1 //===--- CGStmt.cpp - Emit LLVM Code from Statements ----------------------===// 2 // 3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. 4 // See https://llvm.org/LICENSE.txt for license information. 5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception 6 // 7 //===----------------------------------------------------------------------===// 8 // 9 // This contains code to emit Stmt nodes as LLVM code. 10 // 11 //===----------------------------------------------------------------------===// 12 13 #include "CGDebugInfo.h" 14 #include "CGOpenMPRuntime.h" 15 #include "CodeGenFunction.h" 16 #include "CodeGenModule.h" 17 #include "TargetInfo.h" 18 #include "clang/AST/Attr.h" 19 #include "clang/AST/StmtVisitor.h" 20 #include "clang/Basic/Builtins.h" 21 #include "clang/Basic/DiagnosticSema.h" 22 #include "clang/Basic/PrettyStackTrace.h" 23 #include "clang/Basic/SourceManager.h" 24 #include "clang/Basic/TargetInfo.h" 25 #include "llvm/ADT/SmallSet.h" 26 #include "llvm/ADT/StringExtras.h" 27 #include "llvm/IR/DataLayout.h" 28 #include "llvm/IR/InlineAsm.h" 29 #include "llvm/IR/Intrinsics.h" 30 #include "llvm/IR/MDBuilder.h" 31 #include "llvm/Support/SaveAndRestore.h" 32 33 using namespace clang; 34 using namespace CodeGen; 35 36 //===----------------------------------------------------------------------===// 37 // Statement Emission 38 //===----------------------------------------------------------------------===// 39 40 void CodeGenFunction::EmitStopPoint(const Stmt *S) { 41 if (CGDebugInfo *DI = getDebugInfo()) { 42 SourceLocation Loc; 43 Loc = S->getBeginLoc(); 44 DI->EmitLocation(Builder, Loc); 45 46 LastStopPoint = Loc; 47 } 48 } 49 50 void CodeGenFunction::EmitStmt(const Stmt *S, ArrayRef<const Attr *> Attrs) { 51 assert(S && "Null statement?"); 52 PGO.setCurrentStmt(S); 53 54 // These statements have their own debug info handling. 55 if (EmitSimpleStmt(S, Attrs)) 56 return; 57 58 // Check if we are generating unreachable code. 59 if (!HaveInsertPoint()) { 60 // If so, and the statement doesn't contain a label, then we do not need to 61 // generate actual code. This is safe because (1) the current point is 62 // unreachable, so we don't need to execute the code, and (2) we've already 63 // handled the statements which update internal data structures (like the 64 // local variable map) which could be used by subsequent statements. 65 if (!ContainsLabel(S)) { 66 // Verify that any decl statements were handled as simple, they may be in 67 // scope of subsequent reachable statements. 68 assert(!isa<DeclStmt>(*S) && "Unexpected DeclStmt!"); 69 return; 70 } 71 72 // Otherwise, make a new block to hold the code. 73 EnsureInsertPoint(); 74 } 75 76 // Generate a stoppoint if we are emitting debug info. 77 EmitStopPoint(S); 78 79 // Ignore all OpenMP directives except for simd if OpenMP with Simd is 80 // enabled. 81 if (getLangOpts().OpenMP && getLangOpts().OpenMPSimd) { 82 if (const auto *D = dyn_cast<OMPExecutableDirective>(S)) { 83 EmitSimpleOMPExecutableDirective(*D); 84 return; 85 } 86 } 87 88 switch (S->getStmtClass()) { 89 case Stmt::NoStmtClass: 90 case Stmt::CXXCatchStmtClass: 91 case Stmt::SEHExceptStmtClass: 92 case Stmt::SEHFinallyStmtClass: 93 case Stmt::MSDependentExistsStmtClass: 94 llvm_unreachable("invalid statement class to emit generically"); 95 case Stmt::NullStmtClass: 96 case Stmt::CompoundStmtClass: 97 case Stmt::DeclStmtClass: 98 case Stmt::LabelStmtClass: 99 case Stmt::AttributedStmtClass: 100 case Stmt::GotoStmtClass: 101 case Stmt::BreakStmtClass: 102 case Stmt::ContinueStmtClass: 103 case Stmt::DefaultStmtClass: 104 case Stmt::CaseStmtClass: 105 case Stmt::SEHLeaveStmtClass: 106 llvm_unreachable("should have emitted these statements as simple"); 107 108 #define STMT(Type, Base) 109 #define ABSTRACT_STMT(Op) 110 #define EXPR(Type, Base) \ 111 case Stmt::Type##Class: 112 #include "clang/AST/StmtNodes.inc" 113 { 114 // Remember the block we came in on. 115 llvm::BasicBlock *incoming = Builder.GetInsertBlock(); 116 assert(incoming && "expression emission must have an insertion point"); 117 118 EmitIgnoredExpr(cast<Expr>(S)); 119 120 llvm::BasicBlock *outgoing = Builder.GetInsertBlock(); 121 assert(outgoing && "expression emission cleared block!"); 122 123 // The expression emitters assume (reasonably!) that the insertion 124 // point is always set. To maintain that, the call-emission code 125 // for noreturn functions has to enter a new block with no 126 // predecessors. We want to kill that block and mark the current 127 // insertion point unreachable in the common case of a call like 128 // "exit();". Since expression emission doesn't otherwise create 129 // blocks with no predecessors, we can just test for that. 130 // However, we must be careful not to do this to our incoming 131 // block, because *statement* emission does sometimes create 132 // reachable blocks which will have no predecessors until later in 133 // the function. This occurs with, e.g., labels that are not 134 // reachable by fallthrough. 135 if (incoming != outgoing && outgoing->use_empty()) { 136 outgoing->eraseFromParent(); 137 Builder.ClearInsertionPoint(); 138 } 139 break; 140 } 141 142 case Stmt::IndirectGotoStmtClass: 143 EmitIndirectGotoStmt(cast<IndirectGotoStmt>(*S)); break; 144 145 case Stmt::IfStmtClass: EmitIfStmt(cast<IfStmt>(*S)); break; 146 case Stmt::WhileStmtClass: EmitWhileStmt(cast<WhileStmt>(*S), Attrs); break; 147 case Stmt::DoStmtClass: EmitDoStmt(cast<DoStmt>(*S), Attrs); break; 148 case Stmt::ForStmtClass: EmitForStmt(cast<ForStmt>(*S), Attrs); break; 149 150 case Stmt::ReturnStmtClass: EmitReturnStmt(cast<ReturnStmt>(*S)); break; 151 152 case Stmt::SwitchStmtClass: EmitSwitchStmt(cast<SwitchStmt>(*S)); break; 153 case Stmt::GCCAsmStmtClass: // Intentional fall-through. 154 case Stmt::MSAsmStmtClass: EmitAsmStmt(cast<AsmStmt>(*S)); break; 155 case Stmt::CoroutineBodyStmtClass: 156 EmitCoroutineBody(cast<CoroutineBodyStmt>(*S)); 157 break; 158 case Stmt::CoreturnStmtClass: 159 EmitCoreturnStmt(cast<CoreturnStmt>(*S)); 160 break; 161 case Stmt::CapturedStmtClass: { 162 const CapturedStmt *CS = cast<CapturedStmt>(S); 163 EmitCapturedStmt(*CS, CS->getCapturedRegionKind()); 164 } 165 break; 166 case Stmt::ObjCAtTryStmtClass: 167 EmitObjCAtTryStmt(cast<ObjCAtTryStmt>(*S)); 168 break; 169 case Stmt::ObjCAtCatchStmtClass: 170 llvm_unreachable( 171 "@catch statements should be handled by EmitObjCAtTryStmt"); 172 case Stmt::ObjCAtFinallyStmtClass: 173 llvm_unreachable( 174 "@finally statements should be handled by EmitObjCAtTryStmt"); 175 case Stmt::ObjCAtThrowStmtClass: 176 EmitObjCAtThrowStmt(cast<ObjCAtThrowStmt>(*S)); 177 break; 178 case Stmt::ObjCAtSynchronizedStmtClass: 179 EmitObjCAtSynchronizedStmt(cast<ObjCAtSynchronizedStmt>(*S)); 180 break; 181 case Stmt::ObjCForCollectionStmtClass: 182 EmitObjCForCollectionStmt(cast<ObjCForCollectionStmt>(*S)); 183 break; 184 case Stmt::ObjCAutoreleasePoolStmtClass: 185 EmitObjCAutoreleasePoolStmt(cast<ObjCAutoreleasePoolStmt>(*S)); 186 break; 187 188 case Stmt::CXXTryStmtClass: 189 EmitCXXTryStmt(cast<CXXTryStmt>(*S)); 190 break; 191 case Stmt::CXXForRangeStmtClass: 192 EmitCXXForRangeStmt(cast<CXXForRangeStmt>(*S), Attrs); 193 break; 194 case Stmt::SEHTryStmtClass: 195 EmitSEHTryStmt(cast<SEHTryStmt>(*S)); 196 break; 197 case Stmt::OMPParallelDirectiveClass: 198 EmitOMPParallelDirective(cast<OMPParallelDirective>(*S)); 199 break; 200 case Stmt::OMPSimdDirectiveClass: 201 EmitOMPSimdDirective(cast<OMPSimdDirective>(*S)); 202 break; 203 case Stmt::OMPForDirectiveClass: 204 EmitOMPForDirective(cast<OMPForDirective>(*S)); 205 break; 206 case Stmt::OMPForSimdDirectiveClass: 207 EmitOMPForSimdDirective(cast<OMPForSimdDirective>(*S)); 208 break; 209 case Stmt::OMPSectionsDirectiveClass: 210 EmitOMPSectionsDirective(cast<OMPSectionsDirective>(*S)); 211 break; 212 case Stmt::OMPSectionDirectiveClass: 213 EmitOMPSectionDirective(cast<OMPSectionDirective>(*S)); 214 break; 215 case Stmt::OMPSingleDirectiveClass: 216 EmitOMPSingleDirective(cast<OMPSingleDirective>(*S)); 217 break; 218 case Stmt::OMPMasterDirectiveClass: 219 EmitOMPMasterDirective(cast<OMPMasterDirective>(*S)); 220 break; 221 case Stmt::OMPCriticalDirectiveClass: 222 EmitOMPCriticalDirective(cast<OMPCriticalDirective>(*S)); 223 break; 224 case Stmt::OMPParallelForDirectiveClass: 225 EmitOMPParallelForDirective(cast<OMPParallelForDirective>(*S)); 226 break; 227 case Stmt::OMPParallelForSimdDirectiveClass: 228 EmitOMPParallelForSimdDirective(cast<OMPParallelForSimdDirective>(*S)); 229 break; 230 case Stmt::OMPParallelMasterDirectiveClass: 231 EmitOMPParallelMasterDirective(cast<OMPParallelMasterDirective>(*S)); 232 break; 233 case Stmt::OMPParallelSectionsDirectiveClass: 234 EmitOMPParallelSectionsDirective(cast<OMPParallelSectionsDirective>(*S)); 235 break; 236 case Stmt::OMPTaskDirectiveClass: 237 EmitOMPTaskDirective(cast<OMPTaskDirective>(*S)); 238 break; 239 case Stmt::OMPTaskyieldDirectiveClass: 240 EmitOMPTaskyieldDirective(cast<OMPTaskyieldDirective>(*S)); 241 break; 242 case Stmt::OMPBarrierDirectiveClass: 243 EmitOMPBarrierDirective(cast<OMPBarrierDirective>(*S)); 244 break; 245 case Stmt::OMPTaskwaitDirectiveClass: 246 EmitOMPTaskwaitDirective(cast<OMPTaskwaitDirective>(*S)); 247 break; 248 case Stmt::OMPTaskgroupDirectiveClass: 249 EmitOMPTaskgroupDirective(cast<OMPTaskgroupDirective>(*S)); 250 break; 251 case Stmt::OMPFlushDirectiveClass: 252 EmitOMPFlushDirective(cast<OMPFlushDirective>(*S)); 253 break; 254 case Stmt::OMPDepobjDirectiveClass: 255 EmitOMPDepobjDirective(cast<OMPDepobjDirective>(*S)); 256 break; 257 case Stmt::OMPScanDirectiveClass: 258 EmitOMPScanDirective(cast<OMPScanDirective>(*S)); 259 break; 260 case Stmt::OMPOrderedDirectiveClass: 261 EmitOMPOrderedDirective(cast<OMPOrderedDirective>(*S)); 262 break; 263 case Stmt::OMPAtomicDirectiveClass: 264 EmitOMPAtomicDirective(cast<OMPAtomicDirective>(*S)); 265 break; 266 case Stmt::OMPTargetDirectiveClass: 267 EmitOMPTargetDirective(cast<OMPTargetDirective>(*S)); 268 break; 269 case Stmt::OMPTeamsDirectiveClass: 270 EmitOMPTeamsDirective(cast<OMPTeamsDirective>(*S)); 271 break; 272 case Stmt::OMPCancellationPointDirectiveClass: 273 EmitOMPCancellationPointDirective(cast<OMPCancellationPointDirective>(*S)); 274 break; 275 case Stmt::OMPCancelDirectiveClass: 276 EmitOMPCancelDirective(cast<OMPCancelDirective>(*S)); 277 break; 278 case Stmt::OMPTargetDataDirectiveClass: 279 EmitOMPTargetDataDirective(cast<OMPTargetDataDirective>(*S)); 280 break; 281 case Stmt::OMPTargetEnterDataDirectiveClass: 282 EmitOMPTargetEnterDataDirective(cast<OMPTargetEnterDataDirective>(*S)); 283 break; 284 case Stmt::OMPTargetExitDataDirectiveClass: 285 EmitOMPTargetExitDataDirective(cast<OMPTargetExitDataDirective>(*S)); 286 break; 287 case Stmt::OMPTargetParallelDirectiveClass: 288 EmitOMPTargetParallelDirective(cast<OMPTargetParallelDirective>(*S)); 289 break; 290 case Stmt::OMPTargetParallelForDirectiveClass: 291 EmitOMPTargetParallelForDirective(cast<OMPTargetParallelForDirective>(*S)); 292 break; 293 case Stmt::OMPTaskLoopDirectiveClass: 294 EmitOMPTaskLoopDirective(cast<OMPTaskLoopDirective>(*S)); 295 break; 296 case Stmt::OMPTaskLoopSimdDirectiveClass: 297 EmitOMPTaskLoopSimdDirective(cast<OMPTaskLoopSimdDirective>(*S)); 298 break; 299 case Stmt::OMPMasterTaskLoopDirectiveClass: 300 EmitOMPMasterTaskLoopDirective(cast<OMPMasterTaskLoopDirective>(*S)); 301 break; 302 case Stmt::OMPMasterTaskLoopSimdDirectiveClass: 303 EmitOMPMasterTaskLoopSimdDirective( 304 cast<OMPMasterTaskLoopSimdDirective>(*S)); 305 break; 306 case Stmt::OMPParallelMasterTaskLoopDirectiveClass: 307 EmitOMPParallelMasterTaskLoopDirective( 308 cast<OMPParallelMasterTaskLoopDirective>(*S)); 309 break; 310 case Stmt::OMPParallelMasterTaskLoopSimdDirectiveClass: 311 EmitOMPParallelMasterTaskLoopSimdDirective( 312 cast<OMPParallelMasterTaskLoopSimdDirective>(*S)); 313 break; 314 case Stmt::OMPDistributeDirectiveClass: 315 EmitOMPDistributeDirective(cast<OMPDistributeDirective>(*S)); 316 break; 317 case Stmt::OMPTargetUpdateDirectiveClass: 318 EmitOMPTargetUpdateDirective(cast<OMPTargetUpdateDirective>(*S)); 319 break; 320 case Stmt::OMPDistributeParallelForDirectiveClass: 321 EmitOMPDistributeParallelForDirective( 322 cast<OMPDistributeParallelForDirective>(*S)); 323 break; 324 case Stmt::OMPDistributeParallelForSimdDirectiveClass: 325 EmitOMPDistributeParallelForSimdDirective( 326 cast<OMPDistributeParallelForSimdDirective>(*S)); 327 break; 328 case Stmt::OMPDistributeSimdDirectiveClass: 329 