LLVMTargetMachine.cpp revision 5669e3009761dff20b67e18a382c334041887928
1//===-- LLVMTargetMachine.cpp - Implement the LLVMTargetMachine class -----===// 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 LLVMTargetMachine class. 11// 12//===----------------------------------------------------------------------===// 13 14#include "llvm/Target/TargetMachine.h" 15#include "llvm/PassManager.h" 16#include "llvm/Pass.h" 17#include "llvm/Assembly/PrintModulePass.h" 18#include "llvm/CodeGen/AsmPrinter.h" 19#include "llvm/CodeGen/Passes.h" 20#include "llvm/CodeGen/GCStrategy.h" 21#include "llvm/CodeGen/MachineFunctionAnalysis.h" 22#include "llvm/Target/TargetOptions.h" 23#include "llvm/MC/MCAsmInfo.h" 24#include "llvm/MC/MCContext.h" 25#include "llvm/MC/MCStreamer.h" 26#include "llvm/Target/TargetData.h" 27#include "llvm/Target/TargetRegistry.h" 28#include "llvm/Transforms/Scalar.h" 29#include "llvm/ADT/OwningPtr.h" 30#include "llvm/Support/CommandLine.h" 31#include "llvm/Support/Debug.h" 32#include "llvm/Support/FormattedStream.h" 33using namespace llvm; 34 35namespace llvm { 36 bool EnableFastISel; 37} 38 39static cl::opt<bool> DisablePostRA("disable-post-ra", cl::Hidden, 40 cl::desc("Disable Post Regalloc")); 41static cl::opt<bool> DisableBranchFold("disable-branch-fold", cl::Hidden, 42 cl::desc("Disable branch folding")); 43static cl::opt<bool> DisableTailDuplicate("disable-tail-duplicate", cl::Hidden, 44 cl::desc("Disable tail duplication")); 45static cl::opt<bool> DisableEarlyTailDup("disable-early-taildup", cl::Hidden, 46 cl::desc("Disable pre-register allocation tail duplication")); 47static cl::opt<bool> DisableCodePlace("disable-code-place", cl::Hidden, 48 cl::desc("Disable code placement")); 49static cl::opt<bool> DisableSSC("disable-ssc", cl::Hidden, 50 cl::desc("Disable Stack Slot Coloring")); 51static cl::opt<bool> DisableMachineLICM("disable-machine-licm", cl::Hidden, 52 cl::desc("Disable Machine LICM")); 53static cl::opt<bool> DisableMachineSink("disable-machine-sink", cl::Hidden, 54 cl::desc("Disable Machine Sinking")); 55static cl::opt<bool> DisableLSR("disable-lsr", cl::Hidden, 56 cl::desc("Disable Loop Strength Reduction Pass")); 57static cl::opt<bool> DisableCGP("disable-cgp", cl::Hidden, 58 cl::desc("Disable Codegen Prepare")); 59static cl::opt<bool> PrintLSR("print-lsr-output", cl::Hidden, 60 cl::desc("Print LLVM IR produced by the loop-reduce pass")); 61static cl::opt<bool> PrintISelInput("print-isel-input", cl::Hidden, 62 cl::desc("Print LLVM IR input to isel pass")); 63static cl::opt<bool> PrintGCInfo("print-gc", cl::Hidden, 64 cl::desc("Dump garbage collector data")); 65static cl::opt<bool> VerifyMachineCode("verify-machineinstrs", cl::Hidden, 66 cl::desc("Verify generated machine code"), 67 cl::init(getenv("LLVM_VERIFY_MACHINEINSTRS")!=NULL)); 68 69static cl::opt<cl::boolOrDefault> 70AsmVerbose("asm-verbose", cl::desc("Add comments to directives."), 71 cl::init(cl::BOU_UNSET)); 72 73static bool getVerboseAsm() { 74 switch (AsmVerbose) { 75 default: 76 case cl::BOU_UNSET: return TargetMachine::getAsmVerbosityDefault(); 77 case cl::BOU_TRUE: return true; 78 case cl::BOU_FALSE: return false; 79 } 80} 81 82// Enable or disable FastISel. Both options are needed, because 83// FastISel is enabled by default with -fast, and we wish to be 84// able to enable or disable fast-isel independently from -O0. 