Compiler.cpp revision 9e0f8f0c8965834de02d007bb30e1dc950cf9762
1/* 2 * Copyright 2010-2012, The Android Open Source Project 3 * 4 * Licensed under the Apache License, Version 2.0 (the "License"); 5 * you may not use this file except in compliance with the License. 6 * You may obtain a copy of the License at 7 * 8 * http://www.apache.org/licenses/LICENSE-2.0 9 * 10 * Unless required by applicable law or agreed to in writing, software 11 * distributed under the License is distributed on an "AS IS" BASIS, 12 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 13 * See the License for the specific language governing permissions and 14 * limitations under the License. 15 */ 16 17#include "bcc/Compiler.h" 18 19#include <llvm/Analysis/Passes.h> 20#include <llvm/Analysis/TargetTransformInfo.h> 21#include <llvm/CodeGen/RegAllocRegistry.h> 22#include <llvm/IR/LegacyPassManager.h> 23#include <llvm/IR/Module.h> 24#include <llvm/Support/TargetRegistry.h> 25#include <llvm/Support/raw_ostream.h> 26#include <llvm/IR/DataLayout.h> 27#include <llvm/Target/TargetSubtargetInfo.h> 28#include <llvm/Target/TargetMachine.h> 29#include <llvm/Transforms/IPO.h> 30#include <llvm/Transforms/IPO/PassManagerBuilder.h> 31#include <llvm/Transforms/Scalar.h> 32#include <llvm/Transforms/Vectorize.h> 33 34#include "bcc/Assert.h" 35#include "bcc/Config/Config.h" 36#include "bcc/Renderscript/RSScript.h" 37#include "bcc/Renderscript/RSTransforms.h" 38#include "bcc/Script.h" 39#include "bcc/Source.h" 40#include "bcc/Support/CompilerConfig.h" 41#include "bcc/Support/Log.h" 42#include "bcc/Support/OutputFile.h" 43#include "bcinfo/MetadataExtractor.h" 44#include "rsDefines.h" 45 46#include <string> 47 48using namespace bcc; 49 50const char *Compiler::GetErrorString(enum ErrorCode pErrCode) { 51 switch (pErrCode) { 52 case kSuccess: 53 return "Successfully compiled."; 54 case kInvalidConfigNoTarget: 55 return "Invalid compiler config supplied (getTarget() returns nullptr.) " 56 "(missing call to CompilerConfig::initialize()?)"; 57 case kErrCreateTargetMachine: 58 return "Failed to create llvm::TargetMachine."; 59 case kErrSwitchTargetMachine: 60 return "Failed to switch llvm::TargetMachine."; 61 case kErrNoTargetMachine: 62 return "Failed to compile the script since there's no available " 63 "TargetMachine. (missing call to Compiler::config()?)"; 64 case kErrMaterialization: 65 return "Failed to materialize the module."; 66 case kErrInvalidOutputFileState: 67 return "Supplied output file was invalid (in the error state.)"; 68 case kErrPrepareOutput: 69 return "Failed to prepare file for output."; 70 case kPrepareCodeGenPass: 71 return "Failed to construct pass list for code-generation."; 72 case kErrCustomPasses: 73 return "Error occurred while adding custom passes."; 74 case kErrInvalidSource: 75 return "Error loading input bitcode"; 76 case kIllegalGlobalFunction: 77 return "Use of undefined external function"; 78 case kErrInvalidTargetMachine: 79 return "Invalid/unexpected llvm::TargetMachine."; 80 } 81 82 // This assert should never be reached as the compiler verifies that the 83 // above switch coveres all enum values. 