EmitOMPDistributeSimdDirective(cast<OMPDistributeSimdDirective>(*S)); 330 break; 331 case Stmt::OMPTargetParallelForSimdDirectiveClass: 332 EmitOMPTargetParallelForSimdDirective( 333 cast<OMPTargetParallelForSimdDirective>(*S)); 334 break; 335 case Stmt::OMPTargetSimdDirectiveClass: 336 EmitOMPTargetSimdDirective(cast<OMPTargetSimdDirective>(*S)); 337 break; 338 case Stmt::OMPTeamsDistributeDirectiveClass: 339 EmitOMPTeamsDistributeDirective(cast<OMPTeamsDistributeDirective>(*S)); 340 break; 341 case Stmt::OMPTeamsDistributeSimdDirectiveClass: 342 EmitOMPTeamsDistributeSimdDirective( 343 cast<OMPTeamsDistributeSimdDirective>(*S)); 344 break; 345 case Stmt::OMPTeamsDistributeParallelForSimdDirectiveClass: 346 EmitOMPTeamsDistributeParallelForSimdDirective( 347 cast<OMPTeamsDistributeParallelForSimdDirective>(*S)); 348 break; 349 case Stmt::OMPTeamsDistributeParallelForDirectiveClass: 350 EmitOMPTeamsDistributeParallelForDirective( 351 cast<OMPTeamsDistributeParallelForDirective>(*S)); 352 break; 353 case Stmt::OMPTargetTeamsDirectiveClass: 354 EmitOMPTargetTeamsDirective(cast<OMPTargetTeamsDirective>(*S)); 355 break; 356 case Stmt::OMPTargetTeamsDistributeDirectiveClass: 357 EmitOMPTargetTeamsDistributeDirective( 358 cast<OMPTargetTeamsDistributeDirective>(*S)); 359 break; 360 case Stmt::OMPTargetTeamsDistributeParallelForDirectiveClass: 361 EmitOMPTargetTeamsDistributeParallelForDirective( 362 cast<OMPTargetTeamsDistributeParallelForDirective>(*S)); 363 break; 364 case Stmt::OMPTargetTeamsDistributeParallelForSimdDirectiveClass: 365 EmitOMPTargetTeamsDistributeParallelForSimdDirective( 366 cast<OMPTargetTeamsDistributeParallelForSimdDirective>(*S)); 367 break; 368 case Stmt::OMPTargetTeamsDistributeSimdDirectiveClass: 369 EmitOMPTargetTeamsDistributeSimdDirective( 370 cast<OMPTargetTeamsDistributeSimdDirective>(*S)); 371 break; 372 } 373 } 374 375 bool CodeGenFunction::EmitSimpleStmt(const Stmt *S, 376 ArrayRef<const Attr *> Attrs) { 377 switch (S->getStmtClass()) { 378 default: 379 return false; 380 case Stmt::NullStmtClass: 381 break; 382 case Stmt::CompoundStmtClass: 383 EmitCompoundStmt(cast<CompoundStmt>(*S)); 384 break; 385 case Stmt::DeclStmtClass: 386 EmitDeclStmt(cast<DeclStmt>(*S)); 387 break; 388 case Stmt::LabelStmtClass: 389 EmitLabelStmt(cast<LabelStmt>(*S)); 390 break; 391 case Stmt::AttributedStmtClass: 392 EmitAttributedStmt(cast<AttributedStmt>(*S)); 393 break; 394 case Stmt::GotoStmtClass: 395 EmitGotoStmt(cast<GotoStmt>(*S)); 396 break; 397 case Stmt::BreakStmtClass: 398 EmitBreakStmt(cast<BreakStmt>(*S)); 399 break; 400 case Stmt::ContinueStmtClass: 401 EmitContinueStmt(cast<ContinueStmt>(*S)); 402 break; 403 case Stmt::DefaultStmtClass: 404 EmitDefaultStmt(cast<DefaultStmt>(*S), Attrs); 405 break; 406 case Stmt::CaseStmtClass: 407 EmitCaseStmt(cast<CaseStmt>(*S), Attrs); 408 break; 409 case Stmt::SEHLeaveStmtClass: 410 EmitSEHLeaveStmt(cast<SEHLeaveStmt>(*S)); 411 break; 412 } 413 return true; 414 } 415 416 /// EmitCompoundStmt - Emit a compound statement {..} node. If GetLast is true, 417 /// this captures the expression result of the last sub-statement and returns it 418 /// (for use by the statement expression extension). 419 Address CodeGenFunction::EmitCompoundStmt(const CompoundStmt &S, bool GetLast, 420 AggValueSlot AggSlot) { 421 PrettyStackTraceLoc CrashInfo(getContext().getSourceManager(),S.getLBracLoc(), 422 "LLVM IR generation of compound statement ('{}')"); 423 424 // Keep track of the current cleanup stack depth, including debug scopes. 425 LexicalScope Scope(*this, S.getSourceRange()); 426 427 return EmitCompoundStmtWithoutScope(S, GetLast, AggSlot); 428 } 429 430 Address 431 CodeGenFunction::EmitCompoundStmtWithoutScope(const CompoundStmt &S, 432 bool GetLast, 433 AggValueSlot AggSlot) { 434 435 const Stmt *ExprResult = S.getStmtExprResult(); 436 assert((!GetLast || (GetLast && ExprResult)) && 437 "If GetLast is true then the CompoundStmt must have a StmtExprResult"); 438 439 Address RetAlloca = Address::invalid(); 440 441 for (auto *CurStmt : S.body()) { 442 if (GetLast && ExprResult == CurStmt) { 443 // We have to special case labels here. They are statements, but when put 444 // at the end of a statement expression, they yield the value of their 445 // subexpression. Handle this by walking through all labels we encounter, 446 // emitting them before we evaluate the subexpr. 447 // Similar issues arise for attributed statements. 448 while (!isa<Expr>(ExprResult)) { 449 if (const auto *LS = dyn_cast<LabelStmt>(ExprResult)) { 450 EmitLabel(LS->getDecl()); 451 ExprResult = LS->getSubStmt(); 452 } else if (const auto *AS = dyn_cast<AttributedStmt>(ExprResult)) { 453 // FIXME: Update this if we ever have attributes that affect the 454 // semantics of an expression. 455 ExprResult = AS->getSubStmt(); 456 } else { 457 llvm_unreachable("unknown value statement"); 458 } 459 } 460 461 EnsureInsertPoint(); 462 463 const Expr *E = cast<Expr>(ExprResult); 464 QualType ExprTy = E->getType(); 465 if (hasAggregateEvaluationKind(ExprTy)) { 466 EmitAggExpr(E, AggSlot); 467 } else { 468 // We can't return an RValue here because there might be cleanups at 469 // the end of the StmtExpr. Because of that, we have to emit the result 470 // here into a temporary alloca. 471 RetAlloca = CreateMemTemp(ExprTy); 472 EmitAnyExprToMem(E, RetAlloca, Qualifiers(), 473 /*IsInit*/ false); 474 } 475 } else { 476 EmitStmt(CurStmt); 477 } 478 } 479 480 return RetAlloca; 481 } 482 483 void CodeGenFunction::SimplifyForwardingBlocks(llvm::BasicBlock *BB) { 484 llvm::BranchInst *BI = dyn_cast<llvm::BranchInst>(BB->getTerminator()); 485 486 // If there is a cleanup stack, then we it isn't worth trying to 487 // simplify this block (we would need to remove it from the scope map 488 // and cleanup entry). 489 if (!EHStack.empty()) 490 return; 491 492 // Can only simplify direct branches. 493 if (!BI || !BI->isUnconditional()) 494 return; 495 496 // Can only simplify empty blocks. 497 if (BI->getIterator() != BB->begin()) 498 return; 499 500 BB->replaceAllUsesWith(BI->getSuccessor(0)); 501 BI->eraseFromParent(); 502 BB->eraseFromParent(); 503 } 504 505 void CodeGenFunction::EmitBlock(llvm::BasicBlock *BB, bool IsFinished) { 506 llvm::BasicBlock *CurBB = Builder.GetInsertBlock(); 507 508 // Fall out of the current block (if necessary). 509 EmitBranch(BB); 510 511 if (IsFinished && BB->use_empty()) { 512 delete BB; 513 return; 514 } 515 516 // Place the block after the current block, if possible, or else at 517 // the end of the function. 518 if (CurBB && CurBB->getParent()) 519 CurFn->getBasicBlockList().insertAfter(CurBB->getIterator(), BB); 520 else 521 CurFn->getBasicBlockList().push_back(BB); 522 Builder.SetInsertPoint(BB); 523 } 524 525 void CodeGenFunction::EmitBranch(llvm::BasicBlock *Target) { 526 // Emit a branch from the current block to the target one if this 527 // was a real block. If this was just a fall-through block after a 528 // terminator, don't emit it. 529 llvm::BasicBlock *CurBB = Builder.GetInsertBlock(); 530 531 if (!CurBB || CurBB->getTerminator()) { 532 // If there is no insert point or the previous block is already 533 // terminated, don't touch it. 534 } else { 535 // Otherwise, create a fall-through branch. 536 Builder.CreateBr(Target); 537 } 538 539 Builder.ClearInsertionPoint(); 540 } 541 542 void CodeGenFunction::EmitBlockAfterUses(llvm::BasicBlock *block) { 543 bool inserted = false; 544 for (llvm::User *u : block->users()) { 545 if (llvm::Instruction *insn = dyn_cast<llvm::Instruction>(u)) { 546 CurFn->getBasicBlockList().insertAfter(insn->getParent()->getIterator(), 547 block); 548 inserted = true; 549 break; 550 } 551 } 552 553 if (!inserted) 554 CurFn->getBasicBlockList().push_back(block); 555 556 Builder.SetInsertPoint(block); 557 } 558 559 CodeGenFunction::JumpDest 560 CodeGenFunction::getJumpDestForLabel(const LabelDecl *D) { 561 JumpDest &Dest = LabelMap[D]; 562 if (Dest.isValid()) return Dest; 563 564 // Create, but don't insert, the new block. 565 Dest = JumpDest(createBasicBlock(D->getName()), 566 EHScopeStack::stable_iterator::invalid(), 567 NextCleanupDestIndex++); 568 return Dest; 569 } 570 571 void CodeGenFunction::EmitLabel(const LabelDecl *D) { 572 // Add this label to the current lexical scope if we're within any 573 // normal cleanups. Jumps "in" to this label --- when permitted by 574 // the language --- may need to be routed around such cleanups. 575 if (EHStack.hasNormalCleanups() && CurLexicalScope) 576 CurLexicalScope->addLabel(D); 577 578 JumpDest &Dest = LabelMap[D]; 579 580 // If we didn't need a forward reference to this label, just go 581 // ahead and create a destination at the current scope. 582 if (!Dest.isValid()) { 583 Dest = getJumpDestInCurrentScope(D->getName()); 584 585 // Otherwise, we need to give this label a target depth and remove 586 // it from the branch-fixups list. 587 } else { 588 assert(!Dest.getScopeDepth().isValid() && "already emitted label!"); 589 Dest.setScopeDepth(EHStack.stable_begin()); 590 ResolveBranchFixups(Dest.getBlock()); 591 } 592 593 EmitBlock(Dest.getBlock()); 594 595 // Emit debug info for labels. 596 if (CGDebugInfo *DI = getDebugInfo()) { 597 if (CGM.getCodeGenOpts().hasReducedDebugInfo()) { 598 DI->setLocation(D->getLocation()); 599 DI->EmitLabel(D, Builder); 600 } 601 } 602 603 incrementProfileCounter(D->getStmt()); 604 } 605 606 /// Change the cleanup scope of the labels in this lexical scope to 607 /// match the scope of the enclosing context. 608 void CodeGenFunction::LexicalScope::rescopeLabels() { 609 assert(!Labels.empty()); 610 EHScopeStack::stable_iterator innermostScope 611 = CGF.EHStack.getInnermostNormalCleanup(); 612 613 // Change the scope depth of all the labels. 614 for (SmallVectorImpl<const LabelDecl*>::const_iterator 615 i = Labels.begin(), e = Labels.end(); i != e; ++i) { 616 assert(CGF.LabelMap.count(*i)); 617 JumpDest &dest = CGF.LabelMap.find(*i)->second; 618 assert(dest.getScopeDepth().isValid()); 619 assert(innermostScope.encloses(dest.getScopeDepth())); 620 dest.setScopeDepth(innermostScope); 621 } 622 623 // Reparent the labels if the new scope also has cleanups. 624 if (innermostScope != EHScopeStack::stable_end() && ParentScope) { 625 ParentScope->Labels.append(Labels.begin(), Labels.end()); 626 } 627 } 628 629 630 void CodeGenFunction::EmitLabelStmt(const LabelStmt &S) { 631 EmitLabel(S.getDecl()); 632 EmitStmt(S.getSubStmt()); 633 } 634 635 void CodeGenFunction::EmitAttributedStmt(const AttributedStmt &S) { 636 bool nomerge = false; 637 for (const auto *A : S.getAttrs()) 638 if (A->getKind() == attr::NoMerge) { 639 nomerge = true; 640 break; 641 } 642 SaveAndRestore<bool> save_nomerge(InNoMergeAttributedStmt, nomerge); 643 EmitStmt(S.getSubStmt(), S.getAttrs()); 644 } 645 646 void CodeGenFunction::EmitGotoStmt(const GotoStmt &S) { 647 // If this code is reachable then emit a stop point (if generating 648 // debug info). We have to do this ourselves because we are on the 649 // "simple" statement path. 650 if (HaveInsertPoint()) 651 EmitStopPoint(&S); 652 653 EmitBranchThroughCleanup(getJumpDestForLabel(S.getLabel())); 654 } 655 656 657 void CodeGenFunction::EmitIndirectGotoStmt(const IndirectGotoStmt &S) { 658 if (const LabelDecl *Target = S.getConstantTarget()) { 659 EmitBranchThroughCleanup(getJumpDestForLabel(Target)); 660 return; 661 } 662 663 // Ensure that we have an i8* for our PHI node. 664 llvm::Value *V = Builder.CreateBitCast(EmitScalarExpr(S.getTarget()), 665 Int8PtrTy, "addr"); 666 llvm::BasicBlock *CurBB = Builder.GetInsertBlock(); 667 668 // Get the basic block for the indirect goto. 669 llvm::BasicBlock *IndGotoBB = GetIndirectGotoBlock(); 670 671 // The first instruction in the block has to be the PHI for the switch dest, 672 // add an entry for this branch. 673 cast<llvm::PHINode>(IndGotoBB->begin())->addIncoming(V, CurBB); 674 675 EmitBranch(IndGotoBB); 676 } 677 678 void CodeGenFunction::EmitIfStmt(const IfStmt &S) { 679 // C99 6.8.4.1: The first substatement is executed if the expression compares 680 // unequal to 0. The condition must be a scalar type. 681 LexicalScope ConditionScope(*this, S.getCond()->getSourceRange()); 682 683 if (S.getInit()) 684 EmitStmt(S.getInit()); 685 686 if (S.getConditionVariable()) 687 EmitDecl(*S.getConditionVariable()); 688 689 // If the condition constant folds and can be elided, try to avoid emitting 690 // the condition and the dead arm of the if/else. 691 bool CondConstant; 692 if (ConstantFoldsToSimpleInteger(S.getCond(), CondConstant, 693 S.isConstexpr())) { 694 // Figure out which block (then or else) is executed. 