1 //===-- X86Subtarget.cpp - X86 Subtarget Information ----------------------===// 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 // 10 // This file implements the X86 specific subclass of TargetSubtargetInfo. 11 // 12 //===----------------------------------------------------------------------===// 13 14 #define DEBUG_TYPE "subtarget" 15 #include "X86Subtarget.h" 16 #include "X86InstrInfo.h" 17 #include "llvm/IR/Attributes.h" 18 #include "llvm/IR/Function.h" 19 #include "llvm/IR/GlobalValue.h" 20 #include "llvm/Support/Debug.h" 21 #include "llvm/Support/ErrorHandling.h" 22 #include "llvm/Support/Host.h" 23 #include "llvm/Support/raw_ostream.h" 24 #include "llvm/Target/TargetMachine.h" 25 #include "llvm/Target/TargetOptions.h" 26 27 #define GET_SUBTARGETINFO_TARGET_DESC 28 #define GET_SUBTARGETINFO_CTOR 29 #include "X86GenSubtargetInfo.inc" 30 31 using namespace llvm; 32 33 #if defined(_MSC_VER) 34 #include <intrin.h> 35 #endif 36 37 /// ClassifyBlockAddressReference - Classify a blockaddress reference for the 38 /// current subtarget according to how we should reference it in a non-pcrel 39 /// context. 40 unsigned char X86Subtarget::ClassifyBlockAddressReference() const { 41 if (isPICStyleGOT()) // 32-bit ELF targets. 42 return X86II::MO_GOTOFF; 43 44 if (isPICStyleStubPIC()) // Darwin/32 in PIC mode. 45 return X86II::MO_PIC_BASE_OFFSET; 46 47 // Direct static reference to label. 48 return X86II::MO_NO_FLAG; 49 } 50 51 /// ClassifyGlobalReference - Classify a global variable reference for the 52 /// current subtarget according to how we should reference it in a non-pcrel 53 /// context. 54 unsigned char X86Subtarget:: 55 ClassifyGlobalReference(const GlobalValue *GV, const TargetMachine &TM) const { 56 // DLLImport only exists on windows, it is implemented as a load from a 57 // DLLIMPORT stub. 58 if (GV->hasDLLImportStorageClass()) 59 return X86II::MO_DLLIMPORT; 60 61 // Determine whether this is a reference to a definition or a declaration. 62 // Materializable GVs (in JIT lazy compilation mode) do not require an extra 63 // load from stub. 64 bool isDecl = GV->hasAvailableExternallyLinkage(); 65 if (GV->isDeclaration() && !GV->isMaterializable()) 66 isDecl = true; 67 68 // X86-64 in PIC mode. 69 if (isPICStyleRIPRel()) { 70 // Large model never uses stubs. 71 if (TM.getCodeModel() == CodeModel::Large) 72 return X86II::MO_NO_FLAG; 73 74 if (isTargetDarwin()) { 75 // If symbol visibility is hidden, the extra load is not needed if 76 // target is x86-64 or the symbol is definitely defined in the current 77 // translation unit. 78 if (GV->hasDefaultVisibility() && 79 (isDecl || GV->isWeakForLinker())) 80 return X86II::MO_GOTPCREL; 81 } else if (!isTargetWin64()) { 82 assert(isTargetELF() && "Unknown rip-relative target"); 83 84 // Extra load is needed for all externally visible. 85 if (!GV->hasLocalLinkage() && GV->hasDefaultVisibility()) 86 return X86II::MO_GOTPCREL; 87 } 88 89 return X86II::MO_NO_FLAG; 90 } 91 92 if (isPICStyleGOT()) { // 32-bit ELF targets. 93 // Extra load is needed for all externally visible. 