Compiler.cpp revision e6795bc257fae07bb8fbf2759bde9332b60a7f78
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/Renderscript/RSScript.h" 36#include "bcc/Renderscript/RSTransforms.h" 37#include "bcc/Script.h" 38#include "bcc/Source.h" 39#include "bcc/Support/CompilerConfig.h" 40#include "bcc/Support/Log.h" 41#include "bcc/Support/OutputFile.h" 42#include "bcinfo/MetadataExtractor.h" 43#include "rsDefines.h" 44 45#include <string> 46 47using namespace bcc; 48 49const char *Compiler::GetErrorString(enum ErrorCode pErrCode) { 50 switch (pErrCode) { 51 case kSuccess: 52 return "Successfully compiled."; 53 case kInvalidConfigNoTarget: 54 return "Invalid compiler config supplied (getTarget() returns nullptr.) " 55 "(missing call to CompilerConfig::initialize()?)"; 56 case kErrCreateTargetMachine: 57 return "Failed to create llvm::TargetMachine."; 58 case kErrSwitchTargetMachine: 59 return "Failed to switch llvm::TargetMachine."; 60 case kErrNoTargetMachine: 61 return "Failed to compile the script since there's no available " 62 "TargetMachine. (missing call to Compiler::config()?)"; 63 case kErrMaterialization: 64 return "Failed to materialize the module."; 65 case kErrInvalidOutputFileState: 66 return "Supplied output file was invalid (in the error state.)"; 67 case kErrPrepareOutput: 68 return "Failed to prepare file for output."; 69 case kPrepareCodeGenPass: 70 return "Failed to construct pass list for code-generation."; 71 case kErrCustomPasses: 72 return "Error occurred while adding custom passes."; 73 case kErrInvalidSource: 74 return "Error loading input bitcode"; 75 case kIllegalGlobalFunction: 76 return "Use of undefined external function"; 77 } 78 79 // This assert should never be reached as the compiler verifies that the 80 // above switch coveres all enum values. 81 assert(false && "Unknown error code encountered"); 82 return ""; 83} 84 85//===----------------------------------------------------------------------===// 86// Instance Methods 87//===----------------------------------------------------------------------===// 88Compiler::Compiler() : mTarget(nullptr), mEnableOpt(true) { 89 return; 90} 91 92Compiler::Compiler(const CompilerConfig &pConfig) : mTarget(nullptr), 93 mEnableOpt(true) { 94 const std::string &triple = pConfig.getTriple(); 95 96 enum ErrorCode err = config(pConfig); 97 if (err != kSuccess) { 98 ALOGE("%s (%s, features: %s)", GetErrorString(err), 99 triple.c_str(), pConfig.getFeatureString().c_str()); 100 return; 101 } 102 103 return; 104} 105 106enum Compiler::ErrorCode Compiler::config(const CompilerConfig &pConfig) { 107 if (pConfig.getTarget() == nullptr) { 108 return kInvalidConfigNoTarget; 109 } 110 111 llvm::TargetMachine *new_target = 112 (pConfig.getTarget())->createTargetMachine(pConfig.getTriple(), 113 pConfig.getCPU(), 114 pConfig.getFeatureString(), 115 pConfig.getTargetOptions(), 116 pConfig.getRelocationModel(), 117 pConfig.getCodeModel(), 118 pConfig.getOptimizationLevel()); 119 120 if (new_target == nullptr) { 121 return ((mTarget != nullptr) ? kErrSwitchTargetMachine : 122 kErrCreateTargetMachine); 123 } 124 125 // Replace the old TargetMachine. 126 delete mTarget; 127 mTarget = new_target; 128 129 // Adjust register allocation policy according to the optimization level. 130 // createFastRegisterAllocator: fast but bad quality 131 // createLinearScanRegisterAllocator: not so fast but good quality 132 if ((pConfig.getOptimizationLevel() == llvm::CodeGenOpt::None)) { 133 llvm::RegisterRegAlloc::setDefault(llvm::createFastRegisterAllocator); 134 } else { 135 llvm::RegisterRegAlloc::setDefault(llvm::createGreedyRegisterAllocator); 136 } 137 138 return kSuccess; 139} 140 141Compiler::~Compiler() { 142 delete mTarget; 143} 144 145 146enum Compiler::ErrorCode Compiler::runPasses(Script &pScript, 147 llvm::raw_pwrite_stream &pResult) { 148 // Pass manager for link-time optimization 149 llvm::legacy::PassManager passes; 150 151 // Empty MCContext. 152 llvm::MCContext *mc_context = nullptr; 153 154 passes.add(createTargetTransformInfoWrapperPass(mTarget->getTargetIRAnalysis())); 155 156 // Add our custom passes. 157 if (!addCustomPasses(pScript, passes)) { 158 return kErrCustomPasses; 159 } 160 161 if (mTarget->getOptLevel() == llvm::CodeGenOpt::None) { 162 passes.add(llvm::createGlobalOptimizerPass()); 163 passes.add(llvm::createConstantMergePass()); 164 165 } else { 166 // FIXME: Figure out which passes should be executed. 