695 const Stmt *Executed = S.getThen(); 696 const Stmt *Skipped = S.getElse(); 697 if (!CondConstant) // Condition false? 698 std::swap(Executed, Skipped); 699 700 // If the skipped block has no labels in it, just emit the executed block. 701 // This avoids emitting dead code and simplifies the CFG substantially. 702 if (S.isConstexpr() || !ContainsLabel(Skipped)) { 703 if (CondConstant) 704 incrementProfileCounter(&S); 705 if (Executed) { 706 RunCleanupsScope ExecutedScope(*this); 707 EmitStmt(Executed); 708 } 709 return; 710 } 711 } 712 713 // Otherwise, the condition did not fold, or we couldn't elide it. Just emit 714 // the conditional branch. 715 llvm::BasicBlock *ThenBlock = createBasicBlock("if.then"); 716 llvm::BasicBlock *ContBlock = createBasicBlock("if.end"); 717 llvm::BasicBlock *ElseBlock = ContBlock; 718 if (S.getElse()) 719 ElseBlock = createBasicBlock("if.else"); 720 721 // Prefer the PGO based weights over the likelihood attribute. 722 // When the build isn't optimized the metadata isn't used, so don't generate 723 // it. 724 Stmt::Likelihood LH = Stmt::LH_None; 725 uint64_t Count = getProfileCount(S.getThen()); 726 if (!Count && CGM.getCodeGenOpts().OptimizationLevel) 727 LH = Stmt::getLikelihood(S.getThen(), S.getElse()); 728 EmitBranchOnBoolExpr(S.getCond(), ThenBlock, ElseBlock, Count, LH); 729 730 // Emit the 'then' code. 731 EmitBlock(ThenBlock); 732 incrementProfileCounter(&S); 733 { 734 RunCleanupsScope ThenScope(*this); 735 EmitStmt(S.getThen()); 736 } 737 EmitBranch(ContBlock); 738 739 // Emit the 'else' code if present. 740 if (const Stmt *Else = S.getElse()) { 741 { 742 // There is no need to emit line number for an unconditional branch. 743 auto NL = ApplyDebugLocation::CreateEmpty(*this); 744 EmitBlock(ElseBlock); 745 } 746 { 747 RunCleanupsScope ElseScope(*this); 748 EmitStmt(Else); 749 } 750 { 751 // There is no need to emit line number for an unconditional branch. 752 auto NL = ApplyDebugLocation::CreateEmpty(*this); 753 EmitBranch(ContBlock); 754 } 755 } 756 757 // Emit the continuation block for code after the if. 758 EmitBlock(ContBlock, true); 759 } 760 761 void CodeGenFunction::EmitWhileStmt(const WhileStmt &S, 762 ArrayRef<const Attr *> WhileAttrs) { 763 // Emit the header for the loop, which will also become 764 // the continue target. 765 JumpDest LoopHeader = getJumpDestInCurrentScope("while.cond"); 766 EmitBlock(LoopHeader.getBlock()); 767 768 const SourceRange &R = S.getSourceRange(); 769 LoopStack.push(LoopHeader.getBlock(), CGM.getContext(), CGM.getCodeGenOpts(), 770 WhileAttrs, SourceLocToDebugLoc(R.getBegin()), 771 SourceLocToDebugLoc(R.getEnd())); 772 773 // Create an exit block for when the condition fails, which will 774 // also become the break target. 775 JumpDest LoopExit = getJumpDestInCurrentScope("while.end"); 776 777 // Store the blocks to use for break and continue. 778 BreakContinueStack.push_back(BreakContinue(LoopExit, LoopHeader)); 779 780 // C++ [stmt.while]p2: 781 // When the condition of a while statement is a declaration, the 782 // scope of the variable that is declared extends from its point 783 // of declaration (3.3.2) to the end of the while statement. 784 // [...] 785 // The object created in a condition is destroyed and created 786 // with each iteration of the loop. 787 RunCleanupsScope ConditionScope(*this); 788 789 if (S.getConditionVariable()) 790 EmitDecl(*S.getConditionVariable()); 791 792 // Evaluate the conditional in the while header. C99 6.8.5.1: The 793 // evaluation of the controlling expression takes place before each 794 // execution of the loop body. 795 llvm::Value *BoolCondVal = EvaluateExprAsBool(S.getCond()); 796 797 // while(1) is common, avoid extra exit blocks. Be sure 798 // to correctly handle break/continue though. 799 bool EmitBoolCondBranch = true; 800 if (llvm::ConstantInt *C = dyn_cast<llvm::ConstantInt>(BoolCondVal)) 801 if (C->isOne()) 802 EmitBoolCondBranch = false; 803 804 // As long as the condition is true, go to the loop body. 805 llvm::BasicBlock *LoopBody = createBasicBlock("while.body"); 806 if (EmitBoolCondBranch) { 807 llvm::BasicBlock *ExitBlock = LoopExit.getBlock(); 808 if (ConditionScope.requiresCleanups()) 809 ExitBlock = createBasicBlock("while.exit"); 810 llvm::MDNode *Weights = createProfileOrBranchWeightsForLoop( 811 S.getCond(), getProfileCount(S.getBody()), S.getBody()); 812 Builder.CreateCondBr(BoolCondVal, LoopBody, ExitBlock, Weights); 813 814 if (ExitBlock != LoopExit.getBlock()) { 815 EmitBlock(ExitBlock); 816 EmitBranchThroughCleanup(LoopExit); 817 } 818 } else if (const Attr *A = Stmt::getLikelihoodAttr(S.getBody())) { 819 CGM.getDiags().Report(A->getLocation(), 820 diag::warn_attribute_has_no_effect_on_infinite_loop) 821 << A << A->getRange(); 822 CGM.getDiags().Report( 823 S.getWhileLoc(), 824 diag::note_attribute_has_no_effect_on_infinite_loop_here) 825 << SourceRange(S.getWhileLoc(), S.getRParenLoc()); 826 } 827 828 // Emit the loop body. We have to emit this in a cleanup scope 829 // because it might be a singleton DeclStmt. 830 { 831 RunCleanupsScope BodyScope(*this); 832 EmitBlock(LoopBody); 833 incrementProfileCounter(&S); 834 EmitStmt(S.getBody()); 835 } 836 837 BreakContinueStack.pop_back(); 838 839 // Immediately force cleanup. 840 ConditionScope.ForceCleanup(); 841 842 EmitStopPoint(&S); 843 // Branch to the loop header again. 844 EmitBranch(LoopHeader.getBlock()); 845 846 LoopStack.pop(); 847 848 // Emit the exit block. 849 EmitBlock(LoopExit.getBlock(), true); 850 851 // The LoopHeader typically is just a branch if we skipped emitting 852 // a branch, try to erase it. 853 if (!EmitBoolCondBranch) 854 SimplifyForwardingBlocks(LoopHeader.getBlock()); 855 } 856 857 void CodeGenFunction::EmitDoStmt(const DoStmt &S, 858 ArrayRef<const Attr *> DoAttrs) { 859 JumpDest LoopExit = getJumpDestInCurrentScope("do.end"); 860 JumpDest LoopCond = getJumpDestInCurrentScope("do.cond"); 861 862 uint64_t ParentCount = getCurrentProfileCount(); 863 864 // Store the blocks to use for break and continue. 865 BreakContinueStack.push_back(BreakContinue(LoopExit, LoopCond)); 866 867 // Emit the body of the loop. 868 llvm::BasicBlock *LoopBody = createBasicBlock("do.body"); 869 870 EmitBlockWithFallThrough(LoopBody, &S); 871 { 872 RunCleanupsScope BodyScope(*this); 873 EmitStmt(S.getBody()); 874 } 875 876 EmitBlock(LoopCond.getBlock()); 877 878 const SourceRange &R = S.getSourceRange(); 879 LoopStack.push(LoopBody, CGM.getContext(), CGM.getCodeGenOpts(), DoAttrs, 880 SourceLocToDebugLoc(R.getBegin()), 881 SourceLocToDebugLoc(R.getEnd())); 882 883 // C99 6.8.5.2: "The evaluation of the controlling expression takes place 884 // after each execution of the loop body." 885 886 // Evaluate the conditional in the while header. 887 // C99 6.8.5p2/p4: The first substatement is executed if the expression 888 // compares unequal to 0. The condition must be a scalar type. 889 llvm::Value *BoolCondVal = EvaluateExprAsBool(S.getCond()); 890 891 BreakContinueStack.pop_back(); 892 893 // "do {} while (0)" is common in macros, avoid extra blocks. Be sure 894 // to correctly handle break/continue though. 895 bool EmitBoolCondBranch = true; 896 if (llvm::ConstantInt *C = dyn_cast<llvm::ConstantInt>(BoolCondVal)) 897 if (C->isZero()) 898 EmitBoolCondBranch = false; 899 900 // As long as the condition is true, iterate the loop. 901 if (EmitBoolCondBranch) { 902 uint64_t BackedgeCount = getProfileCount(S.getBody()) - ParentCount; 903 Builder.CreateCondBr( 904 BoolCondVal, LoopBody, LoopExit.getBlock(), 905 createProfileWeightsForLoop(S.getCond(), BackedgeCount)); 906 } 907 908 LoopStack.pop(); 909 910 // Emit the exit block. 911 EmitBlock(LoopExit.getBlock()); 912 913 // The DoCond block typically is just a branch if we skipped 914 // emitting a branch, try to erase it. 915 if (!EmitBoolCondBranch) 916 SimplifyForwardingBlocks(LoopCond.getBlock()); 917 } 918 919 void CodeGenFunction::EmitForStmt(const ForStmt &S, 920 ArrayRef<const Attr *> ForAttrs) { 921 JumpDest LoopExit = getJumpDestInCurrentScope("for.end"); 922 923 LexicalScope ForScope(*this, S.getSourceRange()); 924 925 // Evaluate the first part before the loop. 926 if (S.getInit()) 927 EmitStmt(S.getInit()); 928 929 // Start the loop with a block that tests the condition. 930 // If there's an increment, the continue scope will be overwritten 931 // later. 932 JumpDest Continue = getJumpDestInCurrentScope("for.cond"); 933 llvm::BasicBlock *CondBlock = Continue.getBlock(); 934 EmitBlock(CondBlock); 935 936 const SourceRange &R = S.getSourceRange(); 937 LoopStack.push(CondBlock, CGM.getContext(), CGM.getCodeGenOpts(), ForAttrs, 938 SourceLocToDebugLoc(R.getBegin()), 939 SourceLocToDebugLoc(R.getEnd())); 940 941 // If the for loop doesn't have an increment we can just use the 942 // condition as the continue block. Otherwise we'll need to create 943 // a block for it (in the current scope, i.e. in the scope of the 944 // condition), and that we will become our continue block. 945 if (S.getInc()) 946 Continue = getJumpDestInCurrentScope("for.inc"); 947 948 // Store the blocks to use for break and continue. 949 BreakContinueStack.push_back(BreakContinue(LoopExit, Continue)); 950 951 // Create a cleanup scope for the condition variable cleanups. 952 LexicalScope ConditionScope(*this, S.getSourceRange()); 953 954 if (S.getCond()) { 955 // If the for statement has a condition scope, emit the local variable 956 // declaration. 957 if (S.getConditionVariable()) { 958 EmitDecl(*S.getConditionVariable()); 959 } 960 961 llvm::BasicBlock *ExitBlock = LoopExit.getBlock(); 962 // If there are any cleanups between here and the loop-exit scope, 963 // create a block to stage a loop exit along. 964 if (ForScope.requiresCleanups()) 965 ExitBlock = createBasicBlock("for.cond.cleanup"); 966 967 // As long as the condition is true, iterate the loop. 968 llvm::BasicBlock *ForBody = createBasicBlock("for.body"); 969 970 // C99 6.8.5p2/p4: The first substatement is executed if the expression 971 // compares unequal to 0. The condition must be a scalar type. 972 llvm::Value *BoolCondVal = EvaluateExprAsBool(S.getCond()); 973 llvm::MDNode *Weights = createProfileOrBranchWeightsForLoop( 974 S.getCond(), getProfileCount(S.getBody()), S.getBody()); 975 Builder.CreateCondBr(BoolCondVal, ForBody, ExitBlock, Weights); 976 977 if (ExitBlock != LoopExit.getBlock()) { 978 EmitBlock(ExitBlock); 979 EmitBranchThroughCleanup(LoopExit); 980 } 981 982 EmitBlock(ForBody); 983 } else { 984 // Treat it as a non-zero constant. Don't even create a new block for the 985 // body, just fall into it. 986 } 987 incrementProfileCounter(&S); 988 989 { 990 // Create a separate cleanup scope for the body, in case it is not 991 // a compound statement. 992 RunCleanupsScope BodyScope(*this); 993 EmitStmt(S.getBody()); 994 } 995 996 // If there is an increment, emit it next. 997 if (S.getInc()) { 998 EmitBlock(Continue.getBlock()); 999 EmitStmt(S.getInc()); 1000 } 1001 1002 BreakContinueStack.pop_back(); 1003 1004 ConditionScope.ForceCleanup(); 1005 1006 EmitStopPoint(&S); 1007 EmitBranch(CondBlock); 1008 1009 ForScope.ForceCleanup(); 1010 1011 LoopStack.pop(); 1012 1013 // Emit the fall-through block. 1014 EmitBlock(LoopExit.getBlock(), true); 1015 } 1016 1017 void 1018 CodeGenFunction::EmitCXXForRangeStmt(const CXXForRangeStmt &S, 1019 ArrayRef<const Attr *> ForAttrs) { 1020 JumpDest LoopExit = getJumpDestInCurrentScope("for.end"); 1021 1022 LexicalScope ForScope(*this, S.getSourceRange()); 1023 1024 // Evaluate the first pieces before the loop. 1025 if (S.getInit()) 1026 EmitStmt(S.getInit()); 1027 EmitStmt(S.getRangeStmt()); 1028 EmitStmt(S.getBeginStmt()); 1029 EmitStmt(S.getEndStmt()); 1030 1031 // Start the loop with a block that tests the condition. 1032 // If there's an increment, the continue scope will be overwritten 1033 // later. 1034 llvm::BasicBlock *CondBlock = createBasicBlock("for.cond"); 1035 EmitBlock(CondBlock); 1036 1037 const SourceRange &R = S.getSourceRange(); 1038 LoopStack.push(CondBlock, CGM.getContext(), CGM.getCodeGenOpts(), ForAttrs, 1039 SourceLocToDebugLoc(R.getBegin()), 1040 SourceLocToDebugLoc(R.getEnd())); 1041 1042 // If there are any cleanups between here and the loop-exit scope, 1043 // create a block to stage a loop exit along. 1044 llvm::BasicBlock *ExitBlock = LoopExit.getBlock(); 1045 if (ForScope.requiresCleanups()) 1046 ExitBlock = createBasicBlock("for.cond.cleanup"); 1047 1048 // The loop body, consisting of the specified body and the loop variable. 