1 //===--- CGStmt.cpp - Emit LLVM Code from Statements ----------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This contains code to emit Stmt nodes as LLVM code.
12 //===----------------------------------------------------------------------===//
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"
31 using namespace clang;
32 using namespace CodeGen;
34 //===----------------------------------------------------------------------===//
36 //===----------------------------------------------------------------------===//
38 void CodeGenFunction::EmitStopPoint(const Stmt *S) {
39 if (CGDebugInfo *DI = getDebugInfo()) {
41 Loc = S->getLocStart();
42 DI->EmitLocation(Builder, Loc);
48 void CodeGenFunction::EmitStmt(const Stmt *S) {
49 assert(S && "Null statement?");
50 PGO.setCurrentStmt(S);
52 // These statements have their own debug info handling.
53 if (EmitSimpleStmt(S))
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!");
70 // Otherwise, make a new block to hold the code.
74 // Generate a stoppoint if we are emitting debug info.
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");
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"
103 // Remember the block we came in on.
104 llvm::BasicBlock *incoming = Builder.GetInsertBlock();
105 assert(incoming && "expression emission must have an insertion point");
107 EmitIgnoredExpr(cast<Expr>(S));
109 llvm::BasicBlock *outgoing = Builder.GetInsertBlock();
110 assert(outgoing && "expression emission cleared block!");
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();
131 case Stmt::IndirectGotoStmtClass:
132 EmitIndirectGotoStmt(cast<IndirectGotoStmt>(*S)); break;
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;
139 case Stmt::ReturnStmtClass: EmitReturnStmt(cast<ReturnStmt>(*S)); break;
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 EmitCoroutineBody(cast<CoroutineBodyStmt>(*S));
147 case Stmt::CoreturnStmtClass:
148 CGM.ErrorUnsupported(S, "coroutine");
150 case Stmt::CapturedStmtClass: {
151 const CapturedStmt *CS = cast<CapturedStmt>(S);
152 EmitCapturedStmt(*CS, CS->getCapturedRegionKind());
155 case Stmt::ObjCAtTryStmtClass:
156 EmitObjCAtTryStmt(cast<ObjCAtTryStmt>(*S));
158 case Stmt::ObjCAtCatchStmtClass:
160 "@catch statements should be handled by EmitObjCAtTryStmt");
161 case Stmt::ObjCAtFinallyStmtClass:
163 "@finally statements should be handled by EmitObjCAtTryStmt");
164 case Stmt::ObjCAtThrowStmtClass:
165 EmitObjCAtThrowStmt(cast<ObjCAtThrowStmt>(*S));
167 case Stmt::ObjCAtSynchronizedStmtClass:
168 EmitObjCAtSynchronizedStmt(cast<ObjCAtSynchronizedStmt>(*S));
170 case Stmt::ObjCForCollectionStmtClass:
171 EmitObjCForCollectionStmt(cast<ObjCForCollectionStmt>(*S));
173 case Stmt::ObjCAutoreleasePoolStmtClass:
174 EmitObjCAutoreleasePoolStmt(cast<ObjCAutoreleasePoolStmt>(*S));
177 case Stmt::CXXTryStmtClass:
178 EmitCXXTryStmt(cast<CXXTryStmt>(*S));
180 case Stmt::CXXForRangeStmtClass:
181 EmitCXXForRangeStmt(cast<CXXForRangeStmt>(*S));
183 case Stmt::SEHTryStmtClass:
184 EmitSEHTryStmt(cast<SEHTryStmt>(*S));
186 case Stmt::OMPParallelDirectiveClass:
187 EmitOMPParallelDirective(cast<OMPParallelDirective>(*S));
189 case Stmt::OMPSimdDirectiveClass:
190 EmitOMPSimdDirective(cast<OMPSimdDirective>(*S));
192 case Stmt::OMPForDirectiveClass:
193 EmitOMPForDirective(cast<OMPForDirective>(*S));
195 case Stmt::OMPForSimdDirectiveClass:
196 EmitOMPForSimdDirective(cast<OMPForSimdDirective>(*S));
198 case Stmt::OMPSectionsDirectiveClass:
199 EmitOMPSectionsDirective(cast<OMPSectionsDirective>(*S));
201 case Stmt::OMPSectionDirectiveClass:
202 EmitOMPSectionDirective(cast<OMPSectionDirective>(*S));
204 case Stmt::OMPSingleDirectiveClass:
205 EmitOMPSingleDirective(cast<OMPSingleDirective>(*S));
207 case Stmt::OMPMasterDirectiveClass:
208 EmitOMPMasterDirective(cast<OMPMasterDirective>(*S));
210 case Stmt::OMPCriticalDirectiveClass:
211 EmitOMPCriticalDirective(cast<OMPCriticalDirective>(*S));
213 case Stmt::OMPParallelForDirectiveClass:
214 EmitOMPParallelForDirective(cast<OMPParallelForDirective>(*S));
216 case Stmt::OMPParallelForSimdDirectiveClass:
217 EmitOMPParallelForSimdDirective(cast<OMPParallelForSimdDirective>(*S));
219 case Stmt::OMPParallelSectionsDirectiveClass:
220 EmitOMPParallelSectionsDirective(cast<OMPParallelSectionsDirective>(*S));
222 case Stmt::OMPTaskDirectiveClass:
223 EmitOMPTaskDirective(cast<OMPTaskDirective>(*S));
225 case Stmt::OMPTaskyieldDirectiveClass:
226 EmitOMPTaskyieldDirective(cast<OMPTaskyieldDirective>(*S));
228 case Stmt::OMPBarrierDirectiveClass:
229 EmitOMPBarrierDirective(cast<OMPBarrierDirective>(*S));
231 case Stmt::OMPTaskwaitDirectiveClass:
232 EmitOMPTaskwaitDirective(cast<OMPTaskwaitDirective>(*S));
234 case Stmt::OMPTaskgroupDirectiveClass:
235 EmitOMPTaskgroupDirective(cast<OMPTaskgroupDirective>(*S));
237 case Stmt::OMPFlushDirectiveClass:
238 EmitOMPFlushDirective(cast<OMPFlushDirective>(*S));
240 case Stmt::OMPOrderedDirectiveClass:
241 EmitOMPOrderedDirective(cast<OMPOrderedDirective>(*S));
243 case Stmt::OMPAtomicDirectiveClass:
244 EmitOMPAtomicDirective(cast<OMPAtomicDirective>(*S));
246 case Stmt::OMPTargetDirectiveClass:
247 EmitOMPTargetDirective(cast<OMPTargetDirective>(*S));
249 case Stmt::OMPTeamsDirectiveClass:
250 EmitOMPTeamsDirective(cast<OMPTeamsDirective>(*S));
252 case Stmt::OMPCancellationPointDirectiveClass:
253 EmitOMPCancellationPointDirective(cast<OMPCancellationPointDirective>(*S));
255 case Stmt::OMPCancelDirectiveClass:
256 EmitOMPCancelDirective(cast<OMPCancelDirective>(*S));
258 case Stmt::OMPTargetDataDirectiveClass:
259 EmitOMPTargetDataDirective(cast<OMPTargetDataDirective>(*S));
261 case Stmt::OMPTargetEnterDataDirectiveClass:
262 EmitOMPTargetEnterDataDirective(cast<OMPTargetEnterDataDirective>(*S));
264 case Stmt::OMPTargetExitDataDirectiveClass:
265 EmitOMPTargetExitDataDirective(cast<OMPTargetExitDataDirective>(*S));
267 case Stmt::OMPTargetParallelDirectiveClass:
268 EmitOMPTargetParallelDirective(cast<OMPTargetParallelDirective>(*S));
270 case Stmt::OMPTargetParallelForDirectiveClass:
271 EmitOMPTargetParallelForDirective(cast<OMPTargetParallelForDirective>(*S));
273 case Stmt::OMPTaskLoopDirectiveClass:
274 EmitOMPTaskLoopDirective(cast<OMPTaskLoopDirective>(*S));
276 case Stmt::OMPTaskLoopSimdDirectiveClass:
277 EmitOMPTaskLoopSimdDirective(cast<OMPTaskLoopSimdDirective>(*S));
279 case Stmt::OMPDistributeDirectiveClass:
280 EmitOMPDistributeDirective(cast<OMPDistributeDirective>(*S));
282 case Stmt::OMPTargetUpdateDirectiveClass:
283 EmitOMPTargetUpdateDirective(cast<OMPTargetUpdateDirective>(*S));
285 case Stmt::OMPDistributeParallelForDirectiveClass:
286 EmitOMPDistributeParallelForDirective(
287 cast<OMPDistributeParallelForDirective>(*S));
289 case Stmt::OMPDistributeParallelForSimdDirectiveClass:
290 EmitOMPDistributeParallelForSimdDirective(
291 cast<OMPDistributeParallelForSimdDirective>(*S));
293 case Stmt::OMPDistributeSimdDirectiveClass:
294 EmitOMPDistributeSimdDirective(cast<OMPDistributeSimdDirective>(*S));
296 case Stmt::OMPTargetParallelForSimdDirectiveClass:
297 EmitOMPTargetParallelForSimdDirective(
298 cast<OMPTargetParallelForSimdDirective>(*S));
300 case Stmt::OMPTargetSimdDirectiveClass:
301 EmitOMPTargetSimdDirective(cast<OMPTargetSimdDirective>(*S));
303 case Stmt::OMPTeamsDistributeDirectiveClass:
304 EmitOMPTeamsDistributeDirective(cast<OMPTeamsDistributeDirective>(*S));
306 case Stmt::OMPTeamsDistributeSimdDirectiveClass:
307 EmitOMPTeamsDistributeSimdDirective(
308 cast<OMPTeamsDistributeSimdDirective>(*S));
310 case Stmt::OMPTeamsDistributeParallelForSimdDirectiveClass:
311 EmitOMPTeamsDistributeParallelForSimdDirective(
312 cast<OMPTeamsDistributeParallelForSimdDirective>(*S));
314 case Stmt::OMPTeamsDistributeParallelForDirectiveClass:
315 EmitOMPTeamsDistributeParallelForDirective(
316 cast<OMPTeamsDistributeParallelForDirective>(*S));
318 case Stmt::OMPTargetTeamsDirectiveClass:
319 EmitOMPTargetTeamsDirective(cast<OMPTargetTeamsDirective>(*S));
321 case Stmt::OMPTargetTeamsDistributeDirectiveClass:
322 EmitOMPTargetTeamsDistributeDirective(
323 cast<OMPTargetTeamsDistributeDirective>(*S));
325 case Stmt::OMPTargetTeamsDistributeParallelForDirectiveClass:
326 EmitOMPTargetTeamsDistributeParallelForDirective(
327 cast<OMPTargetTeamsDistributeParallelForDirective>(*S));
329 case Stmt::OMPTargetTeamsDistributeParallelForSimdDirectiveClass:
330 EmitOMPTargetTeamsDistributeParallelForSimdDirective(
331 cast<OMPTargetTeamsDistributeParallelForSimdDirective>(*S));
336 bool CodeGenFunction::EmitSimpleStmt(const Stmt *S) {
337 switch (S->getStmtClass()) {
338 default: return false;
339 case Stmt::NullStmtClass: break;
340 case Stmt::CompoundStmtClass: EmitCompoundStmt(cast<CompoundStmt>(*S)); break;
341 case Stmt::DeclStmtClass: EmitDeclStmt(cast<DeclStmt>(*S)); break;
342 case Stmt::LabelStmtClass: EmitLabelStmt(cast<LabelStmt>(*S)); break;
343 case Stmt::AttributedStmtClass:
344 EmitAttributedStmt(cast<AttributedStmt>(*S)); break;
345 case Stmt::GotoStmtClass: EmitGotoStmt(cast<GotoStmt>(*S)); break;
346 case Stmt::BreakStmtClass: EmitBreakStmt(cast<BreakStmt>(*S)); break;
347 case Stmt::ContinueStmtClass: EmitContinueStmt(cast<ContinueStmt>(*S)); break;
348 case Stmt::DefaultStmtClass: EmitDefaultStmt(cast<DefaultStmt>(*S)); break;
349 case Stmt::CaseStmtClass: EmitCaseStmt(cast<CaseStmt>(*S)); break;
350 case Stmt::SEHLeaveStmtClass: EmitSEHLeaveStmt(cast<SEHLeaveStmt>(*S)); break;
356 /// EmitCompoundStmt - Emit a compound statement {..} node. If GetLast is true,
357 /// this captures the expression result of the last sub-statement and returns it
358 /// (for use by the statement expression extension).