94 if (GV->hasLocalLinkage() || GV->hasHiddenVisibility()) 95 return X86II::MO_GOTOFF; 96 return X86II::MO_GOT; 97 } 98 99 if (isPICStyleStubPIC()) { // Darwin/32 in PIC mode. 100 // Determine whether we have a stub reference and/or whether the reference 101 // is relative to the PIC base or not. 102 103 // If this is a strong reference to a definition, it is definitely not 104 // through a stub. 105 if (!isDecl && !GV->isWeakForLinker()) 106 return X86II::MO_PIC_BASE_OFFSET; 107 108 // Unless we have a symbol with hidden visibility, we have to go through a 109 // normal $non_lazy_ptr stub because this symbol might be resolved late. 110 if (!GV->hasHiddenVisibility()) // Non-hidden $non_lazy_ptr reference. 111 return X86II::MO_DARWIN_NONLAZY_PIC_BASE; 112 113 // If symbol visibility is hidden, we have a stub for common symbol 114 // references and external declarations. 115 if (isDecl || GV->hasCommonLinkage()) { 116 // Hidden $non_lazy_ptr reference. 117 return X86II::MO_DARWIN_HIDDEN_NONLAZY_PIC_BASE; 118 } 119 120 // Otherwise, no stub. 121 return X86II::MO_PIC_BASE_OFFSET; 122 } 123 124 if (isPICStyleStubNoDynamic()) { // Darwin/32 in -mdynamic-no-pic mode. 125 // Determine whether we have a stub reference. 126 127 // If this is a strong reference to a definition, it is definitely not 128 // through a stub. 129 if (!isDecl && !GV->isWeakForLinker()) 130 return X86II::MO_NO_FLAG; 131 132 // Unless we have a symbol with hidden visibility, we have to go through a 133 // normal $non_lazy_ptr stub because this symbol might be resolved late. 134 if (!GV->hasHiddenVisibility()) // Non-hidden $non_lazy_ptr reference. 135 return X86II::MO_DARWIN_NONLAZY; 136 137 // Otherwise, no stub. 138 return X86II::MO_NO_FLAG; 139 } 140 141 // Direct static reference to global. 142 return X86II::MO_NO_FLAG; 143 } 144 145 146 /// getBZeroEntry - This function returns the name of a function which has an 147 /// interface like the non-standard bzero function, if such a function exists on 148 /// the current subtarget and it is considered prefereable over memset with zero 149 /// passed as the second argument. Otherwise it returns null. 150 const char *X86Subtarget::getBZeroEntry() const { 151 // Darwin 10 has a __bzero entry point for this purpose. 152 if (getTargetTriple().isMacOSX() && 153 !getTargetTriple().isMacOSXVersionLT(10, 6)) 154 return "__bzero"; 155 156 return 0; 157 } 158 159 bool X86Subtarget::hasSinCos() const { 160 return getTargetTriple().isMacOSX() && 161 !getTargetTriple().isMacOSXVersionLT(10, 9) && 162 is64Bit(); 163 } 164 165 /// IsLegalToCallImmediateAddr - Return true if the subtarget allows calls 166 /// to immediate address. 167 bool X86Subtarget::IsLegalToCallImmediateAddr(const TargetMachine &TM) const { 168 if (In64BitMode) 169 return false; 170 return isTargetELF() || TM.getRelocationModel() == Reloc::Static; 171 } 172 173 static bool OSHasAVXSupport() { 174 #if defined(i386) || defined(__i386__) || defined(__x86__) || defined(_M_IX86)\ 175 || defined(__x86_64__) || defined(_M_AMD64) || defined (_M_X64) 176 #if defined(__GNUC__) 177 // Check xgetbv; this uses a .byte sequence instead of the instruction 178 // directly because older assemblers do not