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/CodeGen/GlobalISel/Legalizer.h" 26 #include "llvm/CodeGen/GlobalISel/RegBankSelect.h" 27 #include "llvm/IR/Attributes.h" 28 #include "llvm/IR/ConstantRange.h" 29 #include "llvm/IR/Function.h" 30 #include "llvm/IR/GlobalValue.h" 31 #include "llvm/Support/Casting.h" 32 #include "llvm/Support/CodeGen.h" 33 #include "llvm/Support/CommandLine.h" 34 #include "llvm/Support/Debug.h" 35 #include "llvm/Support/ErrorHandling.h" 36 #include "llvm/Support/raw_ostream.h" 37 #include "llvm/Target/TargetMachine.h" 38 #include <cassert> 39 #include <string> 40 41 #if defined(_MSC_VER) 42 #include <intrin.h> 43 #endif 44 45 using namespace llvm; 46 47 #define DEBUG_TYPE "subtarget" 48 49 #define GET_SUBTARGETINFO_TARGET_DESC 50 #define GET_SUBTARGETINFO_CTOR 51 #include "X86GenSubtargetInfo.inc" 52 53 // Temporary option to control early if-conversion for x86 while adding machine 54 // models. 55 static cl::opt<bool> 56 X86EarlyIfConv("x86-early-ifcvt", cl::Hidden, 57 cl::desc("Enable early if-conversion on X86")); 58 59 60 /// Classify a blockaddress reference for the current subtarget according to how 61 /// we should reference it in a non-pcrel context. 62 unsigned char X86Subtarget::classifyBlockAddressReference() const { 63 return classifyLocalReference(nullptr); 64 } 65 66 /// Classify a global variable reference for the current subtarget according to 67 /// how we should reference it in a non-pcrel context. 68 unsigned char 69 X86Subtarget::classifyGlobalReference(const GlobalValue *GV) const { 70 return classifyGlobalReference(GV, *GV->getParent()); 71 } 72 73 unsigned char 74 X86Subtarget::classifyLocalReference(const GlobalValue *GV) const { 75 // 64 bits can use %rip addressing for anything local. 76 if (is64Bit()) 77 return X86II::MO_NO_FLAG; 78 79 // If this is for a position dependent executable, the static linker can 80 // figure it out. 81 if (!isPositionIndependent()) 82 return X86II::MO_NO_FLAG; 83 84 // The COFF dynamic linker just patches the executable sections. 85 if (isTargetCOFF()) 86 return X86II::MO_NO_FLAG; 87 88 if (isTargetDarwin()) { 89 // 32 bit macho has no relocation for a-b if a is undefined, even if 90 // b is in the section that is being relocated. 91 // This means we have to use o load even for GVs that are known to be 92 // local to the dso. 93 if (GV && (GV->isDeclarationForLinker() || GV->hasCommonLinkage())) 94 return X86II::MO_DARWIN_NONLAZY_PIC_BASE; 95 96 return X86II::MO_PIC_BASE_OFFSET; 97 } 98 99 return X86II::MO_GOTOFF; 100 } 101 102 unsigned char X86Subtarget::classifyGlobalReference(const GlobalValue *GV, 103 const Module &M) const { 104 // Large model never uses stubs. 105 if (TM.getCodeModel() == CodeModel::Large) 106 return X86II::MO_NO_FLAG; 107 108 // Absolute symbols can be referenced directly. 109 if (GV) { 110 if (Optional<ConstantRange> CR = GV->getAbsoluteSymbolRange()) { 111 // See if we can use the 8-bit immediate form. Note that some instructions 112 // will sign extend the immediate operand, so to be conservative we only 113 // accept the range [0,128). 