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 #include "X86.h" 15 16 #include "X86CallLowering.h" 17 #include "X86LegalizerInfo.h" 18 #include "X86RegisterBankInfo.h" 19 #include "X86Subtarget.h" 20 #include "MCTargetDesc/X86BaseInfo.h" 21 #include "X86TargetMachine.h" 22 #include "llvm/ADT/Triple.h" 23 #include "llvm/CodeGen/GlobalISel/CallLowering.h" 24 #include "llvm/CodeGen/GlobalISel/InstructionSelect.h" 25 #include "llvm/IR/Attributes.h" 26 #include "llvm/IR/ConstantRange.h" 27 #include "llvm/IR/Function.h" 28 #include "llvm/IR/GlobalValue.h" 29 #include "llvm/Support/Casting.h" 30 #include "llvm/Support/CodeGen.h" 31 #include "llvm/Support/CommandLine.h" 32 #include "llvm/Support/Debug.h" 33 #include "llvm/Support/ErrorHandling.h" 34 #include "llvm/Support/raw_ostream.h" 35 #include "llvm/Target/TargetMachine.h" 36 37 #if defined(_MSC_VER) 38 #include <intrin.h> 39 #endif 40 41 using namespace llvm; 42 43 #define DEBUG_TYPE "subtarget" 44 45 #define GET_SUBTARGETINFO_TARGET_DESC 46 #define GET_SUBTARGETINFO_CTOR 47 #include "X86GenSubtargetInfo.inc" 48 49 // Temporary option to control early if-conversion for x86 while adding machine 50 // models. 51 static cl::opt<bool> 52 X86EarlyIfConv("x86-early-ifcvt", cl::Hidden, 53 cl::desc("Enable early if-conversion on X86")); 54 55 56 /// Classify a blockaddress reference for the current subtarget according to how 57 /// we should reference it in a non-pcrel context. 58 unsigned char X86Subtarget::classifyBlockAddressReference() const { 59 return classifyLocalReference(nullptr); 60 } 61 62 /// Classify a global variable reference for the current subtarget according to 63 /// how we should reference it in a non-pcrel context. 64 unsigned char 65 X86Subtarget::classifyGlobalReference(const GlobalValue *GV) const { 66 return classifyGlobalReference(GV, *GV->getParent()); 67 } 68 69 unsigned char 70 X86Subtarget::classifyLocalReference(const GlobalValue *GV) const { 71 // 64 bits can use %rip addressing for anything local. 72 if (is64Bit()) 73 return X86II::MO_NO_FLAG; 74 75 // If this is for a position dependent executable, the static linker can 76 // figure it out. 77 if (!isPositionIndependent()) 78 return X86II::MO_NO_FLAG; 79 80 // The COFF dynamic linker just patches the executable sections. 81 if (isTargetCOFF()) 82 return X86II::MO_NO_FLAG; 83 84 if (isTargetDarwin()) { 85 // 32 bit macho has no relocation for a-b if a is undefined, even if 86 // b is in the section that is being relocated. 87 // This means we have to use o load even for GVs that are known to be 88 // local to the dso. 89 if (GV && (GV->isDeclarationForLinker() || GV->hasCommonLinkage())) 90 return X86II::MO_DARWIN_NONLAZY_PIC_BASE; 91 92 return X86II::MO_PIC_BASE_OFFSET; 93 } 94 95 return X86II::MO_GOTOFF; 96 } 97 98 unsigned char X86Subtarget::classifyGlobalReference(const GlobalValue *GV, 99 const Module &M) const { 100 // Large model never uses stubs. 101 if (TM.getCodeModel() == CodeModel::Large) 102 return X86II::MO_NO_FLAG; 103 104 // Absolute symbols can be referenced directly. 105 if (GV) { 106 if (Optional<ConstantRange> CR = GV->getAbsoluteSymbolRange()) { 107 // See if we can use the 8-bit immediate form. Note that some instructions 108 // will sign extend the immediate operand, so to be conservative we only 109 // accept the range [0,128). 