167 llvm::PassManagerBuilder Builder; 168 Builder.Inliner = llvm::createFunctionInliningPass(); 169 Builder.populateLTOPassManager(passes); 170 171 /* FIXME: Reenable autovectorization after rebase. 172 bug 19324423 173 // Add vectorization passes after LTO passes are in 174 // additional flag: -unroll-runtime 175 passes.add(llvm::createLoopUnrollPass(-1, 16, 0, 1)); 176 // Need to pass appropriate flags here: -scalarize-load-store 177 passes.add(llvm::createScalarizerPass()); 178 passes.add(llvm::createCFGSimplificationPass()); 179 passes.add(llvm::createScopedNoAliasAAPass()); 180 passes.add(llvm::createScalarEvolutionAliasAnalysisPass()); 181 // additional flags: -slp-vectorize-hor -slp-vectorize-hor-store (unnecessary?) 182 passes.add(llvm::createSLPVectorizerPass()); 183 passes.add(llvm::createDeadCodeEliminationPass()); 184 passes.add(llvm::createInstructionCombiningPass()); 185 */ 186 } 187 188 // These passes have to come after LTO, since we don't want to examine 189 // functions that are never actually called. 190 if (!addPostLTOCustomPasses(passes)) { 191 return kErrCustomPasses; 192 } 193 194 // RSEmbedInfoPass needs to come after we have scanned for non-threadable 195 // functions. 196 // Script passed to RSCompiler must be a RSScript. 197 RSScript &script = static_cast<RSScript &>(pScript); 198 if (script.getEmbedInfo()) 199 passes.add(createRSEmbedInfoPass()); 200 201 // Add passes to the pass manager to emit machine code through MC layer. 202 if (mTarget->addPassesToEmitMC(passes, mc_context, pResult, 203 /* DisableVerify */false)) { 204 return kPrepareCodeGenPass; 205 } 206 207 // Execute the passes. 208 passes.run(pScript.getSource().getModule()); 209 210 return kSuccess; 211} 212 213enum Compiler::ErrorCode Compiler::compile(Script &pScript, 214 llvm::raw_pwrite_stream &pResult, 215 llvm::raw_ostream *IRStream) { 216 llvm::Module &module = pScript.getSource().getModule(); 217 enum ErrorCode err; 218 219 if (mTarget == nullptr) { 220 return kErrNoTargetMachine; 221 } 222 223 const std::string &triple = module.getTargetTriple(); 224 const llvm::DataLayout *dl = getTargetMachine().getDataLayout(); 225 unsigned int pointerSize = dl->getPointerSizeInBits(); 226 if (triple == "armv7-none-linux-gnueabi") { 227 if (pointerSize != 32) { 228 return kErrInvalidSource; 229 } 230 } else if (triple == "aarch64-none-linux-gnueabi") { 231 if (pointerSize != 64) { 232 return kErrInvalidSource; 233 } 234 } else { 235 return kErrInvalidSource; 236 } 237 238 // Materialize the bitcode module. 239 if (module.getMaterializer() != nullptr) { 240 // A module with non-null materializer means that it is a lazy-load module. 241 // Materialize it now via invoking MaterializeAllPermanently(). This 242 // function returns false when the materialization is successful. 243 std::error_code ec = module.materializeAllPermanently(); 244 if (ec) { 245 ALOGE("Failed to materialize the module `%s'! (%s)", 246 module.getModuleIdentifier().c_str(), ec.message().c_str()); 247 return kErrMaterialization; 248 } 249 } 250 251 if ((err = runPasses(pScript, pResult)) != kSuccess) { 252 return err; 253 } 254 255 if (IRStream) { 256 *IRStream << module; 257 } 258 259 return kSuccess; 260} 261 262enum Compiler::ErrorCode Compiler::compile(Script &pScript, 263 OutputFile &pResult, 264 llvm::raw_ostream *IRStream) { 265 // Check the state of the specified output file. 266 if (pResult.hasError()) { 267 return kErrInvalidOutputFileState; 268 } 269 270 // Open the output file decorated in llvm::raw_ostream. 271 llvm::raw_pwrite_stream *out = pResult.dup(); 272 if (out == nullptr) { 273 return kErrPrepareOutput; 274 } 275 276 // Delegate the request. 277 enum Compiler::ErrorCode err = compile(pScript, *out, IRStream); 278 279 // Close the output before return. 280 delete out; 281 282 return err; 283} 284 285bool Compiler::addInternalizeSymbolsPass(Script &pScript, llvm::legacy::PassManager &pPM) { 286 // Add a pass to internalize the symbols that don't need to have global 287 // visibility. 288 RSScript &script = static_cast<RSScript &>(pScript); 289 llvm::Module &module = script.getSource().getModule(); 290 bcinfo::MetadataExtractor me(&module); 291 if (!me.extract()) { 292 bccAssert(false && "Could not extract metadata for module!"); 293 return false; 294 } 295 296 // The vector contains the symbols that should not be internalized. 