84 bccAssert(false && "Unknown error code encountered"); 85 return ""; 86} 87 88//===----------------------------------------------------------------------===// 89// Instance Methods 90//===----------------------------------------------------------------------===// 91Compiler::Compiler() : mTarget(nullptr), mEnableOpt(true) { 92 return; 93} 94 95Compiler::Compiler(const CompilerConfig &pConfig) : mTarget(nullptr), 96 mEnableOpt(true) { 97 const std::string &triple = pConfig.getTriple(); 98 99 enum ErrorCode err = config(pConfig); 100 if (err != kSuccess) { 101 ALOGE("%s (%s, features: %s)", GetErrorString(err), 102 triple.c_str(), pConfig.getFeatureString().c_str()); 103 return; 104 } 105 106 return; 107} 108 109enum Compiler::ErrorCode Compiler::config(const CompilerConfig &pConfig) { 110 if (pConfig.getTarget() == nullptr) { 111 return kInvalidConfigNoTarget; 112 } 113 114 llvm::TargetMachine *new_target = 115 (pConfig.getTarget())->createTargetMachine(pConfig.getTriple(), 116 pConfig.getCPU(), 117 pConfig.getFeatureString(), 118 pConfig.getTargetOptions(), 119 pConfig.getRelocationModel(), 120 pConfig.getCodeModel(), 121 pConfig.getOptimizationLevel()); 122 123 if (new_target == nullptr) { 124 return ((mTarget != nullptr) ? kErrSwitchTargetMachine : 125 kErrCreateTargetMachine); 126 } 127 128 // Replace the old TargetMachine. 129 delete mTarget; 130 mTarget = new_target; 131 132 // Adjust register allocation policy according to the optimization level. 133 // createFastRegisterAllocator: fast but bad quality 134 // createLinearScanRegisterAllocator: not so fast but good quality 135 if ((pConfig.getOptimizationLevel() == llvm::CodeGenOpt::None)) { 136 llvm::RegisterRegAlloc::setDefault(llvm::createFastRegisterAllocator); 137 } else { 138 llvm::RegisterRegAlloc::setDefault(llvm::createGreedyRegisterAllocator); 139 } 140 141 return kSuccess; 142} 143 144Compiler::~Compiler() { 145 delete mTarget; 146} 147 148 149// This function has complete responsibility for creating and executing the 150// exact list of compiler passes. 151enum Compiler::ErrorCode Compiler::runPasses(Script &pScript, 152 llvm::raw_pwrite_stream &pResult) { 153 // Pass manager for link-time optimization 154 llvm::legacy::PassManager transformPasses; 155 156 // Empty MCContext. 157 llvm::MCContext *mc_context = nullptr; 158 159 transformPasses.add( 160 createTargetTransformInfoWrapperPass(mTarget->getTargetIRAnalysis())); 161 162 // Add some initial custom passes. 163 addInvokeHelperPass(transformPasses); 164 addExpandKernelPass(transformPasses); 165 addDebugInfoPass(pScript, transformPasses); 166 addInvariantPass(transformPasses); 167 if (mTarget->getOptLevel() != llvm::CodeGenOpt::None) { 168 if (!addInternalizeSymbolsPass(pScript, transformPasses)) 169 return kErrCustomPasses; 170 } 171 addGlobalInfoPass(pScript, transformPasses); 172 173 if (mTarget->getOptLevel() == llvm::CodeGenOpt::None) { 174 transformPasses.add(llvm::createGlobalOptimizerPass()); 175 transformPasses.add(llvm::createConstantMergePass()); 176 177 } else { 178 // FIXME: Figure out which passes should be executed. 179 llvm::PassManagerBuilder Builder; 180 Builder.Inliner = llvm::createFunctionInliningPass(); 181 Builder.populateLTOPassManager(transformPasses); 182 183 /* FIXME: Reenable autovectorization after rebase. 184 bug 19324423 185 // Add vectorization passes after LTO passes are in 186 // additional flag: -unroll-runtime 187 transformPasses.add(llvm::createLoopUnrollPass(-1, 16, 0, 1)); 188 // Need to pass appropriate flags here: -scalarize-load-store 189 transformPasses.add(llvm::createScalarizerPass()); 190 transformPasses.add(llvm::createCFGSimplificationPass()); 191 transformPasses.add(llvm::createScopedNoAliasAAPass()); 192 transformPasses.add(llvm::createScalarEvolutionAliasAnalysisPass()); 193 // additional flags: -slp-vectorize-hor -slp-vectorize-hor-store (unnecessary?) 