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