114 if (CR->getUnsignedMax().ult(128)) 115 return X86II::MO_ABS8; 116 else 117 return X86II::MO_NO_FLAG; 118 } 119 } 120 121 if (TM.shouldAssumeDSOLocal(M, GV)) 122 return classifyLocalReference(GV); 123 124 if (isTargetCOFF()) 125 return X86II::MO_DLLIMPORT; 126 127 if (is64Bit()) 128 return X86II::MO_GOTPCREL; 129 130 if (isTargetDarwin()) { 131 if (!isPositionIndependent()) 132 return X86II::MO_DARWIN_NONLAZY; 133 return X86II::MO_DARWIN_NONLAZY_PIC_BASE; 134 } 135 136 return X86II::MO_GOT; 137 } 138 139 unsigned char 140 X86Subtarget::classifyGlobalFunctionReference(const GlobalValue *GV) const { 141 return classifyGlobalFunctionReference(GV, *GV->getParent()); 142 } 143 144 unsigned char 145 X86Subtarget::classifyGlobalFunctionReference(const GlobalValue *GV, 146 const Module &M) const { 147 const Function *F = dyn_cast_or_null<Function>(GV); 148 149 // Do not use the PLT when explicitly told to do so for ELF 64-bit 150 // target. 151 if (isTargetELF() && is64Bit() && F && 152 F->hasFnAttribute(Attribute::NonLazyBind) && 153 GV->isDeclarationForLinker()) 154 return X86II::MO_GOTPCREL; 155 156 if (TM.shouldAssumeDSOLocal(M, GV)) 157 return X86II::MO_NO_FLAG; 158 159 if (isTargetCOFF()) { 160 assert(GV->hasDLLImportStorageClass() && 161 "shouldAssumeDSOLocal gave inconsistent answer"); 162 return X86II::MO_DLLIMPORT; 163 } 164 165 if (isTargetELF()) { 166 if (is64Bit() && F && (CallingConv::X86_RegCall == F->getCallingConv())) 167 // According to psABI, PLT stub clobbers XMM8-XMM15. 168 // In Regcall calling convention those registers are used for passing 169 // parameters. Thus we need to prevent lazy binding in Regcall. 170 return X86II::MO_GOTPCREL; 171 return X86II::MO_PLT; 172 } 173 174 if (is64Bit()) { 175 if (F && F->hasFnAttribute(Attribute::NonLazyBind)) 176 // If the function is marked as non-lazy, generate an indirect call 177 // which loads from the GOT directly. This avoids runtime overhead 178 // at the cost of eager binding (and one extra byte of encoding). 179 return X86II::MO_GOTPCREL; 180 return X86II::MO_NO_FLAG; 181 } 182 183 return X86II::MO_NO_FLAG; 184 } 185 186 /// This function returns the name of a function which has an interface like 187 /// the non-standard bzero function, if such a function exists on the 188 /// current subtarget and it is considered preferable over memset with zero 189 /// passed as the second argument. Otherwise it returns null. 190 const char *X86Subtarget::getBZeroEntry() const { 191 // Darwin 10 has a __bzero entry point for this purpose. 192 if (getTargetTriple().isMacOSX() && 193 !getTargetTriple().isMacOSXVersionLT(10, 6)) 194 return "__bzero"; 195 196 return nullptr; 197 } 198 199 bool X86Subtarget::hasSinCos() const { 200 if (getTargetTriple().isMacOSX()) { 201 return !getTargetTriple().isMacOSXVersionLT(10, 9) && is64Bit(); 202 } else if (getTargetTriple().isOSFuchsia()) { 203 return true; 204 } 205 return false; 206 } 207 208 /// Return true if the subtarget allows calls to immediate address. 209 bool X86Subtarget::isLegalToCallImmediateAddr() const { 210 // FIXME: I386 PE/COFF supports PC relative calls using IMAGE_REL_I386_REL32 211 // but WinCOFFObjectWriter::RecordRelocation cannot emit them. Once it does, 212 // the following check for Win32 should be removed. 213 if (In64BitMode || isTargetWin32()) 214 return false; 215 return isTargetELF() || TM.getRelocationModel() == Reloc::Static; 216 } 217 218 void X86Subtarget::initSubtargetFeatures(StringRef CPU, StringRef FS) { 219 std::string CPUName = CPU; 220 if (CPUName.empty()) 221 CPUName = "generic"; 222 223 // Make sure 64-bit features are available in 64-bit mode. (But make sure 224 // SSE2 can be turned off explicitly.) 