1049 llvm::BasicBlock *ForBody = createBasicBlock("for.body"); 1050 1051 // The body is executed if the expression, contextually converted 1052 // to bool, is true. 1053 llvm::Value *BoolCondVal = EvaluateExprAsBool(S.getCond()); 1054 llvm::MDNode *Weights = createProfileOrBranchWeightsForLoop( 1055 S.getCond(), getProfileCount(S.getBody()), S.getBody()); 1056 Builder.CreateCondBr(BoolCondVal, ForBody, ExitBlock, Weights); 1057 1058 if (ExitBlock != LoopExit.getBlock()) { 1059 EmitBlock(ExitBlock); 1060 EmitBranchThroughCleanup(LoopExit); 1061 } 1062 1063 EmitBlock(ForBody); 1064 incrementProfileCounter(&S); 1065 1066 // Create a block for the increment. In case of a 'continue', we jump there. 1067 JumpDest Continue = getJumpDestInCurrentScope("for.inc"); 1068 1069 // Store the blocks to use for break and continue. 1070 BreakContinueStack.push_back(BreakContinue(LoopExit, Continue)); 1071 1072 { 1073 // Create a separate cleanup scope for the loop variable and body. 1074 LexicalScope BodyScope(*this, S.getSourceRange()); 1075 EmitStmt(S.getLoopVarStmt()); 1076 EmitStmt(S.getBody()); 1077 } 1078 1079 EmitStopPoint(&S); 1080 // If there is an increment, emit it next. 1081 EmitBlock(Continue.getBlock()); 1082 EmitStmt(S.getInc()); 1083 1084 BreakContinueStack.pop_back(); 1085 1086 EmitBranch(CondBlock); 1087 1088 ForScope.ForceCleanup(); 1089 1090 LoopStack.pop(); 1091 1092 // Emit the fall-through block. 1093 EmitBlock(LoopExit.getBlock(), true); 1094 } 1095 1096 void CodeGenFunction::EmitReturnOfRValue(RValue RV, QualType Ty) { 1097 if (RV.isScalar()) { 1098 Builder.CreateStore(RV.getScalarVal(), ReturnValue); 1099 } else if (RV.isAggregate()) { 1100 LValue Dest = MakeAddrLValue(ReturnValue, Ty); 1101 LValue Src = MakeAddrLValue(RV.getAggregateAddress(), Ty); 1102 EmitAggregateCopy(Dest, Src, Ty, getOverlapForReturnValue()); 1103 } else { 1104 EmitStoreOfComplex(RV.getComplexVal(), MakeAddrLValue(ReturnValue, Ty), 1105 /*init*/ true); 1106 } 1107 EmitBranchThroughCleanup(ReturnBlock); 1108 } 1109 1110 namespace { 1111 // RAII struct used to save and restore a return statment's result expression. 1112 struct SaveRetExprRAII { 1113 SaveRetExprRAII(const Expr *RetExpr, CodeGenFunction &CGF) 1114 : OldRetExpr(CGF.RetExpr), CGF(CGF) { 1115 CGF.RetExpr = RetExpr; 1116 } 1117 ~SaveRetExprRAII() { CGF.RetExpr = OldRetExpr; } 1118 const Expr *OldRetExpr; 1119 CodeGenFunction &CGF; 1120 }; 1121 } // namespace 1122 1123 /// EmitReturnStmt - Note that due to GCC extensions, this can have an operand 1124 /// if the function returns void, or may be missing one if the function returns 1125 /// non-void. Fun stuff :). 1126 void CodeGenFunction::EmitReturnStmt(const ReturnStmt &S) { 1127 if (requiresReturnValueCheck()) { 1128 llvm::Constant *SLoc = EmitCheckSourceLocation(S.getBeginLoc()); 1129 auto *SLocPtr = 1130 new llvm::GlobalVariable(CGM.getModule(), SLoc->getType(), false, 1131 llvm::GlobalVariable::PrivateLinkage, SLoc); 1132 SLocPtr->setUnnamedAddr(llvm::GlobalValue::UnnamedAddr::Global); 1133 CGM.getSanitizerMetadata()->disableSanitizerForGlobal(SLocPtr); 1134 assert(ReturnLocation.isValid() && "No valid return location"); 1135 Builder.CreateStore(Builder.CreateBitCast(SLocPtr, Int8PtrTy), 1136 ReturnLocation); 1137 } 1138 1139 // Returning from an outlined SEH helper is UB, and we already warn on it. 1140 if (IsOutlinedSEHHelper) { 1141 Builder.CreateUnreachable(); 1142 Builder.ClearInsertionPoint(); 1143 } 1144 1145 // Emit the result value, even if unused, to evaluate the side effects. 1146 const Expr *RV = S.getRetValue(); 1147 1148 // Record the result expression of the return statement. The recorded 1149 // expression is used to determine whether a block capture's lifetime should 1150 // end at the end of the full expression as opposed to the end of the scope 1151 // enclosing the block expression. 1152 // 1153 // This permits a small, easily-implemented exception to our over-conservative 1154 // rules about not jumping to statements following block literals with 1155 // non-trivial cleanups. 1156 SaveRetExprRAII SaveRetExpr(RV, *this); 1157 1158 RunCleanupsScope cleanupScope(*this); 1159 if (const auto *EWC = dyn_cast_or_null<ExprWithCleanups>(RV)) 1160 RV = EWC->getSubExpr(); 1161 // FIXME: Clean this up by using an LValue for ReturnTemp, 1162 // EmitStoreThroughLValue, and EmitAnyExpr. 1163 // Check if the NRVO candidate was not globalized in OpenMP mode. 1164 if (getLangOpts().ElideConstructors && S.getNRVOCandidate() && 1165 S.getNRVOCandidate()->isNRVOVariable() && 1166 (!getLangOpts().OpenMP || 1167 !CGM.getOpenMPRuntime() 1168 .getAddressOfLocalVariable(*this, S.getNRVOCandidate()) 1169 .isValid())) { 1170 // Apply the named return value optimization for this return statement, 1171 // which means doing nothing: the appropriate result has already been 1172 // constructed into the NRVO variable. 1173 1174 // If there is an NRVO flag for this variable, set it to 1 into indicate 1175 // that the cleanup code should not destroy the variable. 1176 if (llvm::Value *NRVOFlag = NRVOFlags[S.getNRVOCandidate()]) 1177 Builder.CreateFlagStore(Builder.getTrue(), NRVOFlag); 1178 } else if (!ReturnValue.isValid() || (RV && RV->getType()->isVoidType())) { 1179 // Make sure not to return anything, but evaluate the expression 1180 // for side effects. 1181 if (RV) 1182 EmitAnyExpr(RV); 1183 } else if (!RV) { 1184 // Do nothing (return value is left uninitialized) 1185 } else if (FnRetTy->isReferenceType()) { 1186 // If this function returns a reference, take the address of the expression 1187 // rather than the value. 1188 RValue Result = EmitReferenceBindingToExpr(RV); 1189 Builder.CreateStore(Result.getScalarVal(), ReturnValue); 1190 } else { 1191 switch (getEvaluationKind(RV->getType())) { 1192 case TEK_Scalar: 1193 Builder.CreateStore(EmitScalarExpr(RV), ReturnValue); 1194 break; 1195 case TEK_Complex: 1196 EmitComplexExprIntoLValue(RV, MakeAddrLValue(ReturnValue, RV->getType()), 1197 /*isInit*/ true); 1198 break; 1199 case TEK_Aggregate: 1200 EmitAggExpr(RV, AggValueSlot::forAddr( 1201 ReturnValue, Qualifiers(), 1202 AggValueSlot::IsDestructed, 1203 AggValueSlot::DoesNotNeedGCBarriers, 1204 AggValueSlot::IsNotAliased, 1205 getOverlapForReturnValue())); 1206 break; 1207 } 1208 } 1209 1210 ++NumReturnExprs; 1211 if (!RV || RV->isEvaluatable(getContext())) 1212 ++NumSimpleReturnExprs; 1213 1214 cleanupScope.ForceCleanup(); 1215 EmitBranchThroughCleanup(ReturnBlock); 1216 } 1217 1218 void CodeGenFunction::EmitDeclStmt(const DeclStmt &S) { 1219 // As long as debug info is modeled with instructions, we have to ensure we 1220 // have a place to insert here and write the stop point here. 1221 if (HaveInsertPoint()) 1222 EmitStopPoint(&S); 1223 1224 for (const auto *I : S.decls()) 1225 EmitDecl(*I); 1226 } 1227 1228 void CodeGenFunction::EmitBreakStmt(const BreakStmt &S) { 1229 assert(!BreakContinueStack.empty() && "break stmt not in a loop or switch!"); 1230 1231 // If this code is reachable then emit a stop point (if generating 1232 // debug info). We have to do this ourselves because we are on the 1233 // "simple" statement path. 1234 if (HaveInsertPoint()) 1235 EmitStopPoint(&S); 1236 1237 EmitBranchThroughCleanup(BreakContinueStack.back().BreakBlock); 1238 } 1239 1240 void CodeGenFunction::EmitContinueStmt(const ContinueStmt &S) { 1241 assert(!BreakContinueStack.empty() && "continue stmt not in a loop!"); 1242 1243 // If this code is reachable then emit a stop point (if generating 1244 // debug info). We have to do this ourselves because we are on the 1245 // "simple" statement path. 1246 if (HaveInsertPoint()) 1247 EmitStopPoint(&S); 1248 1249 EmitBranchThroughCleanup(BreakContinueStack.back().ContinueBlock); 1250 } 1251 1252 /// EmitCaseStmtRange - If case statement range is not too big then 1253 /// add multiple cases to switch instruction, one for each value within 1254 /// the range. If range is too big then emit "if" condition check. 1255 void CodeGenFunction::EmitCaseStmtRange(const CaseStmt &S, 1256 ArrayRef<const Attr *> Attrs) { 1257 assert(S.getRHS() && "Expected RHS value in CaseStmt"); 1258 1259 llvm::APSInt LHS = S.getLHS()->EvaluateKnownConstInt(getContext()); 1260 llvm::APSInt RHS = S.getRHS()->EvaluateKnownConstInt(getContext()); 1261 1262 // Emit the code for this case. We do this first to make sure it is 1263 // properly chained from our predecessor before generating the 1264 // switch machinery to enter this block. 1265 llvm::BasicBlock *CaseDest = createBasicBlock("sw.bb"); 1266 EmitBlockWithFallThrough(CaseDest, &S); 1267 EmitStmt(S.getSubStmt()); 1268 1269 // If range is empty, do nothing. 1270 if (LHS.isSigned() ? RHS.slt(LHS) : RHS.ult(LHS)) 1271 return; 1272 1273 Stmt::Likelihood LH = Stmt::getLikelihood(Attrs); 1274 llvm::APInt Range = RHS - LHS; 1275 // FIXME: parameters such as this should not be hardcoded. 1276 if (Range.ult(llvm::APInt(Range.getBitWidth(), 64))) { 1277 // Range is small enough to add multiple switch instruction cases. 1278 uint64_t Total = getProfileCount(&S); 1279 unsigned NCases = Range.getZExtValue() + 1; 1280 // We only have one region counter for the entire set of cases here, so we 1281 // need to divide the weights evenly between the generated cases, ensuring 1282 // that the total weight is preserved. E.g., a weight of 5 over three cases 1283 // will be distributed as weights of 2, 2, and 1. 1284 uint64_t Weight = Total / NCases, Rem = Total % NCases; 1285 for (unsigned I = 0; I != NCases; ++I) { 1286 if (SwitchWeights) 1287 SwitchWeights->push_back(Weight + (Rem ? 1 : 0)); 1288 else if (SwitchLikelihood) 1289 SwitchLikelihood->push_back(LH); 1290 1291 if (Rem) 1292 Rem--; 1293 SwitchInsn->addCase(Builder.getInt(LHS), CaseDest); 1294 ++LHS; 1295 } 1296 return; 1297 } 1298 1299 // The range is too big. Emit "if" condition into a new block, 1300 // making sure to save and restore the current insertion point. 1301 llvm::BasicBlock *RestoreBB = Builder.GetInsertBlock(); 1302 1303 // Push this test onto the chain of range checks (which terminates 1304 // in the default basic block). The switch's default will be changed 1305 // to the top of this chain after switch emission is complete. 1306 llvm::BasicBlock *FalseDest = CaseRangeBlock; 1307 CaseRangeBlock = createBasicBlock("sw.caserange"); 1308 1309 CurFn->getBasicBlockList().push_back(CaseRangeBlock); 1310 Builder.SetInsertPoint(CaseRangeBlock); 1311 1312 // Emit range check. 1313 llvm::Value *Diff = 1314 Builder.CreateSub(SwitchInsn->getCondition(), Builder.getInt(LHS)); 1315 llvm::Value *Cond = 1316 Builder.CreateICmpULE(Diff, Builder.getInt(Range), "inbounds"); 1317 1318 llvm::MDNode *Weights = nullptr; 1319 if (SwitchWeights) { 1320 uint64_t ThisCount = getProfileCount(&S); 1321 uint64_t DefaultCount = (*SwitchWeights)[0]; 1322 Weights = createProfileWeights(ThisCount, DefaultCount); 1323 1324 // Since we're chaining the switch default through each large case range, we 1325 // need to update the weight for the default, ie, the first case, to include 1326 // this case. 1327 (*SwitchWeights)[0] += ThisCount; 1328 } else if (SwitchLikelihood) 1329 Weights = createBranchWeights(LH); 1330 1331 Builder.CreateCondBr(Cond, CaseDest, FalseDest, Weights); 1332 1333 // Restore the appropriate insertion point. 1334 if (RestoreBB) 1335 Builder.SetInsertPoint(RestoreBB); 1336 else 1337 Builder.ClearInsertionPoint(); 1338 } 1339 1340 void CodeGenFunction::EmitCaseStmt(const CaseStmt &S, 1341 ArrayRef<const Attr *> Attrs) { 1342 // If there is no enclosing switch instance that we're aware of, then this 1343 // case statement and its block can be elided. This situation only happens 1344 // when we've constant-folded the switch, are emitting the constant case, 1345 // and part of the constant case includes another case statement. For 1346 // instance: switch (4) { case 4: do { case 5: } while (1); } 1347 if (!SwitchInsn) { 1348 EmitStmt(S.getSubStmt()); 1349 return; 1350 } 1351 1352 // Handle case ranges. 1353 if (S.getRHS()) { 1354 EmitCaseStmtRange(S, Attrs); 1355 return; 1356 } 1357 1358 llvm::ConstantInt *CaseVal = 1359 Builder.getInt(S.getLHS()->EvaluateKnownConstInt(getContext())); 1360 if (SwitchLikelihood) 1361 SwitchLikelihood->push_back(Stmt::getLikelihood(Attrs)); 1362 1363 // If the body of the case is just a 'break', try to not emit an empty block. 1364 // If we're profiling or we're not optimizing, leave the block in for better 1365 // debug and coverage analysis. 