slang_backend.cpp revision 352e62d65caa48251d9c69cc8b5a82ddfea6b014
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 "slang_backend.h" 18 19#include <string> 20#include <vector> 21#include <iostream> 22 23#include "clang/AST/ASTContext.h" 24#include "clang/AST/Attr.h" 25#include "clang/AST/Decl.h" 26#include "clang/AST/DeclGroup.h" 27#include "clang/AST/RecordLayout.h" 28 29#include "clang/Basic/Diagnostic.h" 30#include "clang/Basic/TargetInfo.h" 31#include "clang/Basic/TargetOptions.h" 32 33#include "clang/CodeGen/ModuleBuilder.h" 34 35#include "clang/Frontend/CodeGenOptions.h" 36#include "clang/Frontend/FrontendDiagnostic.h" 37 38#include "llvm/ADT/Twine.h" 39#include "llvm/ADT/StringExtras.h" 40 41#include "llvm/Bitcode/ReaderWriter.h" 42 43#include "llvm/CodeGen/RegAllocRegistry.h" 44#include "llvm/CodeGen/SchedulerRegistry.h" 45 46#include "llvm/IR/Constant.h" 47#include "llvm/IR/Constants.h" 48#include "llvm/IR/DataLayout.h" 49#include "llvm/IR/DebugLoc.h" 50#include "llvm/IR/DerivedTypes.h" 51#include "llvm/IR/Function.h" 52#include "llvm/IR/IRBuilder.h" 53#include "llvm/IR/IRPrintingPasses.h" 54#include "llvm/IR/LLVMContext.h" 55#include "llvm/IR/Metadata.h" 56#include "llvm/IR/Module.h" 57 58#include "llvm/Transforms/IPO/PassManagerBuilder.h" 59 60#include "llvm/Target/TargetMachine.h" 61#include "llvm/Target/TargetOptions.h" 62#include "llvm/Support/TargetRegistry.h" 63 64#include "llvm/MC/SubtargetFeature.h" 65 66#include "slang_assert.h" 67#include "slang.h" 68#include "slang_bitcode_gen.h" 69#include "slang_rs_context.h" 70#include "slang_rs_export_foreach.h" 71#include "slang_rs_export_func.h" 72#include "slang_rs_export_reduce.h" 73#include "slang_rs_export_type.h" 74#include "slang_rs_export_var.h" 75#include "slang_rs_metadata.h" 76 77#include "rs_cc_options.h" 78 79#include "strip_unknown_attributes.h" 80 81namespace slang { 82 83void Backend::CreateFunctionPasses() { 84 if (!mPerFunctionPasses) { 85 mPerFunctionPasses = new llvm::legacy::FunctionPassManager(mpModule); 86 87 llvm::PassManagerBuilder PMBuilder; 88 PMBuilder.OptLevel = mCodeGenOpts.OptimizationLevel; 89 PMBuilder.populateFunctionPassManager(*mPerFunctionPasses); 90 } 91} 92 93void Backend::CreateModulePasses() { 94 if (!mPerModulePasses) { 95 mPerModulePasses = new llvm::legacy::PassManager(); 96 97 llvm::PassManagerBuilder PMBuilder; 98 PMBuilder.OptLevel = mCodeGenOpts.OptimizationLevel; 99 PMBuilder.SizeLevel = mCodeGenOpts.OptimizeSize; 100 PMBuilder.DisableUnitAtATime = 0; // TODO Pirama confirm if this is right 101 102 if (mCodeGenOpts.UnrollLoops) { 103 PMBuilder.DisableUnrollLoops = 0; 104 } else { 105 PMBuilder.DisableUnrollLoops = 1; 106 } 107 108 PMBuilder.populateModulePassManager(*mPerModulePasses); 109 // Add a pass to strip off unknown/unsupported attributes. 110 mPerModulePasses->add(createStripUnknownAttributesPass()); 111 } 112} 113 114bool Backend::CreateCodeGenPasses() { 115 if ((mOT != Slang::OT_Assembly) && (mOT != Slang::OT_Object)) 116 return true; 117 118 // Now we add passes for code emitting 119 if (mCodeGenPasses) { 120 return true; 121 } else { 122 mCodeGenPasses = new llvm::legacy::FunctionPassManager(mpModule); 123 } 124 125 // Create the TargetMachine for generating code. 126 std::string Triple = mpModule->getTargetTriple(); 127 128 std::string Error; 129 const llvm::Target* TargetInfo = 130 llvm::TargetRegistry::lookupTarget(Triple, Error); 131 if (TargetInfo == nullptr) { 132 mDiagEngine.Report(clang::diag::err_fe_unable_to_create_target) << Error; 133 return false; 134 } 135 136 // Target Machine Options 137 llvm::TargetOptions Options; 138 139 // Use soft-float ABI for ARM (which is the target used by Slang during code 140 // generation). Codegen still uses hardware FPU by default. To use software 141 // floating point, add 'soft-float' feature to FeaturesStr below. 142 Options.FloatABIType = llvm::FloatABI::Soft; 143 144 // BCC needs all unknown symbols resolved at compilation time. So we don't 145 // need any relocation model. 146 llvm::Reloc::Model RM = llvm::Reloc::Static; 147 148 // This is set for the linker (specify how large of the virtual addresses we 149 // can access for all unknown symbols.) 150 llvm::CodeModel::Model CM; 151 if (mpModule->getDataLayout().getPointerSize() == 4) { 152 CM = llvm::CodeModel::Small; 153 } else { 154 // The target may have pointer size greater than 32 (e.g. x86_64 155 // architecture) may need large data address model 156 CM = llvm::CodeModel::Medium; 157 } 158 159 // Setup feature string 160 std::string FeaturesStr; 161 if (mTargetOpts.CPU.size() || mTargetOpts.Features.size()) { 162 llvm::SubtargetFeatures Features; 163 164 for (std::vector<std::string>::const_iterator 165 I = mTargetOpts.Features.begin(), E = mTargetOpts.Features.end(); 166 I != E; 167 I++) 168 Features.AddFeature(*I); 169 170 FeaturesStr = Features.getString(); 171 } 172 173 llvm::TargetMachine *TM = 174 TargetInfo->createTargetMachine(Triple, mTargetOpts.CPU, FeaturesStr, 175 Options, RM, CM); 176 177 // Register allocation policy: 178 // createFastRegisterAllocator: fast but bad quality 179 // createGreedyRegisterAllocator: not so fast but good quality 180 llvm::RegisterRegAlloc::setDefault((mCodeGenOpts.OptimizationLevel == 0) ? 