1 //===-- X86TargetMachine.cpp - Define TargetMachine for the X86 -----------===// 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 defines the X86 specific subclass of TargetMachine. 11 // 12 //===----------------------------------------------------------------------===// 13 14 #include "X86TargetMachine.h" 15 #include "X86.h" 16 #include "X86TargetObjectFile.h" 17 #include "X86TargetTransformInfo.h" 18 #include "llvm/CodeGen/Passes.h" 19 #include "llvm/IR/Function.h" 20 #include "llvm/IR/LegacyPassManager.h" 21 #include "llvm/Support/CommandLine.h" 22 #include "llvm/Support/FormattedStream.h" 23 #include "llvm/Support/TargetRegistry.h" 24 #include "llvm/Target/TargetOptions.h" 25 using namespace llvm; 26 27 static cl::opt<bool> EnableMachineCombinerPass("x86-machine-combiner", 28 cl::desc("Enable the machine combiner pass"), 29 cl::init(true), cl::Hidden); 30 31 namespace llvm { 32 void initializeWinEHStatePassPass(PassRegistry &); 33 } 34 35 extern "C" void LLVMInitializeX86Target() { 36 // Register the target. 37 RegisterTargetMachine<X86TargetMachine> X(TheX86_32Target); 38 RegisterTargetMachine<X86TargetMachine> Y(TheX86_64Target); 39 40 PassRegistry &PR = *PassRegistry::getPassRegistry(); 41 initializeWinEHStatePassPass(PR); 42 } 43 44 static std::unique_ptr<TargetLoweringObjectFile> createTLOF(const Triple &TT) { 45 if (TT.isOSBinFormatMachO()) { 46 if (TT.getArch() == Triple::x86_64) 47 return make_unique<X86_64MachoTargetObjectFile>(); 48 return make_unique<TargetLoweringObjectFileMachO>(); 49 } 50 51 if (TT.isOSLinux() || TT.isOSNaCl()) 52 return make_unique<X86LinuxNaClTargetObjectFile>(); 53 if (TT.isOSBinFormatELF()) 54 return make_unique<X86ELFTargetObjectFile>(); 55 if (TT.isKnownWindowsMSVCEnvironment() || TT.isWindowsCoreCLREnvironment()) 56 return make_unique<X86WindowsTargetObjectFile>(); 57 if (TT.isOSBinFormatCOFF()) 58 return make_unique<TargetLoweringObjectFileCOFF>(); 59 llvm_unreachable("unknown subtarget type"); 60 } 61 62 static std::string computeDataLayout(const Triple &TT) { 63 // X86 is little endian 64 std::string Ret = "e"; 65 66 Ret += DataLayout::getManglingComponent(TT); 67 // X86 and x32 have 32 bit pointers. 68 if ((TT.isArch64Bit() && 69 (TT.getEnvironment() == Triple::GNUX32 || TT.isOSNaCl())) || 70 !TT.isArch64Bit()) 71 Ret += "-p:32:32"; 72 73 // Some ABIs align 64 bit integers and doubles to 64 bits, others to 32. 74 if (TT.isArch64Bit() || TT.isOSWindows() || TT.isOSNaCl()) 75 Ret += "-i64:64"; 76 else if (TT.isOSIAMCU()) 77 Ret += "-i64:32-f64:32"; 78 else 79 Ret += "-f64:32:64"; 80 81 // Some ABIs align long double to 128 bits, others to 32. 82 if (TT.isOSNaCl() || TT.isOSIAMCU()) 83 ; // No f80 84 else if (TT.isArch64Bit() || TT.isOSDarwin()) 85 Ret += "-f80:128"; 86 else 87 Ret += "-f80:32"; 88 89 if (TT.isOSIAMCU()) 90 Ret += "-f128:32"; 91 92 // The registers can hold 8, 16, 32 or, in x86-64, 64 bits. 93 if (TT.isArch64Bit()) 94 Ret += "-n8:16:32:64"; 95 else 96 Ret += "-n8:16:32"; 97 98 // The stack is aligned to 32 bits on some ABIs and 128 bits on others. 99 if ((!TT.isArch64Bit() && TT.isOSWindows()) || TT.isOSIAMCU()) 100 Ret += "-a:0:32-S32"; 101 else 102 Ret += "-S128"; 103 104 return Ret; 105 } 106 107 /// X86TargetMachine ctor - Create an X86 target. 