1 //===--- CGStmt.cpp - Emit LLVM Code from Statements ----------------------===// 2 // 3 // The LLVM Compiler Infrastructure 4 // 5 // This file is distributed under the University of Illinois Open Source 6 // License. See LICENSE.TXT for details. 7 // 8 //===----------------------------------------------------------------------===// 9 // 10 // This contains code to emit Stmt nodes as LLVM code. 11 // 12 //===----------------------------------------------------------------------===// 13 14 #include "CodeGenFunction.h" 15 #include "CGDebugInfo.h" 16 #include "CodeGenModule.h" 17 #include "TargetInfo.h" 18 #include "clang/AST/StmtVisitor.h" 19 #include "clang/Basic/Builtins.h" 20 #include "clang/Basic/PrettyStackTrace.h" 21 #include "clang/Basic/TargetInfo.h" 22 #include "clang/Sema/LoopHint.h" 23 #include "clang/Sema/SemaDiagnostic.h" 24 #include "llvm/ADT/StringExtras.h" 25 #include "llvm/IR/CallSite.h" 26 #include "llvm/IR/DataLayout.h" 27 #include "llvm/IR/InlineAsm.h" 28 #include "llvm/IR/Intrinsics.h" 29 #include "llvm/IR/MDBuilder.h" 30 31 using namespace clang; 32 using namespace CodeGen; 33 34 //===----------------------------------------------------------------------===// 35 // Statement Emission 36 //===----------------------------------------------------------------------===// 37 38 void CodeGenFunction::EmitStopPoint(const Stmt *S) { 39 if (CGDebugInfo *DI = getDebugInfo()) { 40 SourceLocation Loc; 41 Loc = S->getLocStart(); 42 DI->EmitLocation(Builder, Loc); 43 44 LastStopPoint = Loc; 45 } 46 } 47 48 void CodeGenFunction::EmitStmt(const Stmt *S) { 49 assert(S && "Null statement?"); 50 PGO.setCurrentStmt(S); 51 52 // These statements have their own debug info handling. 53 if (EmitSimpleStmt(S)) 54 return; 55 56 // Check if we are generating unreachable code. 57 if (!HaveInsertPoint()) { 58 // If so, and the statement doesn't contain a label, then we do not need to 59 // generate actual code. This is safe because (1) the current point is 60 // unreachable, so we don't need to execute the code, and (2) we've already 61 // handled the statements which update internal data structures (like the 62 // local variable map) which could be used by subsequent statements. 63 if (!ContainsLabel(S)) { 64 // Verify that any decl statements were handled as simple, they may be in 65 // scope of subsequent reachable statements. 66 assert(!isa<DeclStmt>(*S) && "Unexpected DeclStmt!"); 67 return; 68 } 69 70 // Otherwise, make a new block to hold the code. 71 EnsureInsertPoint(); 72 } 73 74 // Generate a stoppoint if we are emitting debug info. 75 EmitStopPoint(S); 76 77 switch (S->getStmtClass()) { 78 case Stmt::NoStmtClass: 79 case Stmt::CXXCatchStmtClass: 80 case Stmt::SEHExceptStmtClass: 81 case Stmt::SEHFinallyStmtClass: 82 case Stmt::MSDependentExistsStmtClass: 83 llvm_unreachable("invalid statement class to emit generically"); 84 case Stmt::NullStmtClass: 85 case Stmt::CompoundStmtClass: 86 case Stmt::DeclStmtClass: 87 case Stmt::LabelStmtClass: 88 case Stmt::AttributedStmtClass: 89 case Stmt::GotoStmtClass: 90 case Stmt::BreakStmtClass: 91 case Stmt::ContinueStmtClass: 92 case Stmt::DefaultStmtClass: 93 case Stmt::CaseStmtClass: 94 case Stmt::SEHLeaveStmtClass: 95 llvm_unreachable("should have emitted these statements as simple"); 96 97 #define STMT(Type, Base) 98 #define ABSTRACT_STMT(Op) 99 #define EXPR(Type, Base) \ 100 case Stmt::Type##Class: 101 #include "clang/AST/StmtNodes.inc" 102 { 103 // Remember the block we came in on. 104 llvm::BasicBlock *incoming = Builder.GetInsertBlock(); 105 assert(incoming && "expression emission must have an insertion point"); 106 107 EmitIgnoredExpr(cast<Expr>(S)); 108 109 llvm::BasicBlock *outgoing = Builder.GetInsertBlock(); 110 assert(outgoing && "expression emission cleared block!"); 111 112 // The expression emitters assume (reasonably!) that the insertion 113 // point is always set. To maintain that, the call-emission code 114 // for noreturn functions has to enter a new block with no 115 // predecessors. We want to kill that block and mark the current 116 // insertion point unreachable in the common case of a call like 117 // "exit();". Since expression emission doesn't otherwise create 118 // blocks with no predecessors, we can just test for that. 119 // However, we must be careful not to do this to our incoming 120 // block, because *statement* emission does sometimes create 121 // reachable blocks which will have no predecessors until later in 122 // the function. This occurs with, e.g., labels that are not 123 // reachable by fallthrough. 124 if (incoming != outgoing && outgoing->use_empty()) { 125 outgoing->eraseFromParent(); 126 Builder.ClearInsertionPoint(); 127 } 128 break; 129 } 130 131 case Stmt::IndirectGotoStmtClass: 132 EmitIndirectGotoStmt(cast<IndirectGotoStmt>(*S)); break; 133 134 case Stmt::IfStmtClass: EmitIfStmt(cast<IfStmt>(*S)); break; 135 case Stmt::WhileStmtClass: EmitWhileStmt(cast<WhileStmt>(*S)); break; 136 case Stmt::DoStmtClass: EmitDoStmt(cast<DoStmt>(*S)); break; 137 case Stmt::ForStmtClass: EmitForStmt(cast<ForStmt>(*S)); break; 138 139 case Stmt::ReturnStmtClass: EmitReturnStmt(cast<ReturnStmt>(*S)); break; 140 141 case Stmt::SwitchStmtClass: EmitSwitchStmt(cast<SwitchStmt>(*S)); break; 142 case Stmt::GCCAsmStmtClass: // Intentional fall-through. 143 case Stmt::MSAsmStmtClass: EmitAsmStmt(cast<AsmStmt>(*S)); break; 144 case Stmt::CoroutineBodyStmtClass: 145 case Stmt::CoreturnStmtClass: 146 CGM.ErrorUnsupported(S, "coroutine"); 147 break; 148 case Stmt::CapturedStmtClass: { 149 const CapturedStmt *CS = cast<CapturedStmt>(S); 150 EmitCapturedStmt(*CS, CS->getCapturedRegionKind()); 151 } 152 break; 153 case Stmt::ObjCAtTryStmtClass: 154 EmitObjCAtTryStmt(cast<ObjCAtTryStmt>(*S)); 155 break; 156 case Stmt::ObjCAtCatchStmtClass: 157 llvm_unreachable( 158 "@catch statements should be handled by EmitObjCAtTryStmt"); 159 case Stmt::ObjCAtFinallyStmtClass: 160 llvm_unreachable( 161 "@finally statements should be handled by EmitObjCAtTryStmt"); 162 case Stmt::ObjCAtThrowStmtClass: 163 EmitObjCAtThrowStmt(cast<ObjCAtThrowStmt>(*S)); 164 break; 165 case Stmt::ObjCAtSynchronizedStmtClass: 166 EmitObjCAtSynchronizedStmt(cast<ObjCAtSynchronizedStmt>(*S)); 167 break; 168 case Stmt::ObjCForCollectionStmtClass: 169 EmitObjCForCollectionStmt(cast<ObjCForCollectionStmt>(*S)); 170 break; 171 case Stmt::ObjCAutoreleasePoolStmtClass: 172 EmitObjCAutoreleasePoolStmt(cast<ObjCAutoreleasePoolStmt>(*S)); 173 break; 174 175 case Stmt::CXXTryStmtClass: 176 EmitCXXTryStmt(cast<CXXTryStmt>(*S)); 177 break; 178 case Stmt::CXXForRangeStmtClass: 179 EmitCXXForRangeStmt(cast<CXXForRangeStmt>(*S)); 180 break; 181 case Stmt::SEHTryStmtClass: 182 EmitSEHTryStmt(cast<SEHTryStmt>(*S)); 183 break; 184 case Stmt::OMPParallelDirectiveClass: 185 EmitOMPParallelDirective(cast<OMPParallelDirective>(*S)); 186 break; 187 case Stmt::OMPSimdDirectiveClass: 188 EmitOMPSimdDirective(cast<OMPSimdDirective>(*S)); 189 break; 190 case Stmt::OMPForDirectiveClass: 191 EmitOMPForDirective(cast<OMPForDirective>(*S)); 192 break; 193 case Stmt::OMPForSimdDirectiveClass: 194 EmitOMPForSimdDirective(cast<OMPForSimdDirective>(*S)); 195 break; 196 case Stmt::OMPSectionsDirectiveClass: 197 EmitOMPSectionsDirective(cast<OMPSectionsDirective>(*S)); 198 break; 199 case Stmt::OMPSectionDirectiveClass: 200 EmitOMPSectionDirective(cast<OMPSectionDirective>(*S)); 201 break; 202 case Stmt::OMPSingleDirectiveClass: 203 EmitOMPSingleDirective(cast<OMPSingleDirective>(*S)); 204 break; 205 case Stmt::OMPMasterDirectiveClass: 206 EmitOMPMasterDirective(cast<OMPMasterDirective>(*S)); 207 break; 208 case Stmt::OMPCriticalDirectiveClass: 209 EmitOMPCriticalDirective(cast<OMPCriticalDirective>(*S)); 210 break; 211 case Stmt::OMPParallelForDirectiveClass: 212 EmitOMPParallelForDirective(cast<OMPParallelForDirective>(*S)); 213 break; 214 case Stmt::OMPParallelForSimdDirectiveClass: 215 EmitOMPParallelForSimdDirective(cast<OMPParallelForSimdDirective>(*S)); 216 break; 217 case Stmt::OMPParallelSectionsDirectiveClass: 218 EmitOMPParallelSectionsDirective(cast<OMPParallelSectionsDirective>(*S)); 219 break; 220 case Stmt::OMPTaskDirectiveClass: 221 EmitOMPTaskDirective(cast<OMPTaskDirective>(*S)); 222 break; 223 case Stmt::OMPTaskyieldDirectiveClass: 224 EmitOMPTaskyieldDirective(cast<OMPTaskyieldDirective>(*S)); 225 break; 226 case Stmt::OMPBarrierDirectiveClass: 227 EmitOMPBarrierDirective(cast<OMPBarrierDirective>(*S)); 228 break; 229 case Stmt::OMPTaskwaitDirectiveClass: 230 EmitOMPTaskwaitDirective(cast<OMPTaskwaitDirective>(*S)); 231 break; 232 case Stmt::OMPTaskgroupDirectiveClass: 233 EmitOMPTaskgroupDirective(cast<OMPTaskgroupDirective>(*S)); 234 break; 235 case Stmt::OMPFlushDirectiveClass: 236 EmitOMPFlushDirective(cast<OMPFlushDirective>(*S)); 237 break; 238 case Stmt::OMPOrderedDirectiveClass: 239 EmitOMPOrderedDirective(cast<OMPOrderedDirective>(*S)); 240 break; 241 case Stmt::OMPAtomicDirectiveClass: 242 EmitOMPAtomicDirective(cast<OMPAtomicDirective>(*S)); 243 break; 244 case Stmt::OMPTargetDirectiveClass: 245 EmitOMPTargetDirective(cast<OMPTargetDirective>(*S)); 246 break; 247 case Stmt::OMPTeamsDirectiveClass: 248 EmitOMPTeamsDirective(cast<OMPTeamsDirective>(*S)); 249 break; 250 case Stmt::OMPCancellationPointDirectiveClass: 251 EmitOMPCancellationPointDirective(cast<OMPCancellationPointDirective>(*S)); 252 break; 253 case Stmt::OMPCancelDirectiveClass: 254 EmitOMPCancelDirective(cast<OMPCancelDirective>(*S)); 255 break; 256 case Stmt::OMPTargetDataDirectiveClass: 257 EmitOMPTargetDataDirective(cast<OMPTargetDataDirective>(*S)); 258 break; 259 case Stmt::OMPTaskLoopDirectiveClass: 260 EmitOMPTaskLoopDirective(cast<OMPTaskLoopDirective>(*S)); 261 break; 262 case Stmt::OMPTaskLoopSimdDirectiveClass: 263 EmitOMPTaskLoopSimdDirective(cast<OMPTaskLoopSimdDirective>(*S)); 264 break; 265 case Stmt::OMPDistributeDirectiveClass: 266 EmitOMPDistributeDirective(cast<OMPDistributeDirective>(*S)); 267 break; 268 } 269 } 270 271 bool CodeGenFunction::EmitSimpleStmt(const Stmt *S) { 272 switch (S->getStmtClass()) { 273 default: return false; 274 case Stmt::NullStmtClass: break; 275 case Stmt::CompoundStmtClass: EmitCompoundStmt(cast<CompoundStmt>(*S)); break; 276 case Stmt::DeclStmtClass: EmitDeclStmt(cast<DeclStmt>(*S)); break; 277 case Stmt::LabelStmtClass: EmitLabelStmt(cast<LabelStmt>(*S)); break; 278 case Stmt::AttributedStmtClass: 279 EmitAttributedStmt(cast<AttributedStmt>(*S)); break; 280 case Stmt::GotoStmtClass: EmitGotoStmt(cast<GotoStmt>(*S)); break; 281 case Stmt::BreakStmtClass: EmitBreakStmt(cast<BreakStmt>(*S)); break; 282 case Stmt::ContinueStmtClass: EmitContinueStmt(cast<ContinueStmt>(*S)); break; 283 case Stmt::DefaultStmtClass: EmitDefaultStmt(cast<DefaultStmt>(*S)); break; 284 case Stmt::CaseStmtClass: EmitCaseStmt(cast<CaseStmt>(*S)); break; 285 case Stmt::SEHLeaveStmtClass: EmitSEHLeaveStmt(cast<SEHLeaveStmt>(*S)); break; 286 } 287 288 return true; 289 } 290 291 /// EmitCompoundStmt - Emit a compound statement {..