181 llvm::createFastRegisterAllocator : 182 llvm::createGreedyRegisterAllocator); 183 184 llvm::CodeGenOpt::Level OptLevel = llvm::CodeGenOpt::Default; 185 if (mCodeGenOpts.OptimizationLevel == 0) { 186 OptLevel = llvm::CodeGenOpt::None; 187 } else if (mCodeGenOpts.OptimizationLevel == 3) { 188 OptLevel = llvm::CodeGenOpt::Aggressive; 189 } 190 191 llvm::TargetMachine::CodeGenFileType CGFT = 192 llvm::TargetMachine::CGFT_AssemblyFile; 193 if (mOT == Slang::OT_Object) { 194 CGFT = llvm::TargetMachine::CGFT_ObjectFile; 195 } 196 if (TM->addPassesToEmitFile(*mCodeGenPasses, mBufferOutStream, 197 CGFT, OptLevel)) { 198 mDiagEngine.Report(clang::diag::err_fe_unable_to_interface_with_target); 199 return false; 200 } 201 202 return true; 203} 204 205Backend::Backend(RSContext *Context, clang::DiagnosticsEngine *DiagEngine, 206 const RSCCOptions &Opts, 207 const clang::HeaderSearchOptions &HeaderSearchOpts, 208 const clang::PreprocessorOptions &PreprocessorOpts, 209 const clang::CodeGenOptions &CodeGenOpts, 210 const clang::TargetOptions &TargetOpts, PragmaList *Pragmas, 211 llvm::raw_ostream *OS, Slang::OutputType OT, 212 clang::SourceManager &SourceMgr, bool AllowRSPrefix, 213 bool IsFilterscript) 214 : ASTConsumer(), mTargetOpts(TargetOpts), mpModule(nullptr), mpOS(OS), 215 mOT(OT), mGen(nullptr), mPerFunctionPasses(nullptr), 216 mPerModulePasses(nullptr), mCodeGenPasses(nullptr), 217 mBufferOutStream(*mpOS), mContext(Context), 218 mSourceMgr(SourceMgr), mASTPrint(Opts.mASTPrint), mAllowRSPrefix(AllowRSPrefix), 219 mIsFilterscript(IsFilterscript), mExportVarMetadata(nullptr), 220 mExportFuncMetadata(nullptr), mExportForEachNameMetadata(nullptr), 221 mExportForEachSignatureMetadata(nullptr), 222 mExportReduceMetadata(nullptr), 223 mExportTypeMetadata(nullptr), mRSObjectSlotsMetadata(nullptr), 224 mRefCount(mContext->getASTContext()), 225 mASTChecker(Context, Context->getTargetAPI(), IsFilterscript), 226 mForEachHandler(Context), 227 mLLVMContext(slang::getGlobalLLVMContext()), mDiagEngine(*DiagEngine), 228 mCodeGenOpts(CodeGenOpts), mPragmas(Pragmas) { 229 mGen = CreateLLVMCodeGen(mDiagEngine, "", HeaderSearchOpts, PreprocessorOpts, 230 mCodeGenOpts, mLLVMContext); 231} 232 233void Backend::Initialize(clang::ASTContext &Ctx) { 234 mGen->Initialize(Ctx); 235 236 mpModule = mGen->GetModule(); 237} 238 239void Backend::HandleTranslationUnit(clang::ASTContext &Ctx) { 240 HandleTranslationUnitPre(Ctx); 241 242 if (mASTPrint) 243 Ctx.getTranslationUnitDecl()->dump(); 244 245 mGen->HandleTranslationUnit(Ctx); 246 247 // Here, we complete a translation unit (whole translation unit is now in LLVM 248 // IR). Now, interact with LLVM backend to generate actual machine code (asm 249 // or machine code, whatever.) 250 251 // Silently ignore if we weren't initialized for some reason. 252 if (!mpModule) 253 return; 254 255 llvm::Module *M = mGen->ReleaseModule(); 256 if (!M) { 257 // The module has been released by IR gen on failures, do not double free. 258 mpModule = nullptr; 259 return; 260 } 261 262 slangAssert(mpModule == M && 263 "Unexpected module change during LLVM IR generation"); 264 265 // Insert #pragma information into metadata section of module 266 if (!mPragmas->empty()) { 267 llvm::NamedMDNode *PragmaMetadata = 268 mpModule->getOrInsertNamedMetadata(Slang::PragmaMetadataName); 269 for (PragmaList::const_iterator I = mPragmas->begin(), E = mPragmas->end(); 270 I != E; 271 I++) { 272 llvm::SmallVector<llvm::Metadata*, 2> Pragma; 273 // Name goes first 274 Pragma.push_back(llvm::MDString::get(mLLVMContext, I->first)); 275 // And then value 276 Pragma.push_back(llvm::MDString::get(mLLVMContext, I->second)); 277 278 // Create MDNode and insert into PragmaMetadata 279 PragmaMetadata->addOperand( 280 llvm::MDNode::get(mLLVMContext, Pragma)); 281 } 282 } 283 284 HandleTranslationUnitPost(mpModule); 285 286 // Create passes for optimization and code emission 287 288 // Create and run per-function passes 289 CreateFunctionPasses(); 290 if (mPerFunctionPasses) { 291 mPerFunctionPasses->doInitialization(); 292 293 for (llvm::Module::iterator I = mpModule->begin(), E = mpModule->end(); 294 I != E; 295 I++) 296 if (!I->isDeclaration()) 297 mPerFunctionPasses->run(*I); 298 299 mPerFunctionPasses->doFinalization(); 300 } 301 302 // Create and run module passes 303 CreateModulePasses(); 304 if (mPerModulePasses) 305 mPerModulePasses->run(*mpModule); 306 307 switch (mOT) { 308 case Slang::OT_Assembly: 309 case Slang::OT_Object: { 310 if (!CreateCodeGenPasses()) 311 return; 312 313 mCodeGenPasses->doInitialization(); 314 315 for (llvm::Module::iterator I = mpModule->begin(), E = mpModule->end(); 316 I != E; 317 I++) 318 if (!I->isDeclaration()) 319 mCodeGenPasses->run(*I); 320 321 mCodeGenPasses->doFinalization(); 322 break; 323 } 324 case Slang::OT_LLVMAssembly: { 325 llvm::legacy::PassManager *LLEmitPM = new llvm::legacy::PassManager(); 326 LLEmitPM->add(llvm::createPrintModulePass(mBufferOutStream)); 327 LLEmitPM->run(*mpModule); 328 break; 329 } 330 case Slang::OT_Bitcode: { 331 writeBitcode(mBufferOutStream, *mpModule, getTargetAPI(), 332 mCodeGenOpts.OptimizationLevel, mCodeGenOpts.getDebugInfo()); 333 break; 334 } 335 case Slang::OT_Nothing: { 336 return; 337 } 338 default: { 339 slangAssert(false && "Unknown output type"); 340 } 341 } 342} 343 344// Insert explicit padding fields into struct to follow the current layout. 