108 /// 109 X86TargetMachine::X86TargetMachine(const Target &T, const Triple &TT, 110 StringRef CPU, StringRef FS, 111 const TargetOptions &Options, 112 Reloc::Model RM, CodeModel::Model CM, 113 CodeGenOpt::Level OL) 114 : LLVMTargetMachine(T, computeDataLayout(TT), TT, CPU, FS, Options, RM, CM, 115 OL), 116 TLOF(createTLOF(getTargetTriple())), 117 Subtarget(TT, CPU, FS, *this, Options.StackAlignmentOverride) { 118 // Windows stack unwinder gets confused when execution flow "falls through" 119 // after a call to 'noreturn' function. 120 // To prevent that, we emit a trap for 'unreachable' IR instructions. 121 // (which on X86, happens to be the 'ud2' instruction) 122 if (Subtarget.isTargetWin64()) 123 this->Options.TrapUnreachable = true; 124 125 // By default (and when -ffast-math is on), enable estimate codegen for 126 // everything except scalar division. By default, use 1 refinement step for 127 // all operations. Defaults may be overridden by using command-line options. 128 // Scalar division estimates are disabled because they break too much 129 // real-world code. These defaults match GCC behavior. 130 this->Options.Reciprocals.setDefaults("sqrtf", true, 1); 131 this->Options.Reciprocals.setDefaults("divf", false, 1); 132 this->Options.Reciprocals.setDefaults("vec-sqrtf", true, 1); 133 this->Options.Reciprocals.setDefaults("vec-divf", true, 1); 134 135 initAsmInfo(); 136 } 137 138 X86TargetMachine::~X86TargetMachine() {} 139 140 const X86Subtarget * 141 X86TargetMachine::getSubtargetImpl(const Function &F) const { 142 Attribute CPUAttr = F.getFnAttribute("target-cpu"); 143 Attribute FSAttr = F.getFnAttribute("target-features"); 144 145 std::string CPU = !CPUAttr.hasAttribute(Attribute::None) 146 ? CPUAttr.getValueAsString().str() 147 : TargetCPU; 148 std::string FS = !FSAttr.hasAttribute(Attribute::None) 149 ? FSAttr.getValueAsString().str() 150 : TargetFS; 151 152 // FIXME: This is related to the code below to reset the target options, 153 // we need to know whether or not the soft float flag is set on the 154 // function before we can generate a subtarget. We also need to use 155 // it as a key for the subtarget since that can be the only difference 156 // between two functions. 157 bool SoftFloat = 158 F.hasFnAttribute("use-soft-float") && 159 F.getFnAttribute("use-soft-float").getValueAsString() == "true"; 160 // If the soft float attribute is set on the function turn on the soft float 161 // subtarget feature. 162 if (SoftFloat) 163 FS += FS.empty() ? "+soft-float" : ",+soft-float"; 164 165 auto &I = SubtargetMap[CPU + FS]; 166 if (!I) { 167 // This needs to be done before we create a new subtarget since any 168 // creation will depend on the TM and the code generation flags on the 169 // function that reside in TargetOptions. 170 resetTargetOptions(F); 171 I = llvm::make_unique<X86Subtarget>(TargetTriple, CPU, FS, *this, 172 Options.StackAlignmentOverride); 173 } 174 return I.get(); 175 } 176 177 //===----------------------------------------------------------------------===// 178 // Command line options for x86 179 //===----------------------------------------------------------------------===// 180 static cl::opt<bool> 181 UseVZeroUpper("x86-use-vzeroupper", cl::Hidden, 182 cl::desc("Minimize AVX to SSE transition penalty"), 183 cl::init(true)); 184 185 //===----------------------------------------------------------------------===// 186 // X86 TTI query. 