194 transformPasses.add(llvm::createSLPVectorizerPass()); 195 transformPasses.add(llvm::createDeadCodeEliminationPass()); 196 transformPasses.add(llvm::createInstructionCombiningPass()); 197 */ 198 } 199 200 // These passes have to come after LTO, since we don't want to examine 201 // functions that are never actually called. 202 if (llvm::Triple(getTargetMachine().getTargetTriple()).getArch() == llvm::Triple::x86_64) 203 transformPasses.add(createRSX86_64CallConvPass()); // Add pass to correct calling convention for X86-64. 204 transformPasses.add(createRSIsThreadablePass()); // Add pass to mark script as threadable. 205 206 // RSEmbedInfoPass needs to come after we have scanned for non-threadable 207 // functions. 208 // Script passed to RSCompiler must be a RSScript. 209 RSScript &script = static_cast<RSScript &>(pScript); 210 if (script.getEmbedInfo()) 211 transformPasses.add(createRSEmbedInfoPass()); 212 213 // Execute the passes. 214 transformPasses.run(pScript.getSource().getModule()); 215 216 // Run backend separately to avoid interference between debug metadata 217 // generation and backend initialization. 218 llvm::legacy::PassManager codeGenPasses; 219 220 // Add passes to the pass manager to emit machine code through MC layer. 221 if (mTarget->addPassesToEmitMC(codeGenPasses, mc_context, pResult, 222 /* DisableVerify */false)) { 223 return kPrepareCodeGenPass; 224 } 225 226 // Execute the passes. 227 codeGenPasses.run(pScript.getSource().getModule()); 228 229 return kSuccess; 230} 231 232enum Compiler::ErrorCode Compiler::compile(Script &pScript, 233 llvm::raw_pwrite_stream &pResult, 234 llvm::raw_ostream *IRStream) { 235 llvm::Module &module = pScript.getSource().getModule(); 236 enum ErrorCode err; 237 238 if (mTarget == nullptr) { 239 return kErrNoTargetMachine; 240 } 241 242 const std::string &triple = module.getTargetTriple(); 243 const llvm::DataLayout dl = getTargetMachine().createDataLayout(); 244 unsigned int pointerSize = dl.getPointerSizeInBits(); 245 if (triple == "armv7-none-linux-gnueabi") { 246 if (pointerSize != 32) { 247 return kErrInvalidSource; 248 } 249 } else if (triple == "aarch64-none-linux-gnueabi") { 250 if (pointerSize != 64) { 251 return kErrInvalidSource; 252 } 253 } else { 254 return kErrInvalidSource; 255 } 256 257 if (getTargetMachine().getTargetTriple().getArch() == llvm::Triple::x86) { 258 // Detect and fail if TargetMachine datalayout is different than what we 259 // expect. This is to detect changes in default target layout for x86 and 260 // update X86_CUSTOM_DL_STRING in include/bcc/Config/Config.h appropriately. 261 if (dl.getStringRepresentation().compare(X86_DEFAULT_DL_STRING) != 0) { 262 return kErrInvalidTargetMachine; 263 } 264 } 265 266 // Sanitize module's target information. 267 module.setTargetTriple(getTargetMachine().getTargetTriple().str()); 268 module.setDataLayout(getTargetMachine().createDataLayout()); 269 270 // Materialize the bitcode module. 271 if (module.getMaterializer() != nullptr) { 272 // A module with non-null materializer means that it is a lazy-load module. 273 // Materialize it now. This function returns false when the materialization 274 // is successful. 275 std::error_code ec = module.materializeAll(); 276 if (ec) { 277 ALOGE("Failed to materialize the module `%s'! (%s)", 278 module.getModuleIdentifier().c_str(), ec.message().c_str()); 279 return kErrMaterialization; 280 } 281 } 282 283 if ((err = runPasses(pScript, pResult)) != kSuccess) { 284 return err; 285 } 286 287 if (IRStream) { 288 *IRStream << module; 289 } 290 291 return kSuccess; 292} 293 294enum Compiler::ErrorCode Compiler::compile(Script &pScript, 295 OutputFile &pResult, 296 llvm::raw_ostream *IRStream) { 297 // Check the state of the specified output file. 298 if (pResult.hasError()) { 299 return kErrInvalidOutputFileState; 300 } 301 302 // Open the output file decorated in llvm::raw_ostream. 