware SQRTPS/VSQRTPS instructions are at least as fast 551// (throughput) as RSQRTPS/VRSQRTPS followed by a Newton-Raphson iteration. 552def TuningFastVectorFSQRT 553 : SubtargetFeature<"fast-vector-fsqrt", "HasFastVectorFSQRT", 554 "true", "Vector SQRT is fast (disable Newton-Raphson)">; 555 556// If lzcnt has equivalent latency/throughput to most simple integer ops, it can 557// be used to replace test/set sequences. 558def TuningFastLZCNT 559 : SubtargetFeature< 560 "fast-lzcnt", "HasFastLZCNT", "true", 561 "LZCNT instructions are as fast as most simple integer ops">; 562 563// If the target can efficiently decode NOPs upto 7-bytes in length. 564def TuningFast7ByteNOP 565 : SubtargetFeature< 566 "fast-7bytenop", "HasFast7ByteNOP", "true", 567 "Target can quickly decode up to 7 byte NOPs">; 568 569// If the target can efficiently decode NOPs upto 11-bytes in length. 570def TuningFast11ByteNOP 571 : SubtargetFeature< 572 "fast-11bytenop", "HasFast11ByteNOP", "true", 573 "Target can quickly decode up to 11 byte NOPs">; 574 575// If the target can efficiently decode NOPs upto 15-bytes in length. 576def TuningFast15ByteNOP 577 : SubtargetFeature< 578 "fast-15bytenop", "HasFast15ByteNOP", "true", 579 "Target can quickly decode up to 15 byte NOPs">; 580 581// Sandy Bridge and newer processors can use SHLD with the same source on both 582// inputs to implement rotate to avoid the partial flag update of the normal 583// rotate instructions. 584def TuningFastSHLDRotate 585 : SubtargetFeature< 586 "fast-shld-rotate", "HasFastSHLDRotate", "true", 587 "SHLD can be used as a faster rotate">; 588 589// Bulldozer and newer processors can merge CMP/TEST (but not other 590// instructions) with conditional branches. 591def TuningBranchFusion 592 : SubtargetFeature<"branchfusion", "HasBranchFusion", "true", 593 "CMP/TEST can be fused with conditional branches">; 594 595// Sandy Bridge and newer processors have many instructions that can be 596// fused with conditional branches and pass through the CPU as a single 597// operation. 598def TuningMacroFusion 599 : SubtargetFeature<"macrofusion", "HasMacroFusion", "true", 600 "Various instructions can be fused with conditional branches">; 601 602// Gather is available since Haswell (AVX2 set). So technically, we can 603// generate Gathers on all AVX2 processors. But the overhead on HSW is high. 604// Skylake Client processor has faster Gathers than HSW and performance is 605// similar to Skylake Server (AVX-512). 606def TuningFastGather 607 : SubtargetFeature<"fast-gather", "HasFastGather", "true", 608 "Indicates if gather is reasonably fast (this is true for Skylake client and all AVX-512 CPUs)">; 609 610def TuningPrefer128Bit 611 : SubtargetFeature<"prefer-128-bit", "Prefer128Bit", "true", 612 "Prefer 128-bit AVX instructions">; 613 614def TuningPrefer256Bit 615 : SubtargetFeature<"prefer-256-bit", "Prefer256Bit", "true", 616 "Prefer 256-bit AVX instructions">; 617 618def TuningAllowLight256Bit 619 : SubtargetFeature<"allow-light-256-bit", "AllowLight256Bit", "true", 620 "Enable generation of 256-bit load/stores even if we prefer 128-bit">; 621 622def TuningPreferMaskRegisters 623 : SubtargetFeature<"prefer-mask-registers", "PreferMaskRegisters", "true", 624 "Prefer AVX512 mask registers over PTEST/MOVMSK">; 625 626def TuningFastBEXTR : SubtargetFeature<"fast-bextr", "HasFastBEXTR", "true", 627 "Indicates that the BEXTR instruction is implemented as a single uop " 628 "with good throughput">; 629 630// Combine vector math operations with shuffles into horizontal math 631// instructions if a CPU implements horizontal operations (introduced with 632// SSE3) with better latency/throughput than the alternative sequence. 633def TuningFastHorizontalOps 634 : SubtargetFeature< 635 "fast-hops", "HasFastHorizontalOps", "true", 636 "Prefer horizontal vector math instructions (haddp, phsub, etc.) over " 637 "normal vector instructions with shuffles">; 638 639def TuningFastScalarShiftMasks 640 : SubtargetFeature< 641 "fast-scalar-shift-masks", "HasFastScalarShiftMasks", "true", 642 "Prefer a left/right scalar logical shift pair over a shift+and pair">; 643 644def TuningFastVectorShiftMasks 645 : SubtargetFeature< 646 "fast-vector-shift-masks", "HasFastVectorShiftMasks", "true", 647 "Prefer a left/right vector logical shift pair over a shift+and pair">; 648 649def TuningFastMOVBE 650 : SubtargetFeature<"fast-movbe", "HasFastMOVBE", "true", 651 "Prefer a movbe over a single-use load + bswap / single-use bswap + store">; 652 653def TuningUseSLMArithCosts 654 : SubtargetFeature<"use-slm-arith-costs", "UseSLMArithCosts", "true", 655 "Use Silvermont specific arithmetic costs">; 656 657def TuningUseGLMDivSqrtCosts 658 : SubtargetFeature<"use-glm-div-sqrt-costs", "UseGLMDivSqrtCosts", "true", 659 "Use Goldmont specific floating point div/sqrt costs">; 660 661//===----------------------------------------------------------------------===// 662// X86 CPU Families 663// TODO: Remove these - use general tuning features to determine codegen. 