85static cl::opt<cl::boolOrDefault> 86EnableFastISelOption("fast-isel", cl::Hidden, 87 cl::desc("Enable the \"fast\" instruction selector")); 88 89// Enable or disable an experimental optimization to split GEPs 90// and run a special GVN pass which does not examine loads, in 91// an effort to factor out redundancy implicit in complex GEPs. 92static cl::opt<bool> EnableSplitGEPGVN("split-gep-gvn", cl::Hidden, 93 cl::desc("Split GEPs and run no-load GVN")); 94 95LLVMTargetMachine::LLVMTargetMachine(const Target &T, 96 const std::string &TargetTriple) 97 : TargetMachine(T) { 98 AsmInfo = T.createAsmInfo(TargetTriple); 99} 100 101// Set the default code model for the JIT for a generic target. 102// FIXME: Is small right here? or .is64Bit() ? Large : Small? 103void 104LLVMTargetMachine::setCodeModelForJIT() { 105 setCodeModel(CodeModel::Small); 106} 107 108// Set the default code model for static compilation for a generic target. 109void 110LLVMTargetMachine::setCodeModelForStatic() { 111 setCodeModel(CodeModel::Small); 112} 113 114bool LLVMTargetMachine::addPassesToEmitFile(PassManagerBase &PM, 115 formatted_raw_ostream &Out, 116 CodeGenFileType FileType, 117 CodeGenOpt::Level OptLevel) { 118 // Add common CodeGen passes. 119 if (addCommonCodeGenPasses(PM, OptLevel)) 120 return true; 121 122 OwningPtr<MCContext> Context(new MCContext()); 123 OwningPtr<MCStreamer> AsmStreamer; 124 125 formatted_raw_ostream *LegacyOutput; 126 switch (FileType) { 127 default: return true; 128 case CGFT_AssemblyFile: { 129 const MCAsmInfo &MAI = *getMCAsmInfo(); 130 MCInstPrinter *InstPrinter = 131 getTarget().createMCInstPrinter(MAI.getAssemblerDialect(), MAI, Out); 132 AsmStreamer.reset(createAsmStreamer(*Context, Out, MAI, 133 getTargetData()->isLittleEndian(), 134 getVerboseAsm(), InstPrinter, 135 /*codeemitter*/0)); 136 // Set the AsmPrinter's "O" to the output file. 137 LegacyOutput = &Out; 138 break; 139 } 140 case CGFT_ObjectFile: { 141 // Create the code emitter for the target if it exists. If not, .o file 142 // emission fails. 143 MCCodeEmitter *MCE = getTarget().createCodeEmitter(*this); 144 if (MCE == 0) 145 return true; 146 147 AsmStreamer.reset(createMachOStreamer(*Context, Out, MCE)); 148 149 // Any output to the asmprinter's "O" stream is bad and needs to be fixed, 150 // force it to come out stderr. 151 // FIXME: this is horrible and leaks, eventually remove the raw_ostream from 152 // asmprinter. 153 LegacyOutput = new formatted_raw_ostream(errs()); 154 break; 155 } 156 case CGFT_Null: 157 // The Null output is intended for use for performance analysis and testing, 158 // not real users. 159 AsmStreamer.reset(createNullStreamer(*Context)); 160 // Any output to the asmprinter's "O" stream is bad and needs to be fixed, 161 // force it to come out stderr. 162 // FIXME: this is horrible and leaks, eventually remove the raw_ostream from 163 // asmprinter. 164 LegacyOutput = new formatted_raw_ostream(errs()); 165 break; 166 } 167 168 // Create the AsmPrinter, which takes ownership of Context and AsmStreamer 169 // if successful. 170 FunctionPass *Printer = 171 getTarget().createAsmPrinter(*LegacyOutput, *this, *Context, *AsmStreamer, 172 getMCAsmInfo()); 173 if (Printer == 0) 174 return true; 175 176 // If successful, createAsmPrinter took ownership of AsmStreamer and Context. 