1 //===--- CGStmt.cpp - Emit LLVM Code from Statements ----------------------===// 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 contains code to emit Stmt nodes as LLVM code. 11 // 12 //===----------------------------------------------------------------------===// 13 14 #include "CodeGenFunction.h" 15 #include "CGDebugInfo.h" 16 #include "CodeGenModule.h" 17 #include "TargetInfo.h" 18 #include "clang/AST/StmtVisitor.h" 19 #include "clang/Basic/Builtins.h" 20 #include "clang/Basic/PrettyStackTrace.h" 21 #include "clang/Basic/TargetInfo.h" 22 #include "clang/Sema/LoopHint.h" 23 #include "clang/Sema/SemaDiagnostic.h" 24 #include "llvm/ADT/StringExtras.h" 25 #include "llvm/IR/CallSite.h" 26 #include "llvm/IR/DataLayout.h" 27 #include "llvm/IR/InlineAsm.h" 28 #include "llvm/IR/Intrinsics.h" 29 #include "llvm/IR/MDBuilder.h" 30 31 using namespace clang; 32 using namespace CodeGen; 33 34 //===----------------------------------------------------------------------===// 35 // Statement Emission 36 //===----------------------------------------------------------------------===// 37 38 void CodeGenFunction::EmitStopPoint(const Stmt *S) { 39 if (CGDebugInfo *DI = getDebugInfo()) { 40 SourceLocation Loc; 41 Loc = S->getLocStart(); 42 DI->EmitLocation(Builder, Loc); 43 44 LastStopPoint = Loc; 45 } 46 } 47 48 void CodeGenFunction::EmitStmt(const Stmt *S, ArrayRef<const Attr *> Attrs) { 49 assert(S && "Null statement?"); 50 PGO.setCurrentStmt(S); 51 52 // These statements have their own debug info handling. 53 if (EmitSimpleStmt(S)) 54 return; 55 56 // Check if we are generating unreachable code. 57 if (!HaveInsertPoint()) { 58 // If so, and the statement doesn't contain a label, then we do not need to 59 // generate actual code. This is safe because (1) the current point is 60 // unreachable, so we don't need to execute the code, and (2) we've already 61 // handled the statements which update internal data structures (like the 62 // local variable map) which could be used by subsequent statements. 63 if (!ContainsLabel(S)) { 64 // Verify that any decl statements were handled as simple, they may be in 65 // scope of subsequent reachable statements. 66 assert(!isa<DeclStmt>(*S) && "Unexpected DeclStmt!"); 67 return; 68 } 69 70 // Otherwise, make a new block to hold the code. 71 EnsureInsertPoint(); 72 } 73 74 // Generate a stoppoint if we are emitting debug info. 75 EmitStopPoint(S); 76 77 // Ignore all OpenMP directives except for simd if OpenMP with Simd is 78 // enabled. 79 if (getLangOpts().OpenMP && getLangOpts().OpenMPSimd) { 80 if (const auto *D = dyn_cast<OMPExecutableDirective>(S)) { 81 EmitSimpleOMPExecutableDirective(*D); 82 return; 83 } 84 } 85 86 switch (S->getStmtClass()) { 87 case Stmt::NoStmtClass: 88 case Stmt::CXXCatchStmtClass: 89 case Stmt::SEHExceptStmtClass: 90 case Stmt::SEHFinallyStmtClass: 91 case Stmt::MSDependentExistsStmtClass: 92 llvm_unreachable("invalid statement class to emit generically"); 93 case Stmt::NullStmtClass: 94 case Stmt::CompoundStmtClass: 95 case Stmt::DeclStmtClass: 96 case Stmt::LabelStmtClass: 97 case Stmt::AttributedStmtClass: 98 case Stmt::GotoStmtClass: 99 case Stmt::BreakStmtClass: 100 case Stmt::ContinueStmtClass: 101 case Stmt::DefaultStmtClass: 102 case Stmt::CaseStmtClass: 103 case Stmt::SEHLeaveStmtClass: 104 llvm_unreachable("should have emitted these statements as simple"); 105 106 #define STMT(Type, Base) 107 #define ABSTRACT_STMT(Op) 108 #define EXPR(Type, Base) \ 109 case Stmt::Type##Class: 110 #include "clang/AST/StmtNodes.inc" 111 { 112 // Remember the block we came in on. 113 llvm::BasicBlock *incoming = Builder.GetInsertBlock(); 114 assert(incoming && "expression emission must have an insertion point"); 115 116 EmitIgnoredExpr(cast<Expr>(S)); 117 118 llvm::BasicBlock *outgoing = Builder.GetInsertBlock(); 119 assert(outgoing && "expression emission cleared block!"); 120 121 // The expression emitters assume (reasonably!) that the insertion 122 // point is always set. To maintain that, the call-emission code 123 // for noreturn functions has to enter a new block with no 124 // predecessors. We want to kill that block and mark the current 125 // insertion point unreachable in the common case of a call like 126 // "exit();". Since expression emission doesn't otherwise create 127 // blocks with no predecessors, we can just test for that. 128 // However, we must be careful not to do this to our incoming 129 // block, because *statement* emission does sometimes create 130 // reachable blocks which will have no predecessors until later in 131 // the function. This occurs with, e.g., labels that are not 132 // reachable by fallthrough. 133 if (incoming != outgoing && outgoing->use_empty()) { 134 outgoing->eraseFromParent(); 135 Builder.ClearInsertionPoint(); 136 } 137 break; 138 } 139 140 case Stmt::IndirectGotoStmtClass: 141 EmitIndirectGotoStmt(cast<IndirectGotoStmt>(*S)); break; 142 143 case Stmt::IfStmtClass: EmitIfStmt(cast<IfStmt>(*S)); break; 144 case Stmt::WhileStmtClass: EmitWhileStmt(cast<WhileStmt>(*S), Attrs); break; 145 case Stmt::DoStmtClass: EmitDoStmt(cast<DoStmt>(*S), Attrs); break; 146 case Stmt::ForStmtClass: EmitForStmt(cast<ForStmt>(*S), Attrs); break; 147 148 case Stmt::ReturnStmtClass: EmitReturnStmt(cast<ReturnStmt>(*S)); break; 149 150 case Stmt::SwitchStmtClass: EmitSwitchStmt(cast<SwitchStmt>(*S)); break; 151 case Stmt::GCCAsmStmtClass: // Intentional fall-through. 152 case Stmt::MSAsmStmtClass: EmitAsmStmt(cast<AsmStmt>(*S)); break; 153 case Stmt::CoroutineBodyStmtClass: 154 EmitCoroutineBody(cast<CoroutineBodyStmt>(*S)); 155 break; 156 case Stmt::CoreturnStmtClass: 157 EmitCoreturnStmt(cast<CoreturnStmt>(*S)); 158 break; 159 case Stmt::CapturedStmtClass: { 160 const CapturedStmt *CS = cast<CapturedStmt>(S); 161 EmitCapturedStmt(*CS, CS->getCapturedRegionKind()); 162 } 163 break; 164 case Stmt::ObjCAtTryStmtClass: 165 EmitObjCAtTryStmt(cast<ObjCAtTryStmt>(*S)); 166 break; 167 case Stmt::ObjCAtCatchStmtClass: 168 llvm_unreachable( 169 "@catch statements should be handled by EmitObjCAtTryStmt"); 170 case Stmt::ObjCAtFinallyStmtClass: 171 llvm_unreachable( 172 "@finally statements should be handled by EmitObjCAtTryStmt"); 173 case Stmt::ObjCAtThrowStmtClass: 174 EmitObjCAtThrowStmt(cast<ObjCAtThrowStmt>(*S)); 175 break; 176 case Stmt::ObjCAtSynchronizedStmtClass: 177 EmitObjCAtSynchronizedStmt(cast<ObjCAtSynchronizedStmt>(*S)); 178 break; 179 case Stmt::ObjCForCollectionStmtClass: 180 EmitObjCForCollectionStmt(cast<ObjCForCollectionStmt>(*S)); 181 break; 182 case Stmt::ObjCAutoreleasePoolStmtClass: 183 EmitObjCAutoreleasePoolStmt(cast<ObjCAutoreleasePoolStmt>(*S)); 184 break; 185 186 case Stmt::CXXTryStmtClass: 187 EmitCXXTryStmt(cast<CXXTryStmt>(*S)); 188 break; 189 case Stmt::CXXForRangeStmtClass: 190 EmitCXXForRangeStmt(cast<CXXForRangeStmt>(*S), Attrs); 191 break; 192 case Stmt::SEHTryStmtClass: 193 EmitSEHTryStmt(cast<SEHTryStmt>(*S)); 194 break; 195 case Stmt::OMPParallelDirectiveClass: 196 EmitOMPParallelDirective(cast<OMPParallelDirective>(*S)); 197 break; 198 case Stmt::OMPSimdDirectiveClass: 199 EmitOMPSimdDirective(cast<OMPSimdDirective>(*S)); 200 break; 201 case Stmt::OMPForDirectiveClass: 202 EmitOMPForDirective(cast<OMPForDirective>(*S)); 203 break; 204 case Stmt::OMPForSimdDirectiveClass: 205 EmitOMPForSimdDirective(cast<OMPForSimdDirective>(*S)); 206 break; 207 case Stmt::OMPSectionsDirectiveClass: 208 EmitOMPSectionsDirective(cast<OMPSectionsDirective>(*S)); 209 break; 210 case Stmt::OMPSectionDirectiveClass: 211 EmitOMPSectionDirective(cast<OMPSectionDirective>(*S)); 212 break; 213 case Stmt::OMPSingleDirectiveClass: 214 EmitOMPSingleDirective(cast<OMPSingleDirective>(*S)); 215 break; 216 case Stmt::OMPMasterDirectiveClass: 217 EmitOMPMasterDirective(cast<OMPMasterDirective>(*S)); 218 break; 219 case Stmt::OMPCriticalDirectiveClass: 220 EmitOMPCriticalDirective(cast<OMPCriticalDirective>(*S)); 221 break; 222 case Stmt::OMPParallelForDirectiveClass: 223 EmitOMPParallelForDirective(cast<OMPParallelForDirective>(*S)); 224 break; 225 case Stmt::OMPParallelForSimdDirectiveClass: 226 EmitOMPParallelForSimdDirective(cast<OMPParallelForSimdDirective>(*S)); 227 break; 228 case Stmt::OMPParallelSectionsDirectiveClass: 229 EmitOMPParallelSectionsDirective(cast<OMPParallelSectionsDirective>(*S)); 230 break; 231 case Stmt::OMPTaskDirectiveClass: 232 EmitOMPTaskDirective(cast<OMPTaskDirective>(*S)); 233 break; 234 case Stmt::OMPTaskyieldDirectiveClass: 235 EmitOMPTaskyieldDirective(cast<OMPTaskyieldDirective>(*S)); 236 break; 237 case Stmt::OMPBarrierDirectiveClass: 238 EmitOMPBarrierDirective(cast<OMPBarrierDirective>(*S)); 239 break; 240 case Stmt::OMPTaskwaitDirectiveClass: 241 EmitOMPTaskwaitDirective(cast<OMPTaskwaitDirective>(*S)); 242 break; 243 case Stmt::OMPTaskgroupDirectiveClass: 244 EmitOMPTaskgroupDirective(cast<OMPTaskgroupDirective>(*S)); 245 break; 246 case Stmt::OMPFlushDirectiveClass: 247 EmitOMPFlushDirective(cast<OMPFlushDirective>(*S)); 248 break; 249 case Stmt::OMPOrderedDirectiveClass: 250 EmitOMPOrderedDirective(cast<OMPOrderedDirective>(*S)); 251 break; 252 case Stmt::OMPAtomicDirectiveClass: 253 EmitOMPAtomicDirective(cast<OMPAtomicDirective>(*S)); 254 break; 255 case Stmt::OMPTargetDirectiveClass: 256 EmitOMPTargetDirective(cast<OMPTargetDirective>(*S)); 257 break; 258 case Stmt::OMPTeamsDirectiveClass: 259 EmitOMPTeamsDirective(cast<OMPTeamsDirective>(*S)); 260 break; 261 case Stmt::OMPCancellationPointDirectiveClass: 262 EmitOMPCancellationPointDirective(cast<OMPCancellationPointDirective>(*S)); 263 break; 264 case Stmt::OMPCancelDirectiveClass: 265 EmitOMPCancelDirective(cast<OMPCancelDirective>(*S)); 266 break; 267 case Stmt::OMPTargetDataDirectiveClass: 268 EmitOMPTargetDataDirective(cast<OMPTargetDataDirective>(*S)); 269 break; 270 case Stmt::OMPTargetEnterDataDirectiveClass: 271 EmitOMPTargetEnterDataDirective(cast<OMPTargetEnterDataDirective>(*S)); 272 break; 273 case Stmt::OMPTargetExitDataDirectiveClass: 274 EmitOMPTargetExitDataDirective(cast<OMPTargetExitDataDirective>(*S)); 275 break; 276 case Stmt::OMPTargetParallelDirectiveClass: 277 EmitOMPTargetParallelDirective(cast<OMPTargetParallelDirective>(*S)); 278 break; 279 case Stmt::OMPTargetParallelForDirectiveClass: 280 EmitOMPTargetParallelForDirective(cast<OMPTargetParallelForDirective>(*S)); 281 break; 282 case Stmt::OMPTaskLoopDirectiveClass: 283 EmitOMPTaskLoopDirective(cast<OMPTaskLoopDirective>(*S)); 284 break; 285 case Stmt::OMPTaskLoopSimdDirectiveClass: 286 EmitOMPTaskLoopSimdDirective(cast<OMPTaskLoopSimdDirective>(*S)); 287 break; 288 case Stmt::OMPDistributeDirectiveClass: 289 EmitOMPDistributeDirective(cast<OMPDistributeDirective>(*S)); 290 break; 291 case Stmt::OMPTargetUpdateDirectiveClass: 292 EmitOMPTargetUpdateDirective(cast<OMPTargetUpdateDirective>(*S)); 293 break; 294 case Stmt::OMPDistributeParallelForDirectiveClass: 295 EmitOMPDistributeParallelForDirective( 296 cast<OMPDistributeParallelForDirective>(*S)); 