} node. If GetLast is true, 292 /// this captures the expression result of the last sub-statement and returns it 293 /// (for use by the statement expression extension). 294 Address CodeGenFunction::EmitCompoundStmt(const CompoundStmt &S, bool GetLast, 295 AggValueSlot AggSlot) { 296 PrettyStackTraceLoc CrashInfo(getContext().getSourceManager(),S.getLBracLoc(), 297 "LLVM IR generation of compound statement ('{}')"); 298 299 // Keep track of the current cleanup stack depth, including debug scopes. 300 LexicalScope Scope(*this, S.getSourceRange()); 301 302 return EmitCompoundStmtWithoutScope(S, GetLast, AggSlot); 303 } 304 305 Address 306 CodeGenFunction::EmitCompoundStmtWithoutScope(const CompoundStmt &S, 307 bool GetLast, 308 AggValueSlot AggSlot) { 309 310 for (CompoundStmt::const_body_iterator I = S.body_begin(), 311 E = S.body_end()-GetLast; I != E; ++I) 312 EmitStmt(*I); 313 314 Address RetAlloca = Address::invalid(); 315 if (GetLast) { 316 // We have to special case labels here. They are statements, but when put 317 // at the end of a statement expression, they yield the value of their 318 // subexpression. Handle this by walking through all labels we encounter, 319 // emitting them before we evaluate the subexpr. 320 const Stmt *LastStmt = S.body_back(); 321 while (const LabelStmt *LS = dyn_cast<LabelStmt>(LastStmt)) { 322 EmitLabel(LS->getDecl()); 323 LastStmt = LS->getSubStmt(); 324 } 325 326 EnsureInsertPoint(); 327 328 QualType ExprTy = cast<Expr>(LastStmt)->getType(); 329 if (hasAggregateEvaluationKind(ExprTy)) { 330 EmitAggExpr(cast<Expr>(LastStmt), AggSlot); 331 } else { 332 // We can't return an RValue here because there might be cleanups at 333 // the end of the StmtExpr. Because of that, we have to emit the result 334 // here into a temporary alloca. 335 RetAlloca = CreateMemTemp(ExprTy); 336 EmitAnyExprToMem(cast<Expr>(LastStmt), RetAlloca, Qualifiers(), 337 /*IsInit*/false); 338 } 339 340 } 341 342 return RetAlloca; 343 } 344 345 void CodeGenFunction::SimplifyForwardingBlocks(llvm::BasicBlock *BB) { 346 llvm::BranchInst *BI = dyn_cast<llvm::BranchInst>(BB->getTerminator()); 347 348 // If there is a cleanup stack, then we it isn't worth trying to 349 // simplify this block (we would need to remove it from the scope map 350 // and cleanup entry). 351 if (!EHStack.empty()) 352 return; 353 354 // Can only simplify direct branches. 355 if (!BI || !BI->isUnconditional()) 356 return; 357 358 // Can only simplify empty blocks. 359 if (BI->getIterator() != BB->begin()) 360 return; 361 362 BB->replaceAllUsesWith(BI->getSuccessor(0)); 363 BI->eraseFromParent(); 364 BB->eraseFromParent(); 365 } 366 367 void CodeGenFunction::EmitBlock(llvm::BasicBlock *BB, bool IsFinished) { 368 llvm::BasicBlock *CurBB = Builder.GetInsertBlock(); 369 370 // Fall out of the current block (if necessary). 371 EmitBranch(BB); 372 373 if (IsFinished && BB->use_empty()) { 374 delete BB; 375 return; 376 } 377 378 // Place the block after the current block, if possible, or else at 379 // the end of the function. 380 if (CurBB && CurBB->getParent()) 381 CurFn->getBasicBlockList().insertAfter(CurBB->getIterator(), BB); 382 else 383 CurFn->getBasicBlockList().push_back(BB); 384 Builder.SetInsertPoint(BB); 385 } 386 387 void CodeGenFunction::EmitBranch(llvm::BasicBlock *Target) { 388 // Emit a branch from the current block to the target one if this 389 // was a real block. If this was just a fall-through block after a 390 // terminator, don't emit it. 391 llvm::BasicBlock *CurBB = Builder.GetInsertBlock(); 392 393 if (!CurBB || CurBB->getTerminator()) { 394 // If there is no insert point or the previous block is already 395 // terminated, don't touch it. 396 } else { 397 // Otherwise, create a fall-through branch. 398 Builder.CreateBr(Target); 399 } 400 401 Builder.ClearInsertionPoint(); 402 } 403 404 void CodeGenFunction::EmitBlockAfterUses(llvm::BasicBlock *block) { 405 bool inserted = false; 406 for (llvm::User *u : block->users()) { 407 if (llvm::Instruction *insn = dyn_cast<llvm::Instruction>(u)) { 408 CurFn->getBasicBlockList().insertAfter(insn->getParent()->getIterator(), 409 block); 410 inserted = true; 411 break; 412 } 413 } 414 415 if (!inserted) 416 CurFn->getBasicBlockList().push_back(block); 417 418 Builder.SetInsertPoint(block); 419 } 420 421 CodeGenFunction::JumpDest 422 CodeGenFunction::getJumpDestForLabel(const LabelDecl *D) { 423 JumpDest &Dest = LabelMap[D]; 424 if (Dest.isValid()) return Dest; 425 426 // Create, but don't insert, the new block. 427 Dest = JumpDest(createBasicBlock(D->getName()), 428 EHScopeStack::stable_iterator::invalid(), 429 NextCleanupDestIndex++); 430 return Dest; 431 } 432 433 void CodeGenFunction::EmitLabel(const LabelDecl *D) { 434 // Add this label to the current lexical scope if we're within any 435 // normal cleanups. Jumps "in" to this label --- when permitted by 436 // the language --- may need to be routed around such cleanups. 437 if (EHStack.hasNormalCleanups() && CurLexicalScope) 438 CurLexicalScope->addLabel(D); 439 440 JumpDest &Dest = LabelMap[D]; 441 442 // If we didn't need a forward reference to this label, just go 443 // ahead and create a destination at the current scope. 444 if (!Dest.isValid()) { 445 Dest = getJumpDestInCurrentScope(D->getName()); 446 447 // Otherwise, we need to give this label a target depth and remove 448 // it from the branch-fixups list. 449 } else { 450 assert(!Dest.getScopeDepth().isValid() && "already emitted label!"); 451 Dest.setScopeDepth(EHStack.stable_begin()); 452 ResolveBranchFixups(Dest.getBlock()); 453 } 454 455 EmitBlock(Dest.getBlock()); 456 incrementProfileCounter(D->getStmt()); 457 } 458 459 /// Change the cleanup scope of the labels in this lexical scope to 460 /// match the scope of the enclosing context. 461 void CodeGenFunction::LexicalScope::rescopeLabels() { 462 assert(!Labels.empty()); 463 EHScopeStack::stable_iterator innermostScope 464 = CGF.EHStack.getInnermostNormalCleanup(); 465 466 // Change the scope depth of all the labels. 467 for (SmallVectorImpl<const LabelDecl*>::const_iterator 468 i = Labels.begin(), e = Labels.end(); i != e; ++i) { 469 assert(CGF.LabelMap.count(*i)); 470 JumpDest &dest = CGF.LabelMap.find(*i)->second; 471 assert(dest.getScopeDepth().isValid()); 472 assert(innermostScope.encloses(dest.getScopeDepth())); 473 dest.setScopeDepth(innermostScope); 474 } 475 476 // Reparent the labels if the new scope also has cleanups. 477 if (innermostScope != EHScopeStack::stable_end() && ParentScope) { 478 ParentScope->Labels.append(Labels.begin(), Labels.end()); 479 } 480 } 481 482 483 void CodeGenFunction::EmitLabelStmt(const LabelStmt &S) { 484 EmitLabel(S.getDecl()); 485 EmitStmt(S.getSubStmt()); 486 } 487 488 void CodeGenFunction::EmitAttributedStmt(const AttributedStmt &S) { 489 const Stmt *SubStmt = S.getSubStmt(); 490 switch (SubStmt->getStmtClass()) { 491 case Stmt::DoStmtClass: 492 EmitDoStmt(cast<DoStmt>(*SubStmt), S.getAttrs()); 493 break; 494 case Stmt::ForStmtClass: 495 EmitForStmt(cast<ForStmt>(*SubStmt), S.getAttrs()); 496 break; 497 case Stmt::WhileStmtClass: 498 EmitWhileStmt(cast<WhileStmt>(*SubStmt), S.getAttrs()); 499 break; 500 case Stmt::CXXForRangeStmtClass: 501 EmitCXXForRangeStmt(cast<CXXForRangeStmt>(*SubStmt), S.getAttrs()); 502 break; 503 default: 504 EmitStmt(SubStmt); 505 } 506 } 507 508 void CodeGenFunction::EmitGotoStmt(const GotoStmt &S) { 509 // If this code is reachable then emit a stop point (if generating 510 // debug info). We have to do this ourselves because we are on the 511 // "simple" statement path. 512 if (HaveInsertPoint()) 513 EmitStopPoint(&S); 514 515 EmitBranchThroughCleanup(getJumpDestForLabel(S.getLabel())); 516 } 517 518 519 void CodeGenFunction::EmitIndirectGotoStmt(const IndirectGotoStmt &S) { 520 if (const LabelDecl *Target = S.getConstantTarget()) { 521 EmitBranchThroughCleanup(getJumpDestForLabel(Target)); 522 return; 523 } 524 525 // Ensure that we have an i8* for our PHI node. 526 llvm::Value *V = Builder.CreateBitCast(EmitScalarExpr(S.getTarget()), 527 Int8PtrTy, "addr"); 528 llvm::BasicBlock *CurBB = Builder.GetInsertBlock(); 529 530 // Get the basic block for the indirect goto. 531 llvm::BasicBlock *IndGotoBB = GetIndirectGotoBlock(); 532 533 // The first instruction in the block has to be the PHI for the switch dest, 534 // add an entry for this branch. 535 cast<llvm::PHINode>(IndGotoBB->begin())->addIncoming(V, CurBB); 536 537 EmitBranch(IndGotoBB); 538 } 539 540 void CodeGenFunction::EmitIfStmt(const IfStmt &S) { 541 // C99 6.8.4.1: The first substatement is executed if the expression compares 542 // unequal to 0. The condition must be a scalar type. 543 LexicalScope ConditionScope(*this, S.getCond()->getSourceRange()); 544 545 if (S.getConditionVariable()) 546 EmitAutoVarDecl(*S.getConditionVariable()); 547 548 // If the condition constant folds and can be elided, try to avoid emitting 549 // the condition and the dead arm of the if/else. 550 bool CondConstant; 551 if (ConstantFoldsToSimpleInteger(S.getCond(), CondConstant)) { 552 // Figure out which block (then or else) is executed. 553 const Stmt *Executed = S.getThen(); 554 const Stmt *Skipped = S.getElse(); 555 if (!CondConstant) // Condition false? 556 std::swap(Executed, Skipped); 557 558 // If the skipped block has no labels in it, just emit the executed block. 559 // This avoids emitting dead code and simplifies the CFG substantially. 