359 Address CodeGenFunction::EmitCompoundStmt(const CompoundStmt &S, bool GetLast,
360 AggValueSlot AggSlot) {
361 PrettyStackTraceLoc CrashInfo(getContext().getSourceManager(),S.getLBracLoc(),
362 "LLVM IR generation of compound statement ('{}')");
364 // Keep track of the current cleanup stack depth, including debug scopes.
365 LexicalScope Scope(*this, S.getSourceRange());
367 return EmitCompoundStmtWithoutScope(S, GetLast, AggSlot);
371 CodeGenFunction::EmitCompoundStmtWithoutScope(const CompoundStmt &S,
373 AggValueSlot AggSlot) {
375 for (CompoundStmt::const_body_iterator I = S.body_begin(),
376 E = S.body_end()-GetLast; I != E; ++I)
379 Address RetAlloca = Address::invalid();
381 // We have to special case labels here. They are statements, but when put
382 // at the end of a statement expression, they yield the value of their
383 // subexpression. Handle this by walking through all labels we encounter,
384 // emitting them before we evaluate the subexpr.
385 const Stmt *LastStmt = S.body_back();
386 while (const LabelStmt *LS = dyn_cast<LabelStmt>(LastStmt)) {
387 EmitLabel(LS->getDecl());
388 LastStmt = LS->getSubStmt();
393 QualType ExprTy = cast<Expr>(LastStmt)->getType();
394 if (hasAggregateEvaluationKind(ExprTy)) {
395 EmitAggExpr(cast<Expr>(LastStmt), AggSlot);
397 // We can't return an RValue here because there might be cleanups at
398 // the end of the StmtExpr. Because of that, we have to emit the result
399 // here into a temporary alloca.
400 RetAlloca = CreateMemTemp(ExprTy);
401 EmitAnyExprToMem(cast<Expr>(LastStmt), RetAlloca, Qualifiers(),
410 void CodeGenFunction::SimplifyForwardingBlocks(llvm::BasicBlock *BB) {
411 llvm::BranchInst *BI = dyn_cast<llvm::BranchInst>(BB->getTerminator());
413 // If there is a cleanup stack, then we it isn't worth trying to
414 // simplify this block (we would need to remove it from the scope map
415 // and cleanup entry).
416 if (!EHStack.empty())
419 // Can only simplify direct branches.
420 if (!BI || !BI->isUnconditional())
423 // Can only simplify empty blocks.
424 if (BI->getIterator() != BB->begin())
427 BB->replaceAllUsesWith(BI->getSuccessor(0));
428 BI->eraseFromParent();
429 BB->eraseFromParent();
432 void CodeGenFunction::EmitBlock(llvm::BasicBlock *BB, bool IsFinished) {
433 llvm::BasicBlock *CurBB = Builder.GetInsertBlock();
435 // Fall out of the current block (if necessary).
438 if (IsFinished && BB->use_empty()) {
443 // Place the block after the current block, if possible, or else at
444 // the end of the function.
445 if (CurBB && CurBB->getParent())
446 CurFn->getBasicBlockList().insertAfter(CurBB->getIterator(), BB);
448 CurFn->getBasicBlockList().push_back(BB);
449 Builder.SetInsertPoint(BB);
452 void CodeGenFunction::EmitBranch(llvm::BasicBlock *Target) {
453 // Emit a branch from the current block to the target one if this
454 // was a real block. If this was just a fall-through block after a
455 // terminator, don't emit it.
456 llvm::BasicBlock *CurBB = Builder.GetInsertBlock();
458 if (!CurBB || CurBB->getTerminator()) {
459 // If there is no insert point or the previous block is already
460 // terminated, don't touch it.
462 // Otherwise, create a fall-through branch.
463 Builder.CreateBr(Target);
466 Builder.ClearInsertionPoint();
469 void CodeGenFunction::EmitBlockAfterUses(llvm::BasicBlock *block) {
470 bool inserted = false;
471 for (llvm::User *u : block->users()) {
472 if (llvm::Instruction *insn = dyn_cast<llvm::Instruction>(u)) {
473 CurFn->getBasicBlockList().insertAfter(insn->getParent()->getIterator(),
481 CurFn->getBasicBlockList().push_back(block);
483 Builder.SetInsertPoint(block);
486 CodeGenFunction::JumpDest
487 CodeGenFunction::getJumpDestForLabel(const LabelDecl *D) {
488 JumpDest &Dest = LabelMap[D];
489 if (Dest.isValid()) return Dest;
491 // Create, but don't insert, the new block.
492 Dest = JumpDest(createBasicBlock(D->getName()),
493 EHScopeStack::stable_iterator::invalid(),
494 NextCleanupDestIndex++);
498 void CodeGenFunction::EmitLabel(const LabelDecl *D) {
499 // Add this label to the current lexical scope if we're within any
500 // normal cleanups. Jumps "in" to this label --- when permitted by
501 // the language --- may need to be routed around such cleanups.
502 if (EHStack.hasNormalCleanups() && CurLexicalScope)
503 CurLexicalScope->addLabel(D);
505 JumpDest &Dest = LabelMap[D];
507 // If we didn't need a forward reference to this label, just go
508 // ahead and create a destination at the current scope.
509 if (!Dest.isValid()) {
510 Dest = getJumpDestInCurrentScope(D->getName());
512 // Otherwise, we need to give this label a target depth and remove
513 // it from the branch-fixups list.
515 assert(!Dest.getScopeDepth().isValid() && "already emitted label!");
516 Dest.setScopeDepth(EHStack.stable_begin());
517 ResolveBranchFixups(Dest.getBlock());
520 EmitBlock(Dest.getBlock());
521 incrementProfileCounter(D->getStmt());
524 /// Change the cleanup scope of the labels in this lexical scope to
525 /// match the scope of the enclosing context.
526 void CodeGenFunction::LexicalScope::rescopeLabels() {
527 assert(!Labels.empty());
528 EHScopeStack::stable_iterator innermostScope
529 = CGF.EHStack.getInnermostNormalCleanup();
531 // Change the scope depth of all the labels.
532 for (SmallVectorImpl<const LabelDecl*>::const_iterator
533 i = Labels.begin(), e = Labels.end(); i != e; ++i) {
534 assert(CGF.LabelMap.count(*i));
535 JumpDest &dest = CGF.LabelMap.find(*i)->second;
536 assert(dest.getScopeDepth().isValid());
537 assert(innermostScope.encloses(dest.getScopeDepth()));
538 dest.setScopeDepth(innermostScope);
541 // Reparent the labels if the new scope also has cleanups.
542 if (innermostScope != EHScopeStack::stable_end() && ParentScope) {
543 ParentScope->Labels.append(Labels.begin(), Labels.end());
548 void CodeGenFunction::EmitLabelStmt(const LabelStmt &S) {
549 EmitLabel(S.getDecl());
550 EmitStmt(S.getSubStmt());
553 void CodeGenFunction::EmitAttributedStmt(const AttributedStmt &S) {
554 const Stmt *SubStmt = S.getSubStmt();
555 switch (SubStmt->getStmtClass()) {
556 case Stmt::DoStmtClass:
557 EmitDoStmt(cast<DoStmt>(*SubStmt), S.getAttrs());
559 case Stmt::ForStmtClass:
560 EmitForStmt(cast<ForStmt>(*SubStmt), S.getAttrs());
562 case Stmt::WhileStmtClass:
563 EmitWhileStmt(cast<WhileStmt>(*SubStmt), S.getAttrs());
565 case Stmt::CXXForRangeStmtClass:
566 EmitCXXForRangeStmt(cast<CXXForRangeStmt>(*SubStmt), S.getAttrs());
573 void CodeGenFunction::EmitGotoStmt(const GotoStmt &S) {
574 // If this code is reachable then emit a stop point (if generating
575 // debug info). We have to do this ourselves because we are on the
576 // "simple" statement path.
577 if (HaveInsertPoint())
580 EmitBranchThroughCleanup(getJumpDestForLabel(S.getLabel()));
584 void CodeGenFunction::EmitIndirectGotoStmt(const IndirectGotoStmt &S) {
585 if (const LabelDecl *Target = S.getConstantTarget()) {
586 EmitBranchThroughCleanup(getJumpDestForLabel(Target));
590 // Ensure that we have an i8* for our PHI node.
591 llvm::Value *V = Builder.CreateBitCast(EmitScalarExpr(S.getTarget()),
593 llvm::BasicBlock *CurBB = Builder.GetInsertBlock();
595 // Get the basic block for the indirect goto.
596 llvm::BasicBlock *IndGotoBB = GetIndirectGotoBlock();
598 // The first instruction in the block has to be the PHI for the switch dest,
599 // add an entry for this branch.
600 cast<llvm::PHINode>(IndGotoBB->begin())->addIncoming(V, CurBB);
602 EmitBranch(IndGotoBB);
605 void CodeGenFunction::EmitIfStmt(const IfStmt &S) {
606 // C99 6.8.4.1: The first substatement is executed if the expression compares
607 // unequal to 0. The condition must be a scalar type.
608 LexicalScope ConditionScope(*this, S.getCond()->getSourceRange());
611 EmitStmt(S.getInit());
613 if (S.getConditionVariable())
614 EmitAutoVarDecl(*S.getConditionVariable());
616 // If the condition constant folds and can be elided, try to avoid emitting
617 // the condition and the dead arm of the if/else.
619 if (ConstantFoldsToSimpleInteger(S.getCond(), CondConstant,
621 // Figure out which block (then or else) is executed.
622 const Stmt *Executed = S.getThen();
623 const Stmt *Skipped = S.getElse();
624 if (!CondConstant) // Condition false?
625 std::swap(Executed, Skipped);
627 // If the skipped block has no labels in it, just emit the executed block.
628 // This avoids emitting dead code and simplifies the CFG substantially.