345// 346// A similar algorithm is present in PadHelperFunctionStruct(). 347void Backend::PadStruct(clang::RecordDecl* RD) { 348 // Example of padding: 349 // 350 // // ORIGINAL CODE // TRANSFORMED CODE 351 // struct foo { struct foo { 352 // int a; int a; 353 // // 4 bytes of padding char <RS_PADDING_FIELD_NAME>[4]; 354 // long b; long b; 355 // int c; int c; 356 // // 4 bytes of (tail) padding char <RS_PADDING_FIELD_NAME>[4]; 357 // }; }; 358 359 // We collect all of RD's fields in a vector FieldsInfo. We 360 // represent tail padding as an entry in the FieldsInfo vector with a 361 // null FieldDecl. 362 typedef std::pair<size_t, clang::FieldDecl*> FieldInfoType; // (pre-field padding bytes, field) 363 std::vector<FieldInfoType> FieldsInfo; 364 365 // RenderScript is C99-based, so we only expect to see fields. We 366 // could iterate over fields, but instead let's iterate over 367 // everything, to verify that there are only fields. 368 for (clang::Decl* D : RD->decls()) { 369 clang::FieldDecl* FD = clang::dyn_cast<clang::FieldDecl>(D); 370 slangAssert(FD && "found a non field declaration within a struct"); 371 FieldsInfo.push_back(std::make_pair(size_t(0), FD)); 372 } 373 374 clang::ASTContext& ASTC = mContext->getASTContext(); 375 376 // ASTContext caches record layout. We may transform the record in a way 377 // that would render this cached information incorrect. clang does 378 // not provide any way to invalidate this cached information. We 379 // take the following approach: 380 // 381 // 1. ASSUME that record layout has not yet been computed for RD. 382 // 383 // 2. Create a temporary clone of RD, and compute its layout. 384 // ASSUME that we know how to clone RD in a way that copies all the 385 // properties that are relevant to its layout. 386 // 387 // 3. Use the layout information from the temporary clone to 388 // transform RD. 389 // 390 // NOTE: ASTContext also caches TypeInfo (see 391 // ASTContext::getTypeInfo()). ASSUME that inserting padding 392 // fields doesn't change the type in any way that affects 393 // TypeInfo. 394 // 395 // NOTE: A RecordType knows its associated RecordDecl -- so even 396 // while we're manipulating RD, the associated RecordType 397 // still recognizes RD as its RecordDecl. ASSUME that we 398 // don't do anything during our manipulation that would cause 399 // the RecordType to be followed to RD while RD is in a 400 // partially transformed state. 401 402 // The assumptions above may be brittle, and if they are incorrect, 403 // we may get mysterious failures. 404 405 // create a temporary clone 406 clang::RecordDecl* RDForLayout = 407 clang::RecordDecl::Create(ASTC, clang::TTK_Struct, RD->getDeclContext(), 408 clang::SourceLocation(), clang::SourceLocation(), 409 nullptr /* IdentifierInfo */); 410 RDForLayout->startDefinition(); 411 RDForLayout->setTypeForDecl(RD->getTypeForDecl()); 412 if (RD->hasAttrs()) 413 RDForLayout->setAttrs(RD->getAttrs()); 414 RDForLayout->completeDefinition(); 415 416 // move all fields from RD to RDForLayout 417 for (const auto &info : FieldsInfo) { 418 RD->removeDecl(info.second); 419 info.second->setLexicalDeclContext(RDForLayout); 420 RDForLayout->addDecl(info.second); 421 } 422 423 const clang::ASTRecordLayout& RL = ASTC.getASTRecordLayout(RDForLayout); 424 425 // An exportable type cannot contain a bitfield. However, it's 426 // possible that this current type might have a bitfield and yet 427 // share a common initial sequence with an exportable type, so even 428 // if the current type has a bitfield, the current type still 429 // needs to have explicit padding inserted (in case the two types 430 // under discussion are members of a union). We don't need to 431 // insert any padding after the bitfield, however, because that 432 // would be beyond the common initial sequence. 433 bool foundBitField = false; 434 435 // Is there any padding in this struct? 