560 if (!ContainsLabel(Skipped)) { 561 if (CondConstant) 562 incrementProfileCounter(&S); 563 if (Executed) { 564 RunCleanupsScope ExecutedScope(*this); 565 EmitStmt(Executed); 566 } 567 return; 568 } 569 } 570 571 // Otherwise, the condition did not fold, or we couldn't elide it. Just emit 572 // the conditional branch. 573 llvm::BasicBlock *ThenBlock = createBasicBlock("if.then"); 574 llvm::BasicBlock *ContBlock = createBasicBlock("if.end"); 575 llvm::BasicBlock *ElseBlock = ContBlock; 576 if (S.getElse()) 577 ElseBlock = createBasicBlock("if.else"); 578 579 EmitBranchOnBoolExpr(S.getCond(), ThenBlock, ElseBlock, 580 getProfileCount(S.getThen())); 581 582 // Emit the 'then' code. 583 EmitBlock(ThenBlock); 584 incrementProfileCounter(&S); 585 { 586 RunCleanupsScope ThenScope(*this); 587 EmitStmt(S.getThen()); 588 } 589 EmitBranch(ContBlock); 590 591 // Emit the 'else' code if present. 592 if (const Stmt *Else = S.getElse()) { 593 { 594 // There is no need to emit line number for an unconditional branch. 595 auto NL = ApplyDebugLocation::CreateEmpty(*this); 596 EmitBlock(ElseBlock); 597 } 598 { 599 RunCleanupsScope ElseScope(*this); 600 EmitStmt(Else); 601 } 602 { 603 // There is no need to emit line number for an unconditional branch. 604 auto NL = ApplyDebugLocation::CreateEmpty(*this); 605 EmitBranch(ContBlock); 606 } 607 } 608 609 // Emit the continuation block for code after the if. 610 EmitBlock(ContBlock, true); 611 } 612 613 void CodeGenFunction::EmitWhileStmt(const WhileStmt &S, 614 ArrayRef<const Attr *> WhileAttrs) { 615 // Emit the header for the loop, which will also become 616 // the continue target. 617 JumpDest LoopHeader = getJumpDestInCurrentScope("while.cond"); 618 EmitBlock(LoopHeader.getBlock()); 619 620 LoopStack.push(LoopHeader.getBlock(), CGM.getContext(), WhileAttrs); 621 622 // Create an exit block for when the condition fails, which will 623 // also become the break target. 624 JumpDest LoopExit = getJumpDestInCurrentScope("while.end"); 625 626 // Store the blocks to use for break and continue. 627 BreakContinueStack.push_back(BreakContinue(LoopExit, LoopHeader)); 628 629 // C++ [stmt.while]p2: 630 // When the condition of a while statement is a declaration, the 631 // scope of the variable that is declared extends from its point 632 // of declaration (3.3.2) to the end of the while statement. 633 // [...] 634 // The object created in a condition is destroyed and created 635 // with each iteration of the loop. 636 RunCleanupsScope ConditionScope(*this); 637 638 if (S.getConditionVariable()) 639 EmitAutoVarDecl(*S.getConditionVariable()); 640 641 // Evaluate the conditional in the while header. C99 6.8.5.1: The 642 // evaluation of the controlling expression takes place before each 643 // execution of the loop body. 644 llvm::Value *BoolCondVal = EvaluateExprAsBool(S.getCond()); 645 646 // while(1) is common, avoid extra exit blocks. Be sure 647 // to correctly handle break/continue though. 648 bool EmitBoolCondBranch = true; 649 if (llvm::ConstantInt *C = dyn_cast<llvm::ConstantInt>(BoolCondVal)) 650 if (C->isOne()) 651 EmitBoolCondBranch = false; 652 653 // As long as the condition is true, go to the loop body. 654 llvm::BasicBlock *LoopBody = createBasicBlock("while.body"); 655 if (EmitBoolCondBranch) { 656 llvm::BasicBlock *ExitBlock = LoopExit.getBlock(); 657 if (ConditionScope.requiresCleanups()) 658 ExitBlock = createBasicBlock("while.exit"); 659 Builder.CreateCondBr( 660 BoolCondVal, LoopBody, ExitBlock, 661 createProfileWeightsForLoop(S.getCond(), getProfileCount(S.getBody()))); 662 663 if (ExitBlock != LoopExit.getBlock()) { 664 EmitBlock(ExitBlock); 665 EmitBranchThroughCleanup(LoopExit); 666 } 667 } 668 669 // Emit the loop body. We have to emit this in a cleanup scope 670 // because it might be a singleton DeclStmt. 671 { 672 RunCleanupsScope BodyScope(*this); 673 EmitBlock(LoopBody); 674 incrementProfileCounter(&S); 675 EmitStmt(S.getBody()); 676 } 677 678 BreakContinueStack.pop_back(); 679 680 // Immediately force cleanup. 681 ConditionScope.ForceCleanup(); 682 683 EmitStopPoint(&S); 684 // Branch to the loop header again. 685 EmitBranch(LoopHeader.getBlock()); 686 687 LoopStack.pop(); 688 689 // Emit the exit block. 690 EmitBlock(LoopExit.getBlock(), true); 691 692 // The LoopHeader typically is just a branch if we skipped emitting 693 // a branch, try to erase it. 694 if (!EmitBoolCondBranch) 695 SimplifyForwardingBlocks(LoopHeader.getBlock()); 696 } 697 698 void CodeGenFunction::EmitDoStmt(const DoStmt &S, 699 ArrayRef<const Attr *> DoAttrs) { 700 JumpDest LoopExit = getJumpDestInCurrentScope("do.end"); 701 JumpDest LoopCond = getJumpDestInCurrentScope("do.cond"); 702 703 uint64_t ParentCount = getCurrentProfileCount(); 704 705 // Store the blocks to use for break and continue. 706 BreakContinueStack.push_back(BreakContinue(LoopExit, LoopCond)); 707 708 // Emit the body of the loop. 709 llvm::BasicBlock *LoopBody = createBasicBlock("do.body"); 710 711 LoopStack.push(LoopBody, CGM.getContext(), DoAttrs); 712 713 EmitBlockWithFallThrough(LoopBody, &S); 714 { 715 RunCleanupsScope BodyScope(*this); 716 EmitStmt(S.getBody()); 717 } 718 719 EmitBlock(LoopCond.getBlock()); 720 721 // C99 6.8.5.2: "The evaluation of the controlling expression takes place 722 // after each execution of the loop body." 723 724 // Evaluate the conditional in the while header. 725 // C99 6.8.5p2/p4: The first substatement is executed if the expression 726 // compares unequal to 0. The condition must be a scalar type. 727 llvm::Value *BoolCondVal = EvaluateExprAsBool(S.getCond()); 728 729 BreakContinueStack.pop_back(); 730 731 // "do {} while (0)" is common in macros, avoid extra blocks. Be sure 732 // to correctly handle break/continue though. 733 bool EmitBoolCondBranch = true; 734 if (llvm::ConstantInt *C = dyn_cast<llvm::ConstantInt>(BoolCondVal)) 735 if (C->isZero()) 736 EmitBoolCondBranch = false; 737 738 // As long as the condition is true, iterate the loop. 739 if (EmitBoolCondBranch) { 740 uint64_t BackedgeCount = getProfileCount(S.getBody()) - ParentCount; 741 Builder.CreateCondBr( 742 BoolCondVal, LoopBody, LoopExit.getBlock(), 743 createProfileWeightsForLoop(S.getCond(), BackedgeCount)); 744 } 745 746 LoopStack.pop(); 747 748 // Emit the exit block. 749 EmitBlock(LoopExit.getBlock()); 750 751 // The DoCond block typically is just a branch if we skipped 752 // emitting a branch, try to erase it. 753 if (!EmitBoolCondBranch) 754 SimplifyForwardingBlocks(LoopCond.getBlock()); 755 } 756 757 void CodeGenFunction::EmitForStmt(const ForStmt &S, 758 ArrayRef<const Attr *> ForAttrs) { 759 JumpDest LoopExit = getJumpDestInCurrentScope("for.end"); 760 761 LexicalScope ForScope(*this, S.getSourceRange()); 762 763 // Evaluate the first part before the loop. 764 if (S.getInit()) 765 EmitStmt(S.getInit()); 766 767 // Start the loop with a block that tests the condition. 768 // If there's an increment, the continue scope will be overwritten 769 // later. 770 JumpDest Continue = getJumpDestInCurrentScope("for.cond"); 771 llvm::BasicBlock *CondBlock = Continue.getBlock(); 772 EmitBlock(CondBlock); 773 774 LoopStack.push(CondBlock, CGM.getContext(), ForAttrs); 775 776 // If the for loop doesn't have an increment we can just use the 777 // condition as the continue block. Otherwise we'll need to create 778 // a block for it (in the current scope, i.e. in the scope of the 779 // condition), and that we will become our continue block. 780 if (S.getInc()) 781 Continue = getJumpDestInCurrentScope("for.inc"); 782 783 // Store the blocks to use for break and continue. 784 BreakContinueStack.push_back(BreakContinue(LoopExit, Continue)); 785 786 // Create a cleanup scope for the condition variable cleanups. 787 LexicalScope ConditionScope(*this, S.getSourceRange()); 788 789 if (S.getCond()) { 790 // If the for statement has a condition scope, emit the local variable 791 // declaration. 792 if (S.getConditionVariable()) { 793 EmitAutoVarDecl(*S.getConditionVariable()); 794 } 795 796 llvm::BasicBlock *ExitBlock = LoopExit.getBlock(); 797 // If there are any cleanups between here and the loop-exit scope, 798 // create a block to stage a loop exit along. 799 if (ForScope.requiresCleanups()) 800 ExitBlock = createBasicBlock("for.cond.cleanup"); 801 802 // As long as the condition is true, iterate the loop. 803 llvm::BasicBlock *ForBody = createBasicBlock("for.body"); 804 805 // C99 6.8.5p2/p4: The first substatement is executed if the expression 806 // compares unequal to 0. The condition must be a scalar type. 807 llvm::Value *BoolCondVal = EvaluateExprAsBool(S.getCond()); 808 Builder.CreateCondBr( 809 BoolCondVal, ForBody, ExitBlock, 810 createProfileWeightsForLoop(S.getCond(), getProfileCount(S.getBody()))); 811 812 if (ExitBlock != LoopExit.getBlock()) { 813 EmitBlock(ExitBlock); 814 EmitBranchThroughCleanup(LoopExit); 815 } 816 817 EmitBlock(ForBody); 818 } else { 819 // Treat it as a non-zero constant. Don't even create a new block for the 820 // body, just fall into it. 821 } 822 incrementProfileCounter(&S); 823 824 { 825 // Create a separate cleanup scope for the body, in case it is not 826 // a compound statement. 827 RunCleanupsScope BodyScope(*this); 828 EmitStmt(S.getBody()); 829 } 830 831 // If there is an increment, emit it next. 832 if (S.getInc()) { 833 EmitBlock(Continue.getBlock()); 834 EmitStmt(S.getInc()); 835 } 836 837 BreakContinueStack.pop_back(); 838 839 ConditionScope.ForceCleanup(); 840 841 EmitStopPoint(&S); 842 EmitBranch(CondBlock); 843 844 ForScope.ForceCleanup(); 845 846 LoopStack.pop(); 847 848 // Emit the fall-through block. 849 EmitBlock(LoopExit.getBlock(), true); 850 } 851 852 void 853 CodeGenFunction::EmitCXXForRangeStmt(const CXXForRangeStmt &S, 854 ArrayRef<const Attr *> ForAttrs) { 855 JumpDest LoopExit = getJumpDestInCurrentScope("for.end"); 856 857 LexicalScope ForScope(*this, S.getSourceRange()); 858 859 // Evaluate the first pieces before the loop. 860 EmitStmt(S.getRangeStmt()); 861 EmitStmt(S.getBeginEndStmt()); 862 863 // Start the loop with a block that tests the condition. 864 // If there's an increment, the continue scope will be overwritten 865 // later. 866 llvm::BasicBlock *CondBlock = createBasicBlock("for.cond"); 867 EmitBlock(CondBlock); 868 869 LoopStack.push(CondBlock, CGM.getContext(), ForAttrs); 870 871 // If there are any cleanups between here and the loop-exit scope, 872 // create a block to stage a loop exit along. 