629 if (S.isConstexpr() || !ContainsLabel(Skipped)) {
631 incrementProfileCounter(&S);
633 RunCleanupsScope ExecutedScope(*this);
640 // Otherwise, the condition did not fold, or we couldn't elide it. Just emit
641 // the conditional branch.
642 llvm::BasicBlock *ThenBlock = createBasicBlock("if.then");
643 llvm::BasicBlock *ContBlock = createBasicBlock("if.end");
644 llvm::BasicBlock *ElseBlock = ContBlock;
646 ElseBlock = createBasicBlock("if.else");
648 EmitBranchOnBoolExpr(S.getCond(), ThenBlock, ElseBlock,
649 getProfileCount(S.getThen()));
651 // Emit the 'then' code.
652 EmitBlock(ThenBlock);
653 incrementProfileCounter(&S);
655 RunCleanupsScope ThenScope(*this);
656 EmitStmt(S.getThen());
658 EmitBranch(ContBlock);
660 // Emit the 'else' code if present.
661 if (const Stmt *Else = S.getElse()) {
663 // There is no need to emit line number for an unconditional branch.
664 auto NL = ApplyDebugLocation::CreateEmpty(*this);
665 EmitBlock(ElseBlock);
668 RunCleanupsScope ElseScope(*this);
672 // There is no need to emit line number for an unconditional branch.
673 auto NL = ApplyDebugLocation::CreateEmpty(*this);
674 EmitBranch(ContBlock);
678 // Emit the continuation block for code after the if.
679 EmitBlock(ContBlock, true);
682 void CodeGenFunction::EmitWhileStmt(const WhileStmt &S,
683 ArrayRef<const Attr *> WhileAttrs) {
684 // Emit the header for the loop, which will also become
685 // the continue target.
686 JumpDest LoopHeader = getJumpDestInCurrentScope("while.cond");
687 EmitBlock(LoopHeader.getBlock());
689 const SourceRange &R = S.getSourceRange();
690 LoopStack.push(LoopHeader.getBlock(), CGM.getContext(), WhileAttrs,
691 SourceLocToDebugLoc(R.getBegin()),
692 SourceLocToDebugLoc(R.getEnd()));
694 // Create an exit block for when the condition fails, which will
695 // also become the break target.
696 JumpDest LoopExit = getJumpDestInCurrentScope("while.end");
698 // Store the blocks to use for break and continue.
699 BreakContinueStack.push_back(BreakContinue(LoopExit, LoopHeader));
701 // C++ [stmt.while]p2:
702 // When the condition of a while statement is a declaration, the
703 // scope of the variable that is declared extends from its point
704 // of declaration (3.3.2) to the end of the while statement.
706 // The object created in a condition is destroyed and created
707 // with each iteration of the loop.
708 RunCleanupsScope ConditionScope(*this);
710 if (S.getConditionVariable())
711 EmitAutoVarDecl(*S.getConditionVariable());
713 // Evaluate the conditional in the while header. C99 6.8.5.1: The
714 // evaluation of the controlling expression takes place before each
715 // execution of the loop body.
716 llvm::Value *BoolCondVal = EvaluateExprAsBool(S.getCond());
718 // while(1) is common, avoid extra exit blocks. Be sure
719 // to correctly handle break/continue though.
720 bool EmitBoolCondBranch = true;
721 if (llvm::ConstantInt *C = dyn_cast<llvm::ConstantInt>(BoolCondVal))
723 EmitBoolCondBranch = false;
725 // As long as the condition is true, go to the loop body.
726 llvm::BasicBlock *LoopBody = createBasicBlock("while.body");
727 if (EmitBoolCondBranch) {
728 llvm::BasicBlock *ExitBlock = LoopExit.getBlock();
729 if (ConditionScope.requiresCleanups())
730 ExitBlock = createBasicBlock("while.exit");
731 Builder.CreateCondBr(
732 BoolCondVal, LoopBody, ExitBlock,
733 createProfileWeightsForLoop(S.getCond(), getProfileCount(S.getBody())));
735 if (ExitBlock != LoopExit.getBlock()) {
736 EmitBlock(ExitBlock);
737 EmitBranchThroughCleanup(LoopExit);
741 // Emit the loop body. We have to emit this in a cleanup scope
742 // because it might be a singleton DeclStmt.
744 RunCleanupsScope BodyScope(*this);
746 incrementProfileCounter(&S);
747 EmitStmt(S.getBody());
750 BreakContinueStack.pop_back();
752 // Immediately force cleanup.
753 ConditionScope.ForceCleanup();
756 // Branch to the loop header again.
757 EmitBranch(LoopHeader.getBlock());
761 // Emit the exit block.
762 EmitBlock(LoopExit.getBlock(), true);
764 // The LoopHeader typically is just a branch if we skipped emitting
765 // a branch, try to erase it.
766 if (!EmitBoolCondBranch)
767 SimplifyForwardingBlocks(LoopHeader.getBlock());
770 void CodeGenFunction::EmitDoStmt(const DoStmt &S,
771 ArrayRef<const Attr *> DoAttrs) {
772 JumpDest LoopExit = getJumpDestInCurrentScope("do.end");
773 JumpDest LoopCond = getJumpDestInCurrentScope("do.cond");
775 uint64_t ParentCount = getCurrentProfileCount();
777 // Store the blocks to use for break and continue.
778 BreakContinueStack.push_back(BreakContinue(LoopExit, LoopCond));
780 // Emit the body of the loop.
781 llvm::BasicBlock *LoopBody = createBasicBlock("do.body");
783 const SourceRange &R = S.getSourceRange();
784 LoopStack.push(LoopBody, CGM.getContext(), DoAttrs,
785 SourceLocToDebugLoc(R.getBegin()),
786 SourceLocToDebugLoc(R.getEnd()));
788 EmitBlockWithFallThrough(LoopBody, &S);
790 RunCleanupsScope BodyScope(*this);
791 EmitStmt(S.getBody());
794 EmitBlock(LoopCond.getBlock());
796 // C99 6.8.5.2: "The evaluation of the controlling expression takes place
797 // after each execution of the loop body."
799 // Evaluate the conditional in the while header.
800 // C99 6.8.5p2/p4: The first substatement is executed if the expression
801 // compares unequal to 0. The condition must be a scalar type.
802 llvm::Value *BoolCondVal = EvaluateExprAsBool(S.getCond());
804 BreakContinueStack.pop_back();
806 // "do {} while (0)" is common in macros, avoid extra blocks. Be sure
807 // to correctly handle break/continue though.
808 bool EmitBoolCondBranch = true;
809 if (llvm::ConstantInt *C = dyn_cast<llvm::ConstantInt>(BoolCondVal))
811 EmitBoolCondBranch = false;
813 // As long as the condition is true, iterate the loop.
814 if (EmitBoolCondBranch) {
815 uint64_t BackedgeCount = getProfileCount(S.getBody()) - ParentCount;
816 Builder.CreateCondBr(
817 BoolCondVal, LoopBody, LoopExit.getBlock(),
818 createProfileWeightsForLoop(S.getCond(), BackedgeCount));
823 // Emit the exit block.
824 EmitBlock(LoopExit.getBlock());
826 // The DoCond block typically is just a branch if we skipped
827 // emitting a branch, try to erase it.
828 if (!EmitBoolCondBranch)
829 SimplifyForwardingBlocks(LoopCond.getBlock());
832 void CodeGenFunction::EmitForStmt(const ForStmt &S,
833 ArrayRef<const Attr *> ForAttrs) {
834 JumpDest LoopExit = getJumpDestInCurrentScope("for.end");
836 LexicalScope ForScope(*this, S.getSourceRange());
838 // Evaluate the first part before the loop.
840 EmitStmt(S.getInit());
842 // Start the loop with a block that tests the condition.
843 // If there's an increment, the continue scope will be overwritten
845 JumpDest Continue = getJumpDestInCurrentScope("for.cond");
846 llvm::BasicBlock *CondBlock = Continue.getBlock();
847 EmitBlock(CondBlock);
849 const SourceRange &R = S.getSourceRange();
850 LoopStack.push(CondBlock, CGM.getContext(), ForAttrs,
851 SourceLocToDebugLoc(R.getBegin()),
852 SourceLocToDebugLoc(R.getEnd()));
854 // If the for loop doesn't have an increment we can just use the
855 // condition as the continue block. Otherwise we'll need to create
856 // a block for it (in the current scope, i.e. in the scope of the
857 // condition), and that we will become our continue block.
859 Continue = getJumpDestInCurrentScope("for.inc");
861 // Store the blocks to use for break and continue.
862 BreakContinueStack.push_back(BreakContinue(LoopExit, Continue));
864 // Create a cleanup scope for the condition variable cleanups.
865 LexicalScope ConditionScope(*this, S.getSourceRange());
868 // If the for statement has a condition scope, emit the local variable
870 if (S.getConditionVariable()) {
871 EmitAutoVarDecl(*S.getConditionVariable());
874 llvm::BasicBlock *ExitBlock = LoopExit.getBlock();
875 // If there are any cleanups between here and the loop-exit scope,
876 // create a block to stage a loop exit along.
877 if (ForScope.requiresCleanups())
878 ExitBlock = createBasicBlock("for.cond.cleanup");
880 // As long as the condition is true, iterate the loop.
881 llvm::BasicBlock *ForBody = createBasicBlock("for.body");
883 // C99 6.8.5p2/p4: The first substatement is executed if the expression
884 // compares unequal to 0. The condition must be a scalar type.
885 llvm::Value *BoolCondVal = EvaluateExprAsBool(S.getCond());
886 Builder.CreateCondBr(
887 BoolCondVal, ForBody, ExitBlock,
888 createProfileWeightsForLoop(S.getCond(), getProfileCount(S.getBody())));
890 if (ExitBlock != LoopExit.getBlock()) {
891 EmitBlock(ExitBlock);
892 EmitBranchThroughCleanup(LoopExit);
897 // Treat it as a non-zero constant. Don't even create a new block for the
898 // body, just fall into it.
900 incrementProfileCounter(&S);
903 // Create a separate cleanup scope for the body, in case it is not
904 // a compound statement.
905 RunCleanupsScope BodyScope(*this);
906 EmitStmt(S.getBody());
909 // If there is an increment, emit it next.
911 EmitBlock(Continue.getBlock());
912 EmitStmt(S.getInc());
915 BreakContinueStack.pop_back();
917 ConditionScope.ForceCleanup();
920 EmitBranch(CondBlock);
922 ForScope.ForceCleanup();
926 // Emit the fall-through block.
927 EmitBlock(LoopExit.getBlock(), true);
931 CodeGenFunction::EmitCXXForRangeStmt(const CXXForRangeStmt &S,
932 ArrayRef<const Attr *> ForAttrs) {
933 JumpDest LoopExit = getJumpDestInCurrentScope("for.end");
935 LexicalScope ForScope(*this, S.getSourceRange());
937 // Evaluate the first pieces before the loop.
938 EmitStmt(S.getRangeStmt());
939 EmitStmt(S.getBeginStmt());
940 EmitStmt(S.getEndStmt());
942 // Start the loop with a block that tests the condition.