664//===----------------------------------------------------------------------===// 665 666// Bonnell 667def ProcIntelAtom : SubtargetFeature<"", "IsAtom", "true", "Is Intel Atom processor">; 668 669//===----------------------------------------------------------------------===// 670// Register File Description 671//===----------------------------------------------------------------------===// 672 673include "X86RegisterInfo.td" 674include "X86RegisterBanks.td" 675 676//===----------------------------------------------------------------------===// 677// Instruction Descriptions 678//===----------------------------------------------------------------------===// 679 680include "X86Schedule.td" 681include "X86InstrInfo.td" 682include "X86SchedPredicates.td" 683 684def X86InstrInfo : InstrInfo; 685 686//===----------------------------------------------------------------------===// 687// X86 Scheduler Models 688//===----------------------------------------------------------------------===// 689 690include "X86ScheduleAtom.td" 691include "X86SchedSandyBridge.td" 692include "X86SchedHaswell.td" 693include "X86SchedBroadwell.td" 694include "X86ScheduleSLM.td" 695include "X86ScheduleZnver1.td" 696include "X86ScheduleZnver2.td" 697include "X86ScheduleZnver3.td" 698include "X86ScheduleZnver4.td" 699include "X86ScheduleBdVer2.td" 700include "X86ScheduleBtVer2.td" 701include "X86SchedSkylakeClient.td" 702include "X86SchedSkylakeServer.td" 703include "X86SchedIceLake.td" 704include "X86SchedAlderlakeP.td" 705 706//===----------------------------------------------------------------------===// 707// X86 Processor Feature Lists 708//===----------------------------------------------------------------------===// 709 710def ProcessorFeatures { 711 // x86-64 and x86-64-v[234] 712 list<SubtargetFeature> X86_64V1Features = [ 713 FeatureX87, FeatureCX8, FeatureCMOV, FeatureMMX, FeatureSSE2, 714 FeatureFXSR, FeatureNOPL, FeatureX86_64, 715 ]; 716 list<SubtargetFeature> X86_64V2Features = !listconcat(X86_64V1Features, [ 717 FeatureCX16, FeatureLAHFSAHF64, FeatureCRC32, FeaturePOPCNT, 718 FeatureSSE42 719 ]); 720 list<SubtargetFeature> X86_64V3Features = !listconcat(X86_64V2Features, [ 721 FeatureAVX2, FeatureBMI, FeatureBMI2, FeatureF16C, FeatureFMA, FeatureLZCNT, 722 FeatureMOVBE, FeatureXSAVE 723 ]); 724 list<SubtargetFeature> X86_64V4Features = !listconcat(X86_64V3Features, [ 725 FeatureBWI, 726 FeatureCDI, 727 FeatureDQI, 728 FeatureVLX, 729 ]); 730 731 // Nehalem 732 list<SubtargetFeature> NHMFeatures = X86_64V2Features; 733 list<SubtargetFeature> NHMTuning = [TuningMacroFusion, 734 TuningInsertVZEROUPPER]; 735 736 // Westmere 737 list<SubtargetFeature> WSMAdditionalFeatures = [FeaturePCLMUL]; 738 list<SubtargetFeature> WSMTuning = NHMTuning; 739 list<SubtargetFeature> WSMFeatures = 740 !listconcat(NHMFeatures, WSMAdditionalFeatures); 741 742 // Sandybridge 743 list<SubtargetFeature> SNBAdditionalFeatures = [FeatureAVX, 744 FeatureXSAVE, 745 FeatureXSAVEOPT]; 746 list<SubtargetFeature> SNBTuning = [TuningMacroFusion, 747 TuningSlow3OpsLEA, 748 TuningSlowDivide64, 749 TuningSlowUAMem32, 750 TuningFastScalarFSQRT, 751 TuningFastSHLDRotate, 752 TuningFast15ByteNOP, 753 TuningPOPCNTFalseDeps, 754 TuningInsertVZEROUPPER]; 755 list<SubtargetFeature> SNBFeatures = 756 !listconcat(WSMFeatures, SNBAdditionalFeatures); 757 758 // Ivybridge 759 list<SubtargetFeature> IVBAdditionalFeatures = [FeatureRDRAND, 760 FeatureF16C, 761 FeatureFSGSBase]; 762 list<SubtargetFeature> IVBTuning = SNBTuning; 763 list<SubtargetFeature> IVBFeatures = 764 !listconcat(SNBFeatures, IVBAdditionalFeatures); 765 766 // Haswell 767 list<SubtargetFeature> HSWAdditionalFeatures = [FeatureAVX2, 768 FeatureBMI, 769 FeatureBMI2, 770 FeatureERMSB, 771 FeatureFMA, 772 FeatureINVPCID, 773 FeatureLZCNT, 774 FeatureMOVBE]; 775 list<SubtargetFeature> HSWTuning = [TuningMacroFusion, 776 TuningSlow3OpsLEA, 777 TuningSlowDivide64, 778 TuningFastScalarFSQRT, 779 TuningFastSHLDRotate, 780 TuningFast15ByteNOP, 781 TuningFastVariableCrossLaneShuffle, 782 TuningFastVariablePerLaneShuffle, 783 TuningPOPCNTFalseDeps, 784 TuningLZCNTFalseDeps, 785 TuningInsertVZEROUPPER, 786 TuningAllowLight256Bit]; 787 list<SubtargetFeature> HSWFeatures = 788 !listconcat(IVBFeatures, HSWAdditionalFeatures); 789 790 // Broadwell 791 list<SubtargetFeature> BDWAdditionalFeatures = [FeatureADX, 792 FeatureRDSEED, 793 FeaturePRFCHW]; 794 list<SubtargetFeature> BDWTuning = HSWTuning; 795 list<SubtargetFeature> BDWFeatures = 796 !listconcat(HSWFeatures, BDWAdditionalFeatures); 797 798 // Skylake 799 list<SubtargetFeature> SKLAdditionalFeatures = [FeatureAES, 800 FeatureXSAVEC, 801 FeatureXSAVES, 802 FeatureCLFLUSHOPT]; 803 list<SubtargetFeature> SKLTuning = [TuningFastGather, 804 TuningMacroFusion, 805 TuningSlow3OpsLEA, 806 TuningSlowDivide64, 807 TuningFastScalarFSQRT, 808 TuningFastVectorFSQRT, 809 TuningFastSHLDRotate, 810 TuningFast15ByteNOP, 811 TuningFastVariableCrossLaneShuffle, 812 TuningFastVariablePerLaneShuffle, 813 TuningPOPCNTFalseDeps, 814 TuningInsertVZEROUPPER, 815 TuningAllowLight256Bit]; 816 list<SubtargetFeature> SKLFeatures = 817 !listconcat(BDWFeatures, SKLAdditionalFeatures); 818 819 // Skylake-AVX512 820 list<SubtargetFeature> SKXAdditionalFeatures = [FeatureAES, 821 FeatureXSAVEC, 822 FeatureXSAVES, 823 FeatureCLFLUSHOPT, 824 FeatureAVX512, 825 FeatureCDI, 826 FeatureDQI, 827 FeatureBWI, 828 FeatureVLX, 829 FeaturePKU, 830 FeatureCLWB]; 831 list<SubtargetFeature> SKXTuning = [TuningFastGather, 832 TuningMacroFusion, 833 TuningSlow3OpsLEA, 834 TuningSlowDivide64, 835 TuningFastScalarFSQRT, 836 TuningFastVectorFSQRT, 837 TuningFastSHLDRotate, 838 TuningFast15ByteNOP, 839 TuningFastVariableCrossLaneShuffle, 840 TuningFastVariablePerLaneShuffle, 841 TuningPrefer256Bit, 842 TuningPOPCNTFalseDeps, 843 TuningInsertVZEROUPPER, 844 TuningAllowLight256Bit]; 845 list<SubtargetFeature> SKXFeatures = 846 !listconcat(BDWFeatures, SKXAdditionalFeatures); 847 848 // Cascadelake 849 list<SubtargetFeature> CLXAdditionalFeatures = [FeatureVNNI]; 850 list<SubtargetFeature> CLXTuning = SKXTuning; 851 list<SubtargetFeature> CLXFeatures = 852 !listconcat(SKXFeatures, CLXAdditionalFeatures); 853 854 // Cooperlake 855 list<SubtargetFeature> CPXAdditionalFeatures = [FeatureBF16]; 856 list<SubtargetFeature> CPXTuning = SKXTuning; 857 list<SubtargetFeature> CPXFeatures = 858 !listconcat(CLXFeatures, CPXAdditionalFeatures); 859 860 // Cannonlake 861 list<SubtargetFeature> CNLAdditionalFeatures = [FeatureAVX512, 862 FeatureCDI, 863 FeatureDQI, 864 FeatureBWI, 865 FeatureVLX, 866 FeaturePKU, 867 FeatureVBMI, 868 FeatureIFMA, 869 FeatureSHA]; 870 list<SubtargetFeature> CNLTuning = [TuningFastGather, 871 TuningMacroFusion, 872 TuningSlow3OpsLEA, 873 TuningSlowDivide64, 874 TuningFastScalarFSQRT, 875 TuningFastVectorFSQRT, 876 TuningFastSHLDRotate, 877 TuningFast15ByteNOP, 878 TuningFastVariableCrossLaneShuffle, 879 TuningFastVariablePerLaneShuffle, 880 TuningPrefer256Bit, 881 TuningInsertVZEROUPPER, 882 TuningAllowLight256Bit]; 883 list<SubtargetFeature> CNLFeatures = 884 !listconcat(SKLFeatures, CNLAdditionalFeatures); 885 886 // Icelake 887 list<SubtargetFeature> ICLAdditionalFeatures = [FeatureBITALG, 888 FeatureVAES, 889 FeatureVBMI2, 890 FeatureVNNI, 891 FeatureVPCLMULQDQ, 892 FeatureVPOPCNTDQ, 893 FeatureGFNI, 894 FeatureRDPID, 895 FeatureFSRM]; 896 list<SubtargetFeature> ICLTuning = [TuningFastGather, 897 TuningMacroFusion, 898 TuningSlowDivide64, 899 TuningFastScalarFSQRT, 900 TuningFastVectorFSQRT, 901 TuningFastSHLDRotate, 902 TuningFast15ByteNOP, 903 TuningFastVariableCrossLaneShuffle, 904 TuningFastVariablePerLaneShuffle, 905 TuningPrefer256Bit, 906 TuningInsertVZEROUPPER, 907 TuningAllowLight256Bit]; 908 list<SubtargetFeature> ICLFeatures = 909 !listconcat(CNLFeatures, ICLAdditionalFeatures); 910 911 // Icelake Server 912 list<SubtargetFeature> ICXAdditionalFeatures = [FeaturePCONFIG, 913 FeatureCLWB, 914 FeatureWBNOINVD]; 915 list<SubtargetFeature> ICXTuning = ICLTuning; 916 list<SubtargetFeature> ICXFeatures = 917 !listconcat(ICLFeatures, ICXAdditionalFeatures); 918 919 // Tigerlake 920 list<SubtargetFeature> TGLAdditionalFeatures = [FeatureVP2INTERSECT, 921 FeatureCLWB, 922 FeatureMOVDIRI, 923 FeatureMOVDIR64B, 924 FeatureSHSTK]; 925 list<SubtargetFeature> TGLTuning = ICLTuning; 926 list<SubtargetFeature> TGLFeatures = 927 !listconcat(ICLFeatures, TGLAdditionalFeatures ); 928 929 // Sapphirerapids 930 list<SubtargetFeature> SPRAdditionalFeatures = [FeatureAMXTILE, 931 FeatureAMXINT8, 932 FeatureAMXBF16, 933 FeatureBF16, 934 FeatureSERIALIZE, 935 FeatureCLDEMOTE, 936 FeatureWAITPKG, 937 FeaturePTWRITE, 938 FeatureFP16, 939 FeatureAVXVNNI, 940 FeatureTSXLDTRK, 941 FeatureENQCMD, 942 FeatureSHSTK, 943 FeatureMOVDIRI, 944 FeatureMOVDIR64B, 945 FeatureUINTR]; 946 list<SubtargetFeature> SPRAdditionalTuning = [TuningMULCFalseDeps, 947 TuningPERMFalseDeps, 948 TuningRANGEFalseDeps, 949 TuningGETMANTFalseDeps, 950 TuningMULLQFalseDeps]; 951 list<SubtargetFeature> SPRTuning = !listconcat(ICXTuning, SPRAdditionalTuning); 952 list<SubtargetFeature> SPRFeatures = 953 !listconcat(ICXFeatures, SPRAdditionalFeatures); 954 955 // Graniterapids 956 list<SubtargetFeature> GNRAdditionalFeatures = [FeatureAMXFP16, 957 FeaturePREFETCHI]; 958 list<SubtargetFeature> GNRFeatures = 959 !listconcat(SPRFeatures, GNRAdditionalFeatures); 960 961 // Atom 962 list<SubtargetFeature> AtomFeatures = [FeatureX87, 963 FeatureCX8, 964 FeatureCMOV, 965 FeatureMMX, 966 FeatureSSSE3, 967 FeatureFXSR, 968 FeatureNOPL, 969 FeatureX86_64, 970 FeatureCX16, 971 FeatureMOVBE, 972 FeatureLAHFSAHF64]; 973 list<SubtargetFeature> AtomTuning = [ProcIntelAtom, 974 TuningSlowUAMem16, 975 TuningLEAForSP, 976 TuningSlowDivide32, 977 TuningSlowDivide64, 978 TuningSlowTwoMemOps, 979 TuningLEAUsesAG, 980 TuningPadShortFunctions, 981 TuningInsertVZEROUPPER]; 982 983 // Silvermont 984 list<SubtargetFeature> SLMAdditionalFeatures = [FeatureSSE42, 985 FeatureCRC32, 986 FeaturePOPCNT, 987 FeaturePCLMUL, 988 FeaturePRFCHW, 989 FeatureRDRAND]; 990 list<SubtargetFeature> SLMTuning = [TuningUseSLMArithCosts, 991 TuningSlowTwoMemOps, 992 TuningSlowLEA, 993 TuningSlowIncDec, 994 