873 llvm::BasicBlock *ExitBlock = LoopExit.getBlock(); 874 if (ForScope.requiresCleanups()) 875 ExitBlock = createBasicBlock("for.cond.cleanup"); 876 877 // The loop body, consisting of the specified body and the loop variable. 878 llvm::BasicBlock *ForBody = createBasicBlock("for.body"); 879 880 // The body is executed if the expression, contextually converted 881 // to bool, is true. 882 llvm::Value *BoolCondVal = EvaluateExprAsBool(S.getCond()); 883 Builder.CreateCondBr( 884 BoolCondVal, ForBody, ExitBlock, 885 createProfileWeightsForLoop(S.getCond(), getProfileCount(S.getBody()))); 886 887 if (ExitBlock != LoopExit.getBlock()) { 888 EmitBlock(ExitBlock); 889 EmitBranchThroughCleanup(LoopExit); 890 } 891 892 EmitBlock(ForBody); 893 incrementProfileCounter(&S); 894 895 // Create a block for the increment. In case of a 'continue', we jump there. 896 JumpDest Continue = getJumpDestInCurrentScope("for.inc"); 897 898 // Store the blocks to use for break and continue. 899 BreakContinueStack.push_back(BreakContinue(LoopExit, Continue)); 900 901 { 902 // Create a separate cleanup scope for the loop variable and body. 903 LexicalScope BodyScope(*this, S.getSourceRange()); 904 EmitStmt(S.getLoopVarStmt()); 905 EmitStmt(S.getBody()); 906 } 907 908 EmitStopPoint(&S); 909 // If there is an increment, emit it next. 910 EmitBlock(Continue.getBlock()); 911 EmitStmt(S.getInc()); 912 913 BreakContinueStack.pop_back(); 914 915 EmitBranch(CondBlock); 916 917 ForScope.ForceCleanup(); 918 919 LoopStack.pop(); 920 921 // Emit the fall-through block. 922 EmitBlock(LoopExit.getBlock(), true); 923 } 924 925 void CodeGenFunction::EmitReturnOfRValue(RValue RV, QualType Ty) { 926 if (RV.isScalar()) { 927 Builder.CreateStore(RV.getScalarVal(), ReturnValue); 928 } else if (RV.isAggregate()) { 929 EmitAggregateCopy(ReturnValue, RV.getAggregateAddress(), Ty); 930 } else { 931 EmitStoreOfComplex(RV.getComplexVal(), MakeAddrLValue(ReturnValue, Ty), 932 /*init*/ true); 933 } 934 EmitBranchThroughCleanup(ReturnBlock); 935 } 936 937 /// EmitReturnStmt - Note that due to GCC extensions, this can have an operand 938 /// if the function returns void, or may be missing one if the function returns 939 /// non-void. Fun stuff :). 940 void CodeGenFunction::EmitReturnStmt(const ReturnStmt &S) { 941 // Returning from an outlined SEH helper is UB, and we already warn on it. 942 if (IsOutlinedSEHHelper) { 943 Builder.CreateUnreachable(); 944 Builder.ClearInsertionPoint(); 945 } 946 947 // Emit the result value, even if unused, to evalute the side effects. 948 const Expr *RV = S.getRetValue(); 949 950 // Treat block literals in a return expression as if they appeared 951 // in their own scope. This permits a small, easily-implemented 952 // exception to our over-conservative rules about not jumping to 953 // statements following block literals with non-trivial cleanups. 954 RunCleanupsScope cleanupScope(*this); 955 if (const ExprWithCleanups *cleanups = 956 dyn_cast_or_null<ExprWithCleanups>(RV)) { 957 enterFullExpression(cleanups); 958 RV = cleanups->getSubExpr(); 959 } 960 961 // FIXME: Clean this up by using an LValue for ReturnTemp, 962 // EmitStoreThroughLValue, and EmitAnyExpr. 963 if (getLangOpts().ElideConstructors && 964 S.getNRVOCandidate() && S.getNRVOCandidate()->isNRVOVariable()) { 965 // Apply the named return value optimization for this return statement, 966 // which means doing nothing: the appropriate result has already been 967 // constructed into the NRVO variable. 968 969 // If there is an NRVO flag for this variable, set it to 1 into indicate 970 // that the cleanup code should not destroy the variable. 971 if (llvm::Value *NRVOFlag = NRVOFlags[S.getNRVOCandidate()]) 972 Builder.CreateFlagStore(Builder.getTrue(), NRVOFlag); 973 } else if (!ReturnValue.isValid() || (RV && RV->getType()->isVoidType())) { 974 // Make sure not to return anything, but evaluate the expression 975 // for side effects. 976 if (RV) 977 EmitAnyExpr(RV); 978 } else if (!RV) { 979 // Do nothing (return value is left uninitialized) 980 } else if (FnRetTy->isReferenceType()) { 981 // If this function returns a reference, take the address of the expression 982 // rather than the value. 983 RValue Result = EmitReferenceBindingToExpr(RV); 984 Builder.CreateStore(Result.getScalarVal(), ReturnValue); 985 } else { 986 switch (getEvaluationKind(RV->getType())) { 987 case TEK_Scalar: 988 Builder.CreateStore(EmitScalarExpr(RV), ReturnValue); 989 break; 990 case TEK_Complex: 991 EmitComplexExprIntoLValue(RV, MakeAddrLValue(ReturnValue, RV->getType()), 992 /*isInit*/ true); 993 break; 994 case TEK_Aggregate: 995 EmitAggExpr(RV, AggValueSlot::forAddr(ReturnValue, 996 Qualifiers(), 997 AggValueSlot::IsDestructed, 998 AggValueSlot::DoesNotNeedGCBarriers, 999 AggValueSlot::IsNotAliased)); 1000 break; 1001 } 1002 } 1003 1004 ++NumReturnExprs; 1005 if (!RV || RV->isEvaluatable(getContext())) 1006 ++NumSimpleReturnExprs; 1007 1008 cleanupScope.ForceCleanup(); 1009 EmitBranchThroughCleanup(ReturnBlock); 1010 } 1011 1012 void CodeGenFunction::EmitDeclStmt(const DeclStmt &S) { 1013 // As long as debug info is modeled with instructions, we have to ensure we 1014 // have a place to insert here and write the stop point here. 1015 if (HaveInsertPoint()) 1016 EmitStopPoint(&S); 1017 1018 for (const auto *I : S.decls()) 1019 EmitDecl(*I); 1020 } 1021 1022 void CodeGenFunction::EmitBreakStmt(const BreakStmt &S) { 1023 assert(!BreakContinueStack.empty() && "break stmt not in a loop or switch!"); 1024 1025 // If this code is reachable then emit a stop point (if generating 1026 // debug info). We have to do this ourselves because we are on the 1027 // "simple" statement path. 1028 if (HaveInsertPoint()) 1029 EmitStopPoint(&S); 1030 1031 EmitBranchThroughCleanup(BreakContinueStack.back().BreakBlock); 1032 } 1033 1034 void CodeGenFunction::EmitContinueStmt(const ContinueStmt &S) { 1035 assert(!BreakContinueStack.empty() && "continue stmt not in a loop!"); 1036 1037 // If this code is reachable then emit a stop point (if generating 1038 // debug info). We have to do this ourselves because we are on the 1039 // "simple" statement path. 1040 if (HaveInsertPoint()) 1041 EmitStopPoint(&S); 1042 1043 EmitBranchThroughCleanup(BreakContinueStack.back().ContinueBlock); 1044 } 1045 1046 /// EmitCaseStmtRange - If case statement range is not too big then 1047 /// add multiple cases to switch instruction, one for each value within 1048 /// the range. If range is too big then emit "if" condition check. 1049 void CodeGenFunction::EmitCaseStmtRange(const CaseStmt &S) { 1050 assert(S.getRHS() && "Expected RHS value in CaseStmt"); 1051 1052 llvm::APSInt LHS = S.getLHS()->EvaluateKnownConstInt(getContext()); 1053 llvm::APSInt RHS = S.getRHS()->EvaluateKnownConstInt(getContext()); 1054 1055 // Emit the code for this case. We do this first to make sure it is 1056 // properly chained from our predecessor before generating the 1057 // switch machinery to enter this block. 1058 llvm::BasicBlock *CaseDest = createBasicBlock("sw.bb"); 1059 EmitBlockWithFallThrough(CaseDest, &S); 1060 EmitStmt(S.getSubStmt()); 1061 1062 // If range is empty, do nothing. 1063 if (LHS.isSigned() ? RHS.slt(LHS) : RHS.ult(LHS)) 1064 return; 1065 1066 llvm::APInt Range = RHS - LHS; 1067 // FIXME: parameters such as this should not be hardcoded. 1068 if (Range.ult(llvm::APInt(Range.getBitWidth(), 64))) { 1069 // Range is small enough to add multiple switch instruction cases. 1070 uint64_t Total = getProfileCount(&S); 1071 unsigned NCases = Range.getZExtValue() + 1; 1072 // We only have one region counter for the entire set of cases here, so we 1073 // need to divide the weights evenly between the generated cases, ensuring 1074 // that the total weight is preserved. E.g., a weight of 5 over three cases 1075 // will be distributed as weights of 2, 2, and 1. 1076 uint64_t Weight = Total / NCases, Rem = Total % NCases; 1077 for (unsigned I = 0; I != NCases; ++I) { 1078 if (SwitchWeights) 1079 SwitchWeights->push_back(Weight + (Rem ? 1 : 0)); 1080 if (Rem) 1081 Rem--; 1082 SwitchInsn->addCase(Builder.getInt(LHS), CaseDest); 1083 LHS++; 1084 } 1085 return; 1086 } 1087 1088 // The range is too big. Emit "if" condition into a new block, 1089 // making sure to save and restore the current insertion point. 1090 llvm::BasicBlock *RestoreBB = Builder.GetInsertBlock(); 1091 1092 // Push this test onto the chain of range checks (which terminates 1093 // in the default basic block). The switch's default will be changed 1094 // to the top of this chain after switch emission is complete. 1095 llvm::BasicBlock *FalseDest = CaseRangeBlock; 1096 CaseRangeBlock = createBasicBlock("sw.caserange"); 1097 1098 CurFn->getBasicBlockList().push_back(CaseRangeBlock); 1099 Builder.SetInsertPoint(CaseRangeBlock); 1100 1101 // Emit range check. 1102 llvm::Value *Diff = 1103 Builder.CreateSub(SwitchInsn->getCondition(), Builder.getInt(LHS)); 1104 llvm::Value *Cond = 1105 Builder.CreateICmpULE(Diff, Builder.getInt(Range), "inbounds"); 1106 1107 llvm::MDNode *Weights = nullptr; 1108 if (SwitchWeights) { 1109 uint64_t ThisCount = getProfileCount(&S); 1110 uint64_t DefaultCount = (*SwitchWeights)[0]; 1111 Weights = createProfileWeights(ThisCount, DefaultCount); 1112 1113 // Since we're chaining the switch default through each large case range, we 1114 // need to update the weight for the default, ie, the first case, to include 1115 // this case. 1116 (*SwitchWeights)[0] += ThisCount; 1117 } 1118 Builder.CreateCondBr(Cond, CaseDest, FalseDest, Weights); 1119 1120 // Restore the appropriate insertion point. 1121 if (RestoreBB) 1122 Builder.SetInsertPoint(RestoreBB); 1123 else 1124 Builder.ClearInsertionPoint(); 1125 } 1126 1127 void CodeGenFunction::EmitCaseStmt(const CaseStmt &S) { 1128 // If there is no enclosing switch instance that we're aware of, then this 1129 // case statement and its block can be elided. This situation only happens 1130 // when we've constant-folded the switch, are emitting the constant case, 1131 // and part of the constant case includes another case statement. For 1132 // instance: switch (4) { case 4: do { case 5: } while (1); } 1133 if (!SwitchInsn) { 1134 EmitStmt(S.getSubStmt()); 1135 return; 1136 } 1137 1138 // Handle case ranges. 