110 if (CR->getUnsignedMax().ult(128)) 111 return X86II::MO_ABS8; 112 else 113 return X86II::MO_NO_FLAG; 114 } 115 } 116 117 if (TM.shouldAssumeDSOLocal(M, GV)) 118 return classifyLocalReference(GV); 119 120 if (isTargetCOFF()) 121 return X86II::MO_DLLIMPORT; 122 123 if (is64Bit()) 124 return X86II::MO_GOTPCREL; 125 126 if (isTargetDarwin()) { 127 if (!isPositionIndependent()) 128 return X86II::MO_DARWIN_NONLAZY; 129 return X86II::MO_DARWIN_NONLAZY_PIC_BASE; 130 } 131 132 return X86II::MO_GOT; 133 } 134 135 unsigned char 136 X86Subtarget::classifyGlobalFunctionReference(const GlobalValue *GV) const { 137 return classifyGlobalFunctionReference(GV, *GV->getParent()); 138 } 139 140 unsigned char 141 X86Subtarget::classifyGlobalFunctionReference(const GlobalValue *GV, 142 const Module &M) const { 143 if (TM.shouldAssumeDSOLocal(M, GV)) 144 return X86II::MO_NO_FLAG; 145 146 if (isTargetCOFF()) { 147 assert(GV->hasDLLImportStorageClass() && 148 "shouldAssumeDSOLocal gave inconsistent answer"); 149 return X86II::MO_DLLIMPORT; 150 } 151 152 const Function *F = dyn_cast_or_null<Function>(GV); 153 154 if (isTargetELF()) { 155 if (is64Bit() && F && (CallingConv::X86_RegCall == F->getCallingConv())) 156 // According to psABI, PLT stub clobbers XMM8-XMM15. 157 // In Regcall calling convention those registers are used for passing 158 // parameters. Thus we need to prevent lazy binding in Regcall. 159 return X86II::MO_GOTPCREL; 160 // If PLT must be avoided then the call should be via GOTPCREL. 161 if (((F && F->hasFnAttribute(Attribute::NonLazyBind)) || 162 (!F && M.getRtLibUseGOT())) && 163 is64Bit()) 164 return X86II::MO_GOTPCREL; 165 return X86II::MO_PLT; 166 } 167 168 if (is64Bit()) { 169 if (F && F->hasFnAttribute(Attribute::NonLazyBind)) 170 // If the function is marked as non-lazy, generate an indirect call 171 // which loads from the GOT directly. This avoids runtime overhead 172 // at the cost of eager binding (and one extra byte of encoding). 173 return X86II::MO_GOTPCREL; 174 return X86II::MO_NO_FLAG; 175 } 176 177 return X86II::MO_NO_FLAG; 178 } 179 180 /// Return true if the subtarget allows calls to immediate address. 181 bool X86Subtarget::isLegalToCallImmediateAddr() const { 182 // FIXME: I386 PE/COFF supports PC relative calls using IMAGE_REL_I386_REL32 183 // but WinCOFFObjectWriter::RecordRelocation cannot emit them. Once it does, 184 // the following check for Win32 should be removed. 185 if (In64BitMode || isTargetWin32()) 186 return false; 187 return isTargetELF() || TM.getRelocationModel() == Reloc::Static; 188 } 189 190 void X86Subtarget::initSubtargetFeatures(StringRef CPU, StringRef FS) { 191 std::string CPUName = CPU; 192 if (CPUName.empty()) 193 CPUName = "generic"; 194 195 // Make sure 64-bit features are available in 64-bit mode. (But make sure 196 // SSE2 can be turned off explicitly.) 197 std::string FullFS = FS; 198 if (In64BitMode) { 199 if (!FullFS.empty()) 200 FullFS = "+64bit,+sse2," + FullFS; 201 else 202 FullFS = "+64bit,+sse2"; 203 } 204 205 // LAHF/SAHF are always supported in non-64-bit mode. 206 if (!In64BitMode) { 207 if (!FullFS.empty()) 208 FullFS = "+sahf," + FullFS; 209 else 210 FullFS = "+sahf"; 211 } 212 213 // Parse features string and set the CPU. 214 ParseSubtargetFeatures(CPUName, FullFS); 215 216 // All CPUs that implement SSE4.2 or SSE4A support unaligned accesses of 217 // 16-bytes and under that are reasonably fast. These features were 218 // introduced with Intel's Nehalem/Silvermont and AMD's Family10h 219 // micro-architectures respectively. 