TuningSlowDivide64, 995 TuningSlowPMULLD, 996 TuningFast7ByteNOP, 997 TuningFastMOVBE, 998 TuningPOPCNTFalseDeps, 999 TuningInsertVZEROUPPER]; 1000 list<SubtargetFeature> SLMFeatures = 1001 !listconcat(AtomFeatures, SLMAdditionalFeatures); 1002 1003 // Goldmont 1004 list<SubtargetFeature> GLMAdditionalFeatures = [FeatureAES, 1005 FeatureSHA, 1006 FeatureRDSEED, 1007 FeatureXSAVE, 1008 FeatureXSAVEOPT, 1009 FeatureXSAVEC, 1010 FeatureXSAVES, 1011 FeatureCLFLUSHOPT, 1012 FeatureFSGSBase]; 1013 list<SubtargetFeature> GLMTuning = [TuningUseGLMDivSqrtCosts, 1014 TuningSlowTwoMemOps, 1015 TuningSlowLEA, 1016 TuningSlowIncDec, 1017 TuningFastMOVBE, 1018 TuningPOPCNTFalseDeps, 1019 TuningInsertVZEROUPPER]; 1020 list<SubtargetFeature> GLMFeatures = 1021 !listconcat(SLMFeatures, GLMAdditionalFeatures); 1022 1023 // Goldmont Plus 1024 list<SubtargetFeature> GLPAdditionalFeatures = [FeaturePTWRITE, 1025 FeatureRDPID]; 1026 list<SubtargetFeature> GLPTuning = [TuningUseGLMDivSqrtCosts, 1027 TuningSlowTwoMemOps, 1028 TuningSlowLEA, 1029 TuningSlowIncDec, 1030 TuningFastMOVBE, 1031 TuningInsertVZEROUPPER]; 1032 list<SubtargetFeature> GLPFeatures = 1033 !listconcat(GLMFeatures, GLPAdditionalFeatures); 1034 1035 // Tremont 1036 list<SubtargetFeature> TRMAdditionalFeatures = [FeatureCLWB, 1037 FeatureGFNI]; 1038 list<SubtargetFeature> TRMTuning = GLPTuning; 1039 list<SubtargetFeature> TRMFeatures = 1040 !listconcat(GLPFeatures, TRMAdditionalFeatures); 1041 1042 // Alderlake 1043 list<SubtargetFeature> ADLAdditionalFeatures = [FeatureSERIALIZE, 1044 FeaturePCONFIG, 1045 FeatureSHSTK, 1046 FeatureWIDEKL, 1047 FeatureINVPCID, 1048 FeatureADX, 1049 FeatureFMA, 1050 FeatureVAES, 1051 FeatureVPCLMULQDQ, 1052 FeatureF16C, 1053 FeatureBMI, 1054 FeatureBMI2, 1055 FeatureLZCNT, 1056 FeatureAVXVNNI, 1057 FeaturePKU, 1058 FeatureHRESET, 1059 FeatureCLDEMOTE, 1060 FeatureMOVDIRI, 1061 FeatureMOVDIR64B, 1062 FeatureWAITPKG]; 1063 list<SubtargetFeature> ADLAdditionalTuning = [TuningPERMFalseDeps]; 1064 list<SubtargetFeature> ADLTuning = !listconcat(SKLTuning, ADLAdditionalTuning); 1065 list<SubtargetFeature> ADLFeatures = 1066 !listconcat(TRMFeatures, ADLAdditionalFeatures); 1067 1068 // Sierraforest 1069 list<SubtargetFeature> SRFAdditionalFeatures = [FeatureCMPCCXADD, 1070 FeatureAVXIFMA, 1071 FeatureAVXNECONVERT, 1072 FeatureAVXVNNIINT8]; 1073 list<SubtargetFeature> SRFFeatures = 1074 !listconcat(ADLFeatures, SRFAdditionalFeatures); 1075 1076 // Grandridge 1077 list<SubtargetFeature> GRRAdditionalFeatures = [FeatureRAOINT]; 1078 list<SubtargetFeature> GRRFeatures = 1079 !listconcat(SRFFeatures, GRRAdditionalFeatures); 1080 1081 // Knights Landing 1082 list<SubtargetFeature> KNLFeatures = [FeatureX87, 1083 FeatureCX8, 1084 FeatureCMOV, 1085 FeatureMMX, 1086 FeatureFXSR, 1087 FeatureNOPL, 1088 FeatureX86_64, 1089 FeatureCX16, 1090 FeatureCRC32, 1091 FeaturePOPCNT, 1092 FeaturePCLMUL, 1093 FeatureXSAVE, 1094 FeatureXSAVEOPT, 1095 FeatureLAHFSAHF64, 1096 FeatureAES, 1097 FeatureRDRAND, 1098 FeatureF16C, 1099 FeatureFSGSBase, 1100 FeatureAVX512, 1101 FeatureERI, 1102 FeatureCDI, 1103 FeaturePFI, 1104 FeaturePREFETCHWT1, 1105 FeatureADX, 1106 FeatureRDSEED, 1107 FeatureMOVBE, 1108 FeatureLZCNT, 1109 FeatureBMI, 1110 FeatureBMI2, 1111 FeatureFMA, 1112 FeaturePRFCHW]; 1113 list<SubtargetFeature> KNLTuning = [TuningSlowDivide64, 1114 TuningSlow3OpsLEA, 1115 TuningSlowIncDec, 1116 TuningSlowTwoMemOps, 1117 TuningPreferMaskRegisters, 1118 TuningFastGather, 1119 TuningFastMOVBE, 1120 TuningSlowPMADDWD]; 1121 // TODO Add AVX5124FMAPS/AVX5124VNNIW features 1122 list<SubtargetFeature> KNMFeatures = 1123 !listconcat(KNLFeatures, [FeatureVPOPCNTDQ]); 1124 1125 // Barcelona 1126 list<SubtargetFeature> BarcelonaFeatures = [FeatureX87, 1127 FeatureCX8, 1128 FeatureSSE4A, 1129 Feature3DNowA, 1130 FeatureFXSR, 1131 FeatureNOPL, 1132 FeatureCX16, 1133 FeaturePRFCHW, 1134 FeatureLZCNT, 1135 FeaturePOPCNT, 1136 FeatureLAHFSAHF64, 1137 FeatureCMOV, 1138 FeatureX86_64]; 1139 list<SubtargetFeature> BarcelonaTuning = [TuningFastScalarShiftMasks, 1140 TuningSlowSHLD, 1141 TuningSBBDepBreaking, 1142 TuningInsertVZEROUPPER]; 1143 1144 // Bobcat 1145 list<SubtargetFeature> BtVer1Features = [FeatureX87, 1146 FeatureCX8, 1147 FeatureCMOV, 1148 FeatureMMX, 1149 FeatureSSSE3, 1150 FeatureSSE4A, 1151 FeatureFXSR, 1152 FeatureNOPL, 1153 FeatureX86_64, 1154 FeatureCX16, 1155 FeaturePRFCHW, 1156 FeatureLZCNT, 1157 FeaturePOPCNT, 1158 FeatureLAHFSAHF64]; 1159 list<SubtargetFeature> BtVer1Tuning = [TuningFast15ByteNOP, 1160 TuningFastScalarShiftMasks, 1161 TuningFastVectorShiftMasks, 1162 TuningSlowSHLD, 1163 TuningSBBDepBreaking, 1164 TuningInsertVZEROUPPER]; 1165 1166 // Jaguar 1167 list<SubtargetFeature> BtVer2AdditionalFeatures = [FeatureAVX, 1168 FeatureAES, 1169 FeatureCRC32, 1170 FeaturePCLMUL, 1171 FeatureBMI, 1172 FeatureF16C, 1173 FeatureMOVBE, 1174 FeatureXSAVE, 1175 FeatureXSAVEOPT]; 1176 list<SubtargetFeature> BtVer2Tuning = [TuningFastLZCNT, 1177 TuningFastBEXTR, 1178 TuningFastHorizontalOps, 1179 TuningFast15ByteNOP, 1180 TuningFastScalarShiftMasks, 1181 