1139 if (S.getRHS()) { 1140 EmitCaseStmtRange(S); 1141 return; 1142 } 1143 1144 llvm::ConstantInt *CaseVal = 1145 Builder.getInt(S.getLHS()->EvaluateKnownConstInt(getContext())); 1146 1147 // If the body of the case is just a 'break', try to not emit an empty block. 1148 // If we're profiling or we're not optimizing, leave the block in for better 1149 // debug and coverage analysis. 1150 if (!CGM.getCodeGenOpts().ProfileInstrGenerate && 1151 CGM.getCodeGenOpts().OptimizationLevel > 0 && 1152 isa<BreakStmt>(S.getSubStmt())) { 1153 JumpDest Block = BreakContinueStack.back().BreakBlock; 1154 1155 // Only do this optimization if there are no cleanups that need emitting. 1156 if (isObviouslyBranchWithoutCleanups(Block)) { 1157 if (SwitchWeights) 1158 SwitchWeights->push_back(getProfileCount(&S)); 1159 SwitchInsn->addCase(CaseVal, Block.getBlock()); 1160 1161 // If there was a fallthrough into this case, make sure to redirect it to 1162 // the end of the switch as well. 1163 if (Builder.GetInsertBlock()) { 1164 Builder.CreateBr(Block.getBlock()); 1165 Builder.ClearInsertionPoint(); 1166 } 1167 return; 1168 } 1169 } 1170 1171 llvm::BasicBlock *CaseDest = createBasicBlock("sw.bb"); 1172 EmitBlockWithFallThrough(CaseDest, &S); 1173 if (SwitchWeights) 1174 SwitchWeights->push_back(getProfileCount(&S)); 1175 SwitchInsn->addCase(CaseVal, CaseDest); 1176 1177 // Recursively emitting the statement is acceptable, but is not wonderful for 1178 // code where we have many case statements nested together, i.e.: 1179 // case 1: 1180 // case 2: 1181 // case 3: etc. 1182 // Handling this recursively will create a new block for each case statement 1183 // that falls through to the next case which is IR intensive. It also causes 1184 // deep recursion which can run into stack depth limitations. Handle 1185 // sequential non-range case statements specially. 1186 const CaseStmt *CurCase = &S; 1187 const CaseStmt *NextCase = dyn_cast<CaseStmt>(S.getSubStmt()); 1188 1189 // Otherwise, iteratively add consecutive cases to this switch stmt. 1190 while (NextCase && NextCase->getRHS() == nullptr) { 1191 CurCase = NextCase; 1192 llvm::ConstantInt *CaseVal = 1193 Builder.getInt(CurCase->getLHS()->EvaluateKnownConstInt(getContext())); 1194 1195 if (SwitchWeights) 1196 SwitchWeights->push_back(getProfileCount(NextCase)); 1197 if (CGM.getCodeGenOpts().ProfileInstrGenerate) { 1198 CaseDest = createBasicBlock("sw.bb"); 1199 EmitBlockWithFallThrough(CaseDest, &S); 1200 } 1201 1202 SwitchInsn->addCase(CaseVal, CaseDest); 1203 NextCase = dyn_cast<CaseStmt>(CurCase->getSubStmt()); 1204 } 1205 1206 // Normal default recursion for non-cases. 1207 EmitStmt(CurCase->getSubStmt()); 1208 } 1209 1210 void CodeGenFunction::EmitDefaultStmt(const DefaultStmt &S) { 1211 llvm::BasicBlock *DefaultBlock = SwitchInsn->getDefaultDest(); 1212 assert(DefaultBlock->empty() && 1213 "EmitDefaultStmt: Default block already defined?"); 1214 1215 EmitBlockWithFallThrough(DefaultBlock, &S); 1216 1217 EmitStmt(S.getSubStmt()); 1218 } 1219 1220 /// CollectStatementsForCase - Given the body of a 'switch' statement and a 1221 /// constant value that is being switched on, see if we can dead code eliminate 1222 /// the body of the switch to a simple series of statements to emit. Basically, 1223 /// on a switch (5) we want to find these statements: 1224 /// case 5: 1225 /// printf(...); <-- 1226 /// ++i; <-- 1227 /// break; 1228 /// 1229 /// and add them to the ResultStmts vector. If it is unsafe to do this 1230 /// transformation (for example, one of the elided statements contains a label 1231 /// that might be jumped to), return CSFC_Failure. If we handled it and 'S' 1232 /// should include statements after it (e.g. the printf() line is a substmt of 1233 /// the case) then return CSFC_FallThrough. If we handled it and found a break 1234 /// statement, then return CSFC_Success. 1235 /// 1236 /// If Case is non-null, then we are looking for the specified case, checking 1237 /// that nothing we jump over contains labels. If Case is null, then we found 1238 /// the case and are looking for the break. 1239 /// 1240 /// If the recursive walk actually finds our Case, then we set FoundCase to 1241 /// true. 1242 /// 1243 enum CSFC_Result { CSFC_Failure, CSFC_FallThrough, CSFC_Success }; 1244 static CSFC_Result CollectStatementsForCase(const Stmt *S, 1245 const SwitchCase *Case, 1246 bool &FoundCase, 1247 SmallVectorImpl<const Stmt*> &ResultStmts) { 1248 // If this is a null statement, just succeed. 1249 if (!S) 1250 return Case ? CSFC_Success : CSFC_FallThrough; 1251 1252 // If this is the switchcase (case 4: or default) that we're looking for, then 1253 // we're in business. Just add the substatement. 1254 if (const SwitchCase *SC = dyn_cast<SwitchCase>(S)) { 1255 if (S == Case) { 1256 FoundCase = true; 1257 return CollectStatementsForCase(SC->getSubStmt(), nullptr, FoundCase, 1258 ResultStmts); 1259 } 1260 1261 // Otherwise, this is some other case or default statement, just ignore it. 1262 return CollectStatementsForCase(SC->getSubStmt(), Case, FoundCase, 1263 ResultStmts); 1264 } 1265 1266 // If we are in the live part of the code and we found our break statement, 1267 // return a success! 1268 if (!Case && isa<BreakStmt>(S)) 1269 return CSFC_Success; 1270 1271 // If this is a switch statement, then it might contain the SwitchCase, the 1272 // break, or neither. 1273 if (const CompoundStmt *CS = dyn_cast<CompoundStmt>(S)) { 1274 // Handle this as two cases: we might be looking for the SwitchCase (if so 1275 // the skipped statements must be skippable) or we might already have it. 1276 CompoundStmt::const_body_iterator I = CS->body_begin(), E = CS->body_end(); 1277 if (Case) { 1278 // Keep track of whether we see a skipped declaration. The code could be 1279 // using the declaration even if it is skipped, so we can't optimize out 1280 // the decl if the kept statements might refer to it. 1281 bool HadSkippedDecl = false; 1282 1283 // If we're looking for the case, just see if we can skip each of the 1284 // substatements. 1285 for (; Case && I != E; ++I) { 1286 HadSkippedDecl |= isa<DeclStmt>(*I); 1287 1288 switch (CollectStatementsForCase(*I, Case, FoundCase, ResultStmts)) { 1289 case CSFC_Failure: return CSFC_Failure; 1290 case CSFC_Success: 1291 // A successful result means that either 1) that the statement doesn't 1292 // have the case and is skippable, or 2) does contain the case value 1293 // and also contains the break to exit the switch. In the later case, 1294 // we just verify the rest of the statements are elidable. 1295 if (FoundCase) { 1296 // If we found the case and skipped declarations, we can't do the 1297 // optimization. 1298 if (HadSkippedDecl) 1299 return CSFC_Failure; 1300 1301 for (++I; I != E; ++I) 1302 if (CodeGenFunction::ContainsLabel(*I, true)) 1303 return CSFC_Failure; 1304 return CSFC_Success; 1305 } 1306 break; 1307 case CSFC_FallThrough: 1308 // If we have a fallthrough condition, then we must have found the 1309 // case started to include statements. Consider the rest of the 1310 // statements in the compound statement as candidates for inclusion. 1311 assert(FoundCase && "Didn't find case but returned fallthrough?"); 1312 // We recursively found Case, so we're not looking for it anymore. 1313 Case = nullptr; 1314 1315 // If we found the case and skipped declarations, we can't do the 1316 // optimization. 1317 if (HadSkippedDecl) 1318 return CSFC_Failure; 1319 break; 1320 } 1321 } 1322 } 1323 1324 // If we have statements in our range, then we know that the statements are 1325 // live and need to be added to the set of statements we're tracking. 1326 for (; I != E; ++I) { 1327 switch (CollectStatementsForCase(*I, nullptr, FoundCase, ResultStmts)) { 1328 case CSFC_Failure: return CSFC_Failure; 1329 case CSFC_FallThrough: 1330 // A fallthrough result means that the statement was simple and just 1331 // included in ResultStmt, keep adding them afterwards. 1332 break; 1333 case CSFC_Success: 1334 // A successful result means that we found the break statement and 1335 // stopped statement inclusion. We just ensure that any leftover stmts 1336 // are skippable and return success ourselves. 1337 for (++I; I != E; ++I) 1338 if (CodeGenFunction::ContainsLabel(*I, true)) 1339 return CSFC_Failure; 1340 return CSFC_Success; 1341 } 1342 } 1343 1344 return Case ? CSFC_Success : CSFC_FallThrough; 1345 } 1346 1347 // Okay, this is some other statement that we don't handle explicitly, like a 1348 // for statement or increment etc. If we are skipping over this statement, 1349 // just verify it doesn't have labels, which would make it invalid to elide. 1350 if (Case) { 1351 if (CodeGenFunction::ContainsLabel(S, true)) 1352 return CSFC_Failure; 1353 return CSFC_Success; 1354 } 1355 1356 // Otherwise, we want to include this statement. Everything is cool with that 1357 // so long as it doesn't contain a break out of the switch we're in. 1358 if (CodeGenFunction::containsBreak(S)) return CSFC_Failure; 1359 1360 // Otherwise, everything is great. Include the statement and tell the caller 1361 // that we fall through and include the next statement as well. 1362 ResultStmts.push_back(S); 1363 return CSFC_FallThrough; 1364 } 1365 1366 /// FindCaseStatementsForValue - Find the case statement being jumped to and 1367 /// then invoke CollectStatementsForCase to find the list of statements to emit 1368 /// for a switch on constant. See the comment above CollectStatementsForCase 1369 /// for more details. 1370 static bool FindCaseStatementsForValue(const SwitchStmt &S, 1371 const llvm::APSInt &ConstantCondValue, 1372 SmallVectorImpl<const Stmt*> &ResultStmts, 1373 ASTContext &C, 1374 const SwitchCase *&ResultCase) { 1375 // First step, find the switch case that is being branched to. We can do this 1376 // efficiently by scanning the SwitchCase list. 1377 const SwitchCase *Case = S.getSwitchCaseList(); 1378 const DefaultStmt *DefaultCase = nullptr; 1379 1380 for (; Case; Case = Case->getNextSwitchCase()) { 1381 // It's either a default or case. Just remember the default statement in 1382 // case we're not jumping to any numbered cases. 