943 // If there's an increment, the continue scope will be overwritten
945 llvm::BasicBlock *CondBlock = createBasicBlock("for.cond");
946 EmitBlock(CondBlock);
948 const SourceRange &R = S.getSourceRange();
949 LoopStack.push(CondBlock, CGM.getContext(), ForAttrs,
950 SourceLocToDebugLoc(R.getBegin()),
951 SourceLocToDebugLoc(R.getEnd()));
953 // If there are any cleanups between here and the loop-exit scope,
954 // create a block to stage a loop exit along.
955 llvm::BasicBlock *ExitBlock = LoopExit.getBlock();
956 if (ForScope.requiresCleanups())
957 ExitBlock = createBasicBlock("for.cond.cleanup");
959 // The loop body, consisting of the specified body and the loop variable.
960 llvm::BasicBlock *ForBody = createBasicBlock("for.body");
962 // The body is executed if the expression, contextually converted
964 llvm::Value *BoolCondVal = EvaluateExprAsBool(S.getCond());
965 Builder.CreateCondBr(
966 BoolCondVal, ForBody, ExitBlock,
967 createProfileWeightsForLoop(S.getCond(), getProfileCount(S.getBody())));
969 if (ExitBlock != LoopExit.getBlock()) {
970 EmitBlock(ExitBlock);
971 EmitBranchThroughCleanup(LoopExit);
975 incrementProfileCounter(&S);
977 // Create a block for the increment. In case of a 'continue', we jump there.
978 JumpDest Continue = getJumpDestInCurrentScope("for.inc");
980 // Store the blocks to use for break and continue.
981 BreakContinueStack.push_back(BreakContinue(LoopExit, Continue));
984 // Create a separate cleanup scope for the loop variable and body.
985 LexicalScope BodyScope(*this, S.getSourceRange());
986 EmitStmt(S.getLoopVarStmt());
987 EmitStmt(S.getBody());
991 // If there is an increment, emit it next.
992 EmitBlock(Continue.getBlock());
993 EmitStmt(S.getInc());
995 BreakContinueStack.pop_back();
997 EmitBranch(CondBlock);
999 ForScope.ForceCleanup();
1003 // Emit the fall-through block.
1004 EmitBlock(LoopExit.getBlock(), true);
1007 void CodeGenFunction::EmitReturnOfRValue(RValue RV, QualType Ty) {
1008 if (RV.isScalar()) {
1009 Builder.CreateStore(RV.getScalarVal(), ReturnValue);
1010 } else if (RV.isAggregate()) {
1011 EmitAggregateCopy(ReturnValue, RV.getAggregateAddress(), Ty);
1013 EmitStoreOfComplex(RV.getComplexVal(), MakeAddrLValue(ReturnValue, Ty),
1016 EmitBranchThroughCleanup(ReturnBlock);
1019 /// EmitReturnStmt - Note that due to GCC extensions, this can have an operand
1020 /// if the function returns void, or may be missing one if the function returns
1021 /// non-void. Fun stuff :).
1022 void CodeGenFunction::EmitReturnStmt(const ReturnStmt &S) {
1023 // Returning from an outlined SEH helper is UB, and we already warn on it.
1024 if (IsOutlinedSEHHelper) {
1025 Builder.CreateUnreachable();
1026 Builder.ClearInsertionPoint();
1029 // Emit the result value, even if unused, to evalute the side effects.
1030 const Expr *RV = S.getRetValue();
1032 // Treat block literals in a return expression as if they appeared
1033 // in their own scope. This permits a small, easily-implemented
1034 // exception to our over-conservative rules about not jumping to
1035 // statements following block literals with non-trivial cleanups.
1036 RunCleanupsScope cleanupScope(*this);
1037 if (const ExprWithCleanups *cleanups =
1038 dyn_cast_or_null<ExprWithCleanups>(RV)) {
1039 enterFullExpression(cleanups);
1040 RV = cleanups->getSubExpr();
1043 // FIXME: Clean this up by using an LValue for ReturnTemp,
1044 // EmitStoreThroughLValue, and EmitAnyExpr.
1045 if (getLangOpts().ElideConstructors &&
1046 S.getNRVOCandidate() && S.getNRVOCandidate()->isNRVOVariable()) {
1047 // Apply the named return value optimization for this return statement,
1048 // which means doing nothing: the appropriate result has already been
1049 // constructed into the NRVO variable.
1051 // If there is an NRVO flag for this variable, set it to 1 into indicate
1052 // that the cleanup code should not destroy the variable.
1053 if (llvm::Value *NRVOFlag = NRVOFlags[S.getNRVOCandidate()])
1054 Builder.CreateFlagStore(Builder.getTrue(), NRVOFlag);
1055 } else if (!ReturnValue.isValid() || (RV && RV->getType()->isVoidType())) {
1056 // Make sure not to return anything, but evaluate the expression
1057 // for side effects.
1061 // Do nothing (return value is left uninitialized)
1062 } else if (FnRetTy->isReferenceType()) {
1063 // If this function returns a reference, take the address of the expression
1064 // rather than the value.
1065 RValue Result = EmitReferenceBindingToExpr(RV);
1066 Builder.CreateStore(Result.getScalarVal(), ReturnValue);
1068 switch (getEvaluationKind(RV->getType())) {
1070 Builder.CreateStore(EmitScalarExpr(RV), ReturnValue);
1073 EmitComplexExprIntoLValue(RV, MakeAddrLValue(ReturnValue, RV->getType()),
1077 EmitAggExpr(RV, AggValueSlot::forAddr(ReturnValue,
1079 AggValueSlot::IsDestructed,
1080 AggValueSlot::DoesNotNeedGCBarriers,
1081 AggValueSlot::IsNotAliased));
1087 if (!RV || RV->isEvaluatable(getContext()))
1088 ++NumSimpleReturnExprs;
1090 cleanupScope.ForceCleanup();
1091 EmitBranchThroughCleanup(ReturnBlock);
1094 void CodeGenFunction::EmitDeclStmt(const DeclStmt &S) {
1095 // As long as debug info is modeled with instructions, we have to ensure we
1096 // have a place to insert here and write the stop point here.
1097 if (HaveInsertPoint())
1100 for (const auto *I : S.decls())
1104 void CodeGenFunction::EmitBreakStmt(const BreakStmt &S) {
1105 assert(!BreakContinueStack.empty() && "break stmt not in a loop or switch!");
1107 // If this code is reachable then emit a stop point (if generating
1108 // debug info). We have to do this ourselves because we are on the
1109 // "simple" statement path.
1110 if (HaveInsertPoint())
1113 EmitBranchThroughCleanup(BreakContinueStack.back().BreakBlock);
1116 void CodeGenFunction::EmitContinueStmt(const ContinueStmt &S) {
1117 assert(!BreakContinueStack.empty() && "continue stmt not in a loop!");
1119 // If this code is reachable then emit a stop point (if generating
1120 // debug info). We have to do this ourselves because we are on the
1121 // "simple" statement path.
1122 if (HaveInsertPoint())
1125 EmitBranchThroughCleanup(BreakContinueStack.back().ContinueBlock);
1128 /// EmitCaseStmtRange - If case statement range is not too big then
1129 /// add multiple cases to switch instruction, one for each value within
1130 /// the range. If range is too big then emit "if" condition check.
1131 void CodeGenFunction::EmitCaseStmtRange(const CaseStmt &S) {
1132 assert(S.getRHS() && "Expected RHS value in CaseStmt");
1134 llvm::APSInt LHS = S.getLHS()->EvaluateKnownConstInt(getContext());
1135 llvm::APSInt RHS = S.getRHS()->EvaluateKnownConstInt(getContext());
1137 // Emit the code for this case. We do this first to make sure it is
1138 // properly chained from our predecessor before generating the
1139 // switch machinery to enter this block.
1140 llvm::BasicBlock *CaseDest = createBasicBlock("sw.bb");
1141 EmitBlockWithFallThrough(CaseDest, &S);
1142 EmitStmt(S.getSubStmt());
1144 // If range is empty, do nothing.
1145 if (LHS.isSigned() ? RHS.slt(LHS) : RHS.ult(LHS))
1148 llvm::APInt Range = RHS - LHS;
1149 // FIXME: parameters such as this should not be hardcoded.
1150 if (Range.ult(llvm::APInt(Range.getBitWidth(), 64))) {
1151 // Range is small enough to add multiple switch instruction cases.
1152 uint64_t Total = getProfileCount(&S);
1153 unsigned NCases = Range.getZExtValue() + 1;
1154 // We only have one region counter for the entire set of cases here, so we
1155 // need to divide the weights evenly between the generated cases, ensuring
1156 // that the total weight is preserved. E.g., a weight of 5 over three cases
1157 // will be distributed as weights of 2, 2, and 1.
1158 uint64_t Weight = Total / NCases, Rem = Total % NCases;
1159 for (unsigned I = 0; I != NCases; ++I) {
1161 SwitchWeights->push_back(Weight + (Rem ? 1 : 0));
1164 SwitchInsn->addCase(Builder.getInt(LHS), CaseDest);
1170 // The range is too big. Emit "if" condition into a new block,
1171 // making sure to save and restore the current insertion point.
1172 llvm::BasicBlock *RestoreBB = Builder.GetInsertBlock();
1174 // Push this test onto the chain of range checks (which terminates
1175 // in the default basic block). The switch's default will be changed
1176 // to the top of this chain after switch emission is complete.
1177 llvm::BasicBlock *FalseDest = CaseRangeBlock;
1178 CaseRangeBlock = createBasicBlock("sw.caserange");
1180 CurFn->getBasicBlockList().push_back(CaseRangeBlock);
1181 Builder.SetInsertPoint(CaseRangeBlock);
1183 // Emit range check.
1185 Builder.CreateSub(SwitchInsn->getCondition(), Builder.getInt(LHS));
1187 Builder.CreateICmpULE(Diff, Builder.getInt(Range), "inbounds");
1189 llvm::MDNode *Weights = nullptr;
1190 if (SwitchWeights) {
1191 uint64_t ThisCount = getProfileCount(&S);
1192 uint64_t DefaultCount = (*SwitchWeights)[0];
1193 Weights = createProfileWeights(ThisCount, DefaultCount);
1195 // Since we're chaining the switch default through each large case range, we
1196 // need to update the weight for the default, ie, the first case, to include
1198 (*SwitchWeights)[0] += ThisCount;
1200 Builder.CreateCondBr(Cond, CaseDest, FalseDest, Weights);
1202 // Restore the appropriate insertion point.
1204 Builder.SetInsertPoint(RestoreBB);
1206 Builder.ClearInsertionPoint();
1209 void CodeGenFunction::EmitCaseStmt(const CaseStmt &S) {
1210 // If there is no enclosing switch instance that we're aware of, then this
1211 // case statement and its block can be elided. This situation only happens
1212 // when we've constant-folded the switch, are emitting the constant case,
1213 // and part of the constant case includes another case statement. For
1214 // instance: switch (4) { case 4: do { case 5: } while (1); }
1216 EmitStmt(S.getSubStmt());
1220 // Handle case ranges.
1222 EmitCaseStmtRange(S);
1226 llvm::ConstantInt *CaseVal =
1227 Builder.getInt(S.getLHS()->EvaluateKnownConstInt(getContext()));
1229 // If the body of the case is just a 'break', try to not emit an empty block.
1230 // If we're profiling or we're not optimizing, leave the block in for better
1231 // debug and coverage analysis.