include support for xgetbv and 179 // there is no easy way to conditionally compile based on the assembler used. 180 int rEAX, rEDX; 181 __asm__ (".byte 0x0f, 0x01, 0xd0" : "=a" (rEAX), "=d" (rEDX) : "c" (0)); 182 #elif defined(_MSC_FULL_VER) && defined(_XCR_XFEATURE_ENABLED_MASK) 183 unsigned long long rEAX = _xgetbv(_XCR_XFEATURE_ENABLED_MASK); 184 #else 185 int rEAX = 0; // Ensures we return false 186 #endif 187 return (rEAX & 6) == 6; 188 #else 189 return false; 190 #endif 191 } 192 193 void X86Subtarget::AutoDetectSubtargetFeatures() { 194 unsigned EAX = 0, EBX = 0, ECX = 0, EDX = 0; 195 unsigned MaxLevel; 196 union { 197 unsigned u[3]; 198 char c[12]; 199 } text; 200 201 if (X86_MC::GetCpuIDAndInfo(0, &MaxLevel, text.u+0, text.u+2, text.u+1) || 202 MaxLevel < 1) 203 return; 204 205 X86_MC::GetCpuIDAndInfo(0x1, &EAX, &EBX, &ECX, &EDX); 206 207 if ((EDX >> 15) & 1) { HasCMov = true; ToggleFeature(X86::FeatureCMOV); } 208 if ((EDX >> 23) & 1) { X86SSELevel = MMX; ToggleFeature(X86::FeatureMMX); } 209 if ((EDX >> 25) & 1) { X86SSELevel = SSE1; ToggleFeature(X86::FeatureSSE1); } 210 if ((EDX >> 26) & 1) { X86SSELevel = SSE2; ToggleFeature(X86::FeatureSSE2); } 211 if (ECX & 0x1) { X86SSELevel = SSE3; ToggleFeature(X86::FeatureSSE3); } 212 if ((ECX >> 9) & 1) { X86SSELevel = SSSE3; ToggleFeature(X86::FeatureSSSE3);} 213 if ((ECX >> 19) & 1) { X86SSELevel = SSE41; ToggleFeature(X86::FeatureSSE41);} 214 if ((ECX >> 20) & 1) { X86SSELevel = SSE42; ToggleFeature(X86::FeatureSSE42);} 215 if (((ECX >> 27) & 1) && ((ECX >> 28) & 1) && OSHasAVXSupport()) { 216 X86SSELevel = AVX; ToggleFeature(X86::FeatureAVX); 217 } 218 219 bool IsIntel = memcmp(text.c, "GenuineIntel", 12) == 0; 220 bool IsAMD = !IsIntel && memcmp(text.c, "AuthenticAMD", 12) == 0; 221 222 if ((ECX >> 1) & 0x1) { 223 HasPCLMUL = true; 224 ToggleFeature(X86::FeaturePCLMUL); 225 } 226 if ((ECX >> 12) & 0x1) { 227 HasFMA = true; 228 ToggleFeature(X86::FeatureFMA); 229 } 230 if (IsIntel && ((ECX >> 22) & 0x1)) { 231 HasMOVBE = true; 232 ToggleFeature(X86::FeatureMOVBE); 233 } 234 if ((ECX >> 23) & 0x1) { 235 HasPOPCNT = true; 236 ToggleFeature(X86::FeaturePOPCNT); 237 } 238 if ((ECX >> 25) & 0x1) { 239 HasAES = true; 240 ToggleFeature(X86::FeatureAES); 241 } 242 if ((ECX >> 29) & 0x1) { 243 HasF16C = true; 244 ToggleFeature(X86::FeatureF16C); 245 } 246 if (IsIntel && ((ECX >> 30) & 0x1)) { 247 HasRDRAND = true; 248 ToggleFeature(X86::FeatureRDRAND); 249 } 250 251 if ((ECX >> 13) & 0x1) { 252 HasCmpxchg16b = true; 253 ToggleFeature(X86::FeatureCMPXCHG16B); 254 } 255 256 if (IsIntel || IsAMD) { 257 // Determine if bit test memory instructions are slow. 258 unsigned Family = 0; 259 unsigned Model = 0; 260 X86_MC::DetectFamilyModel(EAX, Family, Model); 261 if (IsAMD || (Family == 6 && Model >= 13)) { 262 IsBTMemSlow = true; 263 ToggleFeature(X86::FeatureSlowBTMem); 264 } 265 266 // Determine if SHLD/SHRD instructions have higher latency then the 267 // equivalent series of shifts/or instructions. 268 // FIXME: Add Intel's processors that have SHLD instructions with very 269 // poor latency. 