220 if (hasSSE42() || hasSSE4A()) 221 IsUAMem16Slow = false; 222 223 InstrItins = getInstrItineraryForCPU(CPUName); 224 225 // It's important to keep the MCSubtargetInfo feature bits in sync with 226 // target data structure which is shared with MC code emitter, etc. 227 if (In64BitMode) 228 ToggleFeature(X86::Mode64Bit); 229 else if (In32BitMode) 230 ToggleFeature(X86::Mode32Bit); 231 else if (In16BitMode) 232 ToggleFeature(X86::Mode16Bit); 233 else 234 llvm_unreachable("Not 16-bit, 32-bit or 64-bit mode!"); 235 236 DEBUG(dbgs() << "Subtarget features: SSELevel " << X86SSELevel 237 << ", 3DNowLevel " << X863DNowLevel 238 << ", 64bit " << HasX86_64 << "\n"); 239 assert((!In64BitMode || HasX86_64) && 240 "64-bit code requested on a subtarget that doesn't support it!"); 241 242 // Stack alignment is 16 bytes on Darwin, Linux, kFreeBSD and Solaris (both 243 // 32 and 64 bit) and for all 64-bit targets. 244 if (StackAlignOverride) 245 stackAlignment = StackAlignOverride; 246 else if (isTargetDarwin() || isTargetLinux() || isTargetSolaris() || 247 isTargetKFreeBSD() || In64BitMode) 248 stackAlignment = 16; 249 250 // Some CPUs have more overhead for gather. The specified overhead is relative 251 // to the Load operation. "2" is the number provided by Intel architects. This 252 // parameter is used for cost estimation of Gather Op and comparison with 253 // other alternatives. 254 // TODO: Remove the explicit hasAVX512()?, That would mean we would only 255 // enable gather with a -march. 256 if (hasAVX512() || (hasAVX2() && hasFastGather())) 257 GatherOverhead = 2; 258 if (hasAVX512()) 259 ScatterOverhead = 2; 260 261 // Consume the vector width attribute or apply any target specific limit. 262 if (PreferVectorWidthOverride) 263 PreferVectorWidth = PreferVectorWidthOverride; 264 else if (Prefer256Bit) 265 PreferVectorWidth = 256; 266 } 267 268 void X86Subtarget::initializeEnvironment() { 269 X86SSELevel = NoSSE; 270 X863DNowLevel = NoThreeDNow; 271 HasX87 = false; 272 HasNOPL = false; 273 HasCMov = false; 274 HasX86_64 = false; 275 HasPOPCNT = false; 276 HasSSE4A = false; 277 HasAES = false; 278 HasVAES = false; 279 HasFXSR = false; 280 HasXSAVE = false; 281 HasXSAVEOPT = false; 282 HasXSAVEC = false; 283 HasXSAVES = false; 284 HasPCLMUL = false; 285 HasVPCLMULQDQ = false; 286 HasGFNI = false; 287 HasFMA = false; 288 HasFMA4 = false; 289 HasXOP = false; 290 HasTBM = false; 291 HasLWP = false; 292 HasMOVBE = false; 293 HasRDRAND = false; 294 HasF16C = false; 295 HasFSGSBase = false; 296 HasLZCNT = false; 297 HasBMI = false; 298 HasBMI2 = false; 299 HasVBMI = false; 300 HasVBMI2 = false; 301 HasIFMA = false; 302 HasRTM = false; 303 HasERI = false; 304 HasCDI = false; 305 HasPFI = false; 306 HasDQI = false; 307 HasVPOPCNTDQ = false; 308 HasBWI = false; 309 HasVLX = false; 310 HasADX = false; 311 HasPKU = false; 312 HasVNNI = false; 313 HasBITALG = false; 314 HasSHA = false; 315 HasPREFETCHWT1 = false; 316 HasPRFCHW = false; 317 HasRDSEED = false; 318 HasLAHFSAHF = false; 319 HasMWAITX = false; 320 HasCLZERO = false; 321 HasMPX = false; 322 HasSHSTK = false; 323 HasIBT = false; 324 HasSGX = false; 325 HasCLFLUSHOPT = false; 326 HasCLWB = false; 327 HasRDPID = false; 328 UseRetpoline = false; 329 UseRetpolineExternalThunk = false; 330 IsPMULLDSlow = false; 331 IsSHLDSlow = false; 332 IsUAMem16Slow = false; 333 IsUAMem32Slow = false; 334 HasSSEUnalignedMem = false; 335 HasCmpxchg16b = false; 336 UseLeaForSP = false; 337 HasPOPCNTFalseDeps = false; 338 HasLZCNTFalseDeps = false; 339 HasFastVariableShuffle = false; 340 HasFastPartialYMMorZMMWrite = false; 341 HasFast11ByteNOP = false; 342 HasFast15ByteNOP = false; 343 HasFastGather = false; 344 HasFastScalarFSQRT = false; 345 HasFastVectorFSQRT = false; 346 HasFastLZCNT = false; 347 HasFastSHLDRotate = false; 348 HasMacroFusion = false; 349 HasERMSB = false; 350 HasSlowDivide32 = false; 351 HasSlowDivide64 = false; 352 PadShortFunctions = false; 353 SlowTwoMemOps = false; 354 LEAUsesAG = false; 355 SlowLEA = false; 356 Slow3OpsLEA = false; 357 SlowIncDec = false; 358 stackAlignment = 4; 359 // FIXME: this is a known good value for Yonah. How about others? 360 MaxInlineSizeThreshold = 128; 361 UseSoftFloat = false; 362 X86ProcFamily = Others; 363 GatherOverhead = 1024; 364 ScatterOverhead = 1024; 365 PreferVectorWidth = UINT32_MAX; 366 Prefer256Bit = false; 367 } 368 369 X86Subtarget &X86Subtarget::initializeSubtargetDependencies(StringRef CPU, 370 StringRef FS) { 371 initializeEnvironment(); 372 initSubtargetFeatures(CPU, FS); 373 return *this; 374 } 375 376 X86Subtarget::X86Subtarget(const Triple &TT, StringRef CPU, StringRef FS, 377 const X86TargetMachine &TM, 378 unsigned StackAlignOverride, 379 unsigned PreferVectorWidthOverride, 380 unsigned RequiredVectorWidth) 381 : X86GenSubtargetInfo(TT, CPU, FS), X86ProcFamily(Others), 382 PICStyle(PICStyles::None), TM(TM), TargetTriple(TT), 383 StackAlignOverride(StackAlignOverride), 384 PreferVectorWidthOverride(PreferVectorWidthOverride), 385 RequiredVectorWidth(RequiredVectorWidth), 386 In64BitMode(TargetTriple.getArch() == Triple::x86_64), 387 In32BitMode(TargetTriple.getArch() == Triple::x86 && 388 TargetTriple.getEnvironment() != Triple::CODE16), 389 In16BitMode(TargetTriple.getArch() == Triple::x86 && 390 TargetTriple.getEnvironment() == Triple::CODE16), 391 InstrInfo(initializeSubtargetDependencies(CPU, FS)), TLInfo(TM, *this), 392 FrameLowering(*this, getStackAlignment()) { 393 // Determine the PICStyle based on the target selected. 394 if (!isPositionIndependent()) 395 setPICStyle(PICStyles::None); 396 else if (is64Bit()) 397 setPICStyle(PICStyles::RIPRel); 398 else if (isTargetCOFF()) 399 setPICStyle(PICStyles::None); 400 else if (isTargetDarwin()) 401 setPICStyle(PICStyles::StubPIC); 402 else if (isTargetELF()) 403 setPICStyle(PICStyles::GOT); 404 405 CallLoweringInfo.reset(new X86CallLowering(*getTargetLowering())); 406 Legalizer.reset(new X86LegalizerInfo(*this, TM)); 407 408 auto *RBI = new X86RegisterBankInfo(*getRegisterInfo()); 409 RegBankInfo.reset(RBI); 410 InstSelector.reset(createX86InstructionSelector(TM, *this, *RBI)); 411 } 412 413 const CallLowering *X86Subtarget::getCallLowering() const { 414 return CallLoweringInfo.get(); 415 } 416 417 const InstructionSelector *X86Subtarget::getInstructionSelector() const { 418 return InstSelector.get(); 419 } 420 421 const LegalizerInfo *X86Subtarget::getLegalizerInfo() const { 422 return Legalizer.get(); 423 } 424 425 const RegisterBankInfo *X86Subtarget::getRegBankInfo() const { 426 return RegBankInfo.get(); 427 } 428 429 bool X86Subtarget::enableEarlyIfConversion() const { 430 return hasCMov() && X86EarlyIfConv; 431 } 432