303 llvm::raw_pwrite_stream *out = pResult.dup(); 304 if (out == nullptr) { 305 return kErrPrepareOutput; 306 } 307 308 // Delegate the request. 309 enum Compiler::ErrorCode err = compile(pScript, *out, IRStream); 310 311 // Close the output before return. 312 delete out; 313 314 return err; 315} 316 317bool Compiler::addInternalizeSymbolsPass(Script &pScript, llvm::legacy::PassManager &pPM) { 318 // Add a pass to internalize the symbols that don't need to have global 319 // visibility. 320 RSScript &script = static_cast<RSScript &>(pScript); 321 llvm::Module &module = script.getSource().getModule(); 322 bcinfo::MetadataExtractor me(&module); 323 if (!me.extract()) { 324 bccAssert(false && "Could not extract metadata for module!"); 325 return false; 326 } 327 328 // The vector contains the symbols that should not be internalized. 329 std::vector<const char *> export_symbols; 330 331 const char *sf[] = { 332 kRoot, // Graphics drawing function or compute kernel. 333 kInit, // Initialization routine called implicitly on startup. 334 kRsDtor, // Static global destructor for a script instance. 335 kRsInfo, // Variable containing string of RS metadata info. 336 kRsGlobalEntries, // Optional number of global variables. 337 kRsGlobalNames, // Optional global variable name info. 338 kRsGlobalAddresses, // Optional global variable address info. 339 kRsGlobalSizes, // Optional global variable size info. 340 kRsGlobalProperties, // Optional global variable properties. 341 nullptr // Must be nullptr-terminated. 342 }; 343 const char **special_functions = sf; 344 // Special RS functions should always be global symbols. 345 while (*special_functions != nullptr) { 346 export_symbols.push_back(*special_functions); 347 special_functions++; 348 } 349 350 // Visibility of symbols appeared in rs_export_var and rs_export_func should 351 // also be preserved. 352 size_t exportVarCount = me.getExportVarCount(); 353 size_t exportFuncCount = me.getExportFuncCount(); 354 size_t exportForEachCount = me.getExportForEachSignatureCount(); 355 size_t exportReduceCount = me.getExportReduceCount(); 356 size_t exportReduceNewCount = me.getExportReduceNewCount(); 357 const char **exportVarNameList = me.getExportVarNameList(); 358 const char **exportFuncNameList = me.getExportFuncNameList(); 359 const char **exportForEachNameList = me.getExportForEachNameList(); 360 const char **exportReduceNameList = me.getExportReduceNameList(); 361 const bcinfo::MetadataExtractor::ReduceNew *exportReduceNewList = me.getExportReduceNewList(); 362 size_t i; 363 364 for (i = 0; i < exportVarCount; ++i) { 365 export_symbols.push_back(exportVarNameList[i]); 366 } 367 368 for (i = 0; i < exportFuncCount; ++i) { 369 export_symbols.push_back(exportFuncNameList[i]); 370 } 371 372 // Expanded foreach and reduce functions should not be internalized; 373 // nor should general reduction initializer, combiner, and 374 // outconverter functions. keep_funcs keeps the names of these 375 // functions around until createInternalizePass() is finished making 376 // its own copy of the visible symbols. 