225 std::string FullFS = FS; 226 if (In64BitMode) { 227 if (!FullFS.empty()) 228 FullFS = "+64bit,+sse2," + FullFS; 229 else 230 FullFS = "+64bit,+sse2"; 231 } 232 233 // LAHF/SAHF are always supported in non-64-bit mode. 234 if (!In64BitMode) { 235 if (!FullFS.empty()) 236 FullFS = "+sahf," + FullFS; 237 else 238 FullFS = "+sahf"; 239 } 240 241 // Parse features string and set the CPU. 242 ParseSubtargetFeatures(CPUName, FullFS); 243 244 // All CPUs that implement SSE4.2 or SSE4A support unaligned accesses of 245 // 16-bytes and under that are reasonably fast. These features were 246 // introduced with Intel's Nehalem/Silvermont and AMD's Family10h 247 // micro-architectures respectively. 248 if (hasSSE42() || hasSSE4A()) 249 IsUAMem16Slow = false; 250 251 InstrItins = getInstrItineraryForCPU(CPUName); 252 253 // It's important to keep the MCSubtargetInfo feature bits in sync with 254 // target data structure which is shared with MC code emitter, etc. 255 if (In64BitMode) 256 ToggleFeature(X86::Mode64Bit); 257 else if (In32BitMode) 258 ToggleFeature(X86::Mode32Bit); 259 else if (In16BitMode) 260 ToggleFeature(X86::Mode16Bit); 261 else 262 llvm_unreachable("Not 16-bit, 32-bit or 64-bit mode!"); 263 264 DEBUG(dbgs() << "Subtarget features: SSELevel " << X86SSELevel 265 << ", 3DNowLevel " << X863DNowLevel 266 << ", 64bit " << HasX86_64 << "\n"); 267 assert((!In64BitMode || HasX86_64) && 268 "64-bit code requested on a subtarget that doesn't support it!"); 269 270 // Stack alignment is 16 bytes on Darwin, Linux, kFreeBSD and Solaris (both 271 // 32 and 64 bit) and for all 64-bit targets. 272 if (StackAlignOverride) 273 stackAlignment = StackAlignOverride; 274 else if (isTargetDarwin() || isTargetLinux() || isTargetSolaris() || 275 isTargetKFreeBSD() || In64BitMode) 276 stackAlignment = 16; 277 278 // Gather is available since Haswell (AVX2 set). So technically, we can generate Gathers 279 // on all AVX2 processors. But the overhead on HSW is high. Skylake Client processor has 280 // faster Gathers than HSW and performance is similar to Skylake Server (AVX-512). 281 // The specified overhead is relative to the Load operation."2" is the number provided 282 // by Intel architects, This parameter is used for cost estimation of Gather Op and 283 // comparison with other alternatives. 284 if (X86ProcFamily == IntelSkylake || hasAVX512()) 285 GatherOverhead = 2; 286 if (hasAVX512()) 287 ScatterOverhead = 2; 288 } 289 290 void X86Subtarget::initializeEnvironment() { 291 X86SSELevel = NoSSE; 292 X863DNowLevel = NoThreeDNow; 293 HasX87 = false; 294 HasCMov = false; 295 HasX86_64 = false; 296 HasPOPCNT = false; 297 HasSSE4A = false; 298 HasAES = false; 299 HasFXSR = false; 300 HasXSAVE = false; 301 HasXSAVEOPT = false; 302 HasXSAVEC = false; 303 HasXSAVES = false; 304 HasPCLMUL = false; 305 HasFMA = false; 306 HasFMA4 = false; 307 HasXOP = false; 308 HasTBM = false; 309 HasLWP = false; 310 HasMOVBE = false; 311 HasRDRAND = false; 312 HasF16C = false; 313 HasFSGSBase = false; 314 HasLZCNT = false; 315 HasBMI = false; 316 HasBMI2 = false; 317 HasVBMI = false; 318 HasIFMA = false; 319 HasRTM = false; 320 HasERI = false; 321 HasCDI = false; 322 HasPFI = false; 