TuningFastVectorShiftMasks, 1182 TuningFastMOVBE, 1183 TuningSBBDepBreaking, 1184 TuningSlowSHLD]; 1185 list<SubtargetFeature> BtVer2Features = 1186 !listconcat(BtVer1Features, BtVer2AdditionalFeatures); 1187 1188 // Bulldozer 1189 list<SubtargetFeature> BdVer1Features = [FeatureX87, 1190 FeatureCX8, 1191 FeatureCMOV, 1192 FeatureXOP, 1193 FeatureX86_64, 1194 FeatureCX16, 1195 FeatureAES, 1196 FeatureCRC32, 1197 FeaturePRFCHW, 1198 FeaturePCLMUL, 1199 FeatureMMX, 1200 FeatureFXSR, 1201 FeatureNOPL, 1202 FeatureLZCNT, 1203 FeaturePOPCNT, 1204 FeatureXSAVE, 1205 FeatureLWP, 1206 FeatureLAHFSAHF64]; 1207 list<SubtargetFeature> BdVer1Tuning = [TuningSlowSHLD, 1208 TuningFast11ByteNOP, 1209 TuningFastScalarShiftMasks, 1210 TuningBranchFusion, 1211 TuningSBBDepBreaking, 1212 TuningInsertVZEROUPPER]; 1213 1214 // PileDriver 1215 list<SubtargetFeature> BdVer2AdditionalFeatures = [FeatureF16C, 1216 FeatureBMI, 1217 FeatureTBM, 1218 FeatureFMA]; 1219 list<SubtargetFeature> BdVer2AdditionalTuning = [TuningFastBEXTR, 1220 TuningFastMOVBE]; 1221 list<SubtargetFeature> BdVer2Tuning = 1222 !listconcat(BdVer1Tuning, BdVer2AdditionalTuning); 1223 list<SubtargetFeature> BdVer2Features = 1224 !listconcat(BdVer1Features, BdVer2AdditionalFeatures); 1225 1226 // Steamroller 1227 list<SubtargetFeature> BdVer3AdditionalFeatures = [FeatureXSAVEOPT, 1228 FeatureFSGSBase]; 1229 list<SubtargetFeature> BdVer3Tuning = BdVer2Tuning; 1230 list<SubtargetFeature> BdVer3Features = 1231 !listconcat(BdVer2Features, BdVer3AdditionalFeatures); 1232 1233 // Excavator 1234 list<SubtargetFeature> BdVer4AdditionalFeatures = [FeatureAVX2, 1235 FeatureBMI2, 1236 FeatureMOVBE, 1237 FeatureRDRAND, 1238 FeatureMWAITX]; 1239 list<SubtargetFeature> BdVer4Tuning = BdVer3Tuning; 1240 list<SubtargetFeature> BdVer4Features = 1241 !listconcat(BdVer3Features, BdVer4AdditionalFeatures); 1242 1243 1244 // AMD Zen Processors common ISAs 1245 list<SubtargetFeature> ZNFeatures = [FeatureADX, 1246 FeatureAES, 1247 FeatureAVX2, 1248 FeatureBMI, 1249 FeatureBMI2, 1250 FeatureCLFLUSHOPT, 1251 FeatureCLZERO, 1252 FeatureCMOV, 1253 FeatureX86_64, 1254 FeatureCX16, 1255 FeatureCRC32, 1256 FeatureF16C, 1257 FeatureFMA, 1258 FeatureFSGSBase, 1259 FeatureFXSR, 1260 FeatureNOPL, 1261 FeatureLAHFSAHF64, 1262 FeatureLZCNT, 1263 FeatureMMX, 1264 FeatureMOVBE, 1265 FeatureMWAITX, 1266 FeaturePCLMUL, 1267 FeaturePOPCNT, 1268 FeaturePRFCHW, 1269 FeatureRDRAND, 1270 FeatureRDSEED, 1271 FeatureSHA, 1272 FeatureSSE4A, 1273 FeatureX87, 1274 FeatureXSAVE, 1275 FeatureXSAVEC, 1276 FeatureXSAVEOPT, 1277 FeatureXSAVES]; 1278 list<SubtargetFeature> ZNTuning = [TuningFastLZCNT, 1279 TuningFastBEXTR, 1280 TuningFast15ByteNOP, 1281 TuningBranchFusion, 1282 TuningFastScalarFSQRT, 1283 TuningFastVectorFSQRT, 1284 TuningFastScalarShiftMasks, 1285 TuningFastVariablePerLaneShuffle, 1286 TuningFastMOVBE, 1287 TuningSlowSHLD, 1288 TuningSBBDepBreaking, 1289 TuningInsertVZEROUPPER, 1290 TuningAllowLight256Bit]; 1291 list<SubtargetFeature> ZN2AdditionalFeatures = [FeatureCLWB, 1292 FeatureRDPID, 1293 FeatureRDPRU, 1294 FeatureWBNOINVD]; 1295 list<SubtargetFeature> ZN2Tuning = ZNTuning; 1296 list<SubtargetFeature> ZN2Features = 1297 !listconcat(ZNFeatures, ZN2AdditionalFeatures); 1298 list<SubtargetFeature> ZN3AdditionalFeatures = [FeatureFSRM, 1299 FeatureINVPCID, 1300 FeaturePKU, 1301 FeatureVAES, 1302 FeatureVPCLMULQDQ]; 1303 list<SubtargetFeature> ZN3AdditionalTuning = [TuningMacroFusion]; 1304 list<SubtargetFeature> ZN3Tuning = 1305 !listconcat(ZN2Tuning, ZN3AdditionalTuning); 1306 list<SubtargetFeature> ZN3Features = 1307 !listconcat(ZN2Features, ZN3AdditionalFeatures); 1308 list<SubtargetFeature> ZN4Tuning = ZN3Tuning; 1309 list<SubtargetFeature> ZN4AdditionalFeatures = [FeatureAVX512, 1310 FeatureCDI, 1311 FeatureDQI, 1312 FeatureBWI, 1313 FeatureVLX, 1314 FeatureVBMI, 1315 FeatureVBMI2, 1316 FeatureIFMA, 1317 FeatureVNNI, 1318 FeatureBITALG, 1319 FeatureGFNI, 1320 FeatureBF16, 1321 FeatureSHSTK, 1322 FeatureVPOPCNTDQ]; 1323 list<SubtargetFeature> ZN4Features = 1324 !listconcat(ZN3Features, ZN4AdditionalFeatures); 1325} 1326 1327//===----------------------------------------------------------------------===// 1328// X86 processors supported. 1329//===----------------------------------------------------------------------===// 1330 1331class Proc<string Name, list<SubtargetFeature> Features, 1332 list<SubtargetFeature> TuneFeatures> 1333 : ProcessorModel<Name, GenericModel, Features, TuneFeatures>; 1334 1335class ProcModel<string Name, SchedMachineModel Model, 1336 list<SubtargetFeature> Features, 1337 list<SubtargetFeature> TuneFeatures> 1338 : ProcessorModel<Name, Model, Features, TuneFeatures>; 1339 1340// NOTE: CMPXCHG8B is here for legacy compatibility so that it is only disabled 1341// if i386/i486 is specifically requested. 1342// NOTE: 64Bit is here as "generic" is the default llc CPU. The X86Subtarget 1343// constructor checks that any CPU used in 64-bit mode has FeatureX86_64 1344// enabled. It has no effect on code generation. 