1366 if (!CGM.getCodeGenOpts().hasProfileClangInstr() && 1367 CGM.getCodeGenOpts().OptimizationLevel > 0 && 1368 isa<BreakStmt>(S.getSubStmt())) { 1369 JumpDest Block = BreakContinueStack.back().BreakBlock; 1370 1371 // Only do this optimization if there are no cleanups that need emitting. 1372 if (isObviouslyBranchWithoutCleanups(Block)) { 1373 if (SwitchWeights) 1374 SwitchWeights->push_back(getProfileCount(&S)); 1375 SwitchInsn->addCase(CaseVal, Block.getBlock()); 1376 1377 // If there was a fallthrough into this case, make sure to redirect it to 1378 // the end of the switch as well. 1379 if (Builder.GetInsertBlock()) { 1380 Builder.CreateBr(Block.getBlock()); 1381 Builder.ClearInsertionPoint(); 1382 } 1383 return; 1384 } 1385 } 1386 1387 llvm::BasicBlock *CaseDest = createBasicBlock("sw.bb"); 1388 EmitBlockWithFallThrough(CaseDest, &S); 1389 if (SwitchWeights) 1390 SwitchWeights->push_back(getProfileCount(&S)); 1391 SwitchInsn->addCase(CaseVal, CaseDest); 1392 1393 // Recursively emitting the statement is acceptable, but is not wonderful for 1394 // code where we have many case statements nested together, i.e.: 1395 // case 1: 1396 // case 2: 1397 // case 3: etc. 1398 // Handling this recursively will create a new block for each case statement 1399 // that falls through to the next case which is IR intensive. It also causes 1400 // deep recursion which can run into stack depth limitations. Handle 1401 // sequential non-range case statements specially. 1402 // 1403 // TODO When the next case has a likelihood attribute the code returns to the 1404 // recursive algorithm. Maybe improve this case if it becomes common practice 1405 // to use a lot of attributes. 1406 const CaseStmt *CurCase = &S; 1407 const CaseStmt *NextCase = dyn_cast<CaseStmt>(S.getSubStmt()); 1408 1409 // Otherwise, iteratively add consecutive cases to this switch stmt. 1410 while (NextCase && NextCase->getRHS() == nullptr) { 1411 CurCase = NextCase; 1412 llvm::ConstantInt *CaseVal = 1413 Builder.getInt(CurCase->getLHS()->EvaluateKnownConstInt(getContext())); 1414 1415 if (SwitchWeights) 1416 SwitchWeights->push_back(getProfileCount(NextCase)); 1417 if (CGM.getCodeGenOpts().hasProfileClangInstr()) { 1418 CaseDest = createBasicBlock("sw.bb"); 1419 EmitBlockWithFallThrough(CaseDest, &S); 1420 } 1421 // Since this loop is only executed when the CaseStmt has no attributes 1422 // use a hard-coded value. 1423 if (SwitchLikelihood) 1424 SwitchLikelihood->push_back(Stmt::LH_None); 1425 1426 SwitchInsn->addCase(CaseVal, CaseDest); 1427 NextCase = dyn_cast<CaseStmt>(CurCase->getSubStmt()); 1428 } 1429 1430 // Normal default recursion for non-cases. 1431 EmitStmt(CurCase->getSubStmt()); 1432 } 1433 1434 void CodeGenFunction::EmitDefaultStmt(const DefaultStmt &S, 1435 ArrayRef<const Attr *> Attrs) { 1436 // If there is no enclosing switch instance that we're aware of, then this 1437 // default statement can be elided. This situation only happens when we've 1438 // constant-folded the switch. 1439 if (!SwitchInsn) { 1440 EmitStmt(S.getSubStmt()); 1441 return; 1442 } 1443 1444 llvm::BasicBlock *DefaultBlock = SwitchInsn->getDefaultDest(); 1445 assert(DefaultBlock->empty() && 1446 "EmitDefaultStmt: Default block already defined?"); 1447 1448 if (SwitchLikelihood) 1449 SwitchLikelihood->front() = Stmt::getLikelihood(Attrs); 1450 1451 EmitBlockWithFallThrough(DefaultBlock, &S); 1452 1453 EmitStmt(S.getSubStmt()); 1454 } 1455 1456 /// CollectStatementsForCase - Given the body of a 'switch' statement and a 1457 /// constant value that is being switched on, see if we can dead code eliminate 1458 /// the body of the switch to a simple series of statements to emit. Basically, 1459 /// on a switch (5) we want to find these statements: 1460 /// case 5: 1461 /// printf(...); <-- 1462 /// ++i; <-- 1463 /// break; 1464 /// 1465 /// and add them to the ResultStmts vector. If it is unsafe to do this 1466 /// transformation (for example, one of the elided statements contains a label 1467 /// that might be jumped to), return CSFC_Failure. If we handled it and 'S' 1468 /// should include statements after it (e.g. the printf() line is a substmt of 1469 /// the case) then return CSFC_FallThrough. If we handled it and found a break 1470 /// statement, then return CSFC_Success. 1471 /// 1472 /// If Case is non-null, then we are looking for the specified case, checking 1473 /// that nothing we jump over contains labels. If Case is null, then we found 1474 /// the case and are looking for the break. 1475 /// 1476 /// If the recursive walk actually finds our Case, then we set FoundCase to 1477 /// true. 1478 /// 1479 enum CSFC_Result { CSFC_Failure, CSFC_FallThrough, CSFC_Success }; 1480 static CSFC_Result CollectStatementsForCase(const Stmt *S, 1481 const SwitchCase *Case, 1482 bool &FoundCase, 1483 SmallVectorImpl<const Stmt*> &ResultStmts) { 1484 // If this is a null statement, just succeed. 1485 if (!S) 1486 return Case ? CSFC_Success : CSFC_FallThrough; 1487 1488 // If this is the switchcase (case 4: or default) that we're looking for, then 1489 // we're in business. Just add the substatement. 1490 if (const SwitchCase *SC = dyn_cast<SwitchCase>(S)) { 1491 if (S == Case) { 1492 FoundCase = true; 1493 return CollectStatementsForCase(SC->getSubStmt(), nullptr, FoundCase, 1494 ResultStmts); 1495 } 1496 1497 // Otherwise, this is some other case or default statement, just ignore it. 1498 return CollectStatementsForCase(SC->getSubStmt(), Case, FoundCase, 1499 ResultStmts); 1500 } 1501 1502 // If we are in the live part of the code and we found our break statement, 1503 // return a success! 1504 if (!Case && isa<BreakStmt>(S)) 1505 return CSFC_Success; 1506 1507 // If this is a switch statement, then it might contain the SwitchCase, the 1508 // break, or neither. 1509 if (const CompoundStmt *CS = dyn_cast<CompoundStmt>(S)) { 1510 // Handle this as two cases: we might be looking for the SwitchCase (if so 1511 // the skipped statements must be skippable) or we might already have it. 1512 CompoundStmt::const_body_iterator I = CS->body_begin(), E = CS->body_end(); 1513 bool StartedInLiveCode = FoundCase; 1514 unsigned StartSize = ResultStmts.size(); 1515 1516 // If we've not found the case yet, scan through looking for it. 1517 if (Case) { 1518 // Keep track of whether we see a skipped declaration. The code could be 1519 // using the declaration even if it is skipped, so we can't optimize out 1520 // the decl if the kept statements might refer to it. 1521 bool HadSkippedDecl = false; 1522 1523 // If we're looking for the case, just see if we can skip each of the 1524 // substatements. 1525 for (; Case && I != E; ++I) { 1526 HadSkippedDecl |= CodeGenFunction::mightAddDeclToScope(*I); 1527 1528 switch (CollectStatementsForCase(*I, Case, FoundCase, ResultStmts)) { 1529 case CSFC_Failure: return CSFC_Failure; 1530 case CSFC_Success: 1531 // A successful result means that either 1) that the statement doesn't 1532 // have the case and is skippable, or 2) does contain the case value 1533 // and also contains the break to exit the switch. In the later case, 1534 // we just verify the rest of the statements are elidable. 1535 if (FoundCase) { 1536 // If we found the case and skipped declarations, we can't do the 1537 // optimization. 1538 if (HadSkippedDecl) 1539 return CSFC_Failure; 1540 1541 for (++I; I != E; ++I) 1542 if (CodeGenFunction::ContainsLabel(*I, true)) 1543 return CSFC_Failure; 1544 return CSFC_Success; 1545 } 1546 break; 1547 case CSFC_FallThrough: 1548 // If we have a fallthrough condition, then we must have found the 1549 // case started to include statements. Consider the rest of the 1550 // statements in the compound statement as candidates for inclusion. 1551 assert(FoundCase && "Didn't find case but returned fallthrough?"); 1552 // We recursively found Case, so we're not looking for it anymore. 1553 Case = nullptr; 1554 1555 // If we found the case and skipped declarations, we can't do the 1556 // optimization. 1557 if (HadSkippedDecl) 1558 return CSFC_Failure; 1559 break; 1560 } 1561 } 1562 1563 if (!FoundCase) 1564 return CSFC_Success; 1565 1566 assert(!HadSkippedDecl && "fallthrough after skipping decl"); 1567 } 1568 1569 // If we have statements in our range, then we know that the statements are 1570 // live and need to be added to the set of statements we're tracking. 1571 bool AnyDecls = false; 1572 for (; I != E; ++I) { 1573 AnyDecls |= CodeGenFunction::mightAddDeclToScope(*I); 1574 1575 switch (CollectStatementsForCase(*I, nullptr, FoundCase, ResultStmts)) { 1576 case CSFC_Failure: return CSFC_Failure; 1577 case CSFC_FallThrough: 1578 // A fallthrough result means that the statement was simple and just 1579 // included in ResultStmt, keep adding them afterwards. 1580 break; 1581 case CSFC_Success: 1582 // A successful result means that we found the break statement and 1583 // stopped statement inclusion. We just ensure that any leftover stmts 1584 // are skippable and return success ourselves. 1585 for (++I; I != E; ++I) 1586 if (CodeGenFunction::ContainsLabel(*I, true)) 1587 return CSFC_Failure; 1588 return CSFC_Success; 1589 } 1590 } 1591 1592 // If we're about to fall out of a scope without hitting a 'break;', we 1593 // can't perform the optimization if there were any decls in that scope 1594 // (we'd lose their end-of-lifetime). 1595 if (AnyDecls) { 1596 // If the entire compound statement was live, there's one more thing we 1597 // can try before giving up: emit the whole thing as a single statement. 1598 // We can do that unless the statement contains a 'break;'. 1599 // FIXME: Such a break must be at the end of a construct within this one. 1600 // We could emit this by just ignoring the BreakStmts entirely. 1601 if (StartedInLiveCode && !CodeGenFunction::containsBreak(S)) { 1602 ResultStmts.resize(StartSize); 1603 ResultStmts.push_back(S); 1604 } else { 1605 return CSFC_Failure; 1606 } 1607 } 1608 1609 return CSFC_FallThrough; 1610 } 1611 1612 // Okay, this is some other statement that we don't handle explicitly, like a 1613 // for statement or increment etc. If we are skipping over this statement, 1614 // just verify it doesn't have labels, which would make it invalid to elide. 1615 if (Case) { 1616 if (CodeGenFunction::ContainsLabel(S, true)) 1617 return CSFC_Failure; 1618 return CSFC_Success; 1619 } 1620 1621 // Otherwise, we want to include this statement. Everything is cool with that 1622 // so long as it doesn't contain a break out of the switch we're in. 1623 if (CodeGenFunction::containsBreak(S)) return CSFC_Failure; 1624 1625 // Otherwise, everything is great. Include the statement and tell the caller 1626 // that we fall through and include the next statement as well. 1627 ResultStmts.push_back(S); 1628 return CSFC_FallThrough; 1629 } 1630 1631 /// FindCaseStatementsForValue - Find the case statement being jumped to and 1632 /// then invoke CollectStatementsForCase to find the list of statements to emit 1633 /// for a switch on constant. See the comment above CollectStatementsForCase 1634 /// for more details. 1635 static bool FindCaseStatementsForValue(const SwitchStmt &S, 1636 const llvm::APSInt &ConstantCondValue, 1637 SmallVectorImpl<const Stmt*> &ResultStmts, 1638 ASTContext &C, 1639 const SwitchCase *&ResultCase) { 1640 // First step, find the switch case that is being branched to. We can do this 1641 // efficiently by scanning the SwitchCase list. 1642 const SwitchCase *Case = S.getSwitchCaseList(); 1643 const DefaultStmt *DefaultCase = nullptr; 1644 1645 for (; Case; Case = Case->getNextSwitchCase()) { 1646 // It's either a default or case. Just remember the default statement in 1647 // case we're not jumping to any numbered cases. 1648 if (const DefaultStmt *DS = dyn_cast<DefaultStmt>(Case)) { 1649 DefaultCase = DS; 1650 continue; 1651 } 1652 1653 // Check to see if this case is the one we're looking for. 1654 const CaseStmt *CS = cast<CaseStmt>(Case); 1655 // Don't handle case ranges yet. 1656 if (CS->getRHS()) return false; 1657 1658 // If we found our case, remember it as 'case'. 1659 if (CS->getLHS()->EvaluateKnownConstInt(C) == ConstantCondValue) 1660 break; 1661 } 1662 1663 // If we didn't find a matching case, we use a default if it exists, or we 1664 // elide the whole switch body! 1665 if (!Case) { 1666 // It is safe to elide the body of the switch if it doesn't contain labels 1667 // etc. If it is safe, return successfully with an empty ResultStmts list. 1668 if (!DefaultCase) 1669 return !CodeGenFunction::ContainsLabel(&S); 1670 Case = DefaultCase; 1671 } 1672 1673 // Ok, we know which case is being jumped to, try to collect all the 1674 // statements that follow it. This can fail for a variety of reasons. Also, 1675 // check to see that the recursive walk actually found our case statement. 