270 if (IsAMD) { 271 IsSHLDSlow = true; 272 ToggleFeature(X86::FeatureSlowSHLD); 273 } 274 275 // If it's an Intel chip since Nehalem and not an Atom chip, unaligned 276 // memory access is fast. We hard code model numbers here because they 277 // aren't strictly increasing for Intel chips it seems. 278 if (IsIntel && 279 ((Family == 6 && Model == 0x1E) || // Nehalem: Clarksfield, Lynnfield, 280 // Jasper Froest 281 (Family == 6 && Model == 0x1A) || // Nehalem: Bloomfield, Nehalem-EP 282 (Family == 6 && Model == 0x2E) || // Nehalem: Nehalem-EX 283 (Family == 6 && Model == 0x25) || // Westmere: Arrandale, Clarksdale 284 (Family == 6 && Model == 0x2C) || // Westmere: Gulftown, Westmere-EP 285 (Family == 6 && Model == 0x2F) || // Westmere: Westmere-EX 286 (Family == 6 && Model == 0x2A) || // SandyBridge 287 (Family == 6 && Model == 0x2D) || // SandyBridge: SandyBridge-E* 288 (Family == 6 && Model == 0x3A) || // IvyBridge 289 (Family == 6 && Model == 0x3E) || // IvyBridge EP 290 (Family == 6 && Model == 0x3C) || // Haswell 291 (Family == 6 && Model == 0x3F) || // ... 292 (Family == 6 && Model == 0x45) || // ... 293 (Family == 6 && Model == 0x46))) { // ... 294 IsUAMemFast = true; 295 ToggleFeature(X86::FeatureFastUAMem); 296 } 297 298 // Set processor type. Currently only Atom or Silvermont (SLM) is detected. 299 if (Family == 6 && 300 (Model == 28 || Model == 38 || Model == 39 || 301 Model == 53 || Model == 54)) { 302 X86ProcFamily = IntelAtom; 303 304 UseLeaForSP = true; 305 ToggleFeature(X86::FeatureLeaForSP); 306 } 307 else if (Family == 6 && 308 (Model == 55 || Model == 74 || Model == 77)) { 309 X86ProcFamily = IntelSLM; 310 } 311 312 unsigned MaxExtLevel; 313 X86_MC::GetCpuIDAndInfo(0x80000000, &MaxExtLevel, &EBX, &ECX, &EDX); 314 315 if (MaxExtLevel >= 0x80000001) { 316 X86_MC::GetCpuIDAndInfo(0x80000001, &EAX, &EBX, &ECX, &EDX); 317 if ((EDX >> 29) & 0x1) { 318 HasX86_64 = true; 319 ToggleFeature(X86::Feature64Bit); 320 } 321 if ((ECX >> 5) & 0x1) { 322 HasLZCNT = true; 323 ToggleFeature(X86::FeatureLZCNT); 324 } 325 if (IsIntel && ((ECX >> 8) & 0x1)) { 326 HasPRFCHW = true; 327 ToggleFeature(X86::FeaturePRFCHW); 328 } 329 if (IsAMD) { 330 if ((ECX >> 6) & 0x1) { 331 HasSSE4A = true; 332 ToggleFeature(X86::FeatureSSE4A); 333 } 334 if ((ECX >> 11) & 0x1) { 335 HasXOP = true; 336 ToggleFeature(X86::FeatureXOP); 337 } 338 if ((ECX >> 16) & 0x1) { 339 HasFMA4 = true; 340 ToggleFeature(X86::FeatureFMA4); 341 } 342 } 343 } 344 } 345 346 if (MaxLevel >= 7) { 347 if (!X86_MC::GetCpuIDAndInfoEx(0x7, 0x0, &EAX, &EBX, &ECX, &EDX)) { 348 if (IsIntel && (EBX & 0x1)) { 349 HasFSGSBase = true; 350 ToggleFeature(X86::FeatureFSGSBase); 351 } 352 if ((EBX >> 3) & 0x1) { 353 HasBMI = true; 354 ToggleFeature(X86::FeatureBMI); 355 } 356 if ((EBX >> 4) & 0x1) { 357 HasHLE = true; 358 ToggleFeature(X86::FeatureHLE); 359 } 360 if (IsIntel && ((EBX >> 5) & 0x1)) { 361 X86SSELevel = AVX2; 362 ToggleFeature(X86::FeatureAVX2); 363 } 364 if (IsIntel && ((EBX >> 8) & 0x1)) { 365 HasBMI2 = true; 366 ToggleFeature(X86::FeatureBMI2); 367 } 368 if (IsIntel && ((EBX >> 11) & 0x1)) { 369 HasRTM = true; 370 ToggleFeature(X86::FeatureRTM); 371 } 372 if (IsIntel && ((EBX >> 16) & 0x1)) { 373 X86SSELevel = AVX512F; 374 ToggleFeature(X86::FeatureAVX512); 375 } 376 if (IsIntel && ((EBX >> 18) & 0x1)) { 377 HasRDSEED = true; 378 ToggleFeature(X86::FeatureRDSEED); 379 } 380 if (IsIntel && ((EBX >> 19) & 0x1)) { 381 HasADX = true; 382 ToggleFeature(X86::FeatureADX); 383 } 384 if (IsIntel && ((EBX >> 26) & 0x1)) { 385 HasPFI = true; 386 ToggleFeature(X86::FeaturePFI); 387 } 388 if (IsIntel && ((EBX >> 27) & 0x1)) { 389 HasERI = true; 390 ToggleFeature(X86::FeatureERI); 391 } 392 if (IsIntel && ((EBX >> 28) & 0x1)) { 393 HasCDI = true; 394 ToggleFeature(X86::FeatureCDI); 395 } 396 if (IsIntel && ((EBX >> 29) & 0x1)) { 397 HasSHA = true; 398 ToggleFeature(X86::FeatureSHA); 399 } 400 } 401 if (IsAMD && ((ECX >> 21) & 0x1)) { 402 HasTBM = true; 403 ToggleFeature(X86::FeatureTBM); 404 } 405 } 406 } 407 408 void X86Subtarget::resetSubtargetFeatures(const MachineFunction *MF) { 409 AttributeSet FnAttrs = MF->getFunction()->getAttributes(); 410 Attribute CPUAttr = FnAttrs.getAttribute(AttributeSet::FunctionIndex, 411 "target-cpu"); 412 Attribute FSAttr = FnAttrs.getAttribute(AttributeSet::FunctionIndex, 413 "target-features"); 414 std::string CPU = 415 !CPUAttr.hasAttribute(Attribute::None) ?CPUAttr.getValueAsString() : ""; 416 std::string FS = 417 !FSAttr.hasAttribute(Attribute::None) ? FSAttr.getValueAsString() : ""; 418 if (!FS.empty()) { 419 initializeEnvironment(); 420 resetSubtargetFeatures(CPU, FS); 421 } 422 } 423 424 void X86Subtarget::resetSubtargetFeatures(StringRef CPU, StringRef FS) { 425 std::string CPUName = CPU; 426 if (!FS.empty() || !CPU.empty()) { 427 if (CPUName.empty()) { 428 #if defined(i386) || defined(__i386__) || defined(__x86__) || defined(_M_IX86)\ 429 || defined(__x86_64__) || defined(_M_AMD64) || defined (_M_X64) 430 CPUName = sys::getHostCPUName(); 431 #else 432 CPUName = "generic"; 433 #endif 434 } 435 436 // Make sure 64-bit features are available in 64-bit mode. (But make sure 437 // SSE2 can be turned off explicitly.) 438 std::string FullFS = FS; 439 if (In64BitMode) { 440 if (!FullFS.empty()) 441 FullFS = "+64bit,+sse2," + FullFS; 442 else 443 FullFS = "+64bit,+sse2"; 444 } 445 446 // If feature string is not empty, parse features string. 447 ParseSubtargetFeatures(CPUName, FullFS); 448 } else { 449 if (CPUName.empty()) { 450 #if defined (__x86_64__) || defined(__i386__) 451 CPUName = sys::getHostCPUName(); 452 #else 453 CPUName = "generic"; 454 #endif 455 } 456 // Otherwise, use CPUID to auto-detect feature set. 457 AutoDetectSubtargetFeatures(); 458 459 // Make sure 64-bit features are available in 64-bit mode. 460 if (In64BitMode) { 461 if (!HasX86_64) { HasX86_64 = true; ToggleFeature(X86::Feature64Bit); } 462 if (!HasCMov) { HasCMov = true; ToggleFeature(X86::FeatureCMOV); } 463 464 if (X86SSELevel < SSE2) { 465 X86SSELevel = SSE2; 466 ToggleFeature(X86::FeatureSSE1); 467 ToggleFeature(X86::FeatureSSE2); 468 } 469 } 470 } 471 472 // CPUName may have been set by the CPU detection code. Make sure the 473 // new MCSchedModel is used. 474 InitCPUSchedModel(CPUName); 475 476 if (X86ProcFamily == IntelAtom || X86ProcFamily == IntelSLM) 