1383 if (const DefaultStmt *DS = dyn_cast<DefaultStmt>(Case)) { 1384 DefaultCase = DS; 1385 continue; 1386 } 1387 1388 // Check to see if this case is the one we're looking for. 1389 const CaseStmt *CS = cast<CaseStmt>(Case); 1390 // Don't handle case ranges yet. 1391 if (CS->getRHS()) return false; 1392 1393 // If we found our case, remember it as 'case'. 1394 if (CS->getLHS()->EvaluateKnownConstInt(C) == ConstantCondValue) 1395 break; 1396 } 1397 1398 // If we didn't find a matching case, we use a default if it exists, or we 1399 // elide the whole switch body! 1400 if (!Case) { 1401 // It is safe to elide the body of the switch if it doesn't contain labels 1402 // etc. If it is safe, return successfully with an empty ResultStmts list. 1403 if (!DefaultCase) 1404 return !CodeGenFunction::ContainsLabel(&S); 1405 Case = DefaultCase; 1406 } 1407 1408 // Ok, we know which case is being jumped to, try to collect all the 1409 // statements that follow it. This can fail for a variety of reasons. Also, 1410 // check to see that the recursive walk actually found our case statement. 1411 // Insane cases like this can fail to find it in the recursive walk since we 1412 // don't handle every stmt kind: 1413 // switch (4) { 1414 // while (1) { 1415 // case 4: ... 1416 bool FoundCase = false; 1417 ResultCase = Case; 1418 return CollectStatementsForCase(S.getBody(), Case, FoundCase, 1419 ResultStmts) != CSFC_Failure && 1420 FoundCase; 1421 } 1422 1423 void CodeGenFunction::EmitSwitchStmt(const SwitchStmt &S) { 1424 // Handle nested switch statements. 1425 llvm::SwitchInst *SavedSwitchInsn = SwitchInsn; 1426 SmallVector<uint64_t, 16> *SavedSwitchWeights = SwitchWeights; 1427 llvm::BasicBlock *SavedCRBlock = CaseRangeBlock; 1428 1429 // See if we can constant fold the condition of the switch and therefore only 1430 // emit the live case statement (if any) of the switch. 1431 llvm::APSInt ConstantCondValue; 1432 if (ConstantFoldsToSimpleInteger(S.getCond(), ConstantCondValue)) { 1433 SmallVector<const Stmt*, 4> CaseStmts; 1434 const SwitchCase *Case = nullptr; 1435 if (FindCaseStatementsForValue(S, ConstantCondValue, CaseStmts, 1436 getContext(), Case)) { 1437 if (Case) 1438 incrementProfileCounter(Case); 1439 RunCleanupsScope ExecutedScope(*this); 1440 1441 // Emit the condition variable if needed inside the entire cleanup scope 1442 // used by this special case for constant folded switches. 1443 if (S.getConditionVariable()) 1444 EmitAutoVarDecl(*S.getConditionVariable()); 1445 1446 // At this point, we are no longer "within" a switch instance, so 1447 // we can temporarily enforce this to ensure that any embedded case 1448 // statements are not emitted. 1449 SwitchInsn = nullptr; 1450 1451 // Okay, we can dead code eliminate everything except this case. Emit the 1452 // specified series of statements and we're good. 1453 for (unsigned i = 0, e = CaseStmts.size(); i != e; ++i) 1454 EmitStmt(CaseStmts[i]); 1455 incrementProfileCounter(&S); 1456 1457 // Now we want to restore the saved switch instance so that nested 1458 // switches continue to function properly 1459 SwitchInsn = SavedSwitchInsn; 1460 1461 return; 1462 } 1463 } 1464 1465 JumpDest SwitchExit = getJumpDestInCurrentScope("sw.epilog"); 1466 1467 RunCleanupsScope ConditionScope(*this); 1468 if (S.getConditionVariable()) 1469 EmitAutoVarDecl(*S.getConditionVariable()); 1470 llvm::Value *CondV = EmitScalarExpr(S.getCond()); 1471 1472 // Create basic block to hold stuff that comes after switch 1473 // statement. We also need to create a default block now so that 1474 // explicit case ranges tests can have a place to jump to on 1475 // failure. 1476 llvm::BasicBlock *DefaultBlock = createBasicBlock("sw.default"); 1477 SwitchInsn = Builder.CreateSwitch(CondV, DefaultBlock); 1478 if (PGO.haveRegionCounts()) { 1479 // Walk the SwitchCase list to find how many there are. 1480 uint64_t DefaultCount = 0; 1481 unsigned NumCases = 0; 1482 for (const SwitchCase *Case = S.getSwitchCaseList(); 1483 Case; 1484 Case = Case->getNextSwitchCase()) { 1485 if (isa<DefaultStmt>(Case)) 1486 DefaultCount = getProfileCount(Case); 1487 NumCases += 1; 1488 } 1489 SwitchWeights = new SmallVector<uint64_t, 16>(); 1490 SwitchWeights->reserve(NumCases); 1491 // The default needs to be first. We store the edge count, so we already 1492 // know the right weight. 1493 SwitchWeights->push_back(DefaultCount); 1494 } 1495 CaseRangeBlock = DefaultBlock; 1496 1497 // Clear the insertion point to indicate we are in unreachable code. 1498 Builder.ClearInsertionPoint(); 1499 1500 // All break statements jump to NextBlock. If BreakContinueStack is non-empty 1501 // then reuse last ContinueBlock. 1502 JumpDest OuterContinue; 1503 if (!BreakContinueStack.empty()) 1504 OuterContinue = BreakContinueStack.back().ContinueBlock; 1505 1506 BreakContinueStack.push_back(BreakContinue(SwitchExit, OuterContinue)); 1507 1508 // Emit switch body. 1509 EmitStmt(S.getBody()); 1510 1511 BreakContinueStack.pop_back(); 1512 1513 // Update the default block in case explicit case range tests have 1514 // been chained on top. 1515 SwitchInsn->setDefaultDest(CaseRangeBlock); 1516 1517 // If a default was never emitted: 1518 if (!DefaultBlock->getParent()) { 1519 // If we have cleanups, emit the default block so that there's a 1520 // place to jump through the cleanups from. 1521 if (ConditionScope.requiresCleanups()) { 1522 EmitBlock(DefaultBlock); 1523 1524 // Otherwise, just forward the default block to the switch end. 1525 } else { 1526 DefaultBlock->replaceAllUsesWith(SwitchExit.getBlock()); 1527 delete DefaultBlock; 1528 } 1529 } 1530 1531 ConditionScope.ForceCleanup(); 1532 1533 // Emit continuation. 1534 EmitBlock(SwitchExit.getBlock(), true); 1535 incrementProfileCounter(&S); 1536 1537 // If the switch has a condition wrapped by __builtin_unpredictable, 1538 // create metadata that specifies that the switch is unpredictable. 1539 // Don't bother if not optimizing because that metadata would not be used. 1540 if (CGM.getCodeGenOpts().OptimizationLevel != 0) { 1541 if (const CallExpr *Call = dyn_cast<CallExpr>(S.getCond())) { 1542 const Decl *TargetDecl = Call->getCalleeDecl(); 1543 if (const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(TargetDecl)) { 1544 if (FD->getBuiltinID() == Builtin::BI__builtin_unpredictable) { 1545 llvm::MDBuilder MDHelper(getLLVMContext()); 1546 SwitchInsn->setMetadata(llvm::LLVMContext::MD_unpredictable, 1547 MDHelper.createUnpredictable()); 1548 } 1549 } 1550 } 1551 } 1552 1553 if (SwitchWeights) { 1554 assert(SwitchWeights->size() == 1 + SwitchInsn->getNumCases() && 1555 "switch weights do not match switch cases"); 1556 // If there's only one jump destination there's no sense weighting it. 1557 if (SwitchWeights->size() > 1) 1558 SwitchInsn->setMetadata(llvm::LLVMContext::MD_prof, 1559 createProfileWeights(*SwitchWeights)); 1560 delete SwitchWeights; 1561 } 1562 SwitchInsn = SavedSwitchInsn; 1563 SwitchWeights = SavedSwitchWeights; 1564 CaseRangeBlock = SavedCRBlock; 1565 } 1566 1567 static std::string 1568 SimplifyConstraint(const char *Constraint, const TargetInfo &Target, 1569 SmallVectorImpl<TargetInfo::ConstraintInfo> *OutCons=nullptr) { 1570 std::string Result; 1571 1572 while (*Constraint) { 1573 switch (*Constraint) { 1574 default: 1575 Result += Target.convertConstraint(Constraint); 1576 break; 1577 // Ignore these 1578 case '*': 1579 case '?': 1580 case '!': 1581 case '=': // Will see this and the following in mult-alt constraints. 1582 case '+': 1583 break; 1584 case '#': // Ignore the rest of the constraint alternative. 1585 while (Constraint[1] && Constraint[1] != ',') 1586 Constraint++; 1587 break; 1588 case '&': 1589 case '%': 1590 Result += *Constraint; 1591 while (Constraint[1] && Constraint[1] == *Constraint) 1592 Constraint++; 1593 break; 1594 case ',': 1595 Result += "|"; 1596 break; 1597 case 'g': 1598 Result += "imr"; 1599 break; 1600 case '[': { 1601 assert(OutCons && 1602 "Must pass output names to constraints with a symbolic name"); 1603 unsigned Index; 1604 bool result = Target.resolveSymbolicName(Constraint, *OutCons, Index); 1605 assert(result && "Could not resolve symbolic name"); (void)result; 1606 Result += llvm::utostr(Index); 1607 break; 1608 } 1609 } 1610 1611 Constraint++; 1612 } 1613 1614 return Result; 1615 } 1616 1617 /// AddVariableConstraints - Look at AsmExpr and if it is a variable declared 1618 /// as using a particular register add that as a constraint that will be used 1619 /// in this asm stmt. 1620 static std::string 1621 AddVariableConstraints(const std::string &Constraint, const Expr &AsmExpr, 1622 const TargetInfo &Target, CodeGenModule &CGM, 1623 const AsmStmt &Stmt, const bool EarlyClobber) { 1624 const DeclRefExpr *AsmDeclRef = dyn_cast<DeclRefExpr>(&AsmExpr); 1625 if (!AsmDeclRef) 1626 return Constraint; 1627 const ValueDecl &Value = *AsmDeclRef->getDecl(); 1628 const VarDecl *Variable = dyn_cast<VarDecl>(&Value); 1629 if (!Variable) 1630 return Constraint; 1631 if (Variable->getStorageClass() != SC_Register) 1632 return Constraint; 1633 AsmLabelAttr *Attr = Variable->getAttr<AsmLabelAttr>(); 1634 if (!Attr) 1635 return Constraint; 1636 StringRef Register = Attr->getLabel(); 1637 assert(Target.isValidGCCRegisterName(Register)); 1638 // We're using validateOutputConstraint here because we only care if 1639 // this is a register constraint. 