297 break; 298 case Stmt::OMPDistributeParallelForSimdDirectiveClass: 299 EmitOMPDistributeParallelForSimdDirective( 300 cast<OMPDistributeParallelForSimdDirective>(*S)); 301 break; 302 case Stmt::OMPDistributeSimdDirectiveClass: 303 EmitOMPDistributeSimdDirective(cast<OMPDistributeSimdDirective>(*S)); 304 break; 305 case Stmt::OMPTargetParallelForSimdDirectiveClass: 306 EmitOMPTargetParallelForSimdDirective( 307 cast<OMPTargetParallelForSimdDirective>(*S)); 308 break; 309 case Stmt::OMPTargetSimdDirectiveClass: 310 EmitOMPTargetSimdDirective(cast<OMPTargetSimdDirective>(*S)); 311 break; 312 case Stmt::OMPTeamsDistributeDirectiveClass: 313 EmitOMPTeamsDistributeDirective(cast<OMPTeamsDistributeDirective>(*S)); 314 break; 315 case Stmt::OMPTeamsDistributeSimdDirectiveClass: 316 EmitOMPTeamsDistributeSimdDirective( 317 cast<OMPTeamsDistributeSimdDirective>(*S)); 318 break; 319 case Stmt::OMPTeamsDistributeParallelForSimdDirectiveClass: 320 EmitOMPTeamsDistributeParallelForSimdDirective( 321 cast<OMPTeamsDistributeParallelForSimdDirective>(*S)); 322 break; 323 case Stmt::OMPTeamsDistributeParallelForDirectiveClass: 324 EmitOMPTeamsDistributeParallelForDirective( 325 cast<OMPTeamsDistributeParallelForDirective>(*S)); 326 break; 327 case Stmt::OMPTargetTeamsDirectiveClass: 328 EmitOMPTargetTeamsDirective(cast<OMPTargetTeamsDirective>(*S)); 329 break; 330 case Stmt::OMPTargetTeamsDistributeDirectiveClass: 331 EmitOMPTargetTeamsDistributeDirective( 332 cast<OMPTargetTeamsDistributeDirective>(*S)); 333 break; 334 case Stmt::OMPTargetTeamsDistributeParallelForDirectiveClass: 335 EmitOMPTargetTeamsDistributeParallelForDirective( 336 cast<OMPTargetTeamsDistributeParallelForDirective>(*S)); 337 break; 338 case Stmt::OMPTargetTeamsDistributeParallelForSimdDirectiveClass: 339 EmitOMPTargetTeamsDistributeParallelForSimdDirective( 340 cast<OMPTargetTeamsDistributeParallelForSimdDirective>(*S)); 341 break; 342 case Stmt::OMPTargetTeamsDistributeSimdDirectiveClass: 343 EmitOMPTargetTeamsDistributeSimdDirective( 344 cast<OMPTargetTeamsDistributeSimdDirective>(*S)); 345 break; 346 } 347 } 348 349 bool CodeGenFunction::EmitSimpleStmt(const Stmt *S) { 350 switch (S->getStmtClass()) { 351 default: return false; 352 case Stmt::NullStmtClass: break; 353 case Stmt::CompoundStmtClass: EmitCompoundStmt(cast<CompoundStmt>(*S)); break; 354 case Stmt::DeclStmtClass: EmitDeclStmt(cast<DeclStmt>(*S)); break; 355 case Stmt::LabelStmtClass: EmitLabelStmt(cast<LabelStmt>(*S)); break; 356 case Stmt::AttributedStmtClass: 357 EmitAttributedStmt(cast<AttributedStmt>(*S)); break; 358 case Stmt::GotoStmtClass: EmitGotoStmt(cast<GotoStmt>(*S)); break; 359 case Stmt::BreakStmtClass: EmitBreakStmt(cast<BreakStmt>(*S)); break; 360 case Stmt::ContinueStmtClass: EmitContinueStmt(cast<ContinueStmt>(*S)); break; 361 case Stmt::DefaultStmtClass: EmitDefaultStmt(cast<DefaultStmt>(*S)); break; 362 case Stmt::CaseStmtClass: EmitCaseStmt(cast<CaseStmt>(*S)); break; 363 case Stmt::SEHLeaveStmtClass: EmitSEHLeaveStmt(cast<SEHLeaveStmt>(*S)); break; 364 } 365 366 return true; 367 } 368 369 /// EmitCompoundStmt - Emit a compound statement {..} node. If GetLast is true, 370 /// this captures the expression result of the last sub-statement and returns it 371 /// (for use by the statement expression extension). 372 Address CodeGenFunction::EmitCompoundStmt(const CompoundStmt &S, bool GetLast, 373 AggValueSlot AggSlot) { 374 PrettyStackTraceLoc CrashInfo(getContext().getSourceManager(),S.getLBracLoc(), 375 "LLVM IR generation of compound statement ('{}')"); 376 377 // Keep track of the current cleanup stack depth, including debug scopes. 378 LexicalScope Scope(*this, S.getSourceRange()); 379 380 return EmitCompoundStmtWithoutScope(S, GetLast, AggSlot); 381 } 382 383 Address 384 CodeGenFunction::EmitCompoundStmtWithoutScope(const CompoundStmt &S, 385 bool GetLast, 386 AggValueSlot AggSlot) { 387 388 for (CompoundStmt::const_body_iterator I = S.body_begin(), 389 E = S.body_end()-GetLast; I != E; ++I) 390 EmitStmt(*I); 391 392 Address RetAlloca = Address::invalid(); 393 if (GetLast) { 394 // We have to special case labels here. They are statements, but when put 395 // at the end of a statement expression, they yield the value of their 396 // subexpression. Handle this by walking through all labels we encounter, 397 // emitting them before we evaluate the subexpr. 398 const Stmt *LastStmt = S.body_back(); 399 while (const LabelStmt *LS = dyn_cast<LabelStmt>(LastStmt)) { 400 EmitLabel(LS->getDecl()); 401 LastStmt = LS->getSubStmt(); 402 } 403 404 EnsureInsertPoint(); 405 406 QualType ExprTy = cast<Expr>(LastStmt)->getType(); 407 if (hasAggregateEvaluationKind(ExprTy)) { 408 EmitAggExpr(cast<Expr>(LastStmt), AggSlot); 409 } else { 410 // We can't return an RValue here because there might be cleanups at 411 // the end of the StmtExpr. Because of that, we have to emit the result 412 // here into a temporary alloca. 413 RetAlloca = CreateMemTemp(ExprTy); 414 EmitAnyExprToMem(cast<Expr>(LastStmt), RetAlloca, Qualifiers(), 415 /*IsInit*/false); 416 } 417 418 } 419 420 return RetAlloca; 421 } 422 423 void CodeGenFunction::SimplifyForwardingBlocks(llvm::BasicBlock *BB) { 424 llvm::BranchInst *BI = dyn_cast<llvm::BranchInst>(BB->getTerminator()); 425 426 // If there is a cleanup stack, then we it isn't worth trying to 427 // simplify this block (we would need to remove it from the scope map 428 // and cleanup entry). 429 if (!EHStack.empty()) 430 return; 431 432 // Can only simplify direct branches. 433 if (!BI || !BI->isUnconditional()) 434 return; 435 436 // Can only simplify empty blocks. 437 if (BI->getIterator() != BB->begin()) 438 return; 439 440 BB->replaceAllUsesWith(BI->getSuccessor(0)); 441 BI->eraseFromParent(); 442 BB->eraseFromParent(); 443 } 444 445 void CodeGenFunction::EmitBlock(llvm::BasicBlock *BB, bool IsFinished) { 446 llvm::BasicBlock *CurBB = Builder.GetInsertBlock(); 447 448 // Fall out of the current block (if necessary). 449 EmitBranch(BB); 450 451 if (IsFinished && BB->use_empty()) { 452 delete BB; 453 return; 454 } 455 456 // Place the block after the current block, if possible, or else at 457 // the end of the function. 458 if (CurBB && CurBB->getParent()) 459 CurFn->getBasicBlockList().insertAfter(CurBB->getIterator(), BB); 460 else 461 CurFn->getBasicBlockList().push_back(BB); 462 Builder.SetInsertPoint(BB); 463 } 464 465 void CodeGenFunction::EmitBranch(llvm::BasicBlock *Target) { 466 // Emit a branch from the current block to the target one if this 467 // was a real block. If this was just a fall-through block after a 468 // terminator, don't emit it. 469 llvm::BasicBlock *CurBB = Builder.GetInsertBlock(); 470 471 if (!CurBB || CurBB->getTerminator()) { 472 // If there is no insert point or the previous block is already 473 // terminated, don't touch it. 474 } else { 475 // Otherwise, create a fall-through branch. 476 Builder.CreateBr(Target); 477 } 478 479 Builder.ClearInsertionPoint(); 480 } 481 482 void CodeGenFunction::EmitBlockAfterUses(llvm::BasicBlock *block) { 483 bool inserted = false; 484 for (llvm::User *u : block->users()) { 485 if (llvm::Instruction *insn = dyn_cast<llvm::Instruction>(u)) { 486 CurFn->getBasicBlockList().insertAfter(insn->getParent()->getIterator(), 487 block); 488 inserted = true; 489 break; 490 } 491 } 492 493 if (!inserted) 494 CurFn->getBasicBlockList().push_back(block); 495 496 Builder.SetInsertPoint(block); 497 } 498 499 CodeGenFunction::JumpDest 500 CodeGenFunction::getJumpDestForLabel(const LabelDecl *D) { 501 JumpDest &Dest = LabelMap[D]; 502 if (Dest.isValid()) return Dest; 503 504 // Create, but don't insert, the new block. 505 Dest = JumpDest(createBasicBlock(D->getName()), 506 EHScopeStack::stable_iterator::invalid(), 507 NextCleanupDestIndex++); 508 return Dest; 509 } 510 511 void CodeGenFunction::EmitLabel(const LabelDecl *D) { 512 // Add this label to the current lexical scope if we're within any 513 // normal cleanups. Jumps "in" to this label --- when permitted by 514 // the language --- may need to be routed around such cleanups. 515 if (EHStack.hasNormalCleanups() && CurLexicalScope) 516 CurLexicalScope->addLabel(D); 517 518 JumpDest &Dest = LabelMap[D]; 519 520 // If we didn't need a forward reference to this label, just go 521 // ahead and create a destination at the current scope. 522 if (!Dest.isValid()) { 523 Dest = getJumpDestInCurrentScope(D->getName()); 524 525 // Otherwise, we need to give this label a target depth and remove 526 // it from the branch-fixups list. 527 } else { 528 assert(!Dest.getScopeDepth().isValid() && "already emitted label!"); 529 Dest.setScopeDepth(EHStack.stable_begin()); 530 ResolveBranchFixups(Dest.getBlock()); 531 } 532 533 EmitBlock(Dest.getBlock()); 534 incrementProfileCounter(D->getStmt()); 535 } 536 537 /// Change the cleanup scope of the labels in this lexical scope to 538 /// match the scope of the enclosing context. 539 void CodeGenFunction::LexicalScope::rescopeLabels() { 540 assert(!Labels.empty()); 541 EHScopeStack::stable_iterator innermostScope 542 = CGF.EHStack.getInnermostNormalCleanup(); 543 544 // Change the scope depth of all the labels. 545 for (SmallVectorImpl<const LabelDecl*>::const_iterator 546 i = Labels.begin(), e = Labels.end(); i != e; ++i) { 547 assert(CGF.LabelMap.count(*i)); 548 JumpDest &dest = CGF.LabelMap.find(*i)->second; 549 assert(dest.getScopeDepth().isValid()); 550 assert(innermostScope.encloses(dest.getScopeDepth())); 551 dest.setScopeDepth(innermostScope); 552 } 553 554 // Reparent the labels if the new scope also has cleanups. 555 if (innermostScope != EHScopeStack::stable_end() && ParentScope) { 556 ParentScope->Labels.append(Labels.begin(), Labels.end()); 557 } 558 } 559 560 561 void CodeGenFunction::EmitLabelStmt(const LabelStmt &S) { 562 EmitLabel(S.getDecl()); 563 EmitStmt(S.getSubStmt()); 564 } 565 566 void CodeGenFunction::EmitAttributedStmt(const AttributedStmt &S) { 567 EmitStmt(S.getSubStmt(), S.getAttrs()); 568 } 569 570 void CodeGenFunction::EmitGotoStmt(const GotoStmt &S) { 571 // If this code is reachable then emit a stop point (if generating 572 // debug info). We have to do this ourselves because we are on the 573 // "simple" statement path. 574 if (HaveInsertPoint()) 575 EmitStopPoint(&S); 576 577 EmitBranchThroughCleanup(getJumpDestForLabel(S.getLabel())); 578 } 579 580 581 void CodeGenFunction::EmitIndirectGotoStmt(const IndirectGotoStmt &S) { 582 if (const LabelDecl *Target = S.getConstantTarget()) { 583 EmitBranchThroughCleanup(getJumpDestForLabel(Target)); 584 return; 585 } 586 587 // Ensure that we have an i8* for our PHI node. 588 llvm::Value *V = Builder.CreateBitCast(EmitScalarExpr(S.getTarget()), 589 Int8PtrTy, "addr"); 590 llvm::BasicBlock *CurBB = Builder.GetInsertBlock(); 591 592 // Get the basic block for the indirect goto. 593 llvm::BasicBlock *IndGotoBB = GetIndirectGotoBlock(); 594 595 // The first instruction in the block has to be the PHI for the switch dest, 596 // add an entry for this branch. 