1676 // Insane cases like this can fail to find it in the recursive walk since we 1677 // don't handle every stmt kind: 1678 // switch (4) { 1679 // while (1) { 1680 // case 4: ... 1681 bool FoundCase = false; 1682 ResultCase = Case; 1683 return CollectStatementsForCase(S.getBody(), Case, FoundCase, 1684 ResultStmts) != CSFC_Failure && 1685 FoundCase; 1686 } 1687 1688 static Optional<SmallVector<uint64_t, 16>> 1689 getLikelihoodWeights(ArrayRef<Stmt::Likelihood> Likelihoods) { 1690 // Are there enough branches to weight them? 1691 if (Likelihoods.size() <= 1) 1692 return None; 1693 1694 uint64_t NumUnlikely = 0; 1695 uint64_t NumNone = 0; 1696 uint64_t NumLikely = 0; 1697 for (const auto LH : Likelihoods) { 1698 switch (LH) { 1699 case Stmt::LH_Unlikely: 1700 ++NumUnlikely; 1701 break; 1702 case Stmt::LH_None: 1703 ++NumNone; 1704 break; 1705 case Stmt::LH_Likely: 1706 ++NumLikely; 1707 break; 1708 } 1709 } 1710 1711 // Is there a likelihood attribute used? 1712 if (NumUnlikely == 0 && NumLikely == 0) 1713 return None; 1714 1715 // When multiple cases share the same code they can be combined during 1716 // optimization. In that case the weights of the branch will be the sum of 1717 // the individual weights. Make sure the combined sum of all neutral cases 1718 // doesn't exceed the value of a single likely attribute. 1719 // The additions both avoid divisions by 0 and make sure the weights of None 1720 // don't exceed the weight of Likely. 1721 const uint64_t Likely = INT32_MAX / (NumLikely + 2); 1722 const uint64_t None = Likely / (NumNone + 1); 1723 const uint64_t Unlikely = 0; 1724 1725 SmallVector<uint64_t, 16> Result; 1726 Result.reserve(Likelihoods.size()); 1727 for (const auto LH : Likelihoods) { 1728 switch (LH) { 1729 case Stmt::LH_Unlikely: 1730 Result.push_back(Unlikely); 1731 break; 1732 case Stmt::LH_None: 1733 Result.push_back(None); 1734 break; 1735 case Stmt::LH_Likely: 1736 Result.push_back(Likely); 1737 break; 1738 } 1739 } 1740 1741 return Result; 1742 } 1743 1744 void CodeGenFunction::EmitSwitchStmt(const SwitchStmt &S) { 1745 // Handle nested switch statements. 1746 llvm::SwitchInst *SavedSwitchInsn = SwitchInsn; 1747 SmallVector<uint64_t, 16> *SavedSwitchWeights = SwitchWeights; 1748 SmallVector<Stmt::Likelihood, 16> *SavedSwitchLikelihood = SwitchLikelihood; 1749 llvm::BasicBlock *SavedCRBlock = CaseRangeBlock; 1750 1751 // See if we can constant fold the condition of the switch and therefore only 1752 // emit the live case statement (if any) of the switch. 1753 llvm::APSInt ConstantCondValue; 1754 if (ConstantFoldsToSimpleInteger(S.getCond(), ConstantCondValue)) { 1755 SmallVector<const Stmt*, 4> CaseStmts; 1756 const SwitchCase *Case = nullptr; 1757 if (FindCaseStatementsForValue(S, ConstantCondValue, CaseStmts, 1758 getContext(), Case)) { 1759 if (Case) 1760 incrementProfileCounter(Case); 1761 RunCleanupsScope ExecutedScope(*this); 1762 1763 if (S.getInit()) 1764 EmitStmt(S.getInit()); 1765 1766 // Emit the condition variable if needed inside the entire cleanup scope 1767 // used by this special case for constant folded switches. 1768 if (S.getConditionVariable()) 1769 EmitDecl(*S.getConditionVariable()); 1770 1771 // At this point, we are no longer "within" a switch instance, so 1772 // we can temporarily enforce this to ensure that any embedded case 1773 // statements are not emitted. 1774 SwitchInsn = nullptr; 1775 1776 // Okay, we can dead code eliminate everything except this case. Emit the 1777 // specified series of statements and we're good. 1778 for (unsigned i = 0, e = CaseStmts.size(); i != e; ++i) 1779 EmitStmt(CaseStmts[i]); 1780 incrementProfileCounter(&S); 1781 1782 // Now we want to restore the saved switch instance so that nested 1783 // switches continue to function properly 1784 SwitchInsn = SavedSwitchInsn; 1785 1786 return; 1787 } 1788 } 1789 1790 JumpDest SwitchExit = getJumpDestInCurrentScope("sw.epilog"); 1791 1792 RunCleanupsScope ConditionScope(*this); 1793 1794 if (S.getInit()) 1795 EmitStmt(S.getInit()); 1796 1797 if (S.getConditionVariable()) 1798 EmitDecl(*S.getConditionVariable()); 1799 llvm::Value *CondV = EmitScalarExpr(S.getCond()); 1800 1801 // Create basic block to hold stuff that comes after switch 1802 // statement. We also need to create a default block now so that 1803 // explicit case ranges tests can have a place to jump to on 1804 // failure. 1805 llvm::BasicBlock *DefaultBlock = createBasicBlock("sw.default"); 1806 SwitchInsn = Builder.CreateSwitch(CondV, DefaultBlock); 1807 if (PGO.haveRegionCounts()) { 1808 // Walk the SwitchCase list to find how many there are. 1809 uint64_t DefaultCount = 0; 1810 unsigned NumCases = 0; 1811 for (const SwitchCase *Case = S.getSwitchCaseList(); 1812 Case; 1813 Case = Case->getNextSwitchCase()) { 1814 if (isa<DefaultStmt>(Case)) 1815 DefaultCount = getProfileCount(Case); 1816 NumCases += 1; 1817 } 1818 SwitchWeights = new SmallVector<uint64_t, 16>(); 1819 SwitchWeights->reserve(NumCases); 1820 // The default needs to be first. We store the edge count, so we already 1821 // know the right weight. 1822 SwitchWeights->push_back(DefaultCount); 1823 } else if (CGM.getCodeGenOpts().OptimizationLevel) { 1824 SwitchLikelihood = new SmallVector<Stmt::Likelihood, 16>(); 1825 // Initialize the default case. 1826 SwitchLikelihood->push_back(Stmt::LH_None); 1827 } 1828 1829 CaseRangeBlock = DefaultBlock; 1830 1831 // Clear the insertion point to indicate we are in unreachable code. 1832 Builder.ClearInsertionPoint(); 1833 1834 // All break statements jump to NextBlock. If BreakContinueStack is non-empty 1835 // then reuse last ContinueBlock. 1836 JumpDest OuterContinue; 1837 if (!BreakContinueStack.empty()) 1838 OuterContinue = BreakContinueStack.back().ContinueBlock; 1839 1840 BreakContinueStack.push_back(BreakContinue(SwitchExit, OuterContinue)); 1841 1842 // Emit switch body. 1843 EmitStmt(S.getBody()); 1844 1845 BreakContinueStack.pop_back(); 1846 1847 // Update the default block in case explicit case range tests have 1848 // been chained on top. 1849 SwitchInsn->setDefaultDest(CaseRangeBlock); 1850 1851 // If a default was never emitted: 1852 if (!DefaultBlock->getParent()) { 1853 // If we have cleanups, emit the default block so that there's a 1854 // place to jump through the cleanups from. 1855 if (ConditionScope.requiresCleanups()) { 1856 EmitBlock(DefaultBlock); 1857 1858 // Otherwise, just forward the default block to the switch end. 1859 } else { 1860 DefaultBlock->replaceAllUsesWith(SwitchExit.getBlock()); 1861 delete DefaultBlock; 1862 } 1863 } 1864 1865 ConditionScope.ForceCleanup(); 1866 1867 // Emit continuation. 1868 EmitBlock(SwitchExit.getBlock(), true); 1869 incrementProfileCounter(&S); 1870 1871 // If the switch has a condition wrapped by __builtin_unpredictable, 1872 // create metadata that specifies that the switch is unpredictable. 1873 // Don't bother if not optimizing because that metadata would not be used. 1874 auto *Call = dyn_cast<CallExpr>(S.getCond()); 1875 if (Call && CGM.getCodeGenOpts().OptimizationLevel != 0) { 1876 auto *FD = dyn_cast_or_null<FunctionDecl>(Call->getCalleeDecl()); 1877 if (FD && FD->getBuiltinID() == Builtin::BI__builtin_unpredictable) { 1878 llvm::MDBuilder MDHelper(getLLVMContext()); 1879 SwitchInsn->setMetadata(llvm::LLVMContext::MD_unpredictable, 1880 MDHelper.createUnpredictable()); 1881 } 1882 } 1883 1884 if (SwitchWeights) { 1885 assert(SwitchWeights->size() == 1 + SwitchInsn->getNumCases() && 1886 "switch weights do not match switch cases"); 1887 // If there's only one jump destination there's no sense weighting it. 1888 if (SwitchWeights->size() > 1) 1889 SwitchInsn->setMetadata(llvm::LLVMContext::MD_prof, 1890 createProfileWeights(*SwitchWeights)); 1891 delete SwitchWeights; 1892 } else if (SwitchLikelihood) { 1893 assert(SwitchLikelihood->size() == 1 + SwitchInsn->getNumCases() && 1894 "switch likelihoods do not match switch cases"); 1895 Optional<SmallVector<uint64_t, 16>> LHW = 1896 getLikelihoodWeights(*SwitchLikelihood); 1897 if (LHW) { 1898 llvm::MDBuilder MDHelper(CGM.getLLVMContext()); 1899 SwitchInsn->setMetadata(llvm::LLVMContext::MD_prof, 1900 createProfileWeights(*LHW)); 1901 } 1902 delete SwitchLikelihood; 1903 } 1904 SwitchInsn = SavedSwitchInsn; 1905 SwitchWeights = SavedSwitchWeights; 1906 SwitchLikelihood = SavedSwitchLikelihood; 1907 CaseRangeBlock = SavedCRBlock; 1908 } 1909 1910 static std::string 1911 SimplifyConstraint(const char *Constraint, const TargetInfo &Target, 1912 SmallVectorImpl<TargetInfo::ConstraintInfo> *OutCons=nullptr) { 1913 std::string Result; 1914 1915 while (*Constraint) { 1916 switch (*Constraint) { 1917 default: 1918 Result += Target.convertConstraint(Constraint); 1919 break; 1920 // Ignore these 1921 case '*': 1922 case '?': 1923 case '!': 1924 case '=': // Will see this and the following in mult-alt constraints. 1925 case '+': 1926 break; 1927 case '#': // Ignore the rest of the constraint alternative. 1928 while (Constraint[1] && Constraint[1] != ',') 1929 Constraint++; 1930 break; 1931 case '&': 1932 case '%': 1933 Result += *Constraint; 1934 while (Constraint[1] && Constraint[1] == *Constraint) 1935 Constraint++; 1936 break; 1937 case ',': 1938 Result += "|"; 1939 break; 1940 case 'g': 1941 Result += "imr"; 1942 break; 1943 case '[': { 1944 assert(OutCons && 1945 "Must pass output names to constraints with a symbolic name"); 1946 unsigned Index; 1947 bool result = Target.resolveSymbolicName(Constraint, *OutCons, Index); 1948 assert(result && "Could not resolve symbolic name"); (void)result; 1949 Result += llvm::utostr(Index); 1950 break; 1951 } 1952 } 1953 1954 Constraint++; 1955 } 1956 1957 return Result; 1958 } 1959 1960 /// AddVariableConstraints - Look at AsmExpr and if it is a variable declared 1961 /// as using a particular register add that as a constraint that will be used 1962 /// in this asm stmt. 1963 static std::string 1964 AddVariableConstraints(const std::string &Constraint, const Expr &AsmExpr, 1965 const TargetInfo &Target, CodeGenModule &CGM, 1966 const AsmStmt &Stmt, const bool EarlyClobber, 1967 std::string *GCCReg = nullptr) { 1968 const DeclRefExpr *AsmDeclRef = dyn_cast<DeclRefExpr>(&AsmExpr); 1969 if (!AsmDeclRef) 1970 return Constraint; 1971 const ValueDecl &Value = *AsmDeclRef->getDecl(); 1972 const VarDecl *Variable = dyn_cast<VarDecl>(&Value); 1973 if (!Variable) 1974 return Constraint; 1975 if (Variable->getStorageClass() != SC_Register) 1976 return Constraint; 1977 AsmLabelAttr *Attr = Variable->getAttr<AsmLabelAttr>(); 1978 if (!Attr) 1979 return Constraint; 1980 StringRef Register = Attr->getLabel(); 1981 assert(Target.isValidGCCRegisterName(Register)); 1982 // We're using validateOutputConstraint here because we only care if 1983 // this is a register constraint. 