436 bool foundPadding = false; 437 438 unsigned fieldNo = 0; 439 uint64_t fieldPrePaddingOffset = 0; // byte offset of pre-field padding within struct 440 for (auto &info : FieldsInfo) { 441 const clang::FieldDecl* FD = info.second; 442 443 if ((foundBitField = FD->isBitField())) 444 break; 445 446 const uint64_t fieldOffset = RL.getFieldOffset(fieldNo) >> 3; 447 const size_t prePadding = fieldOffset - fieldPrePaddingOffset; 448 foundPadding |= (prePadding != 0); 449 info.first = prePadding; 450 451 // get ready for the next field 452 // 453 // assumes that getTypeSize() is the storage size of the Type -- for example, 454 // that it includes a struct's tail padding (if any) 455 // 456 fieldPrePaddingOffset = fieldOffset + (ASTC.getTypeSize(FD->getType()) >> 3); 457 ++fieldNo; 458 } 459 460 if (!foundBitField) { 461 // In order to ensure that the front end (including reflected 462 // code) and back end agree on struct size (not just field 463 // offsets) we may need to add explicit tail padding, just as we'e 464 // added explicit padding between fields. 465 slangAssert(RL.getSize().getQuantity() >= fieldPrePaddingOffset); 466 if (const size_t tailPadding = RL.getSize().getQuantity() - fieldPrePaddingOffset) { 467 foundPadding = true; 468 FieldsInfo.push_back(std::make_pair(tailPadding, nullptr)); 469 } 470 } 471 472 if (false /* change to "true" for extra debugging output */) { 473 if (foundPadding) { 474 std::cout << "PadStruct(" << RD->getNameAsString() << "):" << std::endl; 475 for (const auto &info : FieldsInfo) 476 std::cout << " " << info.first << ", " << (info.second ? info.second->getNameAsString() : "<tail>") << std::endl; 477 } 478 } 479 480 if (foundPadding && Slang::IsLocInRSHeaderFile(RD->getLocation(), mSourceMgr)) { 481 mContext->ReportError(RD->getLocation(), "system structure contains padding: '%0'") 482 << RD->getName(); 483 } 484 485 // now move fields from RDForLayout to RD, and add any necessary 486 // padding fields 487 const clang::QualType byteType = ASTC.getIntTypeForBitwidth(8, false /* not signed */); 488 clang::IdentifierInfo* const paddingIdentifierInfo = &ASTC.Idents.get(RS_PADDING_FIELD_NAME); 489 for (const auto &info : FieldsInfo) { 490 if (info.first != 0) { 491 // Create a padding field: "char <RS_PADDING_FIELD_NAME>[<info.first>];" 492 493 // TODO: Do we need to do anything else to keep this field from being shown in debugger? 494 // There's no source location, and the field is marked as implicit. 495 const clang::QualType paddingType = 496 ASTC.getConstantArrayType(byteType, 497 llvm::APInt(sizeof(info.first) << 3, info.first), 498 clang::ArrayType::Normal, 0 /* IndexTypeQuals */); 499 clang::FieldDecl* const FD = 500 clang::FieldDecl::Create(ASTC, RD, clang::SourceLocation(), clang::SourceLocation(), 501 paddingIdentifierInfo, 502 paddingType, 503 nullptr, // TypeSourceInfo* 504 nullptr, // BW (bitwidth) 505 false, // Mutable = false 506 clang::ICIS_NoInit); 507 FD->setImplicit(true); 508 RD->addDecl(FD); 509 } 510 if (info.second != nullptr) { 511 RDForLayout->removeDecl(info.second); 512 info.second->setLexicalDeclContext(RD); 513 RD->addDecl(info.second); 514 } 515 } 516 517 // There does not appear to be any safe way to delete a RecordDecl 518 // -- for example, there is no RecordDecl destructor to invalidate 519 // cached record layout, and if we were to get unlucky, some future 520 // RecordDecl could be allocated in the same place as a deleted 521 // RDForLayout and "inherit" the cached record layout from 522 // RDForLayout. 523} 524 525void Backend::HandleTagDeclDefinition(clang::TagDecl *D) { 526 // we want to insert explicit padding fields into structs per http://b/29154200 and http://b/28070272 527 switch (D->getTagKind()) { 528 case clang::TTK_Struct: 529 PadStruct(llvm::cast<clang::RecordDecl>(D)); 530 break; 531 532 case clang::TTK_Union: 533 // cannot be part of an exported type 534 break; 535 536 case clang::TTK_Enum: 537 // a scalar 538 break; 539 540 case clang::TTK_Class: 541 case clang::TTK_Interface: 542 default: 543 slangAssert(false && "Unexpected TagTypeKind"); 544 break; 545 } 546 mGen->HandleTagDeclDefinition(D); 547} 548 549void Backend::CompleteTentativeDefinition(clang::VarDecl *D) { 550 mGen->CompleteTentativeDefinition(D); 551} 552 553Backend::~Backend() { 554 delete mpModule; 555 delete mGen; 556 delete mPerFunctionPasses; 557 delete mPerModulePasses; 558 delete mCodeGenPasses; 559} 560 561// 1) Add zero initialization of local RS object types 562void Backend::AnnotateFunction(clang::FunctionDecl *FD) { 563 if (FD && 564 FD->hasBody() && 565 !FD->isImplicit() && 566 !Slang::IsLocInRSHeaderFile(FD->getLocation(), mSourceMgr)) { 567 mRefCount.Init(); 568 mRefCount.SetDeclContext(FD); 569 mRefCount.Visit(FD->getBody()); 570 } 571} 572 573bool Backend::HandleTopLevelDecl(clang::DeclGroupRef D) { 574 // Find and remember the types for rs_allocation and rs_script_call_t so 575 // they can be used later for translating rsForEach() calls. 576 for (clang::DeclGroupRef::iterator I = D.begin(), E = D.end(); 577 (mContext->getAllocationType().isNull() || 578 mContext->getScriptCallType().isNull()) && 579 I != E; I++) { 580 if (clang::TypeDecl* TD = llvm::dyn_cast<clang::TypeDecl>(*I)) { 581 clang::StringRef TypeName = TD->getName(); 582 if (TypeName.equals("rs_allocation")) { 583 mContext->setAllocationType(TD); 584 } else if (TypeName.equals("rs_script_call_t")) { 585 mContext->setScriptCallType(TD); 586 } 587 } 588 } 589 590 // Disallow user-defined functions with prefix "rs" 591 if (!mAllowRSPrefix) { 592 // Iterate all function declarations in the program. 