1640 TargetInfo::ConstraintInfo Info(Constraint, ""); 1641 if (Target.validateOutputConstraint(Info) && 1642 !Info.allowsRegister()) { 1643 CGM.ErrorUnsupported(&Stmt, "__asm__"); 1644 return Constraint; 1645 } 1646 // Canonicalize the register here before returning it. 1647 Register = Target.getNormalizedGCCRegisterName(Register); 1648 return (EarlyClobber ? "&{" : "{") + Register.str() + "}"; 1649 } 1650 1651 llvm::Value* 1652 CodeGenFunction::EmitAsmInputLValue(const TargetInfo::ConstraintInfo &Info, 1653 LValue InputValue, QualType InputType, 1654 std::string &ConstraintStr, 1655 SourceLocation Loc) { 1656 llvm::Value *Arg; 1657 if (Info.allowsRegister() || !Info.allowsMemory()) { 1658 if (CodeGenFunction::hasScalarEvaluationKind(InputType)) { 1659 Arg = EmitLoadOfLValue(InputValue, Loc).getScalarVal(); 1660 } else { 1661 llvm::Type *Ty = ConvertType(InputType); 1662 uint64_t Size = CGM.getDataLayout().getTypeSizeInBits(Ty); 1663 if (Size <= 64 && llvm::isPowerOf2_64(Size)) { 1664 Ty = llvm::IntegerType::get(getLLVMContext(), Size); 1665 Ty = llvm::PointerType::getUnqual(Ty); 1666 1667 Arg = Builder.CreateLoad(Builder.CreateBitCast(InputValue.getAddress(), 1668 Ty)); 1669 } else { 1670 Arg = InputValue.getPointer(); 1671 ConstraintStr += '*'; 1672 } 1673 } 1674 } else { 1675 Arg = InputValue.getPointer(); 1676 ConstraintStr += '*'; 1677 } 1678 1679 return Arg; 1680 } 1681 1682 llvm::Value* CodeGenFunction::EmitAsmInput( 1683 const TargetInfo::ConstraintInfo &Info, 1684 const Expr *InputExpr, 1685 std::string &ConstraintStr) { 1686 // If this can't be a register or memory, i.e., has to be a constant 1687 // (immediate or symbolic), try to emit it as such. 1688 if (!Info.allowsRegister() && !Info.allowsMemory()) { 1689 llvm::APSInt Result; 1690 if (InputExpr->EvaluateAsInt(Result, getContext())) 1691 return llvm::ConstantInt::get(getLLVMContext(), Result); 1692 assert(!Info.requiresImmediateConstant() && 1693 "Required-immediate inlineasm arg isn't constant?"); 1694 } 1695 1696 if (Info.allowsRegister() || !Info.allowsMemory()) 1697 if (CodeGenFunction::hasScalarEvaluationKind(InputExpr->getType())) 1698 return EmitScalarExpr(InputExpr); 1699 if (InputExpr->getStmtClass() == Expr::CXXThisExprClass) 1700 return EmitScalarExpr(InputExpr); 1701 InputExpr = InputExpr->IgnoreParenNoopCasts(getContext()); 1702 LValue Dest = EmitLValue(InputExpr); 1703 return EmitAsmInputLValue(Info, Dest, InputExpr->getType(), ConstraintStr, 1704 InputExpr->getExprLoc()); 1705 } 1706 1707 /// getAsmSrcLocInfo - Return the !srcloc metadata node to attach to an inline 1708 /// asm call instruction. The !srcloc MDNode contains a list of constant 1709 /// integers which are the source locations of the start of each line in the 1710 /// asm. 1711 static llvm::MDNode *getAsmSrcLocInfo(const StringLiteral *Str, 1712 CodeGenFunction &CGF) { 1713 SmallVector<llvm::Metadata *, 8> Locs; 1714 // Add the location of the first line to the MDNode. 1715 Locs.push_back(llvm::ConstantAsMetadata::get(llvm::ConstantInt::get( 1716 CGF.Int32Ty, Str->getLocStart().getRawEncoding()))); 1717 StringRef StrVal = Str->getString(); 1718 if (!StrVal.empty()) { 1719 const SourceManager &SM = CGF.CGM.getContext().getSourceManager(); 1720 const LangOptions &LangOpts = CGF.CGM.getLangOpts(); 1721 unsigned StartToken = 0; 1722 unsigned ByteOffset = 0; 1723 1724 // Add the location of the start of each subsequent line of the asm to the 1725 // MDNode. 1726 for (unsigned i = 0, e = StrVal.size() - 1; i != e; ++i) { 1727 if (StrVal[i] != '\n') continue; 1728 SourceLocation LineLoc = Str->getLocationOfByte( 1729 i + 1, SM, LangOpts, CGF.getTarget(), &StartToken, &ByteOffset); 1730 Locs.push_back(llvm::ConstantAsMetadata::get( 1731 llvm::ConstantInt::get(CGF.Int32Ty, LineLoc.getRawEncoding()))); 1732 } 1733 } 1734 1735 return llvm::MDNode::get(CGF.getLLVMContext(), Locs); 1736 } 1737 1738 void CodeGenFunction::EmitAsmStmt(const AsmStmt &S) { 1739 // Assemble the final asm string. 1740 std::string AsmString = S.generateAsmString(getContext()); 1741 1742 // Get all the output and input constraints together. 1743 SmallVector<TargetInfo::ConstraintInfo, 4> OutputConstraintInfos; 1744 SmallVector<TargetInfo::ConstraintInfo, 4> InputConstraintInfos; 1745 1746 for (unsigned i = 0, e = S.getNumOutputs(); i != e; i++) { 1747 StringRef Name; 1748 if (const GCCAsmStmt *GAS = dyn_cast<GCCAsmStmt>(&S)) 1749 Name = GAS->getOutputName(i); 1750 TargetInfo::ConstraintInfo Info(S.getOutputConstraint(i), Name); 1751 bool IsValid = getTarget().validateOutputConstraint(Info); (void)IsValid; 1752 assert(IsValid && "Failed to parse output constraint"); 1753 OutputConstraintInfos.push_back(Info); 1754 } 1755 1756 for (unsigned i = 0, e = S.getNumInputs(); i != e; i++) { 1757 StringRef Name; 1758 if (const GCCAsmStmt *GAS = dyn_cast<GCCAsmStmt>(&S)) 1759 Name = GAS->getInputName(i); 1760 TargetInfo::ConstraintInfo Info(S.getInputConstraint(i), Name); 1761 bool IsValid = 1762 getTarget().validateInputConstraint(OutputConstraintInfos, Info); 1763 assert(IsValid && "Failed to parse input constraint"); (void)IsValid; 1764 InputConstraintInfos.push_back(Info); 1765 } 1766 1767 std::string Constraints; 1768 1769 std::vector<LValue> ResultRegDests; 1770 std::vector<QualType> ResultRegQualTys; 1771 std::vector<llvm::Type *> ResultRegTypes; 1772 std::vector<llvm::Type *> ResultTruncRegTypes; 1773 std::vector<llvm::Type *> ArgTypes; 1774 std::vector<llvm::Value*> Args; 1775 1776 // Keep track of inout constraints. 1777 std::string InOutConstraints; 1778 std::vector<llvm::Value*> InOutArgs; 1779 std::vector<llvm::Type*> InOutArgTypes; 1780 1781 // An inline asm can be marked readonly if it meets the following conditions: 1782 // - it doesn't have any sideeffects 1783 // - it doesn't clobber memory 1784 // - it doesn't return a value by-reference 1785 // It can be marked readnone if it doesn't have any input memory constraints 1786 // in addition to meeting the conditions listed above. 1787 bool ReadOnly = true, ReadNone = true; 1788 1789 for (unsigned i = 0, e = S.getNumOutputs(); i != e; i++) { 1790 TargetInfo::ConstraintInfo &Info = OutputConstraintInfos[i]; 1791 1792 // Simplify the output constraint. 1793 std::string OutputConstraint(S.getOutputConstraint(i)); 1794 OutputConstraint = SimplifyConstraint(OutputConstraint.c_str() + 1, 1795 getTarget()); 1796 1797 const Expr *OutExpr = S.getOutputExpr(i); 1798 OutExpr = OutExpr->IgnoreParenNoopCasts(getContext()); 1799 1800 OutputConstraint = AddVariableConstraints(OutputConstraint, *OutExpr, 1801 getTarget(), CGM, S, 1802 Info.earlyClobber()); 1803 1804 LValue Dest = EmitLValue(OutExpr); 1805 if (!Constraints.empty()) 1806 Constraints += ','; 1807 1808 // If this is a register output, then make the inline asm return it 1809 // by-value. If this is a memory result, return the value by-reference. 1810 if (!Info.allowsMemory() && hasScalarEvaluationKind(OutExpr->getType())) { 1811 Constraints += "=" + OutputConstraint; 1812 ResultRegQualTys.push_back(OutExpr->getType()); 1813 ResultRegDests.push_back(Dest); 1814 ResultRegTypes.push_back(ConvertTypeForMem(OutExpr->getType())); 1815 ResultTruncRegTypes.push_back(ResultRegTypes.back()); 1816 1817 // If this output is tied to an input, and if the input is larger, then 1818 // we need to set the actual result type of the inline asm node to be the 1819 // same as the input type. 1820 if (Info.hasMatchingInput()) { 1821 unsigned InputNo; 1822 for (InputNo = 0; InputNo != S.getNumInputs(); ++InputNo) { 1823 TargetInfo::ConstraintInfo &Input = InputConstraintInfos[InputNo]; 1824 if (Input.hasTiedOperand() && Input.getTiedOperand() == i) 1825 break; 1826 } 1827 assert(InputNo != S.getNumInputs() && "Didn't find matching input!"); 1828 1829 QualType InputTy = S.getInputExpr(InputNo)->getType(); 1830 QualType OutputType = OutExpr->getType(); 1831 1832 uint64_t InputSize = getContext().getTypeSize(InputTy); 1833 if (getContext().getTypeSize(OutputType) < InputSize) { 1834 // Form the asm to return the value as a larger integer or fp type. 1835 ResultRegTypes.back() = ConvertType(InputTy); 1836 } 1837 } 1838 if (llvm::Type* AdjTy = 1839 getTargetHooks().adjustInlineAsmType(*this, OutputConstraint, 1840 ResultRegTypes.back())) 1841 ResultRegTypes.back() = AdjTy; 1842 else { 1843 CGM.getDiags().Report(S.getAsmLoc(), 1844 diag::err_asm_invalid_type_in_input) 1845 << OutExpr->getType() << OutputConstraint; 1846 } 1847 } else { 1848 ArgTypes.push_back(Dest.getAddress().getType()); 1849 Args.push_back(Dest.getPointer()); 1850 Constraints += "=*"; 1851 Constraints += OutputConstraint; 1852 ReadOnly = ReadNone = false; 1853 } 1854 1855 if (Info.isReadWrite()) { 1856 InOutConstraints += ','; 1857 1858 const Expr *InputExpr = S.getOutputExpr(i); 1859 llvm::Value *Arg = EmitAsmInputLValue(Info, Dest, InputExpr->getType(), 1860 InOutConstraints, 1861 InputExpr->getExprLoc()); 1862 1863 if (llvm::Type* AdjTy = 1864 getTargetHooks().adjustInlineAsmType(*this, OutputConstraint, 1865 Arg->getType())) 1866 Arg = Builder.CreateBitCast(Arg, AdjTy); 1867 1868 if (Info.allowsRegister()) 1869 InOutConstraints += llvm::utostr(i); 1870 else 1871 InOutConstraints += OutputConstraint; 1872 1873 InOutArgTypes.push_back(Arg->getType()); 1874 InOutArgs.push_back(Arg); 1875 } 1876 } 1877 1878 // If this is a Microsoft-style asm blob, store the return registers (EAX:EDX) 1879 // to the return value slot. Only do this when returning in registers. 1880 if (isa<MSAsmStmt>(&S)) { 1881 const ABIArgInfo &RetAI = CurFnInfo->getReturnInfo(); 1882 if (RetAI.isDirect() || RetAI.isExtend()) { 1883 // Make a fake lvalue for the return value slot. 1884 LValue ReturnSlot = MakeAddrLValue(ReturnValue, FnRetTy); 1885 CGM.getTargetCodeGenInfo().addReturnRegisterOutputs( 1886 *this, ReturnSlot, Constraints, ResultRegTypes, ResultTruncRegTypes, 1887 ResultRegDests, AsmString, S.getNumOutputs()); 1888 SawAsmBlock = true; 1889 } 1890 } 1891 1892 for (unsigned i = 0, e = S.getNumInputs(); i != e; i++) { 1893 const Expr *InputExpr = S.getInputExpr(i); 1894 1895 TargetInfo::ConstraintInfo &Info = InputConstraintInfos[i]; 1896 1897 if (Info.allowsMemory()) 1898 ReadNone = false; 1899 1900 if (!Constraints.empty()) 1901 Constraints += ','; 1902 1903 // Simplify the input constraint. 1904 std::string InputConstraint(S.getInputConstraint(i)); 1905 InputConstraint = SimplifyConstraint(InputConstraint.c_str(), getTarget(), 1906 &OutputConstraintInfos); 1907 1908 InputConstraint = AddVariableConstraints( 1909 InputConstraint, *InputExpr->IgnoreParenNoopCasts(getContext()), 1910 getTarget(), CGM, S, false /* No EarlyClobber */); 1911 1912 llvm::Value *Arg = EmitAsmInput(Info, InputExpr, Constraints); 1913 1914 // If this input argument is tied to a larger output result, extend the 1915 // input to be the same size as the output. The LLVM backend wants to see 1916 // the input and output of a matching constraint be the same size. Note 1917 // that GCC does not define what the top bits are here. We use zext because 1918 // that is usually cheaper, but LLVM IR should really get an anyext someday. 1919 if (Info.hasTiedOperand()) { 1920 unsigned Output = Info.getTiedOperand(); 1921 QualType OutputType = S.getOutputExpr(Output)->getType(); 1922 QualType InputTy = InputExpr->getType(); 1923 1924 if (getContext().getTypeSize(OutputType) > 1925 getContext().getTypeSize(InputTy)) { 1926 // Use ptrtoint as appropriate so that we can do our extension. 