1345// NOTE: As a default tuning, "generic" aims to produce code optimized for the 1346// most common X86 processors. The tunings might be changed over time. It is 1347// recommended to use "tune-cpu"="x86-64" in function attribute for consistency. 1348def : ProcModel<"generic", SandyBridgeModel, 1349 [FeatureX87, FeatureCX8, FeatureX86_64], 1350 [TuningSlow3OpsLEA, 1351 TuningSlowDivide64, 1352 TuningMacroFusion, 1353 TuningFastScalarFSQRT, 1354 TuningFast15ByteNOP, 1355 TuningInsertVZEROUPPER]>; 1356 1357def : Proc<"i386", [FeatureX87], 1358 [TuningSlowUAMem16, TuningInsertVZEROUPPER]>; 1359def : Proc<"i486", [FeatureX87], 1360 [TuningSlowUAMem16, TuningInsertVZEROUPPER]>; 1361def : Proc<"i586", [FeatureX87, FeatureCX8], 1362 [TuningSlowUAMem16, TuningInsertVZEROUPPER]>; 1363def : Proc<"pentium", [FeatureX87, FeatureCX8], 1364 [TuningSlowUAMem16, TuningInsertVZEROUPPER]>; 1365def : Proc<"pentium-mmx", [FeatureX87, FeatureCX8, FeatureMMX], 1366 [TuningSlowUAMem16, TuningInsertVZEROUPPER]>; 1367 1368def : Proc<"i686", [FeatureX87, FeatureCX8, FeatureCMOV], 1369 [TuningSlowUAMem16, TuningInsertVZEROUPPER]>; 1370def : Proc<"pentiumpro", [FeatureX87, FeatureCX8, FeatureCMOV, 1371 FeatureNOPL], 1372 [TuningSlowUAMem16, TuningInsertVZEROUPPER]>; 1373 1374def : Proc<"pentium2", [FeatureX87, FeatureCX8, FeatureMMX, FeatureCMOV, 1375 FeatureFXSR, FeatureNOPL], 1376 [TuningSlowUAMem16, TuningInsertVZEROUPPER]>; 1377 1378foreach P = ["pentium3", "pentium3m"] in { 1379 def : Proc<P, [FeatureX87, FeatureCX8, FeatureMMX, 1380 FeatureSSE1, FeatureFXSR, FeatureNOPL, FeatureCMOV], 1381 [TuningSlowUAMem16, TuningInsertVZEROUPPER]>; 1382} 1383 1384// Enable the PostRAScheduler for SSE2 and SSE3 class cpus. 1385// The intent is to enable it for pentium4 which is the current default 1386// processor in a vanilla 32-bit clang compilation when no specific 1387// architecture is specified. This generally gives a nice performance 1388// increase on silvermont, with largely neutral behavior on other 1389// contemporary large core processors. 1390// pentium-m, pentium4m, prescott and nocona are included as a preventative 1391// measure to avoid performance surprises, in case clang's default cpu 1392// changes slightly. 1393 1394def : ProcModel<"pentium-m", GenericPostRAModel, 1395 [FeatureX87, FeatureCX8, FeatureMMX, FeatureSSE2, 1396 FeatureFXSR, FeatureNOPL, FeatureCMOV], 1397 [TuningSlowUAMem16, TuningInsertVZEROUPPER]>; 1398 1399foreach P = ["pentium4", "pentium4m"] in { 1400 def : ProcModel<P, GenericPostRAModel, 1401 [FeatureX87, FeatureCX8, FeatureMMX, FeatureSSE2, 1402 FeatureFXSR, FeatureNOPL, FeatureCMOV], 1403 [TuningSlowUAMem16, TuningInsertVZEROUPPER]>; 1404} 1405 1406// Intel Quark. 1407def : Proc<"lakemont", [FeatureCX8], 1408 [TuningSlowUAMem16, TuningInsertVZEROUPPER]>; 1409 1410// Intel Core Duo. 1411def : ProcModel<"yonah", SandyBridgeModel, 1412 [FeatureX87, FeatureCX8, FeatureMMX, FeatureSSE3, 1413 FeatureFXSR, FeatureNOPL, FeatureCMOV], 1414 [TuningSlowUAMem16, TuningInsertVZEROUPPER]>; 1415 1416// NetBurst. 1417def : ProcModel<"prescott", GenericPostRAModel, 1418 [FeatureX87, FeatureCX8, FeatureMMX, FeatureSSE3, 1419 FeatureFXSR, FeatureNOPL, FeatureCMOV], 1420 [TuningSlowUAMem16, TuningInsertVZEROUPPER]>; 1421def : ProcModel<"nocona", GenericPostRAModel, [ 1422 FeatureX87, 1423 FeatureCX8, 1424 FeatureCMOV, 1425 FeatureMMX, 1426 FeatureSSE3, 1427 FeatureFXSR, 1428 FeatureNOPL, 1429 FeatureX86_64, 1430 FeatureCX16, 1431], 1432[ 1433 TuningSlowUAMem16, 1434 TuningInsertVZEROUPPER 1435]>; 1436 1437// Intel Core 2 Solo/Duo. 1438def : ProcModel<"core2", SandyBridgeModel, [ 1439 FeatureX87, 1440 FeatureCX8, 1441 FeatureCMOV, 1442 FeatureMMX, 1443 FeatureSSSE3, 1444 FeatureFXSR, 1445 FeatureNOPL, 1446 FeatureX86_64, 1447 FeatureCX16, 1448 FeatureLAHFSAHF64 1449], 1450[ 1451 TuningMacroFusion, 1452 TuningSlowUAMem16, 1453 TuningInsertVZEROUPPER 1454]>; 1455def : ProcModel<"penryn", SandyBridgeModel, [ 1456 FeatureX87, 1457 FeatureCX8, 1458 FeatureCMOV, 1459 FeatureMMX, 1460 FeatureSSE41, 1461 FeatureFXSR, 1462 FeatureNOPL, 1463 FeatureX86_64, 1464 FeatureCX16, 1465 FeatureLAHFSAHF64 1466], 1467[ 1468 TuningMacroFusion, 1469 TuningSlowUAMem16, 1470 TuningInsertVZEROUPPER 1471]>; 1472 1473// Atom CPUs. 1474foreach P = ["bonnell", "atom"] in { 1475 def : ProcModel<P, AtomModel, ProcessorFeatures.AtomFeatures, 1476 ProcessorFeatures.AtomTuning>; 1477} 1478 1479foreach P = ["silvermont", "slm"] in { 1480 def : ProcModel<P, SLMModel, ProcessorFeatures.SLMFeatures, 1481 ProcessorFeatures.SLMTuning>; 1482} 1483 1484def : ProcModel<"goldmont", SLMModel, ProcessorFeatures.GLMFeatures, 1485 ProcessorFeatures.GLMTuning>; 1486def : ProcModel<"goldmont-plus", SLMModel, ProcessorFeatures.GLPFeatures, 1487 ProcessorFeatures.GLPTuning>; 1488def : ProcModel<"tremont", SLMModel, ProcessorFeatures.TRMFeatures, 1489 ProcessorFeatures.TRMTuning>; 1490def : ProcModel<"sierraforest", AlderlakePModel, ProcessorFeatures.SRFFeatures, 1491 ProcessorFeatures.TRMTuning>; 1492def : ProcModel<"grandridge", AlderlakePModel, ProcessorFeatures.GRRFeatures, 1493 ProcessorFeatures.TRMTuning>; 