597 cast<llvm::PHINode>(IndGotoBB->begin())->addIncoming(V, CurBB); 598 599 EmitBranch(IndGotoBB); 600 } 601 602 void CodeGenFunction::EmitIfStmt(const IfStmt &S) { 603 // C99 6.8.4.1: The first substatement is executed if the expression compares 604 // unequal to 0. The condition must be a scalar type. 605 LexicalScope ConditionScope(*this, S.getCond()->getSourceRange()); 606 607 if (S.getInit()) 608 EmitStmt(S.getInit()); 609 610 if (S.getConditionVariable()) 611 EmitAutoVarDecl(*S.getConditionVariable()); 612 613 // If the condition constant folds and can be elided, try to avoid emitting 614 // the condition and the dead arm of the if/else. 615 bool CondConstant; 616 if (ConstantFoldsToSimpleInteger(S.getCond(), CondConstant, 617 S.isConstexpr())) { 618 // Figure out which block (then or else) is executed. 619 const Stmt *Executed = S.getThen(); 620 const Stmt *Skipped = S.getElse(); 621 if (!CondConstant) // Condition false? 622 std::swap(Executed, Skipped); 623 624 // If the skipped block has no labels in it, just emit the executed block. 625 // This avoids emitting dead code and simplifies the CFG substantially. 626 if (S.isConstexpr() || !ContainsLabel(Skipped)) { 627 if (CondConstant) 628 incrementProfileCounter(&S); 629 if (Executed) { 630 RunCleanupsScope ExecutedScope(*this); 631 EmitStmt(Executed); 632 } 633 return; 634 } 635 } 636 637 // Otherwise, the condition did not fold, or we couldn't elide it. Just emit 638 // the conditional branch. 639 llvm::BasicBlock *ThenBlock = createBasicBlock("if.then"); 640 llvm::BasicBlock *ContBlock = createBasicBlock("if.end"); 641 llvm::BasicBlock *ElseBlock = ContBlock; 642 if (S.getElse()) 643 ElseBlock = createBasicBlock("if.else"); 644 645 EmitBranchOnBoolExpr(S.getCond(), ThenBlock, ElseBlock, 646 getProfileCount(S.getThen())); 647 648 // Emit the 'then' code. 649 EmitBlock(ThenBlock); 650 incrementProfileCounter(&S); 651 { 652 RunCleanupsScope ThenScope(*this); 653 EmitStmt(S.getThen()); 654 } 655 EmitBranch(ContBlock); 656 657 // Emit the 'else' code if present. 658 if (const Stmt *Else = S.getElse()) { 659 { 660 // There is no need to emit line number for an unconditional branch. 661 auto NL = ApplyDebugLocation::CreateEmpty(*this); 662 EmitBlock(ElseBlock); 663 } 664 { 665 RunCleanupsScope ElseScope(*this); 666 EmitStmt(Else); 667 } 668 { 669 // There is no need to emit line number for an unconditional branch. 670 auto NL = ApplyDebugLocation::CreateEmpty(*this); 671 EmitBranch(ContBlock); 672 } 673 } 674 675 // Emit the continuation block for code after the if. 676 EmitBlock(ContBlock, true); 677 } 678 679 void CodeGenFunction::EmitWhileStmt(const WhileStmt &S, 680 ArrayRef<const Attr *> WhileAttrs) { 681 // Emit the header for the loop, which will also become 682 // the continue target. 683 JumpDest LoopHeader = getJumpDestInCurrentScope("while.cond"); 684 EmitBlock(LoopHeader.getBlock()); 685 686 const SourceRange &R = S.getSourceRange(); 687 LoopStack.push(LoopHeader.getBlock(), CGM.getContext(), WhileAttrs, 688 SourceLocToDebugLoc(R.getBegin()), 689 SourceLocToDebugLoc(R.getEnd())); 690 691 // Create an exit block for when the condition fails, which will 692 // also become the break target. 693 JumpDest LoopExit = getJumpDestInCurrentScope("while.end"); 694 695 // Store the blocks to use for break and continue. 696 BreakContinueStack.push_back(BreakContinue(LoopExit, LoopHeader)); 697 698 // C++ [stmt.while]p2: 699 // When the condition of a while statement is a declaration, the 700 // scope of the variable that is declared extends from its point 701 // of declaration (3.3.2) to the end of the while statement. 702 // [...] 703 // The object created in a condition is destroyed and created 704 // with each iteration of the loop. 705 RunCleanupsScope ConditionScope(*this); 706 707 if (S.getConditionVariable()) 708 EmitAutoVarDecl(*S.getConditionVariable()); 709 710 // Evaluate the conditional in the while header. C99 6.8.5.1: The 711 // evaluation of the controlling expression takes place before each 712 // execution of the loop body. 713 llvm::Value *BoolCondVal = EvaluateExprAsBool(S.getCond()); 714 715 // while(1) is common, avoid extra exit blocks. Be sure 716 // to correctly handle break/continue though. 717 bool EmitBoolCondBranch = true; 718 if (llvm::ConstantInt *C = dyn_cast<llvm::ConstantInt>(BoolCondVal)) 719 if (C->isOne()) 720 EmitBoolCondBranch = false; 721 722 // As long as the condition is true, go to the loop body. 723 llvm::BasicBlock *LoopBody = createBasicBlock("while.body"); 724 if (EmitBoolCondBranch) { 725 llvm::BasicBlock *ExitBlock = LoopExit.getBlock(); 726 if (ConditionScope.requiresCleanups()) 727 ExitBlock = createBasicBlock("while.exit"); 728 Builder.CreateCondBr( 729 BoolCondVal, LoopBody, ExitBlock, 730 createProfileWeightsForLoop(S.getCond(), getProfileCount(S.getBody()))); 731 732 if (ExitBlock != LoopExit.getBlock()) { 733 EmitBlock(ExitBlock); 734 EmitBranchThroughCleanup(LoopExit); 735 } 736 } 737 738 // Emit the loop body. We have to emit this in a cleanup scope 739 // because it might be a singleton DeclStmt. 740 { 741 RunCleanupsScope BodyScope(*this); 742 EmitBlock(LoopBody); 743 incrementProfileCounter(&S); 744 EmitStmt(S.getBody()); 745 } 746 747 BreakContinueStack.pop_back(); 748 749 // Immediately force cleanup. 750 ConditionScope.ForceCleanup(); 751 752 EmitStopPoint(&S); 753 // Branch to the loop header again. 754 EmitBranch(LoopHeader.getBlock()); 755 756 LoopStack.pop(); 757 758 // Emit the exit block. 759 EmitBlock(LoopExit.getBlock(), true); 760 761 // The LoopHeader typically is just a branch if we skipped emitting 762 // a branch, try to erase it. 763 if (!EmitBoolCondBranch) 764 SimplifyForwardingBlocks(LoopHeader.getBlock()); 765 } 766 767 void CodeGenFunction::EmitDoStmt(const DoStmt &S, 768 ArrayRef<const Attr *> DoAttrs) { 769 JumpDest LoopExit = getJumpDestInCurrentScope("do.end"); 770 JumpDest LoopCond = getJumpDestInCurrentScope("do.cond"); 771 772 uint64_t ParentCount = getCurrentProfileCount(); 773 774 // Store the blocks to use for break and continue. 775 BreakContinueStack.push_back(BreakContinue(LoopExit, LoopCond)); 776 777 // Emit the body of the loop. 778 llvm::BasicBlock *LoopBody = createBasicBlock("do.body"); 779 780 const SourceRange &R = S.getSourceRange(); 781 LoopStack.push(LoopBody, CGM.getContext(), DoAttrs, 782 SourceLocToDebugLoc(R.getBegin()), 783 SourceLocToDebugLoc(R.getEnd())); 784 785 EmitBlockWithFallThrough(LoopBody, &S); 786 { 787 RunCleanupsScope BodyScope(*this); 788 EmitStmt(S.getBody()); 789 } 790 791 EmitBlock(LoopCond.getBlock()); 792 793 // C99 6.8.5.2: "The evaluation of the controlling expression takes place 794 // after each execution of the loop body." 795 796 // Evaluate the conditional in the while header. 797 // C99 6.8.5p2/p4: The first substatement is executed if the expression 798 // compares unequal to 0. The condition must be a scalar type. 799 llvm::Value *BoolCondVal = EvaluateExprAsBool(S.getCond()); 800 801 BreakContinueStack.pop_back(); 802 803 // "do {} while (0)" is common in macros, avoid extra blocks. Be sure 804 // to correctly handle break/continue though. 805 bool EmitBoolCondBranch = true; 806 if (llvm::ConstantInt *C = dyn_cast<llvm::ConstantInt>(BoolCondVal)) 807 if (C->isZero()) 808 EmitBoolCondBranch = false; 809 810 // As long as the condition is true, iterate the loop. 811 if (EmitBoolCondBranch) { 812 uint64_t BackedgeCount = getProfileCount(S.getBody()) - ParentCount; 813 Builder.CreateCondBr( 814 BoolCondVal, LoopBody, LoopExit.getBlock(), 815 createProfileWeightsForLoop(S.getCond(), BackedgeCount)); 816 } 817 818 LoopStack.pop(); 819 820 // Emit the exit block. 821 EmitBlock(LoopExit.getBlock()); 822 823 // The DoCond block typically is just a branch if we skipped 824 // emitting a branch, try to erase it. 825 if (!EmitBoolCondBranch) 826 SimplifyForwardingBlocks(LoopCond.getBlock()); 827 } 828 829 void CodeGenFunction::EmitForStmt(const ForStmt &S, 830 ArrayRef<const Attr *> ForAttrs) { 831 JumpDest LoopExit = getJumpDestInCurrentScope("for.end"); 832 833 LexicalScope ForScope(*this, S.getSourceRange()); 834 835 // Evaluate the first part before the loop. 836 if (S.getInit()) 837 EmitStmt(S.getInit()); 838 839 // Start the loop with a block that tests the condition. 840 // If there's an increment, the continue scope will be overwritten 841 // later. 842 JumpDest Continue = getJumpDestInCurrentScope("for.cond"); 843 llvm::BasicBlock *CondBlock = Continue.getBlock(); 844 EmitBlock(CondBlock); 845 846 const SourceRange &R = S.getSourceRange(); 847 LoopStack.push(CondBlock, CGM.getContext(), ForAttrs, 848 SourceLocToDebugLoc(R.getBegin()), 849 SourceLocToDebugLoc(R.getEnd())); 850 851 // If the for loop doesn't have an increment we can just use the 852 // condition as the continue block. Otherwise we'll need to create 853 // a block for it (in the current scope, i.e. in the scope of the 854 // condition), and that we will become our continue block. 855 if (S.getInc()) 856 Continue = getJumpDestInCurrentScope("for.inc"); 857 858 // Store the blocks to use for break and continue. 859 BreakContinueStack.push_back(BreakContinue(LoopExit, Continue)); 860 861 // Create a cleanup scope for the condition variable cleanups. 862 LexicalScope ConditionScope(*this, S.getSourceRange()); 863 864 if (S.getCond()) { 865 // If the for statement has a condition scope, emit the local variable 866 // declaration. 867 if (S.getConditionVariable()) { 868 EmitAutoVarDecl(*S.getConditionVariable()); 869 } 870 871 llvm::BasicBlock *ExitBlock = LoopExit.getBlock(); 872 // If there are any cleanups between here and the loop-exit scope, 873 // create a block to stage a loop exit along. 874 if (ForScope.requiresCleanups()) 875 ExitBlock = createBasicBlock("for.cond.cleanup"); 876 877 // As long as the condition is true, iterate the loop. 878 llvm::BasicBlock *ForBody = createBasicBlock("for.body"); 879 880 // C99 6.8.5p2/p4: The first substatement is executed if the expression 881 // compares unequal to 0. The condition must be a scalar type. 882 llvm::Value *BoolCondVal = EvaluateExprAsBool(S.getCond()); 883 Builder.CreateCondBr( 884 BoolCondVal, ForBody, ExitBlock, 885 createProfileWeightsForLoop(S.getCond(), getProfileCount(S.getBody()))); 886 887 if (ExitBlock != LoopExit.getBlock()) { 888 EmitBlock(ExitBlock); 889 EmitBranchThroughCleanup(LoopExit); 890 } 891 892 EmitBlock(ForBody); 893 } else { 894 // Treat it as a non-zero constant. Don't even create a new block for the 895 // body, just fall into it. 896 } 897 incrementProfileCounter(&S); 898 899 { 900 // Create a separate cleanup scope for the body, in case it is not 901 // a compound statement. 