593 for (clang::DeclGroupRef::iterator I = D.begin(), E = D.end(); 594 I != E; I++) { 595 clang::FunctionDecl *FD = llvm::dyn_cast<clang::FunctionDecl>(*I); 596 if (FD == nullptr) 597 continue; 598 if (!FD->getName().startswith("rs")) // Check prefix 599 continue; 600 if (!Slang::IsLocInRSHeaderFile(FD->getLocation(), mSourceMgr)) 601 mContext->ReportError(FD->getLocation(), 602 "invalid function name prefix, " 603 "\"rs\" is reserved: '%0'") 604 << FD->getName(); 605 } 606 } 607 608 for (clang::DeclGroupRef::iterator I = D.begin(), E = D.end(); I != E; I++) { 609 clang::FunctionDecl *FD = llvm::dyn_cast<clang::FunctionDecl>(*I); 610 if (FD) { 611 // Handle forward reference from pragma (see 612 // RSReducePragmaHandler::HandlePragma for backward reference). 613 mContext->markUsedByReducePragma(FD, RSContext::CheckNameYes); 614 if (FD->isGlobal()) { 615 // Check that we don't have any array parameters being misinterpreted as 616 // kernel pointers due to the C type system's array to pointer decay. 617 size_t numParams = FD->getNumParams(); 618 for (size_t i = 0; i < numParams; i++) { 619 const clang::ParmVarDecl *PVD = FD->getParamDecl(i); 620 clang::QualType QT = PVD->getOriginalType(); 621 if (QT->isArrayType()) { 622 mContext->ReportError( 623 PVD->getTypeSpecStartLoc(), 624 "exported function parameters may not have array type: %0") 625 << QT; 626 } 627 } 628 AnnotateFunction(FD); 629 } 630 } 631 632 if (getTargetAPI() >= SLANG_FEATURE_SINGLE_SOURCE_API) { 633 if (FD && FD->hasBody() && !FD->isImplicit() && 634 !Slang::IsLocInRSHeaderFile(FD->getLocation(), mSourceMgr)) { 635 if (FD->hasAttr<clang::RenderScriptKernelAttr>()) { 636 // Log functions with attribute "kernel" by their names, and assign 637 // them slot numbers. Any other function cannot be used in a 638 // rsForEach() or rsForEachWithOptions() call, including old-style 639 // kernel functions which are defined without the "kernel" attribute. 640 mContext->addForEach(FD); 641 } 642 // Look for any kernel launch calls and translate them into using the 643 // internal API. 644 // Report a compiler error on kernel launches inside a kernel. 645 mForEachHandler.handleForEachCalls(FD, getTargetAPI()); 646 } 647 } 648 } 649 650 return mGen->HandleTopLevelDecl(D); 651} 652 653void Backend::HandleTranslationUnitPre(clang::ASTContext &C) { 654 clang::TranslationUnitDecl *TUDecl = C.getTranslationUnitDecl(); 655 656 if (!mContext->processReducePragmas(this)) 657 return; 658 659 // If we have an invalid RS/FS AST, don't check further. 660 if (!mASTChecker.Validate()) { 661 return; 662 } 663 664 if (mIsFilterscript) { 665 mContext->addPragma("rs_fp_relaxed", ""); 666 } 667 668 int version = mContext->getVersion(); 669 if (version == 0) { 670 // Not setting a version is an error 671 mDiagEngine.Report( 672 mSourceMgr.getLocForEndOfFile(mSourceMgr.getMainFileID()), 673 mDiagEngine.getCustomDiagID( 674 clang::DiagnosticsEngine::Error, 675 "missing pragma for version in source file")); 676 } else { 677 slangAssert(version == 1); 678 } 679 680 if (mContext->getReflectJavaPackageName().empty()) { 681 mDiagEngine.Report( 682 mSourceMgr.getLocForEndOfFile(mSourceMgr.getMainFileID()), 683 mDiagEngine.getCustomDiagID(clang::DiagnosticsEngine::Error, 684 "missing \"#pragma rs " 685 "java_package_name(com.foo.bar)\" " 686 "in source file")); 687 return; 688 } 689 690 // Create a static global destructor if necessary (to handle RS object 691 // runtime cleanup). 692 clang::FunctionDecl *FD = mRefCount.CreateStaticGlobalDtor(); 693 if (FD) { 694 HandleTopLevelDecl(clang::DeclGroupRef(FD)); 695 } 696 697 // Process any static function declarations 698 for (clang::DeclContext::decl_iterator I = TUDecl->decls_begin(), 699 E = TUDecl->decls_end(); I != E; I++) { 700 if ((I->getKind() >= clang::Decl::firstFunction) && 701 (I->getKind() <= clang::Decl::lastFunction)) { 702 clang::FunctionDecl *FD = llvm::dyn_cast<clang::FunctionDecl>(*I); 703 if (FD && !FD->isGlobal()) { 704 AnnotateFunction(FD); 705 } 706 } 707 } 708} 709 710/////////////////////////////////////////////////////////////////////////////// 711void Backend::dumpExportVarInfo(llvm::Module *M) { 712 int slotCount = 0; 713 if (mExportVarMetadata == nullptr) 714 mExportVarMetadata = M->getOrInsertNamedMetadata(RS_EXPORT_VAR_MN); 715 716 llvm::SmallVector<llvm::Metadata *, 2> ExportVarInfo; 717 718 // We emit slot information (#rs_object_slots) for any reference counted 719 // RS type or pointer (which can also be bound). 