1232 if (!CGM.getCodeGenOpts().hasProfileClangInstr() &&
1233 CGM.getCodeGenOpts().OptimizationLevel > 0 &&
1234 isa<BreakStmt>(S.getSubStmt())) {
1235 JumpDest Block = BreakContinueStack.back().BreakBlock;
1237 // Only do this optimization if there are no cleanups that need emitting.
1238 if (isObviouslyBranchWithoutCleanups(Block)) {
1240 SwitchWeights->push_back(getProfileCount(&S));
1241 SwitchInsn->addCase(CaseVal, Block.getBlock());
1243 // If there was a fallthrough into this case, make sure to redirect it to
1244 // the end of the switch as well.
1245 if (Builder.GetInsertBlock()) {
1246 Builder.CreateBr(Block.getBlock());
1247 Builder.ClearInsertionPoint();
1253 llvm::BasicBlock *CaseDest = createBasicBlock("sw.bb");
1254 EmitBlockWithFallThrough(CaseDest, &S);
1256 SwitchWeights->push_back(getProfileCount(&S));
1257 SwitchInsn->addCase(CaseVal, CaseDest);
1259 // Recursively emitting the statement is acceptable, but is not wonderful for
1260 // code where we have many case statements nested together, i.e.:
1264 // Handling this recursively will create a new block for each case statement
1265 // that falls through to the next case which is IR intensive. It also causes
1266 // deep recursion which can run into stack depth limitations. Handle
1267 // sequential non-range case statements specially.
1268 const CaseStmt *CurCase = &S;
1269 const CaseStmt *NextCase = dyn_cast<CaseStmt>(S.getSubStmt());
1271 // Otherwise, iteratively add consecutive cases to this switch stmt.
1272 while (NextCase && NextCase->getRHS() == nullptr) {
1274 llvm::ConstantInt *CaseVal =
1275 Builder.getInt(CurCase->getLHS()->EvaluateKnownConstInt(getContext()));
1278 SwitchWeights->push_back(getProfileCount(NextCase));
1279 if (CGM.getCodeGenOpts().hasProfileClangInstr()) {
1280 CaseDest = createBasicBlock("sw.bb");
1281 EmitBlockWithFallThrough(CaseDest, &S);
1284 SwitchInsn->addCase(CaseVal, CaseDest);
1285 NextCase = dyn_cast<CaseStmt>(CurCase->getSubStmt());
1288 // Normal default recursion for non-cases.
1289 EmitStmt(CurCase->getSubStmt());
1292 void CodeGenFunction::EmitDefaultStmt(const DefaultStmt &S) {
1293 // If there is no enclosing switch instance that we're aware of, then this
1294 // default statement can be elided. This situation only happens when we've
1295 // constant-folded the switch.
1297 EmitStmt(S.getSubStmt());
1301 llvm::BasicBlock *DefaultBlock = SwitchInsn->getDefaultDest();
1302 assert(DefaultBlock->empty() &&
1303 "EmitDefaultStmt: Default block already defined?");
1305 EmitBlockWithFallThrough(DefaultBlock, &S);
1307 EmitStmt(S.getSubStmt());
1310 /// CollectStatementsForCase - Given the body of a 'switch' statement and a
1311 /// constant value that is being switched on, see if we can dead code eliminate
1312 /// the body of the switch to a simple series of statements to emit. Basically,
1313 /// on a switch (5) we want to find these statements:
1315 /// printf(...); <--
1319 /// and add them to the ResultStmts vector. If it is unsafe to do this
1320 /// transformation (for example, one of the elided statements contains a label
1321 /// that might be jumped to), return CSFC_Failure. If we handled it and 'S'
1322 /// should include statements after it (e.g. the printf() line is a substmt of
1323 /// the case) then return CSFC_FallThrough. If we handled it and found a break
1324 /// statement, then return CSFC_Success.
1326 /// If Case is non-null, then we are looking for the specified case, checking
1327 /// that nothing we jump over contains labels. If Case is null, then we found
1328 /// the case and are looking for the break.
1330 /// If the recursive walk actually finds our Case, then we set FoundCase to
1333 enum CSFC_Result { CSFC_Failure, CSFC_FallThrough, CSFC_Success };
1334 static CSFC_Result CollectStatementsForCase(const Stmt *S,
1335 const SwitchCase *Case,
1337 SmallVectorImpl<const Stmt*> &ResultStmts) {
1338 // If this is a null statement, just succeed.
1340 return Case ? CSFC_Success : CSFC_FallThrough;
1342 // If this is the switchcase (case 4: or default) that we're looking for, then
1343 // we're in business. Just add the substatement.
1344 if (const SwitchCase *SC = dyn_cast<SwitchCase>(S)) {
1347 return CollectStatementsForCase(SC->getSubStmt(), nullptr, FoundCase,
1351 // Otherwise, this is some other case or default statement, just ignore it.
1352 return CollectStatementsForCase(SC->getSubStmt(), Case, FoundCase,
1356 // If we are in the live part of the code and we found our break statement,
1357 // return a success!
1358 if (!Case && isa<BreakStmt>(S))
1359 return CSFC_Success;
1361 // If this is a switch statement, then it might contain the SwitchCase, the
1362 // break, or neither.
1363 if (const CompoundStmt *CS = dyn_cast<CompoundStmt>(S)) {
1364 // Handle this as two cases: we might be looking for the SwitchCase (if so
1365 // the skipped statements must be skippable) or we might already have it.
1366 CompoundStmt::const_body_iterator I = CS->body_begin(), E = CS->body_end();
1367 bool StartedInLiveCode = FoundCase;
1368 unsigned StartSize = ResultStmts.size();
1370 // If we've not found the case yet, scan through looking for it.
1372 // Keep track of whether we see a skipped declaration. The code could be
1373 // using the declaration even if it is skipped, so we can't optimize out
1374 // the decl if the kept statements might refer to it.
1375 bool HadSkippedDecl = false;
1377 // If we're looking for the case, just see if we can skip each of the
1379 for (; Case && I != E; ++I) {
1380 HadSkippedDecl |= CodeGenFunction::mightAddDeclToScope(*I);
1382 switch (CollectStatementsForCase(*I, Case, FoundCase, ResultStmts)) {
1383 case CSFC_Failure: return CSFC_Failure;
1385 // A successful result means that either 1) that the statement doesn't
1386 // have the case and is skippable, or 2) does contain the case value
1387 // and also contains the break to exit the switch. In the later case,
1388 // we just verify the rest of the statements are elidable.
1390 // If we found the case and skipped declarations, we can't do the
1393 return CSFC_Failure;
1395 for (++I; I != E; ++I)
1396 if (CodeGenFunction::ContainsLabel(*I, true))
1397 return CSFC_Failure;
1398 return CSFC_Success;
1401 case CSFC_FallThrough:
1402 // If we have a fallthrough condition, then we must have found the
1403 // case started to include statements. Consider the rest of the
1404 // statements in the compound statement as candidates for inclusion.
1405 assert(FoundCase && "Didn't find case but returned fallthrough?");
1406 // We recursively found Case, so we're not looking for it anymore.
1409 // If we found the case and skipped declarations, we can't do the
1412 return CSFC_Failure;
1418 return CSFC_Success;
1420 assert(!HadSkippedDecl && "fallthrough after skipping decl");
1423 // If we have statements in our range, then we know that the statements are
1424 // live and need to be added to the set of statements we're tracking.
1425 bool AnyDecls = false;
1426 for (; I != E; ++I) {
1427 AnyDecls |= CodeGenFunction::mightAddDeclToScope(*I);
1429 switch (CollectStatementsForCase(*I, nullptr, FoundCase, ResultStmts)) {
1430 case CSFC_Failure: return CSFC_Failure;
1431 case CSFC_FallThrough:
1432 // A fallthrough result means that the statement was simple and just
1433 // included in ResultStmt, keep adding them afterwards.
1436 // A successful result means that we found the break statement and
1437 // stopped statement inclusion. We just ensure that any leftover stmts
1438 // are skippable and return success ourselves.
1439 for (++I; I != E; ++I)
1440 if (CodeGenFunction::ContainsLabel(*I, true))
1441 return CSFC_Failure;
1442 return CSFC_Success;
1446 // If we're about to fall out of a scope without hitting a 'break;', we
1447 // can't perform the optimization if there were any decls in that scope
1448 // (we'd lose their end-of-lifetime).
1450 // If the entire compound statement was live, there's one more thing we
1451 // can try before giving up: emit the whole thing as a single statement.
1452 // We can do that unless the statement contains a 'break;'.
1453 // FIXME: Such a break must be at the end of a construct within this one.
1454 // We could emit this by just ignoring the BreakStmts entirely.
1455 if (StartedInLiveCode && !CodeGenFunction::containsBreak(S)) {
1456 ResultStmts.resize(StartSize);
1457 ResultStmts.push_back(S);
1459 return CSFC_Failure;
1463 return CSFC_FallThrough;
1466 // Okay, this is some other statement that we don't handle explicitly, like a
1467 // for statement or increment etc. If we are skipping over this statement,
1468 // just verify it doesn't have labels, which would make it invalid to elide.
1470 if (CodeGenFunction::ContainsLabel(S, true))
1471 return CSFC_Failure;
1472 return CSFC_Success;
1475 // Otherwise, we want to include this statement. Everything is cool with that
1476 // so long as it doesn't contain a break out of the switch we're in.
1477 if (CodeGenFunction::containsBreak(S)) return CSFC_Failure;
1479 // Otherwise, everything is great. Include the statement and tell the caller
1480 // that we fall through and include the next statement as well.
1481 ResultStmts.push_back(S);
1482 return CSFC_FallThrough;
1485 /// FindCaseStatementsForValue - Find the case statement being jumped to and
1486 /// then invoke CollectStatementsForCase to find the list of statements to emit
1487 /// for a switch on constant. See the comment above CollectStatementsForCase
1488 /// for more details.
1489 static bool FindCaseStatementsForValue(const SwitchStmt &S,
1490 const llvm::APSInt &ConstantCondValue,
1491 SmallVectorImpl<const Stmt*> &ResultStmts,
1493 const SwitchCase *&ResultCase) {
1494 // First step, find the switch case that is being branched to. We can do this
1495 // efficiently by scanning the SwitchCase list.
1496 const SwitchCase *Case = S.getSwitchCaseList();
1497 const DefaultStmt *DefaultCase = nullptr;
1499 for (; Case; Case = Case->getNextSwitchCase()) {
1500 // It's either a default or case. Just remember the default statement in
1501 // case we're not jumping to any numbered cases.
1502 if (const DefaultStmt *DS = dyn_cast<DefaultStmt>(Case)) {
1507 // Check to see if this case is the one we're looking for.
1508 const CaseStmt *CS = cast<CaseStmt>(Case);
1509 // Don't handle case ranges yet.
1510 if (CS->getRHS()) return false;
1512 // If we found our case, remember it as 'case'.
1513 if (CS->getLHS()->EvaluateKnownConstInt(C) == ConstantCondValue)
1517 // If we didn't find a matching case, we use a default if it exists, or we
1518 // elide the whole switch body!
1520 // It is safe to elide the body of the switch if it doesn't contain labels
1521 // etc. If it is safe, return successfully with an empty ResultStmts list.
