================
@@ -630,9 +632,17 @@ llvm::Expected<Expr *>
Interpreter::convertExprToValue(Expr *E) {
// It could fail, like printing an array type in C. (not supported)
if (SetValueE.isInvalid())
- return E;
-
- return SetValueE.get();
+ return FullExpr;
+
+ // The temporaries of E must be destroyed at the end of the statement.
+ // Without the cleanups, CodeGen destroys them at the end of the function
+ // running the top-level statements, which it finishes after emitting the
+ // deferred declarations: their destructors would not be emitted.
+ Expr *Result = SetValueE.get();
+ if (EWC && !isa<ExprWithCleanups>(Result))
+ Result = ExprWithCleanups::Create(
+ Ctx, Result, EWC->cleanupsHaveSideEffects(), EWC->getObjects());
----------------
guitargeek wrote:
Good question! I checked: the placement new does kick in. For a printed object
of class type, e.g. `h(S())` with `R h(S)`, the `R` is constructed once
directly into the storage of the `Value`, without any copy. But that only takes
care of the printed object itself; the temporaries of its initializer (here the
`S()` argument) still need a full-expression to be destroyed at. Without this
patch that case hits the same `moveLazyEmissionStates()` assertion as `f(S())`.
The `ActOnFinishFullExpr()` on the placement new path doesn't help, because
`convertExprToValue()` runs in `HandleTopLevelDecl()` after the statement was
already finished, so Sema has no pending cleanups left to attach.
And for the other case in the test, `f(S())` prints an `int`, so no placement
new is involved at all.
This is also how cling did it:
`ValueExtractionSynthesizer::SynthesizeSVRInit()` strips the `ExprWithCleanups`
and puts the synthesized call (including the placement new) back into it: "In
case of ExprWithCleanups we need to extend its 'scope' to the call." I've added
the placement new case to the test.
https://github.com/llvm/llvm-project/pull/229675