902 RunCleanupsScope BodyScope(*this); 903 EmitStmt(S.getBody()); 904 } 905 906 // If there is an increment, emit it next. 907 if (S.getInc()) { 908 EmitBlock(Continue.getBlock()); 909 EmitStmt(S.getInc()); 910 } 911 912 BreakContinueStack.pop_back(); 913 914 ConditionScope.ForceCleanup(); 915 916 EmitStopPoint(&S); 917 EmitBranch(CondBlock); 918 919 ForScope.ForceCleanup(); 920 921 LoopStack.pop(); 922 923 // Emit the fall-through block. 924 EmitBlock(LoopExit.getBlock(), true); 925 } 926 927 void 928 CodeGenFunction::EmitCXXForRangeStmt(const CXXForRangeStmt &S, 929 ArrayRef<const Attr *> ForAttrs) { 930 JumpDest LoopExit = getJumpDestInCurrentScope("for.end"); 931 932 LexicalScope ForScope(*this, S.getSourceRange()); 933 934 // Evaluate the first pieces before the loop. 935 EmitStmt(S.getRangeStmt()); 936 EmitStmt(S.getBeginStmt()); 937 EmitStmt(S.getEndStmt()); 938 939 // Start the loop with a block that tests the condition. 940 // If there's an increment, the continue scope will be overwritten 941 // later. 942 llvm::BasicBlock *CondBlock = createBasicBlock("for.cond"); 943 EmitBlock(CondBlock); 944 945 const SourceRange &R = S.getSourceRange(); 946 LoopStack.push(CondBlock, CGM.getContext(), ForAttrs, 947 SourceLocToDebugLoc(R.getBegin()), 948 SourceLocToDebugLoc(R.getEnd())); 949 950 // If there are any cleanups between here and the loop-exit scope, 951 // create a block to stage a loop exit along. 952 llvm::BasicBlock *ExitBlock = LoopExit.getBlock(); 953 if (ForScope.requiresCleanups()) 954 ExitBlock = createBasicBlock("for.cond.cleanup"); 955 956 // The loop body, consisting of the specified body and the loop variable. 957 llvm::BasicBlock *ForBody = createBasicBlock("for.body"); 958 959 // The body is executed if the expression, contextually converted 960 // to bool, is true. 961 llvm::Value *BoolCondVal = EvaluateExprAsBool(S.getCond()); 962 Builder.CreateCondBr( 963 BoolCondVal, ForBody, ExitBlock, 964 createProfileWeightsForLoop(S.getCond(), getProfileCount(S.getBody()))); 965 966 if (ExitBlock != LoopExit.getBlock()) { 967 EmitBlock(ExitBlock); 968 EmitBranchThroughCleanup(LoopExit); 969 } 970 971 EmitBlock(ForBody); 972 incrementProfileCounter(&S); 973 974 // Create a block for the increment. In case of a 'continue', we jump there. 975 JumpDest Continue = getJumpDestInCurrentScope("for.inc"); 976 977 // Store the blocks to use for break and continue. 978 BreakContinueStack.push_back(BreakContinue(LoopExit, Continue)); 979 980 { 981 // Create a separate cleanup scope for the loop variable and body. 982 LexicalScope BodyScope(*this, S.getSourceRange()); 983 EmitStmt(S.getLoopVarStmt()); 984 EmitStmt(S.getBody()); 985 } 986 987 EmitStopPoint(&S); 988 // If there is an increment, emit it next. 989 EmitBlock(Continue.getBlock()); 990 EmitStmt(S.getInc()); 991 992 BreakContinueStack.pop_back(); 993 994 EmitBranch(CondBlock); 995 996 ForScope.ForceCleanup(); 997 998 LoopStack.pop(); 999 1000 // Emit the fall-through block. 1001 EmitBlock(LoopExit.getBlock(), true); 1002 } 1003 1004 void CodeGenFunction::EmitReturnOfRValue(RValue RV, QualType Ty) { 1005 if (RV.isScalar()) { 1006 Builder.CreateStore(RV.getScalarVal(), ReturnValue); 1007 } else if (RV.isAggregate()) { 1008 LValue Dest = MakeAddrLValue(ReturnValue, Ty); 1009 LValue Src = MakeAddrLValue(RV.getAggregateAddress(), Ty); 1010 EmitAggregateCopy(Dest, Src, Ty); 1011 } else { 1012 EmitStoreOfComplex(RV.getComplexVal(), MakeAddrLValue(ReturnValue, Ty), 1013 /*init*/ true); 1014 } 1015 EmitBranchThroughCleanup(ReturnBlock); 1016 } 1017 1018 /// EmitReturnStmt - Note that due to GCC extensions, this can have an operand 1019 /// if the function returns void, or may be missing one if the function returns 1020 /// non-void. Fun stuff :). 1021 void CodeGenFunction::EmitReturnStmt(const ReturnStmt &S) { 1022 if (requiresReturnValueCheck()) { 1023 llvm::Constant *SLoc = EmitCheckSourceLocation(S.getLocStart()); 1024 auto *SLocPtr = 1025 new llvm::GlobalVariable(CGM.getModule(), SLoc->getType(), false, 1026 llvm::GlobalVariable::PrivateLinkage, SLoc); 1027 SLocPtr->setUnnamedAddr(llvm::GlobalValue::UnnamedAddr::Global); 1028 CGM.getSanitizerMetadata()->disableSanitizerForGlobal(SLocPtr); 1029 assert(ReturnLocation.isValid() && "No valid return location"); 1030 Builder.CreateStore(Builder.CreateBitCast(SLocPtr, Int8PtrTy), 1031 ReturnLocation); 1032 } 1033 1034 // Returning from an outlined SEH helper is UB, and we already warn on it. 1035 if (IsOutlinedSEHHelper) { 1036 Builder.CreateUnreachable(); 1037 Builder.ClearInsertionPoint(); 1038 } 1039 1040 // Emit the result value, even if unused, to evalute the side effects. 1041 const Expr *RV = S.getRetValue(); 1042 1043 // Treat block literals in a return expression as if they appeared 1044 // in their own scope. This permits a small, easily-implemented 1045 // exception to our over-conservative rules about not jumping to 1046 // statements following block literals with non-trivial cleanups. 1047 RunCleanupsScope cleanupScope(*this); 1048 if (const ExprWithCleanups *cleanups = 1049 dyn_cast_or_null<ExprWithCleanups>(RV)) { 1050 enterFullExpression(cleanups); 1051 RV = cleanups->getSubExpr(); 1052 } 1053 1054 // FIXME: Clean this up by using an LValue for ReturnTemp, 1055 // EmitStoreThroughLValue, and EmitAnyExpr. 1056 if (getLangOpts().ElideConstructors && 1057 S.getNRVOCandidate() && S.getNRVOCandidate()->isNRVOVariable()) { 1058 // Apply the named return value optimization for this return statement, 1059 // which means doing nothing: the appropriate result has already been 1060 // constructed into the NRVO variable. 1061 1062 // If there is an NRVO flag for this variable, set it to 1 into indicate 1063 // that the cleanup code should not destroy the variable. 1064 if (llvm::Value *NRVOFlag = NRVOFlags[S.getNRVOCandidate()]) 1065 Builder.CreateFlagStore(Builder.getTrue(), NRVOFlag); 1066 } else if (!ReturnValue.isValid() || (RV && RV->getType()->isVoidType())) { 1067 // Make sure not to return anything, but evaluate the expression 1068 // for side effects. 1069 if (RV) 1070 EmitAnyExpr(RV); 1071 } else if (!RV) { 1072 // Do nothing (return value is left uninitialized) 1073 } else if (FnRetTy->isReferenceType()) { 1074 // If this function returns a reference, take the address of the expression 1075 // rather than the value. 1076 RValue Result = EmitReferenceBindingToExpr(RV); 1077 Builder.CreateStore(Result.getScalarVal(), ReturnValue); 1078 } else { 1079 switch (getEvaluationKind(RV->getType())) { 1080 case TEK_Scalar: 1081 Builder.CreateStore(EmitScalarExpr(RV), ReturnValue); 1082 break; 1083 case TEK_Complex: 1084 EmitComplexExprIntoLValue(RV, MakeAddrLValue(ReturnValue, RV->getType()), 1085 /*isInit*/ true); 1086 break; 1087 case TEK_Aggregate: 1088 EmitAggExpr(RV, AggValueSlot::forAddr(ReturnValue, 1089 Qualifiers(), 1090 AggValueSlot::IsDestructed, 1091 AggValueSlot::DoesNotNeedGCBarriers, 1092 AggValueSlot::IsNotAliased)); 1093 break; 1094 } 1095 } 1096 1097 ++NumReturnExprs; 1098 if (!RV || RV->isEvaluatable(getContext())) 1099 ++NumSimpleReturnExprs; 1100 1101 cleanupScope.ForceCleanup(); 1102 EmitBranchThroughCleanup(ReturnBlock); 1103 } 1104 1105 void CodeGenFunction::EmitDeclStmt(const DeclStmt &S) { 1106 // As long as debug info is modeled with instructions, we have to ensure we 1107 // have a place to insert here and write the stop point here. 1108 if (HaveInsertPoint()) 1109 EmitStopPoint(&S); 1110 1111 for (const auto *I : S.decls()) 1112 EmitDecl(*I); 1113 } 1114 1115 void CodeGenFunction::EmitBreakStmt(const BreakStmt &S) { 1116 assert(!BreakContinueStack.empty() && "break stmt not in a loop or switch!"); 1117 1118 // If this code is reachable then emit a stop point (if generating 1119 // debug info). We have to do this ourselves because we are on the 1120 // "simple" statement path. 1121 if (HaveInsertPoint()) 1122 EmitStopPoint(&S); 1123 1124 EmitBranchThroughCleanup(BreakContinueStack.back().BreakBlock); 1125 } 1126 1127 void CodeGenFunction::EmitContinueStmt(const ContinueStmt &S) { 1128 assert(!BreakContinueStack.empty() && "continue stmt not in a loop!"); 1129 1130 // If this code is reachable then emit a stop point (if generating 1131 // debug info). We have to do this ourselves because we are on the 1132 // "simple" statement path. 1133 if (HaveInsertPoint()) 1134 EmitStopPoint(&S); 1135 1136 EmitBranchThroughCleanup(BreakContinueStack.back().ContinueBlock); 1137 } 1138 1139 /// EmitCaseStmtRange - If case statement range is not too big then 1140 /// add multiple cases to switch instruction, one for each value within 1141 /// the range. If range is too big then emit "if" condition check. 1142 void CodeGenFunction::EmitCaseStmtRange(const CaseStmt &S) { 1143 assert(S.getRHS() && "Expected RHS value in CaseStmt"); 1144 1145 llvm::APSInt LHS = S.getLHS()->EvaluateKnownConstInt(getContext()); 1146 llvm::APSInt RHS = S.getRHS()->EvaluateKnownConstInt(getContext()); 1147 1148 // Emit the code for this case. We do this first to make sure it is 1149 // properly chained from our predecessor before generating the 1150 // switch machinery to enter this block. 1151 llvm::BasicBlock *CaseDest = createBasicBlock("sw.bb"); 1152 EmitBlockWithFallThrough(CaseDest, &S); 1153 EmitStmt(S.getSubStmt()); 1154 1155 // If range is empty, do nothing. 1156 if (LHS.isSigned() ? RHS.slt(LHS) : RHS.ult(LHS)) 1157 return; 1158 1159 llvm::APInt Range = RHS - LHS; 1160 // FIXME: parameters such as this should not be hardcoded. 1161 if (Range.ult(llvm::APInt(Range.getBitWidth(), 64))) { 1162 // Range is small enough to add multiple switch instruction cases. 1163 uint64_t Total = getProfileCount(&S); 1164 unsigned NCases = Range.getZExtValue() + 1; 1165 // We only have one region counter for the entire set of cases here, so we 1166 // need to divide the weights evenly between the generated cases, ensuring 1167 // that the total weight is preserved. E.g., a weight of 5 over three cases 1168 // will be distributed as weights of 2, 2, and 1. 1169 uint64_t Weight = Total / NCases, Rem = Total % NCases; 1170 for (unsigned I = 0; I != NCases; ++I) { 1171 if (SwitchWeights) 1172 SwitchWeights->push_back(Weight + (Rem ? 1 : 0)); 1173 if (Rem) 1174 Rem--; 1175 SwitchInsn->addCase(Builder.getInt(LHS), CaseDest); 1176 ++LHS; 1177 } 1178 return; 1179 } 1180 1181 // The range is too big. Emit "if" condition into a new block, 1182 // making sure to save and restore the current insertion point. 1183 llvm::BasicBlock *RestoreBB = Builder.GetInsertBlock(); 1184 1185 // Push this test onto the chain of range checks (which terminates 1186 // in the default basic block). The switch's default will be changed 1187 // to the top of this chain after switch emission is complete. 1188 llvm::BasicBlock *FalseDest = CaseRangeBlock; 1189 CaseRangeBlock = createBasicBlock("sw.caserange"); 1190 1191 CurFn->getBasicBlockList().push_back(CaseRangeBlock); 1192 Builder.SetInsertPoint(CaseRangeBlock); 1193 1194 // Emit range check. 