1523 return !CodeGenFunction::ContainsLabel(&S);
1527 // Ok, we know which case is being jumped to, try to collect all the
1528 // statements that follow it. This can fail for a variety of reasons. Also,
1529 // check to see that the recursive walk actually found our case statement.
1530 // Insane cases like this can fail to find it in the recursive walk since we
1531 // don't handle every stmt kind:
1535 bool FoundCase = false;
1537 return CollectStatementsForCase(S.getBody(), Case, FoundCase,
1538 ResultStmts) != CSFC_Failure &&
1542 void CodeGenFunction::EmitSwitchStmt(const SwitchStmt &S) {
1543 // Handle nested switch statements.
1544 llvm::SwitchInst *SavedSwitchInsn = SwitchInsn;
1545 SmallVector<uint64_t, 16> *SavedSwitchWeights = SwitchWeights;
1546 llvm::BasicBlock *SavedCRBlock = CaseRangeBlock;
1548 // See if we can constant fold the condition of the switch and therefore only
1549 // emit the live case statement (if any) of the switch.
1550 llvm::APSInt ConstantCondValue;
1551 if (ConstantFoldsToSimpleInteger(S.getCond(), ConstantCondValue)) {
1552 SmallVector<const Stmt*, 4> CaseStmts;
1553 const SwitchCase *Case = nullptr;
1554 if (FindCaseStatementsForValue(S, ConstantCondValue, CaseStmts,
1555 getContext(), Case)) {
1557 incrementProfileCounter(Case);
1558 RunCleanupsScope ExecutedScope(*this);
1561 EmitStmt(S.getInit());
1563 // Emit the condition variable if needed inside the entire cleanup scope
1564 // used by this special case for constant folded switches.
1565 if (S.getConditionVariable())
1566 EmitAutoVarDecl(*S.getConditionVariable());
1568 // At this point, we are no longer "within" a switch instance, so
1569 // we can temporarily enforce this to ensure that any embedded case
1570 // statements are not emitted.
1571 SwitchInsn = nullptr;
1573 // Okay, we can dead code eliminate everything except this case. Emit the
1574 // specified series of statements and we're good.
1575 for (unsigned i = 0, e = CaseStmts.size(); i != e; ++i)
1576 EmitStmt(CaseStmts[i]);
1577 incrementProfileCounter(&S);
1579 // Now we want to restore the saved switch instance so that nested
1580 // switches continue to function properly
1581 SwitchInsn = SavedSwitchInsn;
1587 JumpDest SwitchExit = getJumpDestInCurrentScope("sw.epilog");
1589 RunCleanupsScope ConditionScope(*this);
1592 EmitStmt(S.getInit());
1594 if (S.getConditionVariable())
1595 EmitAutoVarDecl(*S.getConditionVariable());
1596 llvm::Value *CondV = EmitScalarExpr(S.getCond());
1598 // Create basic block to hold stuff that comes after switch
1599 // statement. We also need to create a default block now so that
1600 // explicit case ranges tests can have a place to jump to on
1602 llvm::BasicBlock *DefaultBlock = createBasicBlock("sw.default");
1603 SwitchInsn = Builder.CreateSwitch(CondV, DefaultBlock);
1604 if (PGO.haveRegionCounts()) {
1605 // Walk the SwitchCase list to find how many there are.
1606 uint64_t DefaultCount = 0;
1607 unsigned NumCases = 0;
1608 for (const SwitchCase *Case = S.getSwitchCaseList();
1610 Case = Case->getNextSwitchCase()) {
1611 if (isa<DefaultStmt>(Case))
1612 DefaultCount = getProfileCount(Case);
1615 SwitchWeights = new SmallVector<uint64_t, 16>();
1616 SwitchWeights->reserve(NumCases);
1617 // The default needs to be first. We store the edge count, so we already
1618 // know the right weight.
1619 SwitchWeights->push_back(DefaultCount);
1621 CaseRangeBlock = DefaultBlock;
1623 // Clear the insertion point to indicate we are in unreachable code.
1624 Builder.ClearInsertionPoint();
1626 // All break statements jump to NextBlock. If BreakContinueStack is non-empty
1627 // then reuse last ContinueBlock.
1628 JumpDest OuterContinue;
1629 if (!BreakContinueStack.empty())
1630 OuterContinue = BreakContinueStack.back().ContinueBlock;
1632 BreakContinueStack.push_back(BreakContinue(SwitchExit, OuterContinue));
1634 // Emit switch body.
1635 EmitStmt(S.getBody());
1637 BreakContinueStack.pop_back();
1639 // Update the default block in case explicit case range tests have
1640 // been chained on top.
1641 SwitchInsn->setDefaultDest(CaseRangeBlock);
1643 // If a default was never emitted:
1644 if (!DefaultBlock->getParent()) {
1645 // If we have cleanups, emit the default block so that there's a
1646 // place to jump through the cleanups from.
1647 if (ConditionScope.requiresCleanups()) {
1648 EmitBlock(DefaultBlock);
1650 // Otherwise, just forward the default block to the switch end.
1652 DefaultBlock->replaceAllUsesWith(SwitchExit.getBlock());
1653 delete DefaultBlock;
1657 ConditionScope.ForceCleanup();
1659 // Emit continuation.
1660 EmitBlock(SwitchExit.getBlock(), true);
1661 incrementProfileCounter(&S);
1663 // If the switch has a condition wrapped by __builtin_unpredictable,
1664 // create metadata that specifies that the switch is unpredictable.
1665 // Don't bother if not optimizing because that metadata would not be used.
1666 auto *Call = dyn_cast<CallExpr>(S.getCond());
1667 if (Call && CGM.getCodeGenOpts().OptimizationLevel != 0) {
1668 auto *FD = dyn_cast_or_null<FunctionDecl>(Call->getCalleeDecl());
1669 if (FD && FD->getBuiltinID() == Builtin::BI__builtin_unpredictable) {
1670 llvm::MDBuilder MDHelper(getLLVMContext());
1671 SwitchInsn->setMetadata(llvm::LLVMContext::MD_unpredictable,
1672 MDHelper.createUnpredictable());
1676 if (SwitchWeights) {
1677 assert(SwitchWeights->size() == 1 + SwitchInsn->getNumCases() &&
1678 "switch weights do not match switch cases");
1679 // If there's only one jump destination there's no sense weighting it.
1680 if (SwitchWeights->size() > 1)
1681 SwitchInsn->setMetadata(llvm::LLVMContext::MD_prof,
1682 createProfileWeights(*SwitchWeights));
1683 delete SwitchWeights;
1685 SwitchInsn = SavedSwitchInsn;
1686 SwitchWeights = SavedSwitchWeights;
1687 CaseRangeBlock = SavedCRBlock;
1691 SimplifyConstraint(const char *Constraint, const TargetInfo &Target,
1692 SmallVectorImpl<TargetInfo::ConstraintInfo> *OutCons=nullptr) {
1695 while (*Constraint) {
1696 switch (*Constraint) {
1698 Result += Target.convertConstraint(Constraint);
1704 case '=': // Will see this and the following in mult-alt constraints.
1707 case '#': // Ignore the rest of the constraint alternative.
1708 while (Constraint[1] && Constraint[1] != ',')
1713 Result += *Constraint;
1714 while (Constraint[1] && Constraint[1] == *Constraint)
1725 "Must pass output names to constraints with a symbolic name");
1727 bool result = Target.resolveSymbolicName(Constraint, *OutCons, Index);
1728 assert(result && "Could not resolve symbolic name"); (void)result;
1729 Result += llvm::utostr(Index);
1740 /// AddVariableConstraints - Look at AsmExpr and if it is a variable declared
1741 /// as using a particular register add that as a constraint that will be used
1742 /// in this asm stmt.
1744 AddVariableConstraints(const std::string &Constraint, const Expr &AsmExpr,
1745 const TargetInfo &Target, CodeGenModule &CGM,
1746 const AsmStmt &Stmt, const bool EarlyClobber) {
1747 const DeclRefExpr *AsmDeclRef = dyn_cast<DeclRefExpr>(&AsmExpr);
1750 const ValueDecl &Value = *AsmDeclRef->getDecl();
1751 const VarDecl *Variable = dyn_cast<VarDecl>(&Value);
1754 if (Variable->getStorageClass() != SC_Register)
1756 AsmLabelAttr *Attr = Variable->getAttr<AsmLabelAttr>();
1759 StringRef Register = Attr->getLabel();
1760 assert(Target.isValidGCCRegisterName(Register));
1761 // We're using validateOutputConstraint here because we only care if
1762 // this is a register constraint.
1763 TargetInfo::ConstraintInfo Info(Constraint, "");
1764 if (Target.validateOutputConstraint(Info) &&
1765 !Info.allowsRegister()) {
1766 CGM.ErrorUnsupported(&Stmt, "__asm__");
1769 // Canonicalize the register here before returning it.
1770 Register = Target.getNormalizedGCCRegisterName(Register);
1771 return (EarlyClobber ? "&{" : "{") + Register.str() + "}";
1775 CodeGenFunction::EmitAsmInputLValue(const TargetInfo::ConstraintInfo &Info,
1776 LValue InputValue, QualType InputType,
1777 std::string &ConstraintStr,
1778 SourceLocation Loc) {
1780 if (Info.allowsRegister() || !Info.allowsMemory()) {
1781 if (CodeGenFunction::hasScalarEvaluationKind(InputType)) {
1782 Arg = EmitLoadOfLValue(InputValue, Loc).getScalarVal();
1784 llvm::Type *Ty = ConvertType(InputType);
1785 uint64_t Size = CGM.getDataLayout().getTypeSizeInBits(Ty);
1786 if (Size <= 64 && llvm::isPowerOf2_64(Size)) {
1787 Ty = llvm::IntegerType::get(getLLVMContext(), Size);
1788 Ty = llvm::PointerType::getUnqual(Ty);
1790 Arg = Builder.CreateLoad(Builder.CreateBitCast(InputValue.getAddress(),
1793 Arg = InputValue.getPointer();
1794 ConstraintStr += '*';
1798 Arg = InputValue.getPointer();
1799 ConstraintStr += '*';
1805 llvm::Value* CodeGenFunction::EmitAsmInput(
1806 const TargetInfo::ConstraintInfo &Info,
1807 const Expr *InputExpr,
1808 std::string &ConstraintStr) {
1809 // If this can't be a register or memory, i.e., has to be a constant
1810 // (immediate or symbolic), try to emit it as such.
1811 if (!Info.allowsRegister() && !Info.allowsMemory()) {
1812 llvm::APSInt Result;
1813 if (InputExpr->EvaluateAsInt(Result, getContext()))
1814 return llvm::ConstantInt::get(getLLVMContext(), Result);
1815 assert(!Info.requiresImmediateConstant() &&
1816 "Required-immediate inlineasm arg isn't constant?");
1819 if (Info.allowsRegister() || !Info.allowsMemory())
1820 if (CodeGenFunction::hasScalarEvaluationKind(InputExpr->getType()))
1821 return EmitScalarExpr(InputExpr);
1822 if (InputExpr->getStmtClass() == Expr::CXXThisExprClass)
1823 return EmitScalarExpr(InputExpr);
1824 InputExpr = InputExpr->IgnoreParenNoopCasts(getContext());
1825 LValue Dest = EmitLValue(InputExpr);
1826 return EmitAsmInputLValue(Info, Dest, InputExpr->getType(), ConstraintStr,
1827 InputExpr->getExprLoc());
1830 /// getAsmSrcLocInfo - Return the !srcloc metadata node to attach to an inline
1831 /// asm call instruction. The !srcloc MDNode contains a list of constant
1832 /// integers which are the source locations of the start of each line in the
1834 static llvm::MDNode *getAsmSrcLocInfo(const StringLiteral *Str,
1835 CodeGenFunction &CGF) {
1836 SmallVector<llvm::Metadata *, 8> Locs;
1837 // Add the location of the first line to the MDNode.