720 721 for (RSContext::const_export_var_iterator I = mContext->export_vars_begin(), 722 E = mContext->export_vars_end(); 723 I != E; 724 I++) { 725 const RSExportVar *EV = *I; 726 const RSExportType *ET = EV->getType(); 727 bool countsAsRSObject = false; 728 729 // Variable name 730 ExportVarInfo.push_back( 731 llvm::MDString::get(mLLVMContext, EV->getName().c_str())); 732 733 // Type name 734 switch (ET->getClass()) { 735 case RSExportType::ExportClassPrimitive: { 736 const RSExportPrimitiveType *PT = 737 static_cast<const RSExportPrimitiveType*>(ET); 738 ExportVarInfo.push_back( 739 llvm::MDString::get( 740 mLLVMContext, llvm::utostr(PT->getType()))); 741 if (PT->isRSObjectType()) { 742 countsAsRSObject = true; 743 } 744 break; 745 } 746 case RSExportType::ExportClassPointer: { 747 ExportVarInfo.push_back( 748 llvm::MDString::get( 749 mLLVMContext, ("*" + static_cast<const RSExportPointerType*>(ET) 750 ->getPointeeType()->getName()).c_str())); 751 break; 752 } 753 case RSExportType::ExportClassMatrix: { 754 ExportVarInfo.push_back( 755 llvm::MDString::get( 756 mLLVMContext, llvm::utostr( 757 /* TODO Strange value. This pushes just a number, quite 758 * different than the other cases. What is this used for? 759 * These are the metadata values that some partner drivers 760 * want to reference (for TBAA, etc.). We may want to look 761 * at whether these provide any reasonable value (or have 762 * distinct enough values to actually depend on). 763 */ 764 DataTypeRSMatrix2x2 + 765 static_cast<const RSExportMatrixType*>(ET)->getDim() - 2))); 766 break; 767 } 768 case RSExportType::ExportClassVector: 769 case RSExportType::ExportClassConstantArray: 770 case RSExportType::ExportClassRecord: { 771 ExportVarInfo.push_back( 772 llvm::MDString::get(mLLVMContext, 773 EV->getType()->getName().c_str())); 774 break; 775 } 776 } 777 778 mExportVarMetadata->addOperand( 779 llvm::MDNode::get(mLLVMContext, ExportVarInfo)); 780 ExportVarInfo.clear(); 781 782 if (mRSObjectSlotsMetadata == nullptr) { 783 mRSObjectSlotsMetadata = 784 M->getOrInsertNamedMetadata(RS_OBJECT_SLOTS_MN); 785 } 786 787 if (countsAsRSObject) { 788 mRSObjectSlotsMetadata->addOperand(llvm::MDNode::get(mLLVMContext, 789 llvm::MDString::get(mLLVMContext, llvm::utostr(slotCount)))); 790 } 791 792 slotCount++; 793 } 794} 795 796// A similar algorithm is present in Backend::PadStruct(). 797static void PadHelperFunctionStruct(llvm::Module *M, 798 llvm::StructType **paddedStructType, 799 std::vector<unsigned> *origFieldNumToPaddedFieldNum, 800 llvm::StructType *origStructType) { 801 slangAssert(origFieldNumToPaddedFieldNum->empty()); 802 origFieldNumToPaddedFieldNum->resize(2 * origStructType->getNumElements()); 803 804 llvm::LLVMContext &llvmContext = M->getContext(); 805 806 const llvm::DataLayout *DL = &M->getDataLayout(); 807 const llvm::StructLayout *SL = DL->getStructLayout(origStructType); 808 809 // Field types -- including any padding fields we need to insert. 810 std::vector<llvm::Type *> paddedFieldTypes; 811 paddedFieldTypes.reserve(2 * origStructType->getNumElements()); 812 813 // Is there any padding in this struct? 814 bool foundPadding = false; 815 816 llvm::Type *const byteType = llvm::Type::getInt8Ty(llvmContext); 817 unsigned origFieldNum = 0, paddedFieldNum = 0; 818 uint64_t fieldPrePaddingOffset = 0; // byte offset of pre-field padding within struct 819 for (llvm::Type *fieldType : origStructType->elements()) { 820 const uint64_t fieldOffset = SL->getElementOffset(origFieldNum); 821 const size_t prePadding = fieldOffset - fieldPrePaddingOffset; 822 if (prePadding != 0) { 823 foundPadding = true; 824 paddedFieldTypes.push_back(llvm::ArrayType::get(byteType, prePadding)); 825 ++paddedFieldNum; 826 } 827 paddedFieldTypes.push_back(fieldType); 828 (*origFieldNumToPaddedFieldNum)[origFieldNum] = paddedFieldNum; 829 830 // get ready for the next field 831 fieldPrePaddingOffset = fieldOffset + DL->getTypeAllocSize(fieldType); 832 ++origFieldNum; 833 ++paddedFieldNum; 834 } 835 836 // In order to ensure that the front end (including reflected code) 837 // and back end agree on struct size (not just field offsets) we may 838 // need to add explicit tail padding, just as we'e added explicit 839 // padding between fields. 840 slangAssert(SL->getSizeInBytes() >= fieldPrePaddingOffset); 841 if (const size_t tailPadding = SL->getSizeInBytes() - fieldPrePaddingOffset) { 842 foundPadding = true; 843 paddedFieldTypes.push_back(llvm::ArrayType::get(byteType, tailPadding)); 844 } 845 846 *paddedStructType = (foundPadding 847 ? llvm::StructType::get(llvmContext, paddedFieldTypes) 848 : origStructType); 849} 850 851void Backend::dumpExportFunctionInfo(llvm::Module *M) { 852 if (mExportFuncMetadata == nullptr) 853 mExportFuncMetadata = 854 M->getOrInsertNamedMetadata(RS_EXPORT_FUNC_MN); 855 856 llvm::SmallVector<llvm::Metadata *, 1> ExportFuncInfo; 857 858 for (RSContext::const_export_func_iterator 859 I = mContext->export_funcs_begin(), 860 E = mContext->export_funcs_end(); 861 I != E; 862 I++) { 863 const RSExportFunc *EF = *I; 864 865 // Function name 866 if (!EF->hasParam()) { 867 ExportFuncInfo.push_back(llvm::MDString::get(mLLVMContext, 868 EF->getName().c_str())); 869 } else { 870 