1195 llvm::Value *Diff = 1196 Builder.CreateSub(SwitchInsn->getCondition(), Builder.getInt(LHS)); 1197 llvm::Value *Cond = 1198 Builder.CreateICmpULE(Diff, Builder.getInt(Range), "inbounds"); 1199 1200 llvm::MDNode *Weights = nullptr; 1201 if (SwitchWeights) { 1202 uint64_t ThisCount = getProfileCount(&S); 1203 uint64_t DefaultCount = (*SwitchWeights)[0]; 1204 Weights = createProfileWeights(ThisCount, DefaultCount); 1205 1206 // Since we're chaining the switch default through each large case range, we 1207 // need to update the weight for the default, ie, the first case, to include 1208 // this case. 1209 (*SwitchWeights)[0] += ThisCount; 1210 } 1211 Builder.CreateCondBr(Cond, CaseDest, FalseDest, Weights); 1212 1213 // Restore the appropriate insertion point. 1214 if (RestoreBB) 1215 Builder.SetInsertPoint(RestoreBB); 1216 else 1217 Builder.ClearInsertionPoint(); 1218 } 1219 1220 void CodeGenFunction::EmitCaseStmt(const CaseStmt &S) { 1221 // If there is no enclosing switch instance that we're aware of, then this 1222 // case statement and its block can be elided. This situation only happens 1223 // when we've constant-folded the switch, are emitting the constant case, 1224 // and part of the constant case includes another case statement. For 1225 // instance: switch (4) { case 4: do { case 5: } while (1); } 1226 if (!SwitchInsn) { 1227 EmitStmt(S.getSubStmt()); 1228 return; 1229 } 1230 1231 // Handle case ranges. 1232 if (S.getRHS()) { 1233 EmitCaseStmtRange(S); 1234 return; 1235 } 1236 1237 llvm::ConstantInt *CaseVal = 1238 Builder.getInt(S.getLHS()->EvaluateKnownConstInt(getContext())); 1239 1240 // If the body of the case is just a 'break', try to not emit an empty block. 1241 // If we're profiling or we're not optimizing, leave the block in for better 1242 // debug and coverage analysis. 1243 if (!CGM.getCodeGenOpts().hasProfileClangInstr() && 1244 CGM.getCodeGenOpts().OptimizationLevel > 0 && 1245 isa<BreakStmt>(S.getSubStmt())) { 1246 JumpDest Block = BreakContinueStack.back().BreakBlock; 1247 1248 // Only do this optimization if there are no cleanups that need emitting. 1249 if (isObviouslyBranchWithoutCleanups(Block)) { 1250 if (SwitchWeights) 1251 SwitchWeights->push_back(getProfileCount(&S)); 1252 SwitchInsn->addCase(CaseVal, Block.getBlock()); 1253 1254 // If there was a fallthrough into this case, make sure to redirect it to 1255 // the end of the switch as well. 1256 if (Builder.GetInsertBlock()) { 1257 Builder.CreateBr(Block.getBlock()); 1258 Builder.ClearInsertionPoint(); 1259 } 1260 return; 1261 } 1262 } 1263 1264 llvm::BasicBlock *CaseDest = createBasicBlock("sw.bb"); 1265 EmitBlockWithFallThrough(CaseDest, &S); 1266 if (SwitchWeights) 1267 SwitchWeights->push_back(getProfileCount(&S)); 1268 SwitchInsn->addCase(CaseVal, CaseDest); 1269 1270 // Recursively emitting the statement is acceptable, but is not wonderful for 1271 // code where we have many case statements nested together, i.e.: 1272 // case 1: 1273 // case 2: 1274 // case 3: etc. 1275 // Handling this recursively will create a new block for each case statement 1276 // that falls through to the next case which is IR intensive. It also causes 1277 // deep recursion which can run into stack depth limitations. Handle 1278 // sequential non-range case statements specially. 1279 const CaseStmt *CurCase = &S; 1280 const CaseStmt *NextCase = dyn_cast<CaseStmt>(S.getSubStmt()); 1281 1282 // Otherwise, iteratively add consecutive cases to this switch stmt. 1283 while (NextCase && NextCase->getRHS() == nullptr) { 1284 CurCase = NextCase; 1285 llvm::ConstantInt *CaseVal = 1286 Builder.getInt(CurCase->getLHS()->EvaluateKnownConstInt(getContext())); 1287 1288 if (SwitchWeights) 1289 SwitchWeights->push_back(getProfileCount(NextCase)); 1290 if (CGM.getCodeGenOpts().hasProfileClangInstr()) { 1291 CaseDest = createBasicBlock("sw.bb"); 1292 EmitBlockWithFallThrough(CaseDest, &S); 1293 } 1294 1295 SwitchInsn->addCase(CaseVal, CaseDest); 1296 NextCase = dyn_cast<CaseStmt>(CurCase->getSubStmt()); 1297 } 1298 1299 // Normal default recursion for non-cases. 1300 EmitStmt(CurCase->getSubStmt()); 1301 } 1302 1303 void CodeGenFunction::EmitDefaultStmt(const DefaultStmt &S) { 1304 // If there is no enclosing switch instance that we're aware of, then this 1305 // default statement can be elided. This situation only happens when we've 1306 // constant-folded the switch. 1307 if (!SwitchInsn) { 1308 EmitStmt(S.getSubStmt()); 1309 return; 1310 } 1311 1312 llvm::BasicBlock *DefaultBlock = SwitchInsn->getDefaultDest(); 1313 assert(DefaultBlock->empty() && 1314 "EmitDefaultStmt: Default block already defined?"); 1315 1316 EmitBlockWithFallThrough(DefaultBlock, &S); 1317 1318 EmitStmt(S.getSubStmt()); 1319 } 1320 1321 /// CollectStatementsForCase - Given the body of a 'switch' statement and a 1322 /// constant value that is being switched on, see if we can dead code eliminate 1323 /// the body of the switch to a simple series of statements to emit. Basically, 1324 /// on a switch (5) we want to find these statements: 1325 /// case 5: 1326 /// printf(...); <-- 1327 /// ++i; <-- 1328 /// break; 1329 /// 1330 /// and add them to the ResultStmts vector. If it is unsafe to do this 1331 /// transformation (for example, one of the elided statements contains a label 1332 /// that might be jumped to), return CSFC_Failure. If we handled it and 'S' 1333 /// should include statements after it (e.g. the printf() line is a substmt of 1334 /// the case) then return CSFC_FallThrough. If we handled it and found a break 1335 /// statement, then return CSFC_Success. 1336 /// 1337 /// If Case is non-null, then we are looking for the specified case, checking 1338 /// that nothing we jump over contains labels. If Case is null, then we found 1339 /// the case and are looking for the break. 1340 /// 1341 /// If the recursive walk actually finds our Case, then we set FoundCase to 1342 /// true. 1343 /// 1344 enum CSFC_Result { CSFC_Failure, CSFC_FallThrough, CSFC_Success }; 1345 static CSFC_Result CollectStatementsForCase(const Stmt *S, 1346 const SwitchCase *Case, 1347 bool &FoundCase, 1348 SmallVectorImpl<const Stmt*> &ResultStmts) { 1349 // If this is a null statement, just succeed. 1350 if (!S) 1351 return Case ? CSFC_Success : CSFC_FallThrough; 1352 1353 // If this is the switchcase (case 4: or default) that we're looking for, then 1354 // we're in business. Just add the substatement. 1355 if (const SwitchCase *SC = dyn_cast<SwitchCase>(S)) { 1356 if (S == Case) { 1357 FoundCase = true; 1358 return CollectStatementsForCase(SC->getSubStmt(), nullptr, FoundCase, 1359 ResultStmts); 1360 } 1361 1362 // Otherwise, this is some other case or default statement, just ignore it. 1363 return CollectStatementsForCase(SC->getSubStmt(), Case, FoundCase, 1364 ResultStmts); 1365 } 1366 1367 // If we are in the live part of the code and we found our break statement, 1368 // return a success! 1369 if (!Case && isa<BreakStmt>(S)) 1370 return CSFC_Success; 1371 1372 // If this is a switch statement, then it might contain the SwitchCase, the 1373 // break, or neither. 1374 if (const CompoundStmt *CS = dyn_cast<CompoundStmt>(S)) { 1375 // Handle this as two cases: we might be looking for the SwitchCase (if so 1376 // the skipped statements must be skippable) or we might already have it. 1377 CompoundStmt::const_body_iterator I = CS->body_begin(), E = CS->body_end(); 1378 bool StartedInLiveCode = FoundCase; 1379 unsigned StartSize = ResultStmts.size(); 1380 1381 // If we've not found the case yet, scan through looking for it. 1382 if (Case) { 1383 // Keep track of whether we see a skipped declaration. The code could be 1384 // using the declaration even if it is skipped, so we can't optimize out 1385 // the decl if the kept statements might refer to it. 1386 bool HadSkippedDecl = false; 1387 1388 // If we're looking for the case, just see if we can skip each of the 1389 // substatements. 1390 for (; Case && I != E; ++I) { 1391 HadSkippedDecl |= CodeGenFunction::mightAddDeclToScope(*I); 1392 1393 switch (CollectStatementsForCase(*I, Case, FoundCase, ResultStmts)) { 1394 case CSFC_Failure: return CSFC_Failure; 1395 case CSFC_Success: 1396 // A successful result means that either 1) that the statement doesn't 1397 // have the case and is skippable, or 2) does contain the case value 1398 // and also contains the break to exit the switch. In the later case, 1399 // we just verify the rest of the statements are elidable. 1400 if (FoundCase) { 1401 // If we found the case and skipped declarations, we can't do the 1402 // optimization. 1403 if (HadSkippedDecl) 1404 return CSFC_Failure; 1405 1406 for (++I; I != E; ++I) 1407 if (CodeGenFunction::ContainsLabel(*I, true)) 1408 return CSFC_Failure; 1409 return CSFC_Success; 1410 } 1411 break; 1412 case CSFC_FallThrough: 1413 // If we have a fallthrough condition, then we must have found the 1414 // case started to include statements. Consider the rest of the 1415 // statements in the compound statement as candidates for inclusion. 1416 assert(FoundCase && "Didn't find case but returned fallthrough?"); 1417 // We recursively found Case, so we're not looking for it anymore. 1418 Case = nullptr; 1419 1420 // If we found the case and skipped declarations, we can't do the 1421 // optimization. 1422 if (HadSkippedDecl) 1423 return CSFC_Failure; 1424 break; 1425 } 1426 } 1427 1428 if (!FoundCase) 1429 return CSFC_Success; 1430 1431 assert(!HadSkippedDecl && "fallthrough after skipping decl"); 1432 } 1433 1434 // If we have statements in our range, then we know that the statements are 1435 // live and need to be added to the set of statements we're tracking. 1436 bool AnyDecls = false; 1437 for (; I != E; ++I) { 1438 AnyDecls |= CodeGenFunction::mightAddDeclToScope(*I); 1439 1440 switch (CollectStatementsForCase(*I, nullptr, FoundCase, ResultStmts)) { 1441 case CSFC_Failure: return CSFC_Failure; 1442 case CSFC_FallThrough: 1443 // A fallthrough result means that the statement was simple and just 1444 // included in ResultStmt, keep adding them afterwards. 1445 break; 1446 case CSFC_Success: 1447 // A successful result means that we found the break statement and 1448 // stopped statement inclusion. We just ensure that any leftover stmts 1449 // are skippable and return success ourselves. 1450 for (++I; I != E; ++I) 1451 if (CodeGenFunction::ContainsLabel(*I, true)) 1452 return CSFC_Failure; 1453 return CSFC_Success; 1454 } 1455 } 1456 1457 // If we're about to fall out of a scope without hitting a 'break;', we 1458 // can't perform the optimization if there were any decls in that scope 1459 // (we'd lose their end-of-lifetime). 1460 if (AnyDecls) { 1461 // If the entire compound statement was live, there's one more thing we 1462 // can try before giving up: emit the whole thing as a single statement. 1463 // We can do that unless the statement contains a 'break;'. 1464 // FIXME: Such a break must be at the end of a construct within this one. 1465 // We could emit this by just ignoring the BreakStmts entirely. 