187 //===----------------------------------------------------------------------===// 188 189 TargetIRAnalysis X86TargetMachine::getTargetIRAnalysis() { 190 return TargetIRAnalysis([this](const Function &F) { 191 return TargetTransformInfo(X86TTIImpl(this, F)); 192 }); 193 } 194 195 196 //===----------------------------------------------------------------------===// 197 // Pass Pipeline Configuration 198 //===----------------------------------------------------------------------===// 199 200 namespace { 201 /// X86 Code Generator Pass Configuration Options. 202 class X86PassConfig : public TargetPassConfig { 203 public: 204 X86PassConfig(X86TargetMachine *TM, PassManagerBase &PM) 205 : TargetPassConfig(TM, PM) {} 206 207 X86TargetMachine &getX86TargetMachine() const { 208 return getTM<X86TargetMachine>(); 209 } 210 211 void addIRPasses() override; 212 bool addInstSelector() override; 213 bool addILPOpts() override; 214 bool addPreISel() override; 215 void addPreRegAlloc() override; 216 void addPostRegAlloc() override; 217 void addPreEmitPass() override; 218 void addPreSched2() override; 219 }; 220 } // namespace 221 222 TargetPassConfig *X86TargetMachine::createPassConfig(PassManagerBase &PM) { 223 return new X86PassConfig(this, PM); 224 } 225 226 void X86PassConfig::addIRPasses() { 227 addPass(createAtomicExpandPass(&getX86TargetMachine())); 228 229 TargetPassConfig::addIRPasses(); 230 } 231 232 bool X86PassConfig::addInstSelector() { 233 // Install an instruction selector. 234 addPass(createX86ISelDag(getX86TargetMachine(), getOptLevel())); 235 236 // For ELF, cleanup any local-dynamic TLS accesses. 237 if (TM->getTargetTriple().isOSBinFormatELF() && 238 getOptLevel() != CodeGenOpt::None) 239 addPass(createCleanupLocalDynamicTLSPass()); 240 241 addPass(createX86GlobalBaseRegPass()); 242 243 return false; 244 } 245 246 bool X86PassConfig::addILPOpts() { 247 addPass(&EarlyIfConverterID); 248 if (EnableMachineCombinerPass) 249 addPass(&MachineCombinerID); 250 return true; 251 } 252 253 bool X86PassConfig::addPreISel() { 254 // Only add this pass for 32-bit x86 Windows. 255 const Triple &TT = TM->getTargetTriple(); 256 if (TT.isOSWindows() && TT.getArch() == Triple::x86) 257 addPass(createX86WinEHStatePass()); 258 return true; 259 } 260 261 void X86PassConfig::addPreRegAlloc() { 262 if (getOptLevel() != CodeGenOpt::None) 263 addPass(createX86OptimizeLEAs()); 264 265 addPass(createX86CallFrameOptimization()); 266 } 267 268 void X86PassConfig::addPostRegAlloc() { 269 addPass(createX86FloatingPointStackifierPass()); 270 } 271 272 void X86PassConfig::addPreSched2() { addPass(createX86ExpandPseudoPass()); } 273 274 void X86PassConfig::addPreEmitPass() { 275 if (getOptLevel() != CodeGenOpt::None) 276 addPass(createExecutionDependencyFixPass(&X86::VR128RegClass)); 277 278 if (UseVZeroUpper) 279 addPass(createX86IssueVZeroUpperPass()); 280 281 if (getOptLevel() != CodeGenOpt::None) { 282 addPass(createX86FixupBWInsts()); 283 addPass(createX86PadShortFunctions()); 284 addPass(createX86FixupLEAs()); 285 } 286 } 287