1838 Locs.push_back(llvm::ConstantAsMetadata::get(llvm::ConstantInt::get(
1839 CGF.Int32Ty, Str->getLocStart().getRawEncoding())));
1840 StringRef StrVal = Str->getString();
1841 if (!StrVal.empty()) {
1842 const SourceManager &SM = CGF.CGM.getContext().getSourceManager();
1843 const LangOptions &LangOpts = CGF.CGM.getLangOpts();
1844 unsigned StartToken = 0;
1845 unsigned ByteOffset = 0;
1847 // Add the location of the start of each subsequent line of the asm to the
1849 for (unsigned i = 0, e = StrVal.size() - 1; i != e; ++i) {
1850 if (StrVal[i] != '\n') continue;
1851 SourceLocation LineLoc = Str->getLocationOfByte(
1852 i + 1, SM, LangOpts, CGF.getTarget(), &StartToken, &ByteOffset);
1853 Locs.push_back(llvm::ConstantAsMetadata::get(
1854 llvm::ConstantInt::get(CGF.Int32Ty, LineLoc.getRawEncoding())));
1858 return llvm::MDNode::get(CGF.getLLVMContext(), Locs);
1861 void CodeGenFunction::EmitAsmStmt(const AsmStmt &S) {
1862 // Assemble the final asm string.
1863 std::string AsmString = S.generateAsmString(getContext());
1865 // Get all the output and input constraints together.
1866 SmallVector<TargetInfo::ConstraintInfo, 4> OutputConstraintInfos;
1867 SmallVector<TargetInfo::ConstraintInfo, 4> InputConstraintInfos;
1869 for (unsigned i = 0, e = S.getNumOutputs(); i != e; i++) {
1871 if (const GCCAsmStmt *GAS = dyn_cast<GCCAsmStmt>(&S))
1872 Name = GAS->getOutputName(i);
1873 TargetInfo::ConstraintInfo Info(S.getOutputConstraint(i), Name);
1874 bool IsValid = getTarget().validateOutputConstraint(Info); (void)IsValid;
1875 assert(IsValid && "Failed to parse output constraint");
1876 OutputConstraintInfos.push_back(Info);
1879 for (unsigned i = 0, e = S.getNumInputs(); i != e; i++) {
1881 if (const GCCAsmStmt *GAS = dyn_cast<GCCAsmStmt>(&S))
1882 Name = GAS->getInputName(i);
1883 TargetInfo::ConstraintInfo Info(S.getInputConstraint(i), Name);
1885 getTarget().validateInputConstraint(OutputConstraintInfos, Info);
1886 assert(IsValid && "Failed to parse input constraint"); (void)IsValid;
1887 InputConstraintInfos.push_back(Info);
1890 std::string Constraints;
1892 std::vector<LValue> ResultRegDests;
1893 std::vector<QualType> ResultRegQualTys;
1894 std::vector<llvm::Type *> ResultRegTypes;
1895 std::vector<llvm::Type *> ResultTruncRegTypes;
1896 std::vector<llvm::Type *> ArgTypes;
1897 std::vector<llvm::Value*> Args;
1899 // Keep track of inout constraints.
1900 std::string InOutConstraints;
1901 std::vector<llvm::Value*> InOutArgs;
1902 std::vector<llvm::Type*> InOutArgTypes;
1904 // An inline asm can be marked readonly if it meets the following conditions:
1905 // - it doesn't have any sideeffects
1906 // - it doesn't clobber memory
1907 // - it doesn't return a value by-reference
1908 // It can be marked readnone if it doesn't have any input memory constraints
1909 // in addition to meeting the conditions listed above.
1910 bool ReadOnly = true, ReadNone = true;
1912 for (unsigned i = 0, e = S.getNumOutputs(); i != e; i++) {
1913 TargetInfo::ConstraintInfo &Info = OutputConstraintInfos[i];
1915 // Simplify the output constraint.
1916 std::string OutputConstraint(S.getOutputConstraint(i));
1917 OutputConstraint = SimplifyConstraint(OutputConstraint.c_str() + 1,
1920 const Expr *OutExpr = S.getOutputExpr(i);
1921 OutExpr = OutExpr->IgnoreParenNoopCasts(getContext());
1923 OutputConstraint = AddVariableConstraints(OutputConstraint, *OutExpr,
1924 getTarget(), CGM, S,
1925 Info.earlyClobber());
1927 LValue Dest = EmitLValue(OutExpr);
1928 if (!Constraints.empty())
1931 // If this is a register output, then make the inline asm return it
1932 // by-value. If this is a memory result, return the value by-reference.
1933 if (!Info.allowsMemory() && hasScalarEvaluationKind(OutExpr->getType())) {
1934 Constraints += "=" + OutputConstraint;
1935 ResultRegQualTys.push_back(OutExpr->getType());
1936 ResultRegDests.push_back(Dest);
1937 ResultRegTypes.push_back(ConvertTypeForMem(OutExpr->getType()));
1938 ResultTruncRegTypes.push_back(ResultRegTypes.back());
1940 // If this output is tied to an input, and if the input is larger, then
1941 // we need to set the actual result type of the inline asm node to be the
1942 // same as the input type.
1943 if (Info.hasMatchingInput()) {
1945 for (InputNo = 0; InputNo != S.getNumInputs(); ++InputNo) {
1946 TargetInfo::ConstraintInfo &Input = InputConstraintInfos[InputNo];
1947 if (Input.hasTiedOperand() && Input.getTiedOperand() == i)
1950 assert(InputNo != S.getNumInputs() && "Didn't find matching input!");
1952 QualType InputTy = S.getInputExpr(InputNo)->getType();
1953 QualType OutputType = OutExpr->getType();
1955 uint64_t InputSize = getContext().getTypeSize(InputTy);
1956 if (getContext().getTypeSize(OutputType) < InputSize) {
1957 // Form the asm to return the value as a larger integer or fp type.
1958 ResultRegTypes.back() = ConvertType(InputTy);
1961 if (llvm::Type* AdjTy =
1962 getTargetHooks().adjustInlineAsmType(*this, OutputConstraint,
1963 ResultRegTypes.back()))
1964 ResultRegTypes.back() = AdjTy;
1966 CGM.getDiags().Report(S.getAsmLoc(),
1967 diag::err_asm_invalid_type_in_input)
1968 << OutExpr->getType() << OutputConstraint;
1971 ArgTypes.push_back(Dest.getAddress().getType());
1972 Args.push_back(Dest.getPointer());
1973 Constraints += "=*";
1974 Constraints += OutputConstraint;
1975 ReadOnly = ReadNone = false;
1978 if (Info.isReadWrite()) {
1979 InOutConstraints += ',';
1981 const Expr *InputExpr = S.getOutputExpr(i);
1982 llvm::Value *Arg = EmitAsmInputLValue(Info, Dest, InputExpr->getType(),
1984 InputExpr->getExprLoc());
1986 if (llvm::Type* AdjTy =
1987 getTargetHooks().adjustInlineAsmType(*this, OutputConstraint,
1989 Arg = Builder.CreateBitCast(Arg, AdjTy);
1991 if (Info.allowsRegister())
1992 InOutConstraints += llvm::utostr(i);
1994 InOutConstraints += OutputConstraint;
1996 InOutArgTypes.push_back(Arg->getType());
1997 InOutArgs.push_back(Arg);
2001 // If this is a Microsoft-style asm blob, store the return registers (EAX:EDX)
2002 // to the return value slot. Only do this when returning in registers.
2003 if (isa<MSAsmStmt>(&S)) {
2004 const ABIArgInfo &RetAI = CurFnInfo->getReturnInfo();
2005 if (RetAI.isDirect() || RetAI.isExtend()) {
2006 // Make a fake lvalue for the return value slot.
2007 LValue ReturnSlot = MakeAddrLValue(ReturnValue, FnRetTy);
2008 CGM.getTargetCodeGenInfo().addReturnRegisterOutputs(
2009 *this, ReturnSlot, Constraints, ResultRegTypes, ResultTruncRegTypes,
2010 ResultRegDests, AsmString, S.getNumOutputs());
2015 for (unsigned i = 0, e = S.getNumInputs(); i != e; i++) {
2016 const Expr *InputExpr = S.getInputExpr(i);
2018 TargetInfo::ConstraintInfo &Info = InputConstraintInfos[i];
2020 if (Info.allowsMemory())
2023 if (!Constraints.empty())
2026 // Simplify the input constraint.
2027 std::string InputConstraint(S.getInputConstraint(i));
2028 InputConstraint = SimplifyConstraint(InputConstraint.c_str(), getTarget(),
2029 &OutputConstraintInfos);
2031 InputConstraint = AddVariableConstraints(
2032 InputConstraint, *InputExpr->IgnoreParenNoopCasts(getContext()),
2033 getTarget(), CGM, S, false /* No EarlyClobber */);
2035 llvm::Value *Arg = EmitAsmInput(Info, InputExpr, Constraints);
2037 // If this input argument is tied to a larger output result, extend the
2038 // input to be the same size as the output. The LLVM backend wants to see
2039 // the input and output of a matching constraint be the same size. Note
2040 // that GCC does not define what the top bits are here. We use zext because
2041 // that is usually cheaper, but LLVM IR should really get an anyext someday.
2042 if (Info.hasTiedOperand()) {
2043 unsigned Output = Info.getTiedOperand();
2044 QualType OutputType = S.getOutputExpr(Output)->getType();
2045 QualType InputTy = InputExpr->getType();
2047 if (getContext().getTypeSize(OutputType) >
2048 getContext().getTypeSize(InputTy)) {
2049 // Use ptrtoint as appropriate so that we can do our extension.
2050 if (isa<llvm::PointerType>(Arg->getType()))
2051 Arg = Builder.CreatePtrToInt(Arg, IntPtrTy);
2052 llvm::Type *OutputTy = ConvertType(OutputType);
2053 if (isa<llvm::IntegerType>(OutputTy))
2054 Arg = Builder.CreateZExt(Arg, OutputTy);
2055 else if (isa<llvm::PointerType>(OutputTy))
2056 Arg = Builder.CreateZExt(Arg, IntPtrTy);
2058 assert(OutputTy->isFloatingPointTy() && "Unexpected output type");
2059 Arg = Builder.CreateFPExt(Arg, OutputTy);
2063 if (llvm::Type* AdjTy =
2064 getTargetHooks().adjustInlineAsmType(*this, InputConstraint,
2066 Arg = Builder.CreateBitCast(Arg, AdjTy);
2068 CGM.getDiags().Report(S.getAsmLoc(), diag::err_asm_invalid_type_in_input)
2069 << InputExpr->getType() << InputConstraint;
2071 ArgTypes.push_back(Arg->getType());
2072 Args.push_back(Arg);
2073 Constraints += InputConstraint;
2076 // Append the "input" part of inout constraints last.