1466 if (StartedInLiveCode && !CodeGenFunction::containsBreak(S)) { 1467 ResultStmts.resize(StartSize); 1468 ResultStmts.push_back(S); 1469 } else { 1470 return CSFC_Failure; 1471 } 1472 } 1473 1474 return CSFC_FallThrough; 1475 } 1476 1477 // Okay, this is some other statement that we don't handle explicitly, like a 1478 // for statement or increment etc. If we are skipping over this statement, 1479 // just verify it doesn't have labels, which would make it invalid to elide. 1480 if (Case) { 1481 if (CodeGenFunction::ContainsLabel(S, true)) 1482 return CSFC_Failure; 1483 return CSFC_Success; 1484 } 1485 1486 // Otherwise, we want to include this statement. Everything is cool with that 1487 // so long as it doesn't contain a break out of the switch we're in. 1488 if (CodeGenFunction::containsBreak(S)) return CSFC_Failure; 1489 1490 // Otherwise, everything is great. Include the statement and tell the caller 1491 // that we fall through and include the next statement as well. 1492 ResultStmts.push_back(S); 1493 return CSFC_FallThrough; 1494 } 1495 1496 /// FindCaseStatementsForValue - Find the case statement being jumped to and 1497 /// then invoke CollectStatementsForCase to find the list of statements to emit 1498 /// for a switch on constant. See the comment above CollectStatementsForCase 1499 /// for more details. 1500 static bool FindCaseStatementsForValue(const SwitchStmt &S, 1501 const llvm::APSInt &ConstantCondValue, 1502 SmallVectorImpl<const Stmt*> &ResultStmts, 1503 ASTContext &C, 1504 const SwitchCase *&ResultCase) { 1505 // First step, find the switch case that is being branched to. We can do this 1506 // efficiently by scanning the SwitchCase list. 1507 const SwitchCase *Case = S.getSwitchCaseList(); 1508 const DefaultStmt *DefaultCase = nullptr; 1509 1510 for (; Case; Case = Case->getNextSwitchCase()) { 1511 // It's either a default or case. Just remember the default statement in 1512 // case we're not jumping to any numbered cases. 1513 if (const DefaultStmt *DS = dyn_cast<DefaultStmt>(Case)) { 1514 DefaultCase = DS; 1515 continue; 1516 } 1517 1518 // Check to see if this case is the one we're looking for. 1519 const CaseStmt *CS = cast<CaseStmt>(Case); 1520 // Don't handle case ranges yet. 1521 if (CS->getRHS()) return false; 1522 1523 // If we found our case, remember it as 'case'. 1524 if (CS->getLHS()->EvaluateKnownConstInt(C) == ConstantCondValue) 1525 break; 1526 } 1527 1528 // If we didn't find a matching case, we use a default if it exists, or we 1529 // elide the whole switch body! 1530 if (!Case) { 1531 // It is safe to elide the body of the switch if it doesn't contain labels 1532 // etc. If it is safe, return successfully with an empty ResultStmts list. 1533 if (!DefaultCase) 1534 return !CodeGenFunction::ContainsLabel(&S); 1535 Case = DefaultCase; 1536 } 1537 1538 // Ok, we know which case is being jumped to, try to collect all the 1539 // statements that follow it. This can fail for a variety of reasons. Also, 1540 // check to see that the recursive walk actually found our case statement. 1541 // Insane cases like this can fail to find it in the recursive walk since we 1542 // don't handle every stmt kind: 1543 // switch (4) { 1544 // while (1) { 1545 // case 4: ... 1546 bool FoundCase = false; 1547 ResultCase = Case; 1548 return CollectStatementsForCase(S.getBody(), Case, FoundCase, 1549 ResultStmts) != CSFC_Failure && 1550 FoundCase; 1551 } 1552 1553 void CodeGenFunction::EmitSwitchStmt(const SwitchStmt &S) { 1554 // Handle nested switch statements. 1555 llvm::SwitchInst *SavedSwitchInsn = SwitchInsn; 1556 SmallVector<uint64_t, 16> *SavedSwitchWeights = SwitchWeights; 1557 llvm::BasicBlock *SavedCRBlock = CaseRangeBlock; 1558 1559 // See if we can constant fold the condition of the switch and therefore only 1560 // emit the live case statement (if any) of the switch. 1561 llvm::APSInt ConstantCondValue; 1562 if (ConstantFoldsToSimpleInteger(S.getCond(), ConstantCondValue)) { 1563 SmallVector<const Stmt*, 4> CaseStmts; 1564 const SwitchCase *Case = nullptr; 1565 if (FindCaseStatementsForValue(S, ConstantCondValue, CaseStmts, 1566 getContext(), Case)) { 1567 if (Case) 1568 incrementProfileCounter(Case); 1569 RunCleanupsScope ExecutedScope(*this); 1570 1571 if (S.getInit()) 1572 EmitStmt(S.getInit()); 1573 1574 // Emit the condition variable if needed inside the entire cleanup scope 1575 // used by this special case for constant folded switches. 1576 if (S.getConditionVariable()) 1577 EmitAutoVarDecl(*S.getConditionVariable()); 1578 1579 // At this point, we are no longer "within" a switch instance, so 1580 // we can temporarily enforce this to ensure that any embedded case 1581 // statements are not emitted. 1582 SwitchInsn = nullptr; 1583 1584 // Okay, we can dead code eliminate everything except this case. Emit the 1585 // specified series of statements and we're good. 1586 for (unsigned i = 0, e = CaseStmts.size(); i != e; ++i) 1587 EmitStmt(CaseStmts[i]); 1588 incrementProfileCounter(&S); 1589 1590 // Now we want to restore the saved switch instance so that nested 1591 // switches continue to function properly 1592 SwitchInsn = SavedSwitchInsn; 1593 1594 return; 1595 } 1596 } 1597 1598 JumpDest SwitchExit = getJumpDestInCurrentScope("sw.epilog"); 1599 1600 RunCleanupsScope ConditionScope(*this); 1601 1602 if (S.getInit()) 1603 EmitStmt(S.getInit()); 1604 1605 if (S.getConditionVariable()) 1606 EmitAutoVarDecl(*S.getConditionVariable()); 1607 llvm::Value *CondV = EmitScalarExpr(S.getCond()); 1608 1609 // Create basic block to hold stuff that comes after switch 1610 // statement. We also need to create a default block now so that 1611 // explicit case ranges tests can have a place to jump to on 1612 // failure. 1613 llvm::BasicBlock *DefaultBlock = createBasicBlock("sw.default"); 1614 SwitchInsn = Builder.CreateSwitch(CondV, DefaultBlock); 1615 if (PGO.haveRegionCounts()) { 1616 // Walk the SwitchCase list to find how many there are. 1617 uint64_t DefaultCount = 0; 1618 unsigned NumCases = 0; 1619 for (const SwitchCase *Case = S.getSwitchCaseList(); 1620 Case; 1621 Case = Case->getNextSwitchCase()) { 1622 if (isa<DefaultStmt>(Case)) 1623 DefaultCount = getProfileCount(Case); 1624 NumCases += 1; 1625 } 1626 SwitchWeights = new SmallVector<uint64_t, 16>(); 1627 SwitchWeights->reserve(NumCases); 1628 // The default needs to be first. We store the edge count, so we already 1629 // know the right weight. 1630 SwitchWeights->push_back(DefaultCount); 1631 } 1632 CaseRangeBlock = DefaultBlock; 1633 1634 // Clear the insertion point to indicate we are in unreachable code. 1635 Builder.ClearInsertionPoint(); 1636 1637 // All break statements jump to NextBlock. If BreakContinueStack is non-empty 1638 // then reuse last ContinueBlock. 1639 JumpDest OuterContinue; 1640 if (!BreakContinueStack.empty()) 1641 OuterContinue = BreakContinueStack.back().ContinueBlock; 1642 1643 BreakContinueStack.push_back(BreakContinue(SwitchExit, OuterContinue)); 1644 1645 // Emit switch body. 1646 EmitStmt(S.getBody()); 1647 1648 BreakContinueStack.pop_back(); 1649 1650 // Update the default block in case explicit case range tests have 1651 // been chained on top. 1652 SwitchInsn->setDefaultDest(CaseRangeBlock); 1653 1654 // If a default was never emitted: 1655 if (!DefaultBlock->getParent()) { 1656 // If we have cleanups, emit the default block so that there's a 1657 // place to jump through the cleanups from. 1658 if (ConditionScope.requiresCleanups()) { 1659 EmitBlock(DefaultBlock); 1660 1661 // Otherwise, just forward the default block to the switch end. 1662 } else { 1663 DefaultBlock->replaceAllUsesWith(SwitchExit.getBlock()); 1664 delete DefaultBlock; 1665 } 1666 } 1667 1668 ConditionScope.ForceCleanup(); 1669 1670 // Emit continuation. 1671 EmitBlock(SwitchExit.getBlock(), true); 1672 incrementProfileCounter(&S); 1673 1674 // If the switch has a condition wrapped by __builtin_unpredictable, 1675 // create metadata that specifies that the switch is unpredictable. 1676 // Don't bother if not optimizing because that metadata would not be used. 1677 auto *Call = dyn_cast<CallExpr>(S.getCond()); 1678 if (Call && CGM.getCodeGenOpts().OptimizationLevel != 0) { 1679 auto *FD = dyn_cast_or_null<FunctionDecl>(Call->getCalleeDecl()); 1680 if (FD && FD->getBuiltinID() == Builtin::BI__builtin_unpredictable) { 1681 llvm::MDBuilder MDHelper(getLLVMContext()); 1682 SwitchInsn->setMetadata(llvm::LLVMContext::MD_unpredictable, 1683 MDHelper.createUnpredictable()); 1684 } 1685 } 1686 1687 if (SwitchWeights) { 1688 assert(SwitchWeights->size() == 1 + SwitchInsn->getNumCases() && 1689 "switch weights do not match switch cases"); 1690 // If there's only one jump destination there's no sense weighting it. 1691 if (SwitchWeights->size() > 1) 1692 SwitchInsn->setMetadata(llvm::LLVMContext::MD_prof, 1693 createProfileWeights(*SwitchWeights)); 1694 delete SwitchWeights; 1695 } 1696 SwitchInsn = SavedSwitchInsn; 1697 SwitchWeights = SavedSwitchWeights; 1698 CaseRangeBlock = SavedCRBlock; 1699 } 1700 1701 static std::string 1702 SimplifyConstraint(const char *Constraint, const TargetInfo &Target, 1703 SmallVectorImpl<TargetInfo::ConstraintInfo> *OutCons=nullptr) { 1704 std::string Result; 1705 1706 while (*Constraint) { 1707 switch (*Constraint) { 1708 default: 1709 Result += Target.convertConstraint(Constraint); 1710 break; 1711 // Ignore these 1712 case '*': 1713 case '?': 1714 case '!': 1715 case '=': // Will see this and the following in mult-alt constraints. 1716 case '+': 1717 break; 1718 case '#': // Ignore the rest of the constraint alternative. 1719 while (Constraint[1] && Constraint[1] != ',') 1720 Constraint++; 1721 break; 1722 case '&': 1723 case '%': 1724 Result += *Constraint; 1725 while (Constraint[1] && Constraint[1] == *Constraint) 1726 Constraint++; 1727 break; 1728 case ',': 1729 Result += "|"; 1730 break; 1731 case 'g': 1732 Result += "imr"; 1733 break; 1734 case '[': { 1735 assert(OutCons && 1736 "Must pass output names to constraints with a symbolic name"); 1737 unsigned Index; 1738 bool result = Target.resolveSymbolicName(Constraint, *OutCons, Index); 1739 assert(result && "Could not resolve symbolic name"); (void)result; 1740 Result += llvm::utostr(Index); 1741 break; 1742 } 1743 } 1744 1745 Constraint++; 1746 } 1747 1748 return Result; 1749 } 1750 1751 /// AddVariableConstraints - Look at AsmExpr and if it is a variable declared 1752 /// as using a particular register add that as a constraint that will be used 1753 /// in this asm stmt. 1754 static std::string 1755 AddVariableConstraints(const std::string &Constraint, const Expr &AsmExpr, 1756 const TargetInfo &Target, CodeGenModule &CGM, 1757 const AsmStmt &Stmt, const bool EarlyClobber) { 1758 const DeclRefExpr *AsmDeclRef = dyn_cast<DeclRefExpr>(&AsmExpr); 1759 if (!AsmDeclRef) 1760 return Constraint; 1761 const ValueDecl &Value = *AsmDeclRef->getDecl(); 1762 const VarDecl *Variable = dyn_cast<VarDecl>(&Value); 1763 if (!Variable) 1764 return Constraint; 1765 if (Variable->getStorageClass() != SC_Register) 1766 return Constraint; 1767 AsmLabelAttr *Attr = Variable->getAttr<AsmLabelAttr>(); 1768 if (!Attr) 1769 return Constraint; 1770 StringRef Register = Attr->getLabel(); 1771 assert(Target.isValidGCCRegisterName(Register)); 1772 // We're using validateOutputConstraint here because we only care if 1773 // this is a register constraint. 