1494 1495// "Arrandale" along with corei3 and corei5 1496foreach P = ["nehalem", "corei7"] in { 1497 def : ProcModel<P, SandyBridgeModel, ProcessorFeatures.NHMFeatures, 1498 ProcessorFeatures.NHMTuning>; 1499} 1500 1501// Westmere is the corei3/i5/i7 path from nehalem to sandybridge 1502def : ProcModel<"westmere", SandyBridgeModel, ProcessorFeatures.WSMFeatures, 1503 ProcessorFeatures.WSMTuning>; 1504 1505foreach P = ["sandybridge", "corei7-avx"] in { 1506 def : ProcModel<P, SandyBridgeModel, ProcessorFeatures.SNBFeatures, 1507 ProcessorFeatures.SNBTuning>; 1508} 1509 1510foreach P = ["ivybridge", "core-avx-i"] in { 1511 def : ProcModel<P, SandyBridgeModel, ProcessorFeatures.IVBFeatures, 1512 ProcessorFeatures.IVBTuning>; 1513} 1514 1515foreach P = ["haswell", "core-avx2"] in { 1516 def : ProcModel<P, HaswellModel, ProcessorFeatures.HSWFeatures, 1517 ProcessorFeatures.HSWTuning>; 1518} 1519 1520def : ProcModel<"broadwell", BroadwellModel, ProcessorFeatures.BDWFeatures, 1521 ProcessorFeatures.BDWTuning>; 1522 1523def : ProcModel<"skylake", SkylakeClientModel, ProcessorFeatures.SKLFeatures, 1524 ProcessorFeatures.SKLTuning>; 1525 1526// FIXME: define KNL scheduler model 1527def : ProcModel<"knl", HaswellModel, ProcessorFeatures.KNLFeatures, 1528 ProcessorFeatures.KNLTuning>; 1529def : ProcModel<"knm", HaswellModel, ProcessorFeatures.KNMFeatures, 1530 ProcessorFeatures.KNLTuning>; 1531 1532foreach P = ["skylake-avx512", "skx"] in { 1533 def : ProcModel<P, SkylakeServerModel, ProcessorFeatures.SKXFeatures, 1534 ProcessorFeatures.SKXTuning>; 1535} 1536 1537def : ProcModel<"cascadelake", SkylakeServerModel, 1538 ProcessorFeatures.CLXFeatures, ProcessorFeatures.CLXTuning>; 1539def : ProcModel<"cooperlake", SkylakeServerModel, 1540 ProcessorFeatures.CPXFeatures, ProcessorFeatures.CPXTuning>; 1541def : ProcModel<"cannonlake", SkylakeServerModel, 1542 ProcessorFeatures.CNLFeatures, ProcessorFeatures.CNLTuning>; 1543def : ProcModel<"icelake-client", IceLakeModel, 1544 ProcessorFeatures.ICLFeatures, ProcessorFeatures.ICLTuning>; 1545def : ProcModel<"rocketlake", IceLakeModel, 1546 ProcessorFeatures.ICLFeatures, ProcessorFeatures.ICLTuning>; 1547def : ProcModel<"icelake-server", IceLakeModel, 1548 ProcessorFeatures.ICXFeatures, ProcessorFeatures.ICXTuning>; 1549def : ProcModel<"tigerlake", IceLakeModel, 1550 ProcessorFeatures.TGLFeatures, ProcessorFeatures.TGLTuning>; 1551def : ProcModel<"sapphirerapids", SkylakeServerModel, 1552 ProcessorFeatures.SPRFeatures, ProcessorFeatures.SPRTuning>; 1553def : ProcModel<"alderlake", AlderlakePModel, 1554 ProcessorFeatures.ADLFeatures, ProcessorFeatures.ADLTuning>; 1555def : ProcModel<"raptorlake", AlderlakePModel, 1556 ProcessorFeatures.ADLFeatures, ProcessorFeatures.ADLTuning>; 1557def : ProcModel<"meteorlake", AlderlakePModel, 1558 ProcessorFeatures.ADLFeatures, ProcessorFeatures.ADLTuning>; 1559def : ProcModel<"graniterapids", SkylakeServerModel, 1560 ProcessorFeatures.GNRFeatures, ProcessorFeatures.SPRTuning>; 1561def : ProcModel<"emeraldrapids", SkylakeServerModel, 1562 ProcessorFeatures.SPRFeatures, ProcessorFeatures.SPRTuning>; 1563 1564// AMD CPUs. 1565 1566def : Proc<"k6", [FeatureX87, FeatureCX8, FeatureMMX], 1567 [TuningSlowUAMem16, TuningInsertVZEROUPPER]>; 1568def : Proc<"k6-2", [FeatureX87, FeatureCX8, Feature3DNow], 1569 [TuningSlowUAMem16, TuningInsertVZEROUPPER]>; 1570def : Proc<"k6-3", [FeatureX87, FeatureCX8, Feature3DNow], 1571 [TuningSlowUAMem16, TuningInsertVZEROUPPER]>; 1572 1573foreach P = ["athlon", "athlon-tbird"] in { 1574 def : Proc<P, [FeatureX87, FeatureCX8, FeatureCMOV, Feature3DNowA, 1575 FeatureNOPL], 1576 [TuningSlowSHLD, TuningSlowUAMem16, TuningInsertVZEROUPPER]>; 1577} 1578 1579foreach P = ["athlon-4", "athlon-xp", "athlon-mp"] in { 1580 def : Proc<P, [FeatureX87, FeatureCX8, FeatureCMOV, 1581 FeatureSSE1, Feature3DNowA, FeatureFXSR, FeatureNOPL], 1582 [TuningSlowSHLD, TuningSlowUAMem16, TuningInsertVZEROUPPER]>; 1583} 1584 1585foreach P = ["k8", "opteron", "athlon64", "athlon-fx"] in { 1586 def : Proc<P, [FeatureX87, FeatureCX8, FeatureSSE2, Feature3DNowA, 1587 FeatureFXSR, FeatureNOPL, FeatureX86_64, FeatureCMOV], 1588 [TuningFastScalarShiftMasks, TuningSlowSHLD, TuningSlowUAMem16, 1589 TuningSBBDepBreaking, TuningInsertVZEROUPPER]>; 1590} 1591 1592foreach P = ["k8-sse3", "opteron-sse3", "athlon64-sse3"] in { 1593 def : Proc<P, [FeatureX87, FeatureCX8, FeatureSSE3, Feature3DNowA, 1594 FeatureFXSR, FeatureNOPL, FeatureCX16, FeatureCMOV, 1595 FeatureX86_64], 1596 [TuningFastScalarShiftMasks, TuningSlowSHLD, TuningSlowUAMem16, 1597 TuningSBBDepBreaking, TuningInsertVZEROUPPER]>; 1598} 1599 1600foreach P = ["amdfam10", "barcelona"] in { 1601 def : Proc<P, ProcessorFeatures.BarcelonaFeatures, 1602 ProcessorFeatures.BarcelonaTuning>; 1603} 1604 1605// Bobcat 1606def : Proc<"btver1", ProcessorFeatures.BtVer1Features, 1607 ProcessorFeatures.BtVer1Tuning>; 1608// Jaguar 1609def : ProcModel<"btver2", BtVer2Model, ProcessorFeatures.BtVer2Features, 1610 ProcessorFeatures.BtVer2Tuning>; 1611 1612// Bulldozer 1613def : ProcModel<"bdver1", BdVer2Model, ProcessorFeatures.BdVer1Features, 1614 ProcessorFeatures.BdVer1Tuning>; 1615// Piledriver 1616def : ProcModel<"bdver2", BdVer2Model, ProcessorFeatures.BdVer2Features, 1617 ProcessorFeatures.BdVer2Tuning>; 1618// Steamroller 1619def : Proc<"bdver3", ProcessorFeatures.BdVer3Features, 1620 ProcessorFeatures.BdVer3Tuning>; 1621// Excavator 1622def : Proc<"bdver4", ProcessorFeatures.BdVer4Features, 1623 ProcessorFeatures.BdVer4Tuning>; 1624 1625def : ProcModel<"znver1", Znver1Model, ProcessorFeatures.ZNFeatures, 1626 ProcessorFeatures.ZNTuning>; 1627def : ProcModel<"znver2", Znver2Model, ProcessorFeatures.ZN2Features, 1628 ProcessorFeatures.ZN2Tuning>; 1629def : ProcModel<"znver3", Znver3Model, ProcessorFeatures.ZN3Features, 1630 ProcessorFeatures.ZN3Tuning>; 1631def : ProcModel<"znver4", Znver4Model, ProcessorFeatures.ZN4Features, 1632 ProcessorFeatures.ZN4Tuning>; 1633 1634def : Proc<"geode", [FeatureX87, FeatureCX8, Feature3DNowA], 1635 [TuningSlowUAMem16, TuningInsertVZEROUPPER]>; 1636 1637def : Proc<"winchip-c6", [FeatureX87, FeatureMMX], 1638 [TuningSlowUAMem16, TuningInsertVZEROUPPER]>; 1639def : Proc<"winchip2", [FeatureX87, Feature3DNow], 1640 [TuningSlowUAMem16, TuningInsertVZEROUPPER]>; 1641def : Proc<"c3", [FeatureX87, Feature3DNow], 1642 [TuningSlowUAMem16, TuningInsertVZEROUPPER]>; 1643def : Proc<"c3-2", [FeatureX87, FeatureCX8, FeatureMMX, 1644 FeatureSSE1, FeatureFXSR, FeatureCMOV], 1645 [TuningSlowUAMem16, TuningInsertVZEROUPPER]>; 1646 1647// We also provide a generic 64-bit specific x86 processor model which tries to 1648// be good for modern chips without enabling instruction set encodings past the 1649// basic SSE2 and 64-bit ones. It disables slow things from any mainstream and 1650// modern 64-bit x86 chip, and enables features that are generally beneficial. 1651// 1652// We currently use the Sandy Bridge model as the default scheduling model as 1653// we use it across Nehalem, Westmere, Sandy Bridge, and Ivy Bridge which 1654// covers a huge swath of x86 processors. If there are specific scheduling 1655// knobs which need to be tuned differently for AMD chips, we might consider 1656// forming a common base for them. 1657def : ProcModel<"x86-64", SandyBridgeModel, ProcessorFeatures.X86_64V1Features, 1658[ 1659 TuningSlow3OpsLEA, 1660 TuningSlowDivide64, 1661 TuningSlowIncDec, 1662 TuningMacroFusion, 1663 TuningInsertVZEROUPPER 1664]>; 1665 1666// x86-64 micro-architecture levels. 1667def : ProcModel<"x86-64-v2", SandyBridgeModel, ProcessorFeatures.X86_64V2Features, 1668 ProcessorFeatures.SNBTuning>; 1669// Close to Haswell. 1670def : ProcModel<"x86-64-v3", HaswellModel, ProcessorFeatures.X86_64V3Features, 1671 ProcessorFeatures.HSWTuning>; 1672// Close to the AVX-512 level implemented by Xeon Scalable Processors. 1673def : ProcModel<"x86-64-v4", SkylakeServerModel, ProcessorFeatures.X86_64V4Features, 1674 ProcessorFeatures.SKXTuning>; 1675 1676//===----------------------------------------------------------------------===// 1677// Calling Conventions 1678//===----------------------------------------------------------------------===// 1679 1680include "X86CallingConv.td" 1681 1682 1683//===----------------------------------------------------------------------===// 1684// Assembly Parser 1685//===----------------------------------------------------------------------===// 1686 1687def ATTAsmParserVariant : AsmParserVariant { 1688 int Variant = 0; 1689 1690 // Variant name. 1691 string Name = "att"; 1692 1693 // Discard comments in assembly strings. 1694 string CommentDelimiter = "#"; 1695 1696 // Recognize hard coded registers. 1697 string RegisterPrefix = "%"; 1698} 1699 1700def IntelAsmParserVariant : AsmParserVariant { 1701 int Variant = 1; 1702 1703 // Variant name. 1704 string Name = "intel"; 1705 1706 // Discard comments in assembly strings. 1707 string CommentDelimiter = ";"; 1708 1709 // Recognize hard coded registers. 1710 string RegisterPrefix = ""; 1711} 1712 1713//===----------------------------------------------------------------------===// 1714// Assembly Printers 1715//===----------------------------------------------------------------------===// 1716 1717// The X86 target supports two different syntaxes for emitting machine code. 1718// This is controlled by the -x86-asm-syntax={att|intel} 1719def ATTAsmWriter : AsmWriter { 1720 string AsmWriterClassName = "ATTInstPrinter"; 1721 int Variant = 0; 1722} 1723def IntelAsmWriter : AsmWriter { 1724 string AsmWriterClassName = "IntelInstPrinter"; 1725 int Variant = 1; 1726} 1727 1728def X86 : Target { 1729 // Information about the instructions... 1730 let InstructionSet = X86InstrInfo; 1731 let AssemblyParserVariants = [ATTAsmParserVariant, IntelAsmParserVariant]; 1732 let AssemblyWriters = [ATTAsmWriter, IntelAsmWriter]; 1733 let AllowRegisterRenaming = 1; 1734} 1735 1736//===----------------------------------------------------------------------===// 1737// Pfm Counters 1738//===----------------------------------------------------------------------===// 1739 1740include "X86PfmCounters.td" 1741