2077 for (unsigned i = 0, e = InOutArgs.size(); i != e; i++) {
2078 ArgTypes.push_back(InOutArgTypes[i]);
2079 Args.push_back(InOutArgs[i]);
2081 Constraints += InOutConstraints;
2084 for (unsigned i = 0, e = S.getNumClobbers(); i != e; i++) {
2085 StringRef Clobber = S.getClobber(i);
2087 if (Clobber == "memory")
2088 ReadOnly = ReadNone = false;
2089 else if (Clobber != "cc")
2090 Clobber = getTarget().getNormalizedGCCRegisterName(Clobber);
2092 if (!Constraints.empty())
2095 Constraints += "~{";
2096 Constraints += Clobber;
2100 // Add machine specific clobbers
2101 std::string MachineClobbers = getTarget().getClobbers();
2102 if (!MachineClobbers.empty()) {
2103 if (!Constraints.empty())
2105 Constraints += MachineClobbers;
2108 llvm::Type *ResultType;
2109 if (ResultRegTypes.empty())
2110 ResultType = VoidTy;
2111 else if (ResultRegTypes.size() == 1)
2112 ResultType = ResultRegTypes[0];
2114 ResultType = llvm::StructType::get(getLLVMContext(), ResultRegTypes);
2116 llvm::FunctionType *FTy =
2117 llvm::FunctionType::get(ResultType, ArgTypes, false);
2119 bool HasSideEffect = S.isVolatile() || S.getNumOutputs() == 0;
2120 llvm::InlineAsm::AsmDialect AsmDialect = isa<MSAsmStmt>(&S) ?
2121 llvm::InlineAsm::AD_Intel : llvm::InlineAsm::AD_ATT;
2122 llvm::InlineAsm *IA =
2123 llvm::InlineAsm::get(FTy, AsmString, Constraints, HasSideEffect,
2124 /* IsAlignStack */ false, AsmDialect);
2125 llvm::CallInst *Result = Builder.CreateCall(IA, Args);
2126 Result->addAttribute(llvm::AttributeSet::FunctionIndex,
2127 llvm::Attribute::NoUnwind);
2129 // Attach readnone and readonly attributes.
2130 if (!HasSideEffect) {
2132 Result->addAttribute(llvm::AttributeSet::FunctionIndex,
2133 llvm::Attribute::ReadNone);
2135 Result->addAttribute(llvm::AttributeSet::FunctionIndex,
2136 llvm::Attribute::ReadOnly);
2139 // Slap the source location of the inline asm into a !srcloc metadata on the
2141 if (const GCCAsmStmt *gccAsmStmt = dyn_cast<GCCAsmStmt>(&S)) {
2142 Result->setMetadata("srcloc", getAsmSrcLocInfo(gccAsmStmt->getAsmString(),
2145 // At least put the line number on MS inline asm blobs.
2146 auto Loc = llvm::ConstantInt::get(Int32Ty, S.getAsmLoc().getRawEncoding());
2147 Result->setMetadata("srcloc",
2148 llvm::MDNode::get(getLLVMContext(),
2149 llvm::ConstantAsMetadata::get(Loc)));
2152 if (getLangOpts().CUDA && getLangOpts().CUDAIsDevice) {
2153 // Conservatively, mark all inline asm blocks in CUDA as convergent
2154 // (meaning, they may call an intrinsically convergent op, such as bar.sync,
2155 // and so can't have certain optimizations applied around them).
2156 Result->addAttribute(llvm::AttributeSet::FunctionIndex,
2157 llvm::Attribute::Convergent);
2160 // Extract all of the register value results from the asm.
2161 std::vector<llvm::Value*> RegResults;
2162 if (ResultRegTypes.size() == 1) {
2163 RegResults.push_back(Result);
2165 for (unsigned i = 0, e = ResultRegTypes.size(); i != e; ++i) {
2166 llvm::Value *Tmp = Builder.CreateExtractValue(Result, i, "asmresult");
2167 RegResults.push_back(Tmp);
2171 assert(RegResults.size() == ResultRegTypes.size());
2172 assert(RegResults.size() == ResultTruncRegTypes.size());
2173 assert(RegResults.size() == ResultRegDests.size());
2174 for (unsigned i = 0, e = RegResults.size(); i != e; ++i) {
2175 llvm::Value *Tmp = RegResults[i];
2177 // If the result type of the LLVM IR asm doesn't match the result type of
2178 // the expression, do the conversion.
2179 if (ResultRegTypes[i] != ResultTruncRegTypes[i]) {
2180 llvm::Type *TruncTy = ResultTruncRegTypes[i];
2182 // Truncate the integer result to the right size, note that TruncTy can be
2184 if (TruncTy->isFloatingPointTy())
2185 Tmp = Builder.CreateFPTrunc(Tmp, TruncTy);
2186 else if (TruncTy->isPointerTy() && Tmp->getType()->isIntegerTy()) {
2187 uint64_t ResSize = CGM.getDataLayout().getTypeSizeInBits(TruncTy);
2188 Tmp = Builder.CreateTrunc(Tmp,
2189 llvm::IntegerType::get(getLLVMContext(), (unsigned)ResSize));
2190 Tmp = Builder.CreateIntToPtr(Tmp, TruncTy);
2191 } else if (Tmp->getType()->isPointerTy() && TruncTy->isIntegerTy()) {
2192 uint64_t TmpSize =CGM.getDataLayout().getTypeSizeInBits(Tmp->getType());
2193 Tmp = Builder.CreatePtrToInt(Tmp,
2194 llvm::IntegerType::get(getLLVMContext(), (unsigned)TmpSize));
2195 Tmp = Builder.CreateTrunc(Tmp, TruncTy);
2196 } else if (TruncTy->isIntegerTy()) {
2197 Tmp = Builder.CreateTrunc(Tmp, TruncTy);
2198 } else if (TruncTy->isVectorTy()) {
2199 Tmp = Builder.CreateBitCast(Tmp, TruncTy);
2203 EmitStoreThroughLValue(RValue::get(Tmp), ResultRegDests[i]);
2207 LValue CodeGenFunction::InitCapturedStruct(const CapturedStmt &S) {
2208 const RecordDecl *RD = S.getCapturedRecordDecl();
2209 QualType RecordTy = getContext().getRecordType(RD);
2211 // Initialize the captured struct.
2213 MakeAddrLValue(CreateMemTemp(RecordTy, "agg.captured"), RecordTy);
2215 RecordDecl::field_iterator CurField = RD->field_begin();
2216 for (CapturedStmt::const_capture_init_iterator I = S.capture_init_begin(),
2217 E = S.capture_init_end();
2218 I != E; ++I, ++CurField) {
2219 LValue LV = EmitLValueForFieldInitialization(SlotLV, *CurField);
2220 if (CurField->hasCapturedVLAType()) {
2221 auto VAT = CurField->getCapturedVLAType();
2222 EmitStoreThroughLValue(RValue::get(VLASizeMap[VAT->getSizeExpr()]), LV);
2224 EmitInitializerForField(*CurField, LV, *I);
2231 /// Generate an outlined function for the body of a CapturedStmt, store any
2232 /// captured variables into the captured struct, and call the outlined function.
2234 CodeGenFunction::EmitCapturedStmt(const CapturedStmt &S, CapturedRegionKind K) {
2235 LValue CapStruct = InitCapturedStruct(S);
2237 // Emit the CapturedDecl
2238 CodeGenFunction CGF(CGM, true);
2239 CGCapturedStmtRAII CapInfoRAII(CGF, new CGCapturedStmtInfo(S, K));
2240 llvm::Function *F = CGF.GenerateCapturedStmtFunction(S);
2241 delete CGF.CapturedStmtInfo;
2243 // Emit call to the helper function.
2244 EmitCallOrInvoke(F, CapStruct.getPointer());
2249 Address CodeGenFunction::GenerateCapturedStmtArgument(const CapturedStmt &S) {
2250 LValue CapStruct = InitCapturedStruct(S);
2251 return CapStruct.getAddress();
2254 /// Creates the outlined function for a CapturedStmt.
2256 CodeGenFunction::GenerateCapturedStmtFunction(const CapturedStmt &S) {
2257 assert(CapturedStmtInfo &&
2258 "CapturedStmtInfo should be set when generating the captured function");
2259 const CapturedDecl *CD = S.getCapturedDecl();
2260 const RecordDecl *RD = S.getCapturedRecordDecl();
2261 SourceLocation Loc = S.getLocStart();
2262 assert(CD->hasBody() && "missing CapturedDecl body");
2264 // Build the argument list.
2265 ASTContext &Ctx = CGM.getContext();
2266 FunctionArgList Args;
2267 Args.append(CD->param_begin(), CD->param_end());
2269 // Create the function declaration.
2270 FunctionType::ExtInfo ExtInfo;
2271 const CGFunctionInfo &FuncInfo =
2272 CGM.getTypes().arrangeBuiltinFunctionDeclaration(Ctx.VoidTy, Args);
2273 llvm::FunctionType *FuncLLVMTy = CGM.getTypes().GetFunctionType(FuncInfo);
2276 llvm::Function::Create(FuncLLVMTy, llvm::GlobalValue::InternalLinkage,
2277 CapturedStmtInfo->getHelperName(), &CGM.getModule());
2278 CGM.SetInternalFunctionAttributes(CD, F, FuncInfo);
2279 if (CD->isNothrow())
2280 F->addFnAttr(llvm::Attribute::NoUnwind);
2282 // Generate the function.
2283 StartFunction(CD, Ctx.VoidTy, F, FuncInfo, Args,
2285 CD->getBody()->getLocStart());
2286 // Set the context parameter in CapturedStmtInfo.
2287 Address DeclPtr = GetAddrOfLocalVar(CD->getContextParam());
2288 CapturedStmtInfo->setContextValue(Builder.CreateLoad(DeclPtr));
2290 // Initialize variable-length arrays.
2291 LValue Base = MakeNaturalAlignAddrLValue(CapturedStmtInfo->getContextValue(),
2292 Ctx.getTagDeclType(RD));
2293 for (auto *FD : RD->fields()) {
2294 if (FD->hasCapturedVLAType()) {
2295 auto *ExprArg = EmitLoadOfLValue(EmitLValueForField(Base, FD),
2296 S.getLocStart()).getScalarVal();
2297 auto VAT = FD->getCapturedVLAType();
2298 VLASizeMap[VAT->getSizeExpr()] = ExprArg;
2302 // If 'this' is captured, load it into CXXThisValue.
2303 if (CapturedStmtInfo->isCXXThisExprCaptured()) {
2304 FieldDecl *FD = CapturedStmtInfo->getThisFieldDecl();
2305 LValue ThisLValue = EmitLValueForField(Base, FD);
2306 CXXThisValue = EmitLoadOfLValue(ThisLValue, Loc).getScalarVal();
2309 PGO.assignRegionCounters(GlobalDecl(CD), F);
2310 CapturedStmtInfo->EmitBody(*this, CD->getBody());
2311 FinishFunction(CD->getBodyRBrace());