1774 TargetInfo::ConstraintInfo Info(Constraint, ""); 1775 if (Target.validateOutputConstraint(Info) && 1776 !Info.allowsRegister()) { 1777 CGM.ErrorUnsupported(&Stmt, "__asm__"); 1778 return Constraint; 1779 } 1780 // Canonicalize the register here before returning it. 1781 Register = Target.getNormalizedGCCRegisterName(Register); 1782 return (EarlyClobber ? "&{" : "{") + Register.str() + "}"; 1783 } 1784 1785 llvm::Value* 1786 CodeGenFunction::EmitAsmInputLValue(const TargetInfo::ConstraintInfo &Info, 1787 LValue InputValue, QualType InputType, 1788 std::string &ConstraintStr, 1789 SourceLocation Loc) { 1790 llvm::Value *Arg; 1791 if (Info.allowsRegister() || !Info.allowsMemory()) { 1792 if (CodeGenFunction::hasScalarEvaluationKind(InputType)) { 1793 Arg = EmitLoadOfLValue(InputValue, Loc).getScalarVal(); 1794 } else { 1795 llvm::Type *Ty = ConvertType(InputType); 1796 uint64_t Size = CGM.getDataLayout().getTypeSizeInBits(Ty); 1797 if (Size <= 64 && llvm::isPowerOf2_64(Size)) { 1798 Ty = llvm::IntegerType::get(getLLVMContext(), Size); 1799 Ty = llvm::PointerType::getUnqual(Ty); 1800 1801 Arg = Builder.CreateLoad(Builder.CreateBitCast(InputValue.getAddress(), 1802 Ty)); 1803 } else { 1804 Arg = InputValue.getPointer(); 1805 ConstraintStr += '*'; 1806 } 1807 } 1808 } else { 1809 Arg = InputValue.getPointer(); 1810 ConstraintStr += '*'; 1811 } 1812 1813 return Arg; 1814 } 1815 1816 llvm::Value* CodeGenFunction::EmitAsmInput( 1817 const TargetInfo::ConstraintInfo &Info, 1818 const Expr *InputExpr, 1819 std::string &ConstraintStr) { 1820 // If this can't be a register or memory, i.e., has to be a constant 1821 // (immediate or symbolic), try to emit it as such. 1822 if (!Info.allowsRegister() && !Info.allowsMemory()) { 1823 llvm::APSInt Result; 1824 if (InputExpr->EvaluateAsInt(Result, getContext())) 1825 return llvm::ConstantInt::get(getLLVMContext(), Result); 1826 assert(!Info.requiresImmediateConstant() && 1827 "Required-immediate inlineasm arg isn't constant?"); 1828 } 1829 1830 if (Info.allowsRegister() || !Info.allowsMemory()) 1831 if (CodeGenFunction::hasScalarEvaluationKind(InputExpr->getType())) 1832 return EmitScalarExpr(InputExpr); 1833 if (InputExpr->getStmtClass() == Expr::CXXThisExprClass) 1834 return EmitScalarExpr(InputExpr); 1835 InputExpr = InputExpr->IgnoreParenNoopCasts(getContext()); 1836 LValue Dest = EmitLValue(InputExpr); 1837 return EmitAsmInputLValue(Info, Dest, InputExpr->getType(), ConstraintStr, 1838 InputExpr->getExprLoc()); 1839 } 1840 1841 /// getAsmSrcLocInfo - Return the !srcloc metadata node to attach to an inline 1842 /// asm call instruction. The !srcloc MDNode contains a list of constant 1843 /// integers which are the source locations of the start of each line in the 1844 /// asm. 1845 static llvm::MDNode *getAsmSrcLocInfo(const StringLiteral *Str, 1846 CodeGenFunction &CGF) { 1847 SmallVector<llvm::Metadata *, 8> Locs; 1848 // Add the location of the first line to the MDNode. 1849 Locs.push_back(llvm::ConstantAsMetadata::get(llvm::ConstantInt::get( 1850 CGF.Int32Ty, Str->getLocStart().getRawEncoding()))); 1851 StringRef StrVal = Str->getString(); 1852 if (!StrVal.empty()) { 1853 const SourceManager &SM = CGF.CGM.getContext().getSourceManager(); 1854 const LangOptions &LangOpts = CGF.CGM.getLangOpts(); 1855 unsigned StartToken = 0; 1856 unsigned ByteOffset = 0; 1857 1858 // Add the location of the start of each subsequent line of the asm to the 1859 // MDNode. 1860 for (unsigned i = 0, e = StrVal.size() - 1; i != e; ++i) { 1861 if (StrVal[i] != '\n') continue; 1862 SourceLocation LineLoc = Str->getLocationOfByte( 1863 i + 1, SM, LangOpts, CGF.getTarget(), &StartToken, &ByteOffset); 1864 Locs.push_back(llvm::ConstantAsMetadata::get( 1865 llvm::ConstantInt::get(CGF.Int32Ty, LineLoc.getRawEncoding()))); 1866 } 1867 } 1868 1869 return llvm::MDNode::get(CGF.getLLVMContext(), Locs); 1870 } 1871 1872 void CodeGenFunction::EmitAsmStmt(const AsmStmt &S) { 1873 // Assemble the final asm string. 1874 std::string AsmString = S.generateAsmString(getContext()); 1875 1876 // Get all the output and input constraints together. 1877 SmallVector<TargetInfo::ConstraintInfo, 4> OutputConstraintInfos; 1878 SmallVector<TargetInfo::ConstraintInfo, 4> InputConstraintInfos; 1879 1880 for (unsigned i = 0, e = S.getNumOutputs(); i != e; i++) { 1881 StringRef Name; 1882 if (const GCCAsmStmt *GAS = dyn_cast<GCCAsmStmt>(&S)) 1883 Name = GAS->getOutputName(i); 1884 TargetInfo::ConstraintInfo Info(S.getOutputConstraint(i), Name); 1885 bool IsValid = getTarget().validateOutputConstraint(Info); (void)IsValid; 1886 assert(IsValid && "Failed to parse output constraint"); 1887 OutputConstraintInfos.push_back(Info); 1888 } 1889 1890 for (unsigned i = 0, e = S.getNumInputs(); i != e; i++) { 1891 StringRef Name; 1892 if (const GCCAsmStmt *GAS = dyn_cast<GCCAsmStmt>(&S)) 1893 Name = GAS->getInputName(i); 1894 TargetInfo::ConstraintInfo Info(S.getInputConstraint(i), Name); 1895 bool IsValid = 1896 getTarget().validateInputConstraint(OutputConstraintInfos, Info); 1897 assert(IsValid && "Failed to parse input constraint"); (void)IsValid; 1898 InputConstraintInfos.push_back(Info); 1899 } 1900 1901 std::string Constraints; 1902 1903 std::vector<LValue> ResultRegDests; 1904 std::vector<QualType> ResultRegQualTys; 1905 std::vector<llvm::Type *> ResultRegTypes; 1906 std::vector<llvm::Type *> ResultTruncRegTypes; 1907 std::vector<llvm::Type *> ArgTypes; 1908 std::vector<llvm::Value*> Args; 1909 1910 // Keep track of inout constraints. 1911 std::string InOutConstraints; 1912 std::vector<llvm::Value*> InOutArgs; 1913 std::vector<llvm::Type*> InOutArgTypes; 1914 1915 // An inline asm can be marked readonly if it meets the following conditions: 1916 // - it doesn't have any sideeffects 1917 // - it doesn't clobber memory 1918 // - it doesn't return a value by-reference 1919 // It can be marked readnone if it doesn't have any input memory constraints 1920 // in addition to meeting the conditions listed above. 1921 bool ReadOnly = true, ReadNone = true; 1922 1923 for (unsigned i = 0, e = S.getNumOutputs(); i != e; i++) { 1924 TargetInfo::ConstraintInfo &Info = OutputConstraintInfos[i]; 1925 1926 // Simplify the output constraint. 1927 std::string OutputConstraint(S.getOutputConstraint(i)); 1928 OutputConstraint = SimplifyConstraint(OutputConstraint.c_str() + 1, 1929 getTarget()); 1930 1931 const Expr *OutExpr = S.getOutputExpr(i); 1932 OutExpr = OutExpr->IgnoreParenNoopCasts(getContext()); 1933 1934 OutputConstraint = AddVariableConstraints(OutputConstraint, *OutExpr, 1935 getTarget(), CGM, S, 1936 Info.earlyClobber()); 1937 1938 LValue Dest = EmitLValue(OutExpr); 1939 if (!Constraints.empty()) 1940 Constraints += ','; 1941 1942 // If this is a register output, then make the inline asm return it 1943 // by-value. If this is a memory result, return the value by-reference. 1944 if (!Info.allowsMemory() && hasScalarEvaluationKind(OutExpr->getType())) { 1945 Constraints += "=" + OutputConstraint; 1946 ResultRegQualTys.push_back(OutExpr->getType()); 1947 ResultRegDests.push_back(Dest); 1948 ResultRegTypes.push_back(ConvertTypeForMem(OutExpr->getType())); 1949 ResultTruncRegTypes.push_back(ResultRegTypes.back()); 1950 1951 // If this output is tied to an input, and if the input is larger, then 1952 // we need to set the actual result type of the inline asm node to be the 1953 // same as the input type. 1954 if (Info.hasMatchingInput()) { 1955 unsigned InputNo; 1956 for (InputNo = 0; InputNo != S.getNumInputs(); ++InputNo) { 1957 TargetInfo::ConstraintInfo &Input = InputConstraintInfos[InputNo]; 1958 if (Input.hasTiedOperand() && Input.getTiedOperand() == i) 1959 break; 1960 } 1961 assert(InputNo != S.getNumInputs() && "Didn't find matching input!"); 1962 1963 QualType InputTy = S.getInputExpr(InputNo)->getType(); 1964 QualType OutputType = OutExpr->getType(); 1965 1966 uint64_t InputSize = getContext().getTypeSize(InputTy); 1967 if (getContext().getTypeSize(OutputType) < InputSize) { 1968 // Form the asm to return the value as a larger integer or fp type. 1969 ResultRegTypes.back() = ConvertType(InputTy); 1970 } 1971 } 1972 if (llvm::Type* AdjTy = 1973 getTargetHooks().adjustInlineAsmType(*this, OutputConstraint, 1974 ResultRegTypes.back())) 1975 ResultRegTypes.back() = AdjTy; 1976 else { 1977 CGM.getDiags().Report(S.getAsmLoc(), 1978 diag::err_asm_invalid_type_in_input) 1979 << OutExpr->getType() << OutputConstraint; 1980 } 1981 } else { 1982 ArgTypes.push_back(Dest.getAddress().getType()); 1983 Args.push_back(Dest.getPointer()); 1984 Constraints += "=*"; 1985 Constraints += OutputConstraint; 1986 ReadOnly = ReadNone = false; 1987 } 1988 1989 if (Info.isReadWrite()) { 1990 InOutConstraints += ','; 1991 1992 const Expr *InputExpr = S.getOutputExpr(i); 1993 llvm::Value *Arg = EmitAsmInputLValue(Info, Dest, InputExpr->getType(), 1994 InOutConstraints, 1995 InputExpr->getExprLoc()); 1996 1997 if (llvm::Type* AdjTy = 1998 getTargetHooks().adjustInlineAsmType(*this, OutputConstraint, 1999 Arg->getType())) 2000 Arg = Builder.CreateBitCast(Arg, AdjTy); 2001 2002 if (Info.allowsRegister()) 2003 InOutConstraints += llvm::utostr(i); 2004 else 2005 InOutConstraints += OutputConstraint; 2006 2007 InOutArgTypes.push_back(Arg->getType()); 2008 InOutArgs.push_back(Arg); 2009 } 2010 } 2011 2012 // If this is a Microsoft-style asm blob, store the return registers (EAX:EDX) 2013 // to the return value slot. Only do this when returning in registers. 2014 if (isa<MSAsmStmt>(&S)) { 2015 const ABIArgInfo &RetAI = CurFnInfo->getReturnInfo(); 2016 if (RetAI.isDirect() || RetAI.isExtend()) { 2017 // Make a fake lvalue for the return value slot. 2018 LValue ReturnSlot = MakeAddrLValue(ReturnValue, FnRetTy); 2019 CGM.getTargetCodeGenInfo().addReturnRegisterOutputs( 2020 *this, ReturnSlot, Constraints, ResultRegTypes, ResultTruncRegTypes, 2021 ResultRegDests, AsmString, S.getNumOutputs()); 2022 SawAsmBlock = true; 2023 } 2024 } 2025 2026 for (unsigned i = 0, e = S.getNumInputs(); i != e; i++) { 2027 const Expr *InputExpr = S.getInputExpr(i); 2028 2029 TargetInfo::ConstraintInfo &Info = InputConstraintInfos[i]; 2030 2031 if (Info.allowsMemory()) 2032 ReadNone = false; 2033 2034 if (!Constraints.empty()) 2035 Constraints += ','; 2036 2037 // Simplify the input constraint. 