177 Context.take(); AsmStreamer.take(); 178 179 PM.add(Printer); 180 181 // Make sure the code model is set. 182 setCodeModelForStatic(); 183 PM.add(createGCInfoDeleter()); 184 return false; 185} 186 187/// addPassesToEmitMachineCode - Add passes to the specified pass manager to 188/// get machine code emitted. This uses a JITCodeEmitter object to handle 189/// actually outputting the machine code and resolving things like the address 190/// of functions. This method should returns true if machine code emission is 191/// not supported. 192/// 193bool LLVMTargetMachine::addPassesToEmitMachineCode(PassManagerBase &PM, 194 JITCodeEmitter &JCE, 195 CodeGenOpt::Level OptLevel) { 196 // Make sure the code model is set. 197 setCodeModelForJIT(); 198 199 // Add common CodeGen passes. 200 if (addCommonCodeGenPasses(PM, OptLevel)) 201 return true; 202 203 addCodeEmitter(PM, OptLevel, JCE); 204 PM.add(createGCInfoDeleter()); 205 206 return false; // success! 207} 208 209static void printAndVerify(PassManagerBase &PM, 210 const char *Banner, 211 bool allowDoubleDefs = false) { 212 if (PrintMachineCode) 213 PM.add(createMachineFunctionPrinterPass(dbgs(), Banner)); 214 215 if (VerifyMachineCode) 216 PM.add(createMachineVerifierPass(allowDoubleDefs)); 217} 218 219/// addCommonCodeGenPasses - Add standard LLVM codegen passes used for both 220/// emitting to assembly files or machine code output. 221/// 222bool LLVMTargetMachine::addCommonCodeGenPasses(PassManagerBase &PM, 223 CodeGenOpt::Level OptLevel) { 224 // Standard LLVM-Level Passes. 225 226 // Optionally, tun split-GEPs and no-load GVN. 227 if (EnableSplitGEPGVN) { 228 PM.add(createGEPSplitterPass()); 229 PM.add(createGVNPass(/*NoPRE=*/false, /*NoLoads=*/true)); 230 } 231 232 // Run loop strength reduction before anything else. 233 if (OptLevel != CodeGenOpt::None && !DisableLSR) { 234 PM.add(createLoopStrengthReducePass(getTargetLowering())); 235 if (PrintLSR) 236 PM.add(createPrintFunctionPass("\n\n*** Code after LSR ***\n", &dbgs())); 237 } 238 239 // Turn exception handling constructs into something the code generators can 240 // handle. 241 switch (getMCAsmInfo()->getExceptionHandlingType()) 242 { 243 case ExceptionHandling::SjLj: 244 // SjLj piggy-backs on dwarf for this bit. The cleanups done apply to both 245 // Dwarf EH prepare needs to be run after SjLj prepare. Otherwise, 246 // catch info can get misplaced when a selector ends up more than one block 247 // removed from the parent invoke(s). This could happen when a landing 248 // pad is shared by multiple invokes and is also a target of a normal 249 // edge from elsewhere. 250 PM.add(createSjLjEHPass(getTargetLowering())); 251 PM.add(createDwarfEHPass(getTargetLowering(), OptLevel==CodeGenOpt::None)); 252 break; 253 case ExceptionHandling::Dwarf: 254 PM.add(createDwarfEHPass(getTargetLowering(), OptLevel==CodeGenOpt::None)); 255 break; 256 case ExceptionHandling::None: 257 PM.add(createLowerInvokePass(getTargetLowering())); 258 break; 259 } 260 261 PM.add(createGCLoweringPass()); 262 263 // Make sure that no unreachable blocks are instruction selected. 264 PM.add(createUnreachableBlockEliminationPass()); 265 266 if (OptLevel != CodeGenOpt::None && !DisableCGP) 267 PM.add(createCodeGenPreparePass(getTargetLowering())); 268 269 PM.add(createStackProtectorPass(getTargetLowering())); 270 271 if (PrintISelInput) 272 PM.add(createPrintFunctionPass("\n\n" 273 "*** Final LLVM Code input to ISel ***\n", 274 &dbgs())); 275 276 // Standard Lower-Level Passes. 