297 std::vector<const char *> export_symbols; 298 299 const char *sf[] = { 300 kRoot, // Graphics drawing function or compute kernel. 301 kInit, // Initialization routine called implicitly on startup. 302 kRsDtor, // Static global destructor for a script instance. 303 kRsInfo, // Variable containing string of RS metadata info. 304 kRsGlobalEntries, // Optional number of global variables. 305 kRsGlobalNames, // Optional global variable name info. 306 kRsGlobalAddresses, // Optional global variable address info. 307 kRsGlobalSizes, // Optional global variable size info. 308 kRsGlobalProperties, // Optional global variable properties. 309 nullptr // Must be nullptr-terminated. 310 }; 311 const char **special_functions = sf; 312 // Special RS functions should always be global symbols. 313 while (*special_functions != nullptr) { 314 export_symbols.push_back(*special_functions); 315 special_functions++; 316 } 317 318 // Visibility of symbols appeared in rs_export_var and rs_export_func should 319 // also be preserved. 320 size_t exportVarCount = me.getExportVarCount(); 321 size_t exportFuncCount = me.getExportFuncCount(); 322 size_t exportForEachCount = me.getExportForEachSignatureCount(); 323 const char **exportVarNameList = me.getExportVarNameList(); 324 const char **exportFuncNameList = me.getExportFuncNameList(); 325 const char **exportForEachNameList = me.getExportForEachNameList(); 326 size_t i; 327 328 for (i = 0; i < exportVarCount; ++i) { 329 export_symbols.push_back(exportVarNameList[i]); 330 } 331 332 for (i = 0; i < exportFuncCount; ++i) { 333 export_symbols.push_back(exportFuncNameList[i]); 334 } 335 336 // Expanded foreach functions should not be internalized, too. 337 // expanded_foreach_funcs keeps the .expand version of the kernel names 338 // around until createInternalizePass() is finished making its own 339 // copy of the visible symbols. 340 std::vector<std::string> expanded_foreach_funcs; 341 for (i = 0; i < exportForEachCount; ++i) { 342 expanded_foreach_funcs.push_back( 343 std::string(exportForEachNameList[i]) + ".expand"); 344 } 345 346 for (i = 0; i < exportForEachCount; i++) { 347 export_symbols.push_back(expanded_foreach_funcs[i].c_str()); 348 } 349 350 pPM.add(llvm::createInternalizePass(export_symbols)); 351 352 return true; 353} 354 355bool Compiler::addInvokeHelperPass(llvm::legacy::PassManager &pPM) { 356 llvm::Triple arch(getTargetMachine().getTargetTriple()); 357 if (arch.isArch64Bit()) { 358 pPM.add(createRSInvokeHelperPass()); 359 } 360 return true; 361} 362 363bool Compiler::addExpandForEachPass(Script &pScript, llvm::legacy::PassManager &pPM) { 364 // Expand ForEach on CPU path to reduce launch overhead. 365 bool pEnableStepOpt = true; 366 pPM.add(createRSForEachExpandPass(pEnableStepOpt)); 367 368 return true; 369} 370 371bool Compiler::addGlobalInfoPass(Script &pScript, llvm::legacy::PassManager &pPM) { 372 // Add additional information about RS global variables inside the Module. 373 RSScript &script = static_cast<RSScript &>(pScript); 374 if (script.getEmbedGlobalInfo()) { 375 pPM.add(createRSGlobalInfoPass(script.getEmbedGlobalInfoSkipConstant())); 376 } 377 378 return true; 379} 380 381bool Compiler::addInvariantPass(llvm::legacy::PassManager &pPM) { 382 // Mark Loads from RsExpandKernelDriverInfo as "load.invariant". 383 // Should run after ExpandForEach and before inlining. 384 pPM.add(createRSInvariantPass()); 385 386 return true; 387} 388 389bool Compiler::addCustomPasses(Script &pScript, llvm::legacy::PassManager &pPM) { 390 if (!addInvokeHelperPass(pPM)) 391 return false; 392 393 if (!addExpandForEachPass(pScript, pPM)) 394 return false; 395 396 if (!addInvariantPass(pPM)) 397 return false; 398 399 if (!addInternalizeSymbolsPass(pScript, pPM)) 400 return false; 401 402 if (!addGlobalInfoPass(pScript, pPM)) 403 return false; 404 405 return true; 406} 407 408bool Compiler::addPostLTOCustomPasses(llvm::legacy::PassManager &pPM) { 409 // Add pass to correct calling convention for X86-64. 410 llvm::Triple arch(getTargetMachine().getTargetTriple()); 411 if (arch.getArch() == llvm::Triple::x86_64) 412 pPM.add(createRSX86_64CallConvPass()); 413 414 // Add pass to check for illegal function calls. 415 pPM.add(createRSScreenFunctionsPass()); 416 417 // Add pass to mark script as threadable. 418 pPM.add(createRSIsThreadablePass()); 419 420 return true; 421} 422