1984 TargetInfo::ConstraintInfo Info(Constraint, ""); 1985 if (Target.validateOutputConstraint(Info) && 1986 !Info.allowsRegister()) { 1987 CGM.ErrorUnsupported(&Stmt, "__asm__"); 1988 return Constraint; 1989 } 1990 // Canonicalize the register here before returning it. 1991 Register = Target.getNormalizedGCCRegisterName(Register); 1992 if (GCCReg != nullptr) 1993 *GCCReg = Register.str(); 1994 return (EarlyClobber ? "&{" : "{") + Register.str() + "}"; 1995 } 1996 1997 llvm::Value* 1998 CodeGenFunction::EmitAsmInputLValue(const TargetInfo::ConstraintInfo &Info, 1999 LValue InputValue, QualType InputType, 2000 std::string &ConstraintStr, 2001 SourceLocation Loc) { 2002 llvm::Value *Arg; 2003 if (Info.allowsRegister() || !Info.allowsMemory()) { 2004 if (CodeGenFunction::hasScalarEvaluationKind(InputType)) { 2005 Arg = EmitLoadOfLValue(InputValue, Loc).getScalarVal(); 2006 } else { 2007 llvm::Type *Ty = ConvertType(InputType); 2008 uint64_t Size = CGM.getDataLayout().getTypeSizeInBits(Ty); 2009 if (Size <= 64 && llvm::isPowerOf2_64(Size)) { 2010 Ty = llvm::IntegerType::get(getLLVMContext(), Size); 2011 Ty = llvm::PointerType::getUnqual(Ty); 2012 2013 Arg = Builder.CreateLoad( 2014 Builder.CreateBitCast(InputValue.getAddress(*this), Ty)); 2015 } else { 2016 Arg = InputValue.getPointer(*this); 2017 ConstraintStr += '*'; 2018 } 2019 } 2020 } else { 2021 Arg = InputValue.getPointer(*this); 2022 ConstraintStr += '*'; 2023 } 2024 2025 return Arg; 2026 } 2027 2028 llvm::Value* CodeGenFunction::EmitAsmInput( 2029 const TargetInfo::ConstraintInfo &Info, 2030 const Expr *InputExpr, 2031 std::string &ConstraintStr) { 2032 // If this can't be a register or memory, i.e., has to be a constant 2033 // (immediate or symbolic), try to emit it as such. 2034 if (!Info.allowsRegister() && !Info.allowsMemory()) { 2035 if (Info.requiresImmediateConstant()) { 2036 Expr::EvalResult EVResult; 2037 InputExpr->EvaluateAsRValue(EVResult, getContext(), true); 2038 2039 llvm::APSInt IntResult; 2040 if (EVResult.Val.toIntegralConstant(IntResult, InputExpr->getType(), 2041 getContext())) 2042 return llvm::ConstantInt::get(getLLVMContext(), IntResult); 2043 } 2044 2045 Expr::EvalResult Result; 2046 if (InputExpr->EvaluateAsInt(Result, getContext())) 2047 return llvm::ConstantInt::get(getLLVMContext(), Result.Val.getInt()); 2048 } 2049 2050 if (Info.allowsRegister() || !Info.allowsMemory()) 2051 if (CodeGenFunction::hasScalarEvaluationKind(InputExpr->getType())) 2052 return EmitScalarExpr(InputExpr); 2053 if (InputExpr->getStmtClass() == Expr::CXXThisExprClass) 2054 return EmitScalarExpr(InputExpr); 2055 InputExpr = InputExpr->IgnoreParenNoopCasts(getContext()); 2056 LValue Dest = EmitLValue(InputExpr); 2057 return EmitAsmInputLValue(Info, Dest, InputExpr->getType(), ConstraintStr, 2058 InputExpr->getExprLoc()); 2059 } 2060 2061 /// getAsmSrcLocInfo - Return the !srcloc metadata node to attach to an inline 2062 /// asm call instruction. The !srcloc MDNode contains a list of constant 2063 /// integers which are the source locations of the start of each line in the 2064 /// asm. 2065 static llvm::MDNode *getAsmSrcLocInfo(const StringLiteral *Str, 2066 CodeGenFunction &CGF) { 2067 SmallVector<llvm::Metadata *, 8> Locs; 2068 // Add the location of the first line to the MDNode. 2069 Locs.push_back(llvm::ConstantAsMetadata::get(llvm::ConstantInt::get( 2070 CGF.Int32Ty, Str->getBeginLoc().getRawEncoding()))); 2071 StringRef StrVal = Str->getString(); 2072 if (!StrVal.empty()) { 2073 const SourceManager &SM = CGF.CGM.getContext().getSourceManager(); 2074 const LangOptions &LangOpts = CGF.CGM.getLangOpts(); 2075 unsigned StartToken = 0; 2076 unsigned ByteOffset = 0; 2077 2078 // Add the location of the start of each subsequent line of the asm to the 2079 // MDNode. 2080 for (unsigned i = 0, e = StrVal.size() - 1; i != e; ++i) { 2081 if (StrVal[i] != '\n') continue; 2082 SourceLocation LineLoc = Str->getLocationOfByte( 2083 i + 1, SM, LangOpts, CGF.getTarget(), &StartToken, &ByteOffset); 2084 Locs.push_back(llvm::ConstantAsMetadata::get( 2085 llvm::ConstantInt::get(CGF.Int32Ty, LineLoc.getRawEncoding()))); 2086 } 2087 } 2088 2089 return llvm::MDNode::get(CGF.getLLVMContext(), Locs); 2090 } 2091 2092 static void UpdateAsmCallInst(llvm::CallBase &Result, bool HasSideEffect, 2093 bool ReadOnly, bool ReadNone, bool NoMerge, 2094 const AsmStmt &S, 2095 const std::vector<llvm::Type *> &ResultRegTypes, 2096 CodeGenFunction &CGF, 2097 std::vector<llvm::Value *> &RegResults) { 2098 Result.addAttribute(llvm::AttributeList::FunctionIndex, 2099 llvm::Attribute::NoUnwind); 2100 if (NoMerge) 2101 Result.addAttribute(llvm::AttributeList::FunctionIndex, 2102 llvm::Attribute::NoMerge); 2103 // Attach readnone and readonly attributes. 2104 if (!HasSideEffect) { 2105 if (ReadNone) 2106 Result.addAttribute(llvm::AttributeList::FunctionIndex, 2107 llvm::Attribute::ReadNone); 2108 else if (ReadOnly) 2109 Result.addAttribute(llvm::AttributeList::FunctionIndex, 2110 llvm::Attribute::ReadOnly); 2111 } 2112 2113 // Slap the source location of the inline asm into a !srcloc metadata on the 2114 // call. 2115 if (const auto *gccAsmStmt = dyn_cast<GCCAsmStmt>(&S)) 2116 Result.setMetadata("srcloc", 2117 getAsmSrcLocInfo(gccAsmStmt->getAsmString(), CGF)); 2118 else { 2119 // At least put the line number on MS inline asm blobs. 2120 llvm::Constant *Loc = llvm::ConstantInt::get(CGF.Int32Ty, 2121 S.getAsmLoc().getRawEncoding()); 2122 Result.setMetadata("srcloc", 2123 llvm::MDNode::get(CGF.getLLVMContext(), 2124 llvm::ConstantAsMetadata::get(Loc))); 2125 } 2126 2127 if (CGF.getLangOpts().assumeFunctionsAreConvergent()) 2128 // Conservatively, mark all inline asm blocks in CUDA or OpenCL as 2129 // convergent (meaning, they may call an intrinsically convergent op, such 2130 // as bar.sync, and so can't have certain optimizations applied around 2131 // them). 2132 Result.addAttribute(llvm::AttributeList::FunctionIndex, 2133 llvm::Attribute::Convergent); 2134 // Extract all of the register value results from the asm. 2135 if (ResultRegTypes.size() == 1) { 2136 RegResults.push_back(&Result); 2137 } else { 2138 for (unsigned i = 0, e = ResultRegTypes.size(); i != e; ++i) { 2139 llvm::Value *Tmp = CGF.Builder.CreateExtractValue(&Result, i, "asmresult"); 2140 RegResults.push_back(Tmp); 2141 } 2142 } 2143 } 2144 2145 void CodeGenFunction::EmitAsmStmt(const AsmStmt &S) { 2146 // Assemble the final asm string. 2147 std::string AsmString = S.generateAsmString(getContext()); 2148 2149 // Get all the output and input constraints together. 2150 SmallVector<TargetInfo::ConstraintInfo, 4> OutputConstraintInfos; 2151 SmallVector<TargetInfo::ConstraintInfo, 4> InputConstraintInfos; 2152 2153 for (unsigned i = 0, e = S.getNumOutputs(); i != e; i++) { 2154 StringRef Name; 2155 if (const GCCAsmStmt *GAS = dyn_cast<GCCAsmStmt>(&S)) 2156 Name = GAS->getOutputName(i); 2157 TargetInfo::ConstraintInfo Info(S.getOutputConstraint(i), Name); 2158 bool IsValid = getTarget().validateOutputConstraint(Info); (void)IsValid; 2159 assert(IsValid && "Failed to parse output constraint"); 2160 OutputConstraintInfos.push_back(Info); 2161 } 2162 2163 for (unsigned i = 0, e = S.getNumInputs(); i != e; i++) { 2164 StringRef Name; 2165 if (const GCCAsmStmt *GAS = dyn_cast<GCCAsmStmt>(&S)) 2166 Name = GAS->getInputName(i); 2167 TargetInfo::ConstraintInfo Info(S.getInputConstraint(i), Name); 2168 bool IsValid = 2169 getTarget().validateInputConstraint(OutputConstraintInfos, Info); 2170 assert(IsValid && "Failed to parse input constraint"); (void)IsValid; 2171 InputConstraintInfos.push_back(Info); 2172 } 2173 2174 std::string Constraints; 2175 2176 std::vector<LValue> ResultRegDests; 2177 std::vector<QualType> ResultRegQualTys; 2178 std::vector<llvm::Type *> ResultRegTypes; 2179 std::vector<llvm::Type *> ResultTruncRegTypes; 2180 std::vector<llvm::Type *> ArgTypes; 2181 std::vector<llvm::Value*> Args; 2182 llvm::BitVector ResultTypeRequiresCast; 2183 2184 // Keep track of inout constraints. 2185 std::string InOutConstraints; 2186 std::vector<llvm::Value*> InOutArgs; 2187 std::vector<llvm::Type*> InOutArgTypes; 2188 2189 // Keep track of out constraints for tied input operand. 2190 std::vector<std::string> OutputConstraints; 2191 2192 // Keep track of defined physregs. 2193 llvm::SmallSet<std::string, 8> PhysRegOutputs; 2194 2195 // An inline asm can be marked readonly if it meets the following conditions: 2196 // - it doesn't have any sideeffects 2197 // - it doesn't clobber memory 2198 // - it doesn't return a value by-reference 2199 // It can be marked readnone if it doesn't have any input memory constraints 2200 // in addition to meeting the conditions listed above. 2201 bool ReadOnly = true, ReadNone = true; 2202 2203 for (unsigned i = 0, e = S.getNumOutputs(); i != e; i++) { 2204 TargetInfo::ConstraintInfo &Info = OutputConstraintInfos[i]; 2205 2206 // Simplify the output constraint. 2207 std::string OutputConstraint(S.getOutputConstraint(i)); 2208 OutputConstraint = SimplifyConstraint(OutputConstraint.c_str() + 1, 2209 getTarget(), &OutputConstraintInfos); 2210 2211 const Expr *OutExpr = S.getOutputExpr(i); 2212 OutExpr = OutExpr->IgnoreParenNoopCasts(getContext()); 2213 2214 std::string GCCReg; 2215 OutputConstraint = AddVariableConstraints(OutputConstraint, *OutExpr, 2216 getTarget(), CGM, S, 2217 Info.earlyClobber(), 2218 &GCCReg); 2219 // Give an error on multiple outputs to same physreg. 2220 if (!GCCReg.empty() && !PhysRegOutputs.insert(GCCReg).second) 2221 CGM.Error(S.getAsmLoc(), "multiple outputs to hard register: " + GCCReg); 2222 2223 OutputConstraints.push_back(OutputConstraint); 2224 LValue Dest = EmitLValue(OutExpr); 2225 if (!Constraints.empty()) 2226 Constraints += ','; 2227 2228 // If this is a register output, then make the inline asm return it 2229 // by-value. If this is a memory result, return the value by-reference. 2230 bool isScalarizableAggregate = 2231 hasAggregateEvaluationKind(OutExpr->getType()); 2232 if (!Info.allowsMemory() && (hasScalarEvaluationKind(OutExpr->getType()) || 2233 isScalarizableAggregate)) { 2234 Constraints += "=" + OutputConstraint; 2235 ResultRegQualTys.push_back(OutExpr->getType()); 2236 ResultRegDests.push_back(Dest); 2237 ResultTruncRegTypes.push_back(ConvertTypeForMem(OutExpr->getType())); 2238 if (Info.allowsRegister() && isScalarizableAggregate) { 2239 ResultTypeRequiresCast.push_back(true); 2240 unsigned Size = getContext().getTypeSize(OutExpr->getType()); 2241 llvm::Type *ConvTy = llvm::IntegerType::get(getLLVMContext(), Size); 2242 ResultRegTypes.push_back(ConvTy); 2243 } else { 2244 ResultTypeRequiresCast.push_back(false); 2245 ResultRegTypes.push_back(ResultTruncRegTypes.back()); 2246 } 2247 // If this output is tied to an input, and if the input is larger, then 2248 // we need to set the actual result type of the inline asm node to be the 2249 // same as the input type. 2250 if (Info.hasMatchingInput()) { 2251 unsigned InputNo; 2252 for (InputNo = 0; InputNo != S.getNumInputs(); ++InputNo) { 2253 TargetInfo::ConstraintInfo &Input = InputConstraintInfos[InputNo]; 2254 if (Input.hasTiedOperand() && Input.getTiedOperand() == i) 2255 break; 2256 } 2257 assert(InputNo != S.getNumInputs() && "Didn't find matching input!"); 2258 2259 QualType InputTy = S.getInputExpr(InputNo)->getType(); 2260 QualType OutputType = OutExpr->getType(); 2261 2262 uint64_t InputSize = getContext().getTypeSize(InputTy); 2263 if (getContext().getTypeSize(OutputType) < InputSize) { 2264 // Form the asm to return the value as a larger integer or fp type. 2265 ResultRegTypes.back() = ConvertType(InputTy); 2266 } 2267 } 2268 if (llvm::Type* AdjTy = 2269 getTargetHooks().adjustInlineAsmType(*this, OutputConstraint, 2270 ResultRegTypes.back())) 2271 ResultRegTypes.back() = AdjTy; 2272 else { 2273 CGM.getDiags().Report(S.getAsmLoc(), 2274 diag::err_asm_invalid_type_in_input) 2275 << OutExpr->getType() << OutputConstraint; 2276 } 2277 2278 // Update largest vector width for any vector types. 2279 if (auto *VT = dyn_cast<llvm::VectorType>(ResultRegTypes.back())) 2280 LargestVectorWidth = 2281 std::max((uint64_t)LargestVectorWidth, 2282 VT->getPrimitiveSizeInBits().getKnownMinSize()); 2283 } else { 2284 ArgTypes.push_back(Dest.getAddress(*this).getType()); 2285 Args.push_back(Dest.getPointer(*this)); 2286 Constraints += "=*"; 2287 Constraints += OutputConstraint; 2288 ReadOnly = ReadNone = false; 2289 } 2290 2291 if (Info.isReadWrite()) { 2292 InOutConstraints += ','; 2293 2294 const Expr *InputExpr = S.getOutputExpr(i); 2295 llvm::Value *Arg = EmitAsmInputLValue(Info, Dest, InputExpr->getType(), 2296 InOutConstraints, 2297 InputExpr->getExprLoc()); 2298 2299 if (llvm::Type* AdjTy = 2300 getTargetHooks().adjustInlineAsmType(*this, OutputConstraint, 2301 Arg->getType())) 2302 Arg = Builder.CreateBitCast(Arg, AdjTy); 2303 2304 // Update largest vector width for any vector types. 2305 if (auto *VT = dyn_cast<llvm::VectorType>(Arg->getType())) 2306 LargestVectorWidth = 2307 std::max((uint64_t)LargestVectorWidth, 2308 VT->getPrimitiveSizeInBits().getKnownMinSize()); 2309 // Only tie earlyclobber physregs. 2310 if (Info.allowsRegister() && (GCCReg.empty() || Info.earlyClobber())) 2311 InOutConstraints += llvm::utostr(i); 2312 else 2313 InOutConstraints += OutputConstraint; 2314 2315 InOutArgTypes.push_back(Arg->getType()); 2316 InOutArgs.push_back(Arg); 2317 } 2318 } 2319 2320 // If this is a Microsoft-style asm blob, store the return registers (EAX:EDX) 2321 // to the return value slot. Only do this when returning in registers. 2322 if (isa<MSAsmStmt>(&S)) { 2323 const ABIArgInfo &RetAI = CurFnInfo->getReturnInfo(); 2324 if (RetAI.isDirect() || RetAI.isExtend()) { 2325 // Make a fake lvalue for the return value slot. 2326 LValue ReturnSlot = MakeAddrLValue(ReturnValue, FnRetTy); 2327 CGM.getTargetCodeGenInfo().addReturnRegisterOutputs( 2328 *this, ReturnSlot, Constraints, ResultRegTypes, ResultTruncRegTypes, 2329 ResultRegDests, AsmString, S.getNumOutputs()); 2330 SawAsmBlock = true; 2331 } 2332 } 2333 2334 for (unsigned i = 0, e = S.getNumInputs(); i != e; i++) { 2335 const Expr *InputExpr = S.getInputExpr(i); 2336 2337 TargetInfo::ConstraintInfo &Info = InputConstraintInfos[i]; 2338 2339 if (Info.allowsMemory()) 2340 ReadNone = false; 2341 2342 if (!Constraints.empty()) 2343 Constraints += ','; 2344 2345 // Simplify the input constraint. 