277 278 // Set up a MachineFunction for the rest of CodeGen to work on. 279 PM.add(new MachineFunctionAnalysis(*this, OptLevel)); 280 281 // Enable FastISel with -fast, but allow that to be overridden. 282 if (EnableFastISelOption == cl::BOU_TRUE || 283 (OptLevel == CodeGenOpt::None && EnableFastISelOption != cl::BOU_FALSE)) 284 EnableFastISel = true; 285 286 // Ask the target for an isel. 287 if (addInstSelector(PM, OptLevel)) 288 return true; 289 290 // Print the instruction selected machine code... 291 printAndVerify(PM, "After Instruction Selection", 292 /* allowDoubleDefs= */ true); 293 294 if (OptLevel != CodeGenOpt::None) { 295 PM.add(createOptimizeExtsPass()); 296 if (!DisableMachineLICM) 297 PM.add(createMachineLICMPass()); 298 if (!DisableMachineSink) 299 PM.add(createMachineSinkingPass()); 300 printAndVerify(PM, "After MachineLICM and MachineSinking", 301 /* allowDoubleDefs= */ true); 302 } 303 304 // Pre-ra tail duplication. 305 if (OptLevel != CodeGenOpt::None && !DisableEarlyTailDup) { 306 PM.add(createTailDuplicatePass(true)); 307 printAndVerify(PM, "After Pre-RegAlloc TailDuplicate", 308 /* allowDoubleDefs= */ true); 309 } 310 311 // Run pre-ra passes. 312 if (addPreRegAlloc(PM, OptLevel)) 313 printAndVerify(PM, "After PreRegAlloc passes", 314 /* allowDoubleDefs= */ true); 315 316 // Perform register allocation. 317 PM.add(createRegisterAllocator()); 318 printAndVerify(PM, "After Register Allocation"); 319 320 // Perform stack slot coloring. 321 if (OptLevel != CodeGenOpt::None && !DisableSSC) { 322 // FIXME: Re-enable coloring with register when it's capable of adding 323 // kill markers. 324 PM.add(createStackSlotColoringPass(false)); 325 printAndVerify(PM, "After StackSlotColoring"); 326 } 327 328 // Run post-ra passes. 329 if (addPostRegAlloc(PM, OptLevel)) 330 printAndVerify(PM, "After PostRegAlloc passes"); 331 332 PM.add(createLowerSubregsPass()); 333 printAndVerify(PM, "After LowerSubregs"); 334 335 // Insert prolog/epilog code. Eliminate abstract frame index references... 336 PM.add(createPrologEpilogCodeInserter()); 337 printAndVerify(PM, "After PrologEpilogCodeInserter"); 338 339 // Run pre-sched2 passes. 340 if (addPreSched2(PM, OptLevel)) 341 printAndVerify(PM, "After PreSched2 passes"); 342 343 // Second pass scheduler. 344 if (OptLevel != CodeGenOpt::None && !DisablePostRA) { 345 PM.add(createPostRAScheduler(OptLevel)); 346 printAndVerify(PM, "After PostRAScheduler"); 347 } 348 349 // Branch folding must be run after regalloc and prolog/epilog insertion. 350 if (OptLevel != CodeGenOpt::None && !DisableBranchFold) { 351 PM.add(createBranchFoldingPass(getEnableTailMergeDefault())); 352 printAndVerify(PM, "After BranchFolding"); 353 } 354 355 // Tail duplication. 356 if (OptLevel != CodeGenOpt::None && !DisableTailDuplicate) { 357 PM.add(createTailDuplicatePass(false)); 358 printAndVerify(PM, "After TailDuplicate"); 359 } 360 361 PM.add(createGCMachineCodeAnalysisPass()); 362 363 if (PrintGCInfo) 364 PM.add(createGCInfoPrinter(dbgs())); 365 366 if (OptLevel != CodeGenOpt::None && !DisableCodePlace) { 367 PM.add(createCodePlacementOptPass()); 368 printAndVerify(PM, "After CodePlacementOpt"); 369 } 370 371 if (addPreEmitPass(PM, OptLevel)) 372 printAndVerify(PM, "After PreEmit passes"); 373 374 return false; 375} 376