323 HasDQI = false; 324 HasVPOPCNTDQ = false; 325 HasBWI = false; 326 HasVLX = false; 327 HasADX = false; 328 HasPKU = false; 329 HasSHA = false; 330 HasPRFCHW = false; 331 HasRDSEED = false; 332 HasLAHFSAHF = false; 333 HasMWAITX = false; 334 HasCLZERO = false; 335 HasMPX = false; 336 HasSGX = false; 337 HasCLFLUSHOPT = false; 338 HasCLWB = false; 339 IsPMULLDSlow = false; 340 IsSHLDSlow = false; 341 IsUAMem16Slow = false; 342 IsUAMem32Slow = false; 343 HasSSEUnalignedMem = false; 344 HasCmpxchg16b = false; 345 UseLeaForSP = false; 346 HasFastPartialYMMorZMMWrite = false; 347 HasFastScalarFSQRT = false; 348 HasFastVectorFSQRT = false; 349 HasFastLZCNT = false; 350 HasFastSHLDRotate = false; 351 HasMacroFusion = false; 352 HasERMSB = false; 353 HasSlowDivide32 = false; 354 HasSlowDivide64 = false; 355 PadShortFunctions = false; 356 SlowTwoMemOps = false; 357 LEAUsesAG = false; 358 SlowLEA = false; 359 Slow3OpsLEA = false; 360 SlowIncDec = false; 361 stackAlignment = 4; 362 // FIXME: this is a known good value for Yonah. How about others? 363 MaxInlineSizeThreshold = 128; 364 UseSoftFloat = false; 365 X86ProcFamily = Others; 366 GatherOverhead = 1024; 367 ScatterOverhead = 1024; 368 } 369 370 X86Subtarget &X86Subtarget::initializeSubtargetDependencies(StringRef CPU, 371 StringRef FS) { 372 initializeEnvironment(); 373 initSubtargetFeatures(CPU, FS); 374 return *this; 375 } 376 377 X86Subtarget::X86Subtarget(const Triple &TT, StringRef CPU, StringRef FS, 378 const X86TargetMachine &TM, 379 unsigned StackAlignOverride) 380 : X86GenSubtargetInfo(TT, CPU, FS), X86ProcFamily(Others), 381 PICStyle(PICStyles::None), TM(TM), TargetTriple(TT), 382 StackAlignOverride(StackAlignOverride), 383 In64BitMode(TargetTriple.getArch() == Triple::x86_64), 384 In32BitMode(TargetTriple.getArch() == Triple::x86 && 385 TargetTriple.getEnvironment() != Triple::CODE16), 386 In16BitMode(TargetTriple.getArch() == Triple::x86 && 387 TargetTriple.getEnvironment() == Triple::CODE16), 388 InstrInfo(initializeSubtargetDependencies(CPU, FS)), TLInfo(TM, *this), 389 FrameLowering(*this, getStackAlignment()) { 390 // Determine the PICStyle based on the target selected. 391 if (!isPositionIndependent()) 392 setPICStyle(PICStyles::None); 393 else if (is64Bit()) 394 setPICStyle(PICStyles::RIPRel); 395 else if (isTargetCOFF()) 396 setPICStyle(PICStyles::None); 397 else if (isTargetDarwin()) 398 setPICStyle(PICStyles::StubPIC); 399 else if (isTargetELF()) 400 setPICStyle(PICStyles::GOT); 401 402 CallLoweringInfo.reset(new X86CallLowering(*getTargetLowering())); 403 Legalizer.reset(new X86LegalizerInfo(*this, TM)); 404 405 auto *RBI = new X86RegisterBankInfo(*getRegisterInfo()); 406 RegBankInfo.reset(RBI); 407 InstSelector.reset(createX86InstructionSelector(TM, *this, *RBI)); 408 } 409 410 const CallLowering *X86Subtarget::getCallLowering() const { 411 return CallLoweringInfo.get(); 412 } 413 414 const InstructionSelector *X86Subtarget::getInstructionSelector() const { 415 return InstSelector.get(); 416 } 417 418 const LegalizerInfo *X86Subtarget::getLegalizerInfo() const { 419 return Legalizer.get(); 420 } 421 422 const RegisterBankInfo *X86Subtarget::getRegBankInfo() const { 423 return RegBankInfo.get(); 424 } 425 426 bool X86Subtarget::enableEarlyIfConversion() const { 427 return hasCMov() && X86EarlyIfConv; 428 } 429