2346 std::string InputConstraint(S.getInputConstraint(i)); 2347 InputConstraint = SimplifyConstraint(InputConstraint.c_str(), getTarget(), 2348 &OutputConstraintInfos); 2349 2350 InputConstraint = AddVariableConstraints( 2351 InputConstraint, *InputExpr->IgnoreParenNoopCasts(getContext()), 2352 getTarget(), CGM, S, false /* No EarlyClobber */); 2353 2354 std::string ReplaceConstraint (InputConstraint); 2355 llvm::Value *Arg = EmitAsmInput(Info, InputExpr, Constraints); 2356 2357 // If this input argument is tied to a larger output result, extend the 2358 // input to be the same size as the output. The LLVM backend wants to see 2359 // the input and output of a matching constraint be the same size. Note 2360 // that GCC does not define what the top bits are here. We use zext because 2361 // that is usually cheaper, but LLVM IR should really get an anyext someday. 2362 if (Info.hasTiedOperand()) { 2363 unsigned Output = Info.getTiedOperand(); 2364 QualType OutputType = S.getOutputExpr(Output)->getType(); 2365 QualType InputTy = InputExpr->getType(); 2366 2367 if (getContext().getTypeSize(OutputType) > 2368 getContext().getTypeSize(InputTy)) { 2369 // Use ptrtoint as appropriate so that we can do our extension. 2370 if (isa<llvm::PointerType>(Arg->getType())) 2371 Arg = Builder.CreatePtrToInt(Arg, IntPtrTy); 2372 llvm::Type *OutputTy = ConvertType(OutputType); 2373 if (isa<llvm::IntegerType>(OutputTy)) 2374 Arg = Builder.CreateZExt(Arg, OutputTy); 2375 else if (isa<llvm::PointerType>(OutputTy)) 2376 Arg = Builder.CreateZExt(Arg, IntPtrTy); 2377 else { 2378 assert(OutputTy->isFloatingPointTy() && "Unexpected output type"); 2379 Arg = Builder.CreateFPExt(Arg, OutputTy); 2380 } 2381 } 2382 // Deal with the tied operands' constraint code in adjustInlineAsmType. 2383 ReplaceConstraint = OutputConstraints[Output]; 2384 } 2385 if (llvm::Type* AdjTy = 2386 getTargetHooks().adjustInlineAsmType(*this, ReplaceConstraint, 2387 Arg->getType())) 2388 Arg = Builder.CreateBitCast(Arg, AdjTy); 2389 else 2390 CGM.getDiags().Report(S.getAsmLoc(), diag::err_asm_invalid_type_in_input) 2391 << InputExpr->getType() << InputConstraint; 2392 2393 // Update largest vector width for any vector types. 2394 if (auto *VT = dyn_cast<llvm::VectorType>(Arg->getType())) 2395 LargestVectorWidth = 2396 std::max((uint64_t)LargestVectorWidth, 2397 VT->getPrimitiveSizeInBits().getKnownMinSize()); 2398 2399 ArgTypes.push_back(Arg->getType()); 2400 Args.push_back(Arg); 2401 Constraints += InputConstraint; 2402 } 2403 2404 // Labels 2405 SmallVector<llvm::BasicBlock *, 16> Transfer; 2406 llvm::BasicBlock *Fallthrough = nullptr; 2407 bool IsGCCAsmGoto = false; 2408 if (const auto *GS = dyn_cast<GCCAsmStmt>(&S)) { 2409 IsGCCAsmGoto = GS->isAsmGoto(); 2410 if (IsGCCAsmGoto) { 2411 for (const auto *E : GS->labels()) { 2412 JumpDest Dest = getJumpDestForLabel(E->getLabel()); 2413 Transfer.push_back(Dest.getBlock()); 2414 llvm::BlockAddress *BA = 2415 llvm::BlockAddress::get(CurFn, Dest.getBlock()); 2416 Args.push_back(BA); 2417 ArgTypes.push_back(BA->getType()); 2418 if (!Constraints.empty()) 2419 Constraints += ','; 2420 Constraints += 'X'; 2421 } 2422 Fallthrough = createBasicBlock("asm.fallthrough"); 2423 } 2424 } 2425 2426 // Append the "input" part of inout constraints last. 2427 for (unsigned i = 0, e = InOutArgs.size(); i != e; i++) { 2428 ArgTypes.push_back(InOutArgTypes[i]); 2429 Args.push_back(InOutArgs[i]); 2430 } 2431 Constraints += InOutConstraints; 2432 2433 // Clobbers 2434 for (unsigned i = 0, e = S.getNumClobbers(); i != e; i++) { 2435 StringRef Clobber = S.getClobber(i); 2436 2437 if (Clobber == "memory") 2438 ReadOnly = ReadNone = false; 2439 else if (Clobber != "cc") { 2440 Clobber = getTarget().getNormalizedGCCRegisterName(Clobber); 2441 if (CGM.getCodeGenOpts().StackClashProtector && 2442 getTarget().isSPRegName(Clobber)) { 2443 CGM.getDiags().Report(S.getAsmLoc(), 2444 diag::warn_stack_clash_protection_inline_asm); 2445 } 2446 } 2447 2448 if (!Constraints.empty()) 2449 Constraints += ','; 2450 2451 Constraints += "~{"; 2452 Constraints += Clobber; 2453 Constraints += '}'; 2454 } 2455 2456 // Add machine specific clobbers 2457 std::string MachineClobbers = getTarget().getClobbers(); 2458 if (!MachineClobbers.empty()) { 2459 if (!Constraints.empty()) 2460 Constraints += ','; 2461 Constraints += MachineClobbers; 2462 } 2463 2464 llvm::Type *ResultType; 2465 if (ResultRegTypes.empty()) 2466 ResultType = VoidTy; 2467 else if (ResultRegTypes.size() == 1) 2468 ResultType = ResultRegTypes[0]; 2469 else 2470 ResultType = llvm::StructType::get(getLLVMContext(), ResultRegTypes); 2471 2472 llvm::FunctionType *FTy = 2473 llvm::FunctionType::get(ResultType, ArgTypes, false); 2474 2475 bool HasSideEffect = S.isVolatile() || S.getNumOutputs() == 0; 2476 llvm::InlineAsm::AsmDialect AsmDialect = isa<MSAsmStmt>(&S) ? 2477 llvm::InlineAsm::AD_Intel : llvm::InlineAsm::AD_ATT; 2478 llvm::InlineAsm *IA = 2479 llvm::InlineAsm::get(FTy, AsmString, Constraints, HasSideEffect, 2480 /* IsAlignStack */ false, AsmDialect); 2481 std::vector<llvm::Value*> RegResults; 2482 if (IsGCCAsmGoto) { 2483 llvm::CallBrInst *Result = 2484 Builder.CreateCallBr(IA, Fallthrough, Transfer, Args); 2485 EmitBlock(Fallthrough); 2486 UpdateAsmCallInst(cast<llvm::CallBase>(*Result), HasSideEffect, ReadOnly, 2487 ReadNone, InNoMergeAttributedStmt, S, ResultRegTypes, 2488 *this, RegResults); 2489 } else { 2490 llvm::CallInst *Result = 2491 Builder.CreateCall(IA, Args, getBundlesForFunclet(IA)); 2492 UpdateAsmCallInst(cast<llvm::CallBase>(*Result), HasSideEffect, ReadOnly, 2493 ReadNone, InNoMergeAttributedStmt, S, ResultRegTypes, 2494 *this, RegResults); 2495 } 2496 2497 assert(RegResults.size() == ResultRegTypes.size()); 2498 assert(RegResults.size() == ResultTruncRegTypes.size()); 2499 assert(RegResults.size() == ResultRegDests.size()); 2500 // ResultRegDests can be also populated by addReturnRegisterOutputs() above, 2501 // in which case its size may grow. 2502 assert(ResultTypeRequiresCast.size() <= ResultRegDests.size()); 2503 for (unsigned i = 0, e = RegResults.size(); i != e; ++i) { 2504 llvm::Value *Tmp = RegResults[i]; 2505 2506 // If the result type of the LLVM IR asm doesn't match the result type of 2507 // the expression, do the conversion. 2508 if (ResultRegTypes[i] != ResultTruncRegTypes[i]) { 2509 llvm::Type *TruncTy = ResultTruncRegTypes[i]; 2510 2511 // Truncate the integer result to the right size, note that TruncTy can be 2512 // a pointer. 2513 if (TruncTy->isFloatingPointTy()) 2514 Tmp = Builder.CreateFPTrunc(Tmp, TruncTy); 2515 else if (TruncTy->isPointerTy() && Tmp->getType()->isIntegerTy()) { 2516 uint64_t ResSize = CGM.getDataLayout().getTypeSizeInBits(TruncTy); 2517 Tmp = Builder.CreateTrunc(Tmp, 2518 llvm::IntegerType::get(getLLVMContext(), (unsigned)ResSize)); 2519 Tmp = Builder.CreateIntToPtr(Tmp, TruncTy); 2520 } else if (Tmp->getType()->isPointerTy() && TruncTy->isIntegerTy()) { 2521 uint64_t TmpSize =CGM.getDataLayout().getTypeSizeInBits(Tmp->getType()); 2522 Tmp = Builder.CreatePtrToInt(Tmp, 2523 llvm::IntegerType::get(getLLVMContext(), (unsigned)TmpSize)); 2524 Tmp = Builder.CreateTrunc(Tmp, TruncTy); 2525 } else if (TruncTy->isIntegerTy()) { 2526 Tmp = Builder.CreateZExtOrTrunc(Tmp, TruncTy); 2527 } else if (TruncTy->isVectorTy()) { 2528 Tmp = Builder.CreateBitCast(Tmp, TruncTy); 2529 } 2530 } 2531 2532 LValue Dest = ResultRegDests[i]; 2533 // ResultTypeRequiresCast elements correspond to the first 2534 // ResultTypeRequiresCast.size() elements of RegResults. 2535 if ((i < ResultTypeRequiresCast.size()) && ResultTypeRequiresCast[i]) { 2536 unsigned Size = getContext().getTypeSize(ResultRegQualTys[i]); 2537 Address A = Builder.CreateBitCast(Dest.getAddress(*this), 2538 ResultRegTypes[i]->getPointerTo()); 2539 QualType Ty = getContext().getIntTypeForBitwidth(Size, /*Signed*/ false); 2540 if (Ty.isNull()) { 2541 const Expr *OutExpr = S.getOutputExpr(i); 2542 CGM.Error( 2543 OutExpr->getExprLoc(), 2544 "impossible constraint in asm: can't store value into a register"); 2545 return; 2546 } 2547 Dest = MakeAddrLValue(A, Ty); 2548 } 2549 EmitStoreThroughLValue(RValue::get(Tmp), Dest); 2550 } 2551 } 2552 2553 LValue CodeGenFunction::InitCapturedStruct(const CapturedStmt &S) { 2554 const RecordDecl *RD = S.getCapturedRecordDecl(); 2555 QualType RecordTy = getContext().getRecordType(RD); 2556 2557 // Initialize the captured struct. 2558 LValue SlotLV = 2559 MakeAddrLValue(CreateMemTemp(RecordTy, "agg.captured"), RecordTy); 2560 2561 RecordDecl::field_iterator CurField = RD->field_begin(); 2562 for (CapturedStmt::const_capture_init_iterator I = S.capture_init_begin(), 2563 E = S.capture_init_end(); 2564 I != E; ++I, ++CurField) { 2565 LValue LV = EmitLValueForFieldInitialization(SlotLV, *CurField); 2566 if (CurField->hasCapturedVLAType()) { 2567 EmitLambdaVLACapture(CurField->getCapturedVLAType(), LV); 2568 } else { 2569 EmitInitializerForField(*CurField, LV, *I); 2570 } 2571 } 2572 2573 return SlotLV; 2574 } 2575 2576 /// Generate an outlined function for the body of a CapturedStmt, store any 2577 /// captured variables into the captured struct, and call the outlined function. 2578 llvm::Function * 2579 CodeGenFunction::EmitCapturedStmt(const CapturedStmt &S, CapturedRegionKind K) { 2580 LValue CapStruct = InitCapturedStruct(S); 2581 2582 // Emit the CapturedDecl 2583 CodeGenFunction CGF(CGM, true); 2584 CGCapturedStmtRAII CapInfoRAII(CGF, new CGCapturedStmtInfo(S, K)); 2585 llvm::Function *F = CGF.GenerateCapturedStmtFunction(S); 2586 delete CGF.CapturedStmtInfo; 2587 2588 // Emit call to the helper function. 2589 EmitCallOrInvoke(F, CapStruct.getPointer(*this)); 2590 2591 return F; 2592 } 2593 2594 Address CodeGenFunction::GenerateCapturedStmtArgument(const CapturedStmt &S) { 2595 LValue CapStruct = InitCapturedStruct(S); 2596 return CapStruct.getAddress(*this); 2597 } 2598 2599 /// Creates the outlined function for a CapturedStmt. 2600 llvm::Function * 2601 CodeGenFunction::GenerateCapturedStmtFunction(const CapturedStmt &S) { 2602 assert(CapturedStmtInfo && 2603 "CapturedStmtInfo should be set when generating the captured function"); 2604 const CapturedDecl *CD = S.getCapturedDecl(); 2605 const RecordDecl *RD = S.getCapturedRecordDecl(); 2606 SourceLocation Loc = S.getBeginLoc(); 2607 assert(CD->hasBody() && "missing CapturedDecl body"); 2608 2609 // Build the argument list. 2610 ASTContext &Ctx = CGM.getContext(); 2611 FunctionArgList Args; 2612 Args.append(CD->param_begin(), CD->param_end()); 2613 2614 // Create the function declaration. 2615 const CGFunctionInfo &FuncInfo = 2616 CGM.getTypes().arrangeBuiltinFunctionDeclaration(Ctx.VoidTy, Args); 2617 llvm::FunctionType *FuncLLVMTy = CGM.getTypes().GetFunctionType(FuncInfo); 2618 2619 llvm::Function *F = 2620 llvm::Function::Create(FuncLLVMTy, llvm::GlobalValue::InternalLinkage, 2621 CapturedStmtInfo->getHelperName(), &CGM.getModule()); 2622 CGM.SetInternalFunctionAttributes(CD, F, FuncInfo); 2623 if (CD->isNothrow()) 2624 F->addFnAttr(llvm::Attribute::NoUnwind); 2625 2626 // Generate the function. 2627 StartFunction(CD, Ctx.VoidTy, F, FuncInfo, Args, CD->getLocation(), 2628 CD->getBody()->getBeginLoc()); 2629 // Set the context parameter in CapturedStmtInfo. 2630 Address DeclPtr = GetAddrOfLocalVar(CD->getContextParam()); 2631 CapturedStmtInfo->setContextValue(Builder.CreateLoad(DeclPtr)); 2632 2633 // Initialize variable-length arrays. 2634 LValue Base = MakeNaturalAlignAddrLValue(CapturedStmtInfo->getContextValue(), 2635 Ctx.getTagDeclType(RD)); 2636 for (auto *FD : RD->fields()) { 2637 if (FD->hasCapturedVLAType()) { 2638 auto *ExprArg = 2639 EmitLoadOfLValue(EmitLValueForField(Base, FD), S.getBeginLoc()) 2640 .getScalarVal(); 2641 auto VAT = FD->getCapturedVLAType(); 2642 VLASizeMap[VAT->getSizeExpr()] = ExprArg; 2643 } 2644 } 2645 2646 // If 'this' is captured, load it into CXXThisValue. 2647 if (CapturedStmtInfo->isCXXThisExprCaptured()) { 2648 FieldDecl *FD = CapturedStmtInfo->getThisFieldDecl(); 2649 LValue ThisLValue = EmitLValueForField(Base, FD); 2650 CXXThisValue = EmitLoadOfLValue(ThisLValue, Loc).getScalarVal(); 2651 } 2652 2653 PGO.assignRegionCounters(GlobalDecl(CD), F); 2654 CapturedStmtInfo->EmitBody(*this, CD->getBody()); 2655 FinishFunction(CD->getBodyRBrace()); 2656 2657 return F; 2658 } 2659