2038 std::string InputConstraint(S.getInputConstraint(i)); 2039 InputConstraint = SimplifyConstraint(InputConstraint.c_str(), getTarget(), 2040 &OutputConstraintInfos); 2041 2042 InputConstraint = AddVariableConstraints( 2043 InputConstraint, *InputExpr->IgnoreParenNoopCasts(getContext()), 2044 getTarget(), CGM, S, false /* No EarlyClobber */); 2045 2046 llvm::Value *Arg = EmitAsmInput(Info, InputExpr, Constraints); 2047 2048 // If this input argument is tied to a larger output result, extend the 2049 // input to be the same size as the output. The LLVM backend wants to see 2050 // the input and output of a matching constraint be the same size. Note 2051 // that GCC does not define what the top bits are here. We use zext because 2052 // that is usually cheaper, but LLVM IR should really get an anyext someday. 2053 if (Info.hasTiedOperand()) { 2054 unsigned Output = Info.getTiedOperand(); 2055 QualType OutputType = S.getOutputExpr(Output)->getType(); 2056 QualType InputTy = InputExpr->getType(); 2057 2058 if (getContext().getTypeSize(OutputType) > 2059 getContext().getTypeSize(InputTy)) { 2060 // Use ptrtoint as appropriate so that we can do our extension. 2061 if (isa<llvm::PointerType>(Arg->getType())) 2062 Arg = Builder.CreatePtrToInt(Arg, IntPtrTy); 2063 llvm::Type *OutputTy = ConvertType(OutputType); 2064 if (isa<llvm::IntegerType>(OutputTy)) 2065 Arg = Builder.CreateZExt(Arg, OutputTy); 2066 else if (isa<llvm::PointerType>(OutputTy)) 2067 Arg = Builder.CreateZExt(Arg, IntPtrTy); 2068 else { 2069 assert(OutputTy->isFloatingPointTy() && "Unexpected output type"); 2070 Arg = Builder.CreateFPExt(Arg, OutputTy); 2071 } 2072 } 2073 } 2074 if (llvm::Type* AdjTy = 2075 getTargetHooks().adjustInlineAsmType(*this, InputConstraint, 2076 Arg->getType())) 2077 Arg = Builder.CreateBitCast(Arg, AdjTy); 2078 else 2079 CGM.getDiags().Report(S.getAsmLoc(), diag::err_asm_invalid_type_in_input) 2080 << InputExpr->getType() << InputConstraint; 2081 2082 ArgTypes.push_back(Arg->getType()); 2083 Args.push_back(Arg); 2084 Constraints += InputConstraint; 2085 } 2086 2087 // Append the "input" part of inout constraints last. 2088 for (unsigned i = 0, e = InOutArgs.size(); i != e; i++) { 2089 ArgTypes.push_back(InOutArgTypes[i]); 2090 Args.push_back(InOutArgs[i]); 2091 } 2092 Constraints += InOutConstraints; 2093 2094 // Clobbers 2095 for (unsigned i = 0, e = S.getNumClobbers(); i != e; i++) { 2096 StringRef Clobber = S.getClobber(i); 2097 2098 if (Clobber == "memory") 2099 ReadOnly = ReadNone = false; 2100 else if (Clobber != "cc") 2101 Clobber = getTarget().getNormalizedGCCRegisterName(Clobber); 2102 2103 if (!Constraints.empty()) 2104 Constraints += ','; 2105 2106 Constraints += "~{"; 2107 Constraints += Clobber; 2108 Constraints += '}'; 2109 } 2110 2111 // Add machine specific clobbers 2112 std::string MachineClobbers = getTarget().getClobbers(); 2113 if (!MachineClobbers.empty()) { 2114 if (!Constraints.empty()) 2115 Constraints += ','; 2116 Constraints += MachineClobbers; 2117 } 2118 2119 llvm::Type *ResultType; 2120 if (ResultRegTypes.empty()) 2121 ResultType = VoidTy; 2122 else if (ResultRegTypes.size() == 1) 2123 ResultType = ResultRegTypes[0]; 2124 else 2125 ResultType = llvm::StructType::get(getLLVMContext(), ResultRegTypes); 2126 2127 llvm::FunctionType *FTy = 2128 llvm::FunctionType::get(ResultType, ArgTypes, false); 2129 2130 bool HasSideEffect = S.isVolatile() || S.getNumOutputs() == 0; 2131 llvm::InlineAsm::AsmDialect AsmDialect = isa<MSAsmStmt>(&S) ? 2132 llvm::InlineAsm::AD_Intel : llvm::InlineAsm::AD_ATT; 2133 llvm::InlineAsm *IA = 2134 llvm::InlineAsm::get(FTy, AsmString, Constraints, HasSideEffect, 2135 /* IsAlignStack */ false, AsmDialect); 2136 llvm::CallInst *Result = 2137 Builder.CreateCall(IA, Args, getBundlesForFunclet(IA)); 2138 Result->addAttribute(llvm::AttributeList::FunctionIndex, 2139 llvm::Attribute::NoUnwind); 2140 2141 // Attach readnone and readonly attributes. 2142 if (!HasSideEffect) { 2143 if (ReadNone) 2144 Result->addAttribute(llvm::AttributeList::FunctionIndex, 2145 llvm::Attribute::ReadNone); 2146 else if (ReadOnly) 2147 Result->addAttribute(llvm::AttributeList::FunctionIndex, 2148 llvm::Attribute::ReadOnly); 2149 } 2150 2151 // Slap the source location of the inline asm into a !srcloc metadata on the 2152 // call. 2153 if (const GCCAsmStmt *gccAsmStmt = dyn_cast<GCCAsmStmt>(&S)) { 2154 Result->setMetadata("srcloc", getAsmSrcLocInfo(gccAsmStmt->getAsmString(), 2155 *this)); 2156 } else { 2157 // At least put the line number on MS inline asm blobs. 2158 auto Loc = llvm::ConstantInt::get(Int32Ty, S.getAsmLoc().getRawEncoding()); 2159 Result->setMetadata("srcloc", 2160 llvm::MDNode::get(getLLVMContext(), 2161 llvm::ConstantAsMetadata::get(Loc))); 2162 } 2163 2164 if (getLangOpts().assumeFunctionsAreConvergent()) { 2165 // Conservatively, mark all inline asm blocks in CUDA or OpenCL as 2166 // convergent (meaning, they may call an intrinsically convergent op, such 2167 // as bar.sync, and so can't have certain optimizations applied around 2168 // them). 2169 Result->addAttribute(llvm::AttributeList::FunctionIndex, 2170 llvm::Attribute::Convergent); 2171 } 2172 2173 // Extract all of the register value results from the asm. 2174 std::vector<llvm::Value*> RegResults; 2175 if (ResultRegTypes.size() == 1) { 2176 RegResults.push_back(Result); 2177 } else { 2178 for (unsigned i = 0, e = ResultRegTypes.size(); i != e; ++i) { 2179 llvm::Value *Tmp = Builder.CreateExtractValue(Result, i, "asmresult"); 2180 RegResults.push_back(Tmp); 2181 } 2182 } 2183 2184 assert(RegResults.size() == ResultRegTypes.size()); 2185 assert(RegResults.size() == ResultTruncRegTypes.size()); 2186 assert(RegResults.size() == ResultRegDests.size()); 2187 for (unsigned i = 0, e = RegResults.size(); i != e; ++i) { 2188 llvm::Value *Tmp = RegResults[i]; 2189 2190 // If the result type of the LLVM IR asm doesn't match the result type of 2191 // the expression, do the conversion. 2192 if (ResultRegTypes[i] != ResultTruncRegTypes[i]) { 2193 llvm::Type *TruncTy = ResultTruncRegTypes[i]; 2194 2195 // Truncate the integer result to the right size, note that TruncTy can be 2196 // a pointer. 2197 if (TruncTy->isFloatingPointTy()) 2198 Tmp = Builder.CreateFPTrunc(Tmp, TruncTy); 2199 else if (TruncTy->isPointerTy() && Tmp->getType()->isIntegerTy()) { 2200 uint64_t ResSize = CGM.getDataLayout().getTypeSizeInBits(TruncTy); 2201 Tmp = Builder.CreateTrunc(Tmp, 2202 llvm::IntegerType::get(getLLVMContext(), (unsigned)ResSize)); 2203 Tmp = Builder.CreateIntToPtr(Tmp, TruncTy); 2204 } else if (Tmp->getType()->isPointerTy() && TruncTy->isIntegerTy()) { 2205 uint64_t TmpSize =CGM.getDataLayout().getTypeSizeInBits(Tmp->getType()); 2206 Tmp = Builder.CreatePtrToInt(Tmp, 2207 llvm::IntegerType::get(getLLVMContext(), (unsigned)TmpSize)); 2208 Tmp = Builder.CreateTrunc(Tmp, TruncTy); 2209 } else if (TruncTy->isIntegerTy()) { 2210 Tmp = Builder.CreateZExtOrTrunc(Tmp, TruncTy); 2211 } else if (TruncTy->isVectorTy()) { 2212 Tmp = Builder.CreateBitCast(Tmp, TruncTy); 2213 } 2214 } 2215 2216 EmitStoreThroughLValue(RValue::get(Tmp), ResultRegDests[i]); 2217 } 2218 } 2219 2220 LValue CodeGenFunction::InitCapturedStruct(const CapturedStmt &S) { 2221 const RecordDecl *RD = S.getCapturedRecordDecl(); 2222 QualType RecordTy = getContext().getRecordType(RD); 2223 2224 // Initialize the captured struct. 2225 LValue SlotLV = 2226 MakeAddrLValue(CreateMemTemp(RecordTy, "agg.captured"), RecordTy); 2227 2228 RecordDecl::field_iterator CurField = RD->field_begin(); 2229 for (CapturedStmt::const_capture_init_iterator I = S.capture_init_begin(), 2230 E = S.capture_init_end(); 2231 I != E; ++I, ++CurField) { 2232 LValue LV = EmitLValueForFieldInitialization(SlotLV, *CurField); 2233 if (CurField->hasCapturedVLAType()) { 2234 auto VAT = CurField->getCapturedVLAType(); 2235 EmitStoreThroughLValue(RValue::get(VLASizeMap[VAT->getSizeExpr()]), LV); 2236 } else { 2237 EmitInitializerForField(*CurField, LV, *I); 2238 } 2239 } 2240 2241 return SlotLV; 2242 } 2243 2244 /// Generate an outlined function for the body of a CapturedStmt, store any 2245 /// captured variables into the captured struct, and call the outlined function. 2246 llvm::Function * 2247 CodeGenFunction::EmitCapturedStmt(const CapturedStmt &S, CapturedRegionKind K) { 2248 LValue CapStruct = InitCapturedStruct(S); 2249 2250 // Emit the CapturedDecl 2251 CodeGenFunction CGF(CGM, true); 2252 CGCapturedStmtRAII CapInfoRAII(CGF, new CGCapturedStmtInfo(S, K)); 2253 llvm::Function *F = CGF.GenerateCapturedStmtFunction(S); 2254 delete CGF.CapturedStmtInfo; 2255 2256 // Emit call to the helper function. 2257 EmitCallOrInvoke(F, CapStruct.getPointer()); 2258 2259 return F; 2260 } 2261 2262 Address CodeGenFunction::GenerateCapturedStmtArgument(const CapturedStmt &S) { 2263 LValue CapStruct = InitCapturedStruct(S); 2264 return CapStruct.getAddress(); 2265 } 2266 2267 /// Creates the outlined function for a CapturedStmt. 2268 llvm::Function * 2269 CodeGenFunction::GenerateCapturedStmtFunction(const CapturedStmt &S) { 2270 assert(CapturedStmtInfo && 2271 "CapturedStmtInfo should be set when generating the captured function"); 2272 const CapturedDecl *CD = S.getCapturedDecl(); 2273 const RecordDecl *RD = S.getCapturedRecordDecl(); 2274 SourceLocation Loc = S.getLocStart(); 2275 assert(CD->hasBody() && "missing CapturedDecl body"); 2276 2277 // Build the argument list. 2278 ASTContext &Ctx = CGM.getContext(); 2279 FunctionArgList Args; 2280 Args.append(CD->param_begin(), CD->param_end()); 2281 2282 // Create the function declaration. 2283 const CGFunctionInfo &FuncInfo = 2284 CGM.getTypes().arrangeBuiltinFunctionDeclaration(Ctx.VoidTy, Args); 2285 llvm::FunctionType *FuncLLVMTy = CGM.getTypes().GetFunctionType(FuncInfo); 2286 2287 llvm::Function *F = 2288 llvm::Function::Create(FuncLLVMTy, llvm::GlobalValue::InternalLinkage, 2289 CapturedStmtInfo->getHelperName(), &CGM.getModule()); 2290 CGM.SetInternalFunctionAttributes(CD, F, FuncInfo); 2291 if (CD->isNothrow()) 2292 F->addFnAttr(llvm::Attribute::NoUnwind); 2293 2294 // Generate the function. 2295 StartFunction(CD, Ctx.VoidTy, F, FuncInfo, Args, 2296 CD->getLocation(), 2297 CD->getBody()->getLocStart()); 2298 // Set the context parameter in CapturedStmtInfo. 2299 Address DeclPtr = GetAddrOfLocalVar(CD->getContextParam()); 2300 CapturedStmtInfo->setContextValue(Builder.CreateLoad(DeclPtr)); 2301 2302 // Initialize variable-length arrays. 2303 LValue Base = MakeNaturalAlignAddrLValue(CapturedStmtInfo->getContextValue(), 2304 Ctx.getTagDeclType(RD)); 2305 for (auto *FD : RD->fields()) { 2306 if (FD->hasCapturedVLAType()) { 2307 auto *ExprArg = EmitLoadOfLValue(EmitLValueForField(Base, FD), 2308 S.getLocStart()).getScalarVal(); 2309 auto VAT = FD->getCapturedVLAType(); 2310 VLASizeMap[VAT->getSizeExpr()] = ExprArg; 2311 } 2312 } 2313 2314 // If 'this' is captured, load it into CXXThisValue. 2315 if (CapturedStmtInfo->isCXXThisExprCaptured()) { 2316 FieldDecl *FD = CapturedStmtInfo->getThisFieldDecl(); 2317 LValue ThisLValue = EmitLValueForField(Base, FD); 2318 CXXThisValue = EmitLoadOfLValue(ThisLValue, Loc).getScalarVal(); 2319 } 2320 2321 PGO.assignRegionCounters(GlobalDecl(CD), F); 2322 CapturedStmtInfo->EmitBody(*this, CD->getBody()); 2323 FinishFunction(CD->getBodyRBrace()); 2324 2325 return F; 2326 } 2327