llvm::Function *F = M->getFunction(EF->getName()); 871 llvm::Function *HelperFunction; 872 const std::string HelperFunctionName(".helper_" + EF->getName()); 873 874 slangAssert(F && "Function marked as exported disappeared in Bitcode"); 875 876 // Create helper function 877 { 878 llvm::StructType *OrigHelperFunctionParameterTy = nullptr; 879 llvm::StructType *PaddedHelperFunctionParameterTy = nullptr; 880 881 std::vector<unsigned> OrigFieldNumToPaddedFieldNum; 882 std::vector<bool> isStructInput; 883 884 if (!F->getArgumentList().empty()) { 885 std::vector<llvm::Type*> HelperFunctionParameterTys; 886 for (llvm::Function::arg_iterator AI = F->arg_begin(), 887 AE = F->arg_end(); AI != AE; AI++) { 888 if (AI->getType()->isPointerTy() && AI->getType()->getPointerElementType()->isStructTy()) { 889 HelperFunctionParameterTys.push_back(AI->getType()->getPointerElementType()); 890 isStructInput.push_back(true); 891 } else { 892 HelperFunctionParameterTys.push_back(AI->getType()); 893 isStructInput.push_back(false); 894 } 895 } 896 OrigHelperFunctionParameterTy = 897 llvm::StructType::get(mLLVMContext, HelperFunctionParameterTys); 898 PadHelperFunctionStruct(M, 899 &PaddedHelperFunctionParameterTy, &OrigFieldNumToPaddedFieldNum, 900 OrigHelperFunctionParameterTy); 901 } 902 903 if (!EF->checkParameterPacketType(OrigHelperFunctionParameterTy)) { 904 fprintf(stderr, "Failed to export function %s: parameter type " 905 "mismatch during creation of helper function.\n", 906 EF->getName().c_str()); 907 908 const RSExportRecordType *Expected = EF->getParamPacketType(); 909 if (Expected) { 910 fprintf(stderr, "Expected:\n"); 911 Expected->getLLVMType()->dump(); 912 } 913 if (OrigHelperFunctionParameterTy) { 914 fprintf(stderr, "Got:\n"); 915 OrigHelperFunctionParameterTy->dump(); 916 } 917 } 918 919 std::vector<llvm::Type*> Params; 920 if (PaddedHelperFunctionParameterTy) { 921 llvm::PointerType *HelperFunctionParameterTyP = 922 llvm::PointerType::getUnqual(PaddedHelperFunctionParameterTy); 923 Params.push_back(HelperFunctionParameterTyP); 924 } 925 926 llvm::FunctionType * HelperFunctionType = 927 llvm::FunctionType::get(F->getReturnType(), 928 Params, 929 /* IsVarArgs = */false); 930 931 HelperFunction = 932 llvm::Function::Create(HelperFunctionType, 933 llvm::GlobalValue::ExternalLinkage, 934 HelperFunctionName, 935 M); 936 937 HelperFunction->addFnAttr(llvm::Attribute::NoInline); 938 HelperFunction->setCallingConv(F->getCallingConv()); 939 940 // Create helper function body 941 { 942 llvm::Argument *HelperFunctionParameter = 943 &(*HelperFunction->arg_begin()); 944 llvm::BasicBlock *BB = 945 llvm::BasicBlock::Create(mLLVMContext, "entry", HelperFunction); 946 llvm::IRBuilder<> *IB = new llvm::IRBuilder<>(BB); 947 llvm::SmallVector<llvm::Value*, 6> Params; 948 llvm::Value *Idx[2]; 949 950 Idx[0] = 951 llvm::ConstantInt::get(llvm::Type::getInt32Ty(mLLVMContext), 0); 952 953 // getelementptr and load instruction for all elements in 954 // parameter .p 955 for (size_t origFieldNum = 0; origFieldNum < EF->getNumParameters(); origFieldNum++) { 956 // getelementptr 957 Idx[1] = llvm::ConstantInt::get( 958 llvm::Type::getInt32Ty(mLLVMContext), OrigFieldNumToPaddedFieldNum[origFieldNum]); 959 960 llvm::Value *Ptr = NULL; 961 962 Ptr = IB->CreateInBoundsGEP(HelperFunctionParameter, Idx); 963 964 // Load is only required for non-struct ptrs 965 if (isStructInput[origFieldNum]) { 966 Params.push_back(Ptr); 967 } else { 968 llvm::Value *V = IB->CreateLoad(Ptr); 969 Params.push_back(V); 970 } 971 } 972 973 // Call and pass the all elements as parameter to F 974 llvm::CallInst *CI = IB->CreateCall(F, Params); 975 976 CI->setCallingConv(F->getCallingConv()); 977 978 if (F->getReturnType() == llvm::Type::getVoidTy(mLLVMContext)) { 979 IB->CreateRetVoid(); 980 } else { 981 IB->CreateRet(CI); 982 } 983 984 delete IB; 985 } 986 } 987 988 ExportFuncInfo.push_back( 989 llvm::MDString::get(mLLVMContext, HelperFunctionName.c_str())); 990 } 991 992 mExportFuncMetadata->addOperand( 993 llvm::MDNode::get(mLLVMContext, ExportFuncInfo)); 994 ExportFuncInfo.clear(); 995 } 996} 997 998void Backend::dumpExportForEachInfo(llvm::Module *M) { 999 if (mExportForEachNameMetadata == nullptr) { 1000 mExportForEachNameMetadata = 1001 M->getOrInsertNamedMetadata(RS_EXPORT_FOREACH_NAME_MN); 1002 } 1003 if (mExportForEachSignatureMetadata == nullptr) { 1004 mExportForEachSignatureMetadata = 1005 M->getOrInsertNamedMetadata(RS_EXPORT_FOREACH_MN); 1006 } 1007 1008 llvm::SmallVector<llvm::Metadata *, 1> ExportForEachName; 1009 llvm::SmallVector<llvm::Metadata *, 1> ExportForEachInfo; 1010 1011 for (RSContext::const_export_foreach_iterator 1012 I = mContext->export_foreach_begin(), 1013 E = mContext->export_foreach_end(); 1014 I != E; 1015 I++) { 1016 const RSExportForEach *EFE = *I; 1017 1018 ExportForEachName.push_back( 1019 llvm::MDString::get(mLLVMContext, EFE->getName().c_str())); 1020 1021 mExportForEachNameMetadata->addOperand( 1022 llvm::MDNode::get(mLLVMContext, ExportForEachName)); 1023 ExportForEachName.clear(); 1024 1025 ExportForEachInfo.push_back( 1026 llvm::MDString::get(mLLVMContext, 1027 llvm::utostr(EFE->getSignatureMetadata()))); 1028 1029 mExportForEachSignatureMetadata->addOperand( 1030 llvm::MDNode::get(mLLVMContext, ExportForEachInfo)); 1031 ExportForEachInfo.clear(); 1032 } 1033} 1034 1035void Backend::dumpExportReduceInfo(llvm::Module *M) { 1036 if (!mExportReduceMetadata) { 1037 mExportReduceMetadata = 1038 M->getOrInsertNamedMetadata(RS_EXPORT_REDUCE_MN); 1039 } 1040 1041 llvm::SmallVector<llvm::Metadata *, 6> ExportReduceInfo; 1042 // Add operand to ExportReduceInfo, padding out missing operands with 1043 // nullptr. 