1927 if (isa<llvm::PointerType>(Arg->getType())) 1928 Arg = Builder.CreatePtrToInt(Arg, IntPtrTy); 1929 llvm::Type *OutputTy = ConvertType(OutputType); 1930 if (isa<llvm::IntegerType>(OutputTy)) 1931 Arg = Builder.CreateZExt(Arg, OutputTy); 1932 else if (isa<llvm::PointerType>(OutputTy)) 1933 Arg = Builder.CreateZExt(Arg, IntPtrTy); 1934 else { 1935 assert(OutputTy->isFloatingPointTy() && "Unexpected output type"); 1936 Arg = Builder.CreateFPExt(Arg, OutputTy); 1937 } 1938 } 1939 } 1940 if (llvm::Type* AdjTy = 1941 getTargetHooks().adjustInlineAsmType(*this, InputConstraint, 1942 Arg->getType())) 1943 Arg = Builder.CreateBitCast(Arg, AdjTy); 1944 else 1945 CGM.getDiags().Report(S.getAsmLoc(), diag::err_asm_invalid_type_in_input) 1946 << InputExpr->getType() << InputConstraint; 1947 1948 ArgTypes.push_back(Arg->getType()); 1949 Args.push_back(Arg); 1950 Constraints += InputConstraint; 1951 } 1952 1953 // Append the "input" part of inout constraints last. 1954 for (unsigned i = 0, e = InOutArgs.size(); i != e; i++) { 1955 ArgTypes.push_back(InOutArgTypes[i]); 1956 Args.push_back(InOutArgs[i]); 1957 } 1958 Constraints += InOutConstraints; 1959 1960 // Clobbers 1961 for (unsigned i = 0, e = S.getNumClobbers(); i != e; i++) { 1962 StringRef Clobber = S.getClobber(i); 1963 1964 if (Clobber == "memory") 1965 ReadOnly = ReadNone = false; 1966 else if (Clobber != "cc") 1967 Clobber = getTarget().getNormalizedGCCRegisterName(Clobber); 1968 1969 if (!Constraints.empty()) 1970 Constraints += ','; 1971 1972 Constraints += "~{"; 1973 Constraints += Clobber; 1974 Constraints += '}'; 1975 } 1976 1977 // Add machine specific clobbers 1978 std::string MachineClobbers = getTarget().getClobbers(); 1979 if (!MachineClobbers.empty()) { 1980 if (!Constraints.empty()) 1981 Constraints += ','; 1982 Constraints += MachineClobbers; 1983 } 1984 1985 llvm::Type *ResultType; 1986 if (ResultRegTypes.empty()) 1987 ResultType = VoidTy; 1988 else if (ResultRegTypes.size() == 1) 1989 ResultType = ResultRegTypes[0]; 1990 else 1991 ResultType = llvm::StructType::get(getLLVMContext(), ResultRegTypes); 1992 1993 llvm::FunctionType *FTy = 1994 llvm::FunctionType::get(ResultType, ArgTypes, false); 1995 1996 bool HasSideEffect = S.isVolatile() || S.getNumOutputs() == 0; 1997 llvm::InlineAsm::AsmDialect AsmDialect = isa<MSAsmStmt>(&S) ? 1998 llvm::InlineAsm::AD_Intel : llvm::InlineAsm::AD_ATT; 1999 llvm::InlineAsm *IA = 2000 llvm::InlineAsm::get(FTy, AsmString, Constraints, HasSideEffect, 2001 /* IsAlignStack */ false, AsmDialect); 2002 llvm::CallInst *Result = Builder.CreateCall(IA, Args); 2003 Result->addAttribute(llvm::AttributeSet::FunctionIndex, 2004 llvm::Attribute::NoUnwind); 2005 2006 if (isa<MSAsmStmt>(&S)) { 2007 // If the assembly contains any labels, mark the call noduplicate to prevent 2008 // defining the same ASM label twice (PR23715). This is pretty hacky, but it 2009 // works. 2010 if (AsmString.find("__MSASMLABEL_") != std::string::npos) 2011 Result->addAttribute(llvm::AttributeSet::FunctionIndex, 2012 llvm::Attribute::NoDuplicate); 2013 } 2014 2015 // Attach readnone and readonly attributes. 2016 if (!HasSideEffect) { 2017 if (ReadNone) 2018 Result->addAttribute(llvm::AttributeSet::FunctionIndex, 2019 llvm::Attribute::ReadNone); 2020 else if (ReadOnly) 2021 Result->addAttribute(llvm::AttributeSet::FunctionIndex, 2022 llvm::Attribute::ReadOnly); 2023 } 2024 2025 // Slap the source location of the inline asm into a !srcloc metadata on the 2026 // call. 2027 if (const GCCAsmStmt *gccAsmStmt = dyn_cast<GCCAsmStmt>(&S)) { 2028 Result->setMetadata("srcloc", getAsmSrcLocInfo(gccAsmStmt->getAsmString(), 2029 *this)); 2030 } else { 2031 // At least put the line number on MS inline asm blobs. 2032 auto Loc = llvm::ConstantInt::get(Int32Ty, S.getAsmLoc().getRawEncoding()); 2033 Result->setMetadata("srcloc", 2034 llvm::MDNode::get(getLLVMContext(), 2035 llvm::ConstantAsMetadata::get(Loc))); 2036 } 2037 2038 // Extract all of the register value results from the asm. 2039 std::vector<llvm::Value*> RegResults; 2040 if (ResultRegTypes.size() == 1) { 2041 RegResults.push_back(Result); 2042 } else { 2043 for (unsigned i = 0, e = ResultRegTypes.size(); i != e; ++i) { 2044 llvm::Value *Tmp = Builder.CreateExtractValue(Result, i, "asmresult"); 2045 RegResults.push_back(Tmp); 2046 } 2047 } 2048 2049 assert(RegResults.size() == ResultRegTypes.size()); 2050 assert(RegResults.size() == ResultTruncRegTypes.size()); 2051 assert(RegResults.size() == ResultRegDests.size()); 2052 for (unsigned i = 0, e = RegResults.size(); i != e; ++i) { 2053 llvm::Value *Tmp = RegResults[i]; 2054 2055 // If the result type of the LLVM IR asm doesn't match the result type of 2056 // the expression, do the conversion. 2057 if (ResultRegTypes[i] != ResultTruncRegTypes[i]) { 2058 llvm::Type *TruncTy = ResultTruncRegTypes[i]; 2059 2060 // Truncate the integer result to the right size, note that TruncTy can be 2061 // a pointer. 2062 if (TruncTy->isFloatingPointTy()) 2063 Tmp = Builder.CreateFPTrunc(Tmp, TruncTy); 2064 else if (TruncTy->isPointerTy() && Tmp->getType()->isIntegerTy()) { 2065 uint64_t ResSize = CGM.getDataLayout().getTypeSizeInBits(TruncTy); 2066 Tmp = Builder.CreateTrunc(Tmp, 2067 llvm::IntegerType::get(getLLVMContext(), (unsigned)ResSize)); 2068 Tmp = Builder.CreateIntToPtr(Tmp, TruncTy); 2069 } else if (Tmp->getType()->isPointerTy() && TruncTy->isIntegerTy()) { 2070 uint64_t TmpSize =CGM.getDataLayout().getTypeSizeInBits(Tmp->getType()); 2071 Tmp = Builder.CreatePtrToInt(Tmp, 2072 llvm::IntegerType::get(getLLVMContext(), (unsigned)TmpSize)); 2073 Tmp = Builder.CreateTrunc(Tmp, TruncTy); 2074 } else if (TruncTy->isIntegerTy()) { 2075 Tmp = Builder.CreateTrunc(Tmp, TruncTy); 2076 } else if (TruncTy->isVectorTy()) { 2077 Tmp = Builder.CreateBitCast(Tmp, TruncTy); 2078 } 2079 } 2080 2081 EmitStoreThroughLValue(RValue::get(Tmp), ResultRegDests[i]); 2082 } 2083 } 2084 2085 LValue CodeGenFunction::InitCapturedStruct(const CapturedStmt &S) { 2086 const RecordDecl *RD = S.getCapturedRecordDecl(); 2087 QualType RecordTy = getContext().getRecordType(RD); 2088 2089 // Initialize the captured struct. 2090 LValue SlotLV = 2091 MakeAddrLValue(CreateMemTemp(RecordTy, "agg.captured"), RecordTy); 2092 2093 RecordDecl::field_iterator CurField = RD->field_begin(); 2094 for (CapturedStmt::const_capture_init_iterator I = S.capture_init_begin(), 2095 E = S.capture_init_end(); 2096 I != E; ++I, ++CurField) { 2097 LValue LV = EmitLValueForFieldInitialization(SlotLV, *CurField); 2098 if (CurField->hasCapturedVLAType()) { 2099 auto VAT = CurField->getCapturedVLAType(); 2100 EmitStoreThroughLValue(RValue::get(VLASizeMap[VAT->getSizeExpr()]), LV); 2101 } else { 2102 EmitInitializerForField(*CurField, LV, *I, None); 2103 } 2104 } 2105 2106 return SlotLV; 2107 } 2108 2109 /// Generate an outlined function for the body of a CapturedStmt, store any 2110 /// captured variables into the captured struct, and call the outlined function. 2111 llvm::Function * 2112 CodeGenFunction::EmitCapturedStmt(const CapturedStmt &S, CapturedRegionKind K) { 2113 LValue CapStruct = InitCapturedStruct(S); 2114 2115 // Emit the CapturedDecl 2116 CodeGenFunction CGF(CGM, true); 2117 CGCapturedStmtRAII CapInfoRAII(CGF, new CGCapturedStmtInfo(S, K)); 2118 llvm::Function *F = CGF.GenerateCapturedStmtFunction(S); 2119 delete CGF.CapturedStmtInfo; 2120 2121 // Emit call to the helper function. 2122 EmitCallOrInvoke(F, CapStruct.getPointer()); 2123 2124 return F; 2125 } 2126 2127 Address CodeGenFunction::GenerateCapturedStmtArgument(const CapturedStmt &S) { 2128 LValue CapStruct = InitCapturedStruct(S); 2129 return CapStruct.getAddress(); 2130 } 2131 2132 /// Creates the outlined function for a CapturedStmt. 2133 llvm::Function * 2134 CodeGenFunction::GenerateCapturedStmtFunction(const CapturedStmt &S) { 2135 assert(CapturedStmtInfo && 2136 "CapturedStmtInfo should be set when generating the captured function"); 2137 const CapturedDecl *CD = S.getCapturedDecl(); 2138 const RecordDecl *RD = S.getCapturedRecordDecl(); 2139 SourceLocation Loc = S.getLocStart(); 2140 assert(CD->hasBody() && "missing CapturedDecl body"); 2141 2142 // Build the argument list. 2143 ASTContext &Ctx = CGM.getContext(); 2144 FunctionArgList Args; 2145 Args.append(CD->param_begin(), CD->param_end()); 2146 2147 // Create the function declaration. 2148 FunctionType::ExtInfo ExtInfo; 2149 const CGFunctionInfo &FuncInfo = 2150 CGM.getTypes().arrangeFreeFunctionDeclaration(Ctx.VoidTy, Args, ExtInfo, 2151 /*IsVariadic=*/false); 2152 llvm::FunctionType *FuncLLVMTy = CGM.getTypes().GetFunctionType(FuncInfo); 2153 2154 llvm::Function *F = 2155 llvm::Function::Create(FuncLLVMTy, llvm::GlobalValue::InternalLinkage, 2156 CapturedStmtInfo->getHelperName(), &CGM.getModule()); 2157 CGM.SetInternalFunctionAttributes(CD, F, FuncInfo); 2158 if (CD->isNothrow()) 2159 F->addFnAttr(llvm::Attribute::NoUnwind); 2160 2161 // Generate the function. 2162 StartFunction(CD, Ctx.VoidTy, F, FuncInfo, Args, 2163 CD->getLocation(), 2164 CD->getBody()->getLocStart()); 2165 // Set the context parameter in CapturedStmtInfo. 2166 Address DeclPtr = GetAddrOfLocalVar(CD->getContextParam()); 2167 CapturedStmtInfo->setContextValue(Builder.CreateLoad(DeclPtr)); 2168 2169 // Initialize variable-length arrays. 2170 LValue Base = MakeNaturalAlignAddrLValue(CapturedStmtInfo->getContextValue(), 2171 Ctx.getTagDeclType(RD)); 2172 for (auto *FD : RD->fields()) { 2173 if (FD->hasCapturedVLAType()) { 2174 auto *ExprArg = EmitLoadOfLValue(EmitLValueForField(Base, FD), 2175 S.getLocStart()).getScalarVal(); 2176 auto VAT = FD->getCapturedVLAType(); 2177 VLASizeMap[VAT->getSizeExpr()] = ExprArg; 2178 } 2179 } 2180 2181 // If 'this' is captured, load it into CXXThisValue. 2182 if (CapturedStmtInfo->isCXXThisExprCaptured()) { 2183 FieldDecl *FD = CapturedStmtInfo->getThisFieldDecl(); 2184 LValue ThisLValue = EmitLValueForField(Base, FD); 2185 CXXThisValue = EmitLoadOfLValue(ThisLValue, Loc).getScalarVal(); 2186 } 2187 2188 PGO.assignRegionCounters(GlobalDecl(CD), F); 2189 CapturedStmtInfo->EmitBody(*this, CD->getBody()); 2190 FinishFunction(CD->getBodyRBrace()); 2191 2192 return F; 2193 } 2194