377 std::vector<std::string> keep_funcs; 378 keep_funcs.reserve(exportForEachCount + exportReduceCount + exportReduceNewCount*4); 379 380 for (i = 0; i < exportForEachCount; ++i) { 381 keep_funcs.push_back(std::string(exportForEachNameList[i]) + ".expand"); 382 } 383 for (i = 0; i < exportReduceCount; ++i) { 384 keep_funcs.push_back(std::string(exportReduceNameList[i]) + ".expand"); 385 } 386 auto keepFuncsPushBackIfPresent = [&keep_funcs](const char *Name) { 387 if (Name) keep_funcs.push_back(Name); 388 }; 389 for (i = 0; i < exportReduceNewCount; ++i) { 390 keep_funcs.push_back(std::string(exportReduceNewList[i].mAccumulatorName) + ".expand"); 391 keepFuncsPushBackIfPresent(exportReduceNewList[i].mInitializerName); 392 keepFuncsPushBackIfPresent(exportReduceNewList[i].mCombinerName); 393 keepFuncsPushBackIfPresent(exportReduceNewList[i].mOutConverterName); 394 } 395 396 for (auto &symbol_name : keep_funcs) { 397 export_symbols.push_back(symbol_name.c_str()); 398 } 399 400 // http://b/26165616 - WAR for this bug defines the __truncxfhf2 function in 401 // frameworks/rs/driver/runtime. Don't internalize this function for x86, so 402 // that a script can find and link against it. 403 llvm::Triple triple(getTargetMachine().getTargetTriple()); 404 if (triple.getArch() == llvm::Triple::x86) { 405 export_symbols.push_back("__truncxfhf2"); 406 } 407 408 pPM.add(llvm::createInternalizePass(export_symbols)); 409 410 return true; 411} 412 413void Compiler::addInvokeHelperPass(llvm::legacy::PassManager &pPM) { 414 llvm::Triple arch(getTargetMachine().getTargetTriple()); 415 if (arch.isArch64Bit()) { 416 pPM.add(createRSInvokeHelperPass()); 417 } 418} 419 420void Compiler::addDebugInfoPass(Script &pScript, llvm::legacy::PassManager &pPM) { 421 if (pScript.getSource().getDebugInfoEnabled()) 422 pPM.add(createRSAddDebugInfoPass()); 423} 424 425void Compiler::addExpandKernelPass(llvm::legacy::PassManager &pPM) { 426 // Expand ForEach and reduce on CPU path to reduce launch overhead. 427 bool pEnableStepOpt = true; 428 pPM.add(createRSKernelExpandPass(pEnableStepOpt)); 429} 430 431void Compiler::addGlobalInfoPass(Script &pScript, llvm::legacy::PassManager &pPM) { 432 // Add additional information about RS global variables inside the Module. 433 RSScript &script = static_cast<RSScript &>(pScript); 434 if (script.getEmbedGlobalInfo()) { 435 pPM.add(createRSGlobalInfoPass(script.getEmbedGlobalInfoSkipConstant())); 436 } 437} 438 439void Compiler::addInvariantPass(llvm::legacy::PassManager &pPM) { 440 // Mark Loads from RsExpandKernelDriverInfo as "load.invariant". 441 // Should run after ExpandForEach and before inlining. 442 pPM.add(createRSInvariantPass()); 443} 444 445enum Compiler::ErrorCode Compiler::screenGlobalFunctions(Script &pScript) { 446 llvm::Module &module = pScript.getSource().getModule(); 447 448 // Materialize the bitcode module in case this is a lazy-load module. Do not 449 // clear the materializer by calling materializeAllPermanently since the 450 // runtime library has not been merged into the module yet. 451 if (module.getMaterializer() != nullptr) { 452 std::error_code ec = module.materializeAll(); 453 if (ec) { 454 ALOGE("Failed to materialize module `%s' when screening globals! (%s)", 455 module.getModuleIdentifier().c_str(), ec.message().c_str()); 456 return kErrMaterialization; 457 } 458 } 459 460 // Add pass to check for illegal function calls. 461 llvm::legacy::PassManager pPM; 462 pPM.add(createRSScreenFunctionsPass()); 463 pPM.run(module); 464 465 return kSuccess; 466 467} 468 469void Compiler::translateGEPs(Script &pScript) { 470 llvm::legacy::PassManager pPM; 471 pPM.add(createRSX86TranslateGEPPass()); 472 473 // Materialization done in screenGlobalFunctions above. 474 pPM.run(pScript.getSource().getModule()); 475} 476