1044 auto addOperand = [&ExportReduceInfo](uint32_t Idx, llvm::Metadata *N) { 1045 while (Idx > ExportReduceInfo.size()) 1046 ExportReduceInfo.push_back(nullptr); 1047 ExportReduceInfo.push_back(N); 1048 }; 1049 // Add string operand to ExportReduceInfo, padding out missing operands 1050 // with nullptr. 1051 // If string is empty, then do not add it unless Always is true. 1052 auto addString = [&addOperand, this](uint32_t Idx, const std::string &S, 1053 bool Always = true) { 1054 if (Always || !S.empty()) 1055 addOperand(Idx, llvm::MDString::get(mLLVMContext, S)); 1056 }; 1057 1058 // Add the description of the reduction kernels to the metadata node. 1059 for (auto I = mContext->export_reduce_begin(), 1060 E = mContext->export_reduce_end(); 1061 I != E; ++I) { 1062 ExportReduceInfo.clear(); 1063 1064 int Idx = 0; 1065 1066 addString(Idx++, (*I)->getNameReduce()); 1067 1068 addOperand(Idx++, llvm::MDString::get(mLLVMContext, llvm::utostr((*I)->getAccumulatorTypeSize()))); 1069 1070 llvm::SmallVector<llvm::Metadata *, 2> Accumulator; 1071 Accumulator.push_back( 1072 llvm::MDString::get(mLLVMContext, (*I)->getNameAccumulator())); 1073 Accumulator.push_back(llvm::MDString::get( 1074 mLLVMContext, 1075 llvm::utostr((*I)->getAccumulatorSignatureMetadata()))); 1076 addOperand(Idx++, llvm::MDTuple::get(mLLVMContext, Accumulator)); 1077 1078 addString(Idx++, (*I)->getNameInitializer(), false); 1079 addString(Idx++, (*I)->getNameCombiner(), false); 1080 addString(Idx++, (*I)->getNameOutConverter(), false); 1081 addString(Idx++, (*I)->getNameHalter(), false); 1082 1083 mExportReduceMetadata->addOperand( 1084 llvm::MDTuple::get(mLLVMContext, ExportReduceInfo)); 1085 } 1086} 1087 1088void Backend::dumpExportTypeInfo(llvm::Module *M) { 1089 llvm::SmallVector<llvm::Metadata *, 1> ExportTypeInfo; 1090 1091 for (RSContext::const_export_type_iterator 1092 I = mContext->export_types_begin(), 1093 E = mContext->export_types_end(); 1094 I != E; 1095 I++) { 1096 // First, dump type name list to export 1097 const RSExportType *ET = I->getValue(); 1098 1099 ExportTypeInfo.clear(); 1100 // Type name 1101 ExportTypeInfo.push_back( 1102 llvm::MDString::get(mLLVMContext, ET->getName().c_str())); 1103 1104 if (ET->getClass() == RSExportType::ExportClassRecord) { 1105 const RSExportRecordType *ERT = 1106 static_cast<const RSExportRecordType*>(ET); 1107 1108 if (mExportTypeMetadata == nullptr) 1109 mExportTypeMetadata = 1110 M->getOrInsertNamedMetadata(RS_EXPORT_TYPE_MN); 1111 1112 mExportTypeMetadata->addOperand( 1113 llvm::MDNode::get(mLLVMContext, ExportTypeInfo)); 1114 1115 // Now, export struct field information to %[struct name] 1116 std::string StructInfoMetadataName("%"); 1117 StructInfoMetadataName.append(ET->getName()); 1118 llvm::NamedMDNode *StructInfoMetadata = 1119 M->getOrInsertNamedMetadata(StructInfoMetadataName); 1120 llvm::SmallVector<llvm::Metadata *, 3> FieldInfo; 1121 1122 slangAssert(StructInfoMetadata->getNumOperands() == 0 && 1123 "Metadata with same name was created before"); 1124 for (RSExportRecordType::const_field_iterator FI = ERT->fields_begin(), 1125 FE = ERT->fields_end(); 1126 FI != FE; 1127 FI++) { 1128 const RSExportRecordType::Field *F = *FI; 1129 1130 // 1. field name 1131 FieldInfo.push_back(llvm::MDString::get(mLLVMContext, 1132 F->getName().c_str())); 1133 1134 // 2. field type name 1135 FieldInfo.push_back( 1136 llvm::MDString::get(mLLVMContext, 1137 F->getType()->getName().c_str())); 1138 1139 StructInfoMetadata->addOperand( 1140 llvm::MDNode::get(mLLVMContext, FieldInfo)); 1141 FieldInfo.clear(); 1142 } 1143 } // ET->getClass() == RSExportType::ExportClassRecord 1144 } 1145} 1146 1147void Backend::HandleTranslationUnitPost(llvm::Module *M) { 1148 1149 if (!mContext->is64Bit()) { 1150 M->setDataLayout("e-p:32:32-i64:64-v128:64:128-n32-S64"); 1151 } 1152 1153 if (!mContext->processExports()) 1154 return; 1155 1156 if (mContext->hasExportVar()) 1157 dumpExportVarInfo(M); 1158 1159 if (mContext->hasExportFunc()) 1160 dumpExportFunctionInfo(M); 1161 1162 if (mContext->hasExportForEach()) 1163 dumpExportForEachInfo(M); 1164 1165 if (mContext->hasExportReduce()) 1166 dumpExportReduceInfo(M); 1167 1168 if (mContext->hasExportType()) 1169 dumpExportTypeInfo(M); 1170} 1171 1172} // namespace slang 1173