477 PostRAScheduler = true; 478 479 InstrItins = getInstrItineraryForCPU(CPUName); 480 481 // It's important to keep the MCSubtargetInfo feature bits in sync with 482 // target data structure which is shared with MC code emitter, etc. 483 if (In64BitMode) 484 ToggleFeature(X86::Mode64Bit); 485 else if (In32BitMode) 486 ToggleFeature(X86::Mode32Bit); 487 else if (In16BitMode) 488 ToggleFeature(X86::Mode16Bit); 489 else 490 llvm_unreachable("Not 16-bit, 32-bit or 64-bit mode!"); 491 492 DEBUG(dbgs() << "Subtarget features: SSELevel " << X86SSELevel 493 << ", 3DNowLevel " << X863DNowLevel 494 << ", 64bit " << HasX86_64 << "\n"); 495 assert((!In64BitMode || HasX86_64) && 496 "64-bit code requested on a subtarget that doesn't support it!"); 497 498 // Stack alignment is 16 bytes on Darwin, Linux and Solaris (both 499 // 32 and 64 bit) and for all 64-bit targets. 500 if (StackAlignOverride) 501 stackAlignment = StackAlignOverride; 502 else if (isTargetDarwin() || isTargetLinux() || isTargetSolaris() || 503 In64BitMode) 504 stackAlignment = 16; 505 } 506 507 void X86Subtarget::initializeEnvironment() { 508 X86SSELevel = NoMMXSSE; 509 X863DNowLevel = NoThreeDNow; 510 HasCMov = false; 511 HasX86_64 = false; 512 HasPOPCNT = false; 513 HasSSE4A = false; 514 HasAES = false; 515 HasPCLMUL = false; 516 HasFMA = false; 517 HasFMA4 = false; 518 HasXOP = false; 519 HasTBM = false; 520 HasMOVBE = false; 521 HasRDRAND = false; 522 HasF16C = false; 523 HasFSGSBase = false; 524 HasLZCNT = false; 525 HasBMI = false; 526 HasBMI2 = false; 527 HasRTM = false; 528 HasHLE = false; 529 HasERI = false; 530 HasCDI = false; 531 HasPFI = false; 532 HasADX = false; 533 HasSHA = false; 534 HasPRFCHW = false; 535 HasRDSEED = false; 536 IsBTMemSlow = false; 537 IsSHLDSlow = false; 538 IsUAMemFast = false; 539 HasVectorUAMem = false; 540 HasCmpxchg16b = false; 541 UseLeaForSP = false; 542 HasSlowDivide = false; 543 PostRAScheduler = false; 544 PadShortFunctions = false; 545 CallRegIndirect = false; 546 LEAUsesAG = false; 547 stackAlignment = 4; 548 // FIXME: this is a known good value for Yonah. How about others? 549 MaxInlineSizeThreshold = 128; 550 } 551 552 X86Subtarget::X86Subtarget(const std::string &TT, const std::string &CPU, 553 const std::string &FS, 554 unsigned StackAlignOverride) 555 : X86GenSubtargetInfo(TT, CPU, FS) 556 , X86ProcFamily(Others) 557 , PICStyle(PICStyles::None) 558 , TargetTriple(TT) 559 , StackAlignOverride(StackAlignOverride) 560 , In64BitMode(TargetTriple.getArch() == Triple::x86_64) 561 , In32BitMode(TargetTriple.getArch() == Triple::x86 && 562 TargetTriple.getEnvironment() != Triple::CODE16) 563 , In16BitMode(TargetTriple.getArch() == Triple::x86 && 564 TargetTriple.getEnvironment() == Triple::CODE16) { 565 initializeEnvironment(); 566 resetSubtargetFeatures(CPU, FS); 567 } 568 569 bool X86Subtarget::enablePostRAScheduler( 570 CodeGenOpt::Level OptLevel, 571 TargetSubtargetInfo::AntiDepBreakMode& Mode, 572 RegClassVector& CriticalPathRCs) const { 573 Mode = TargetSubtargetInfo::ANTIDEP_CRITICAL; 574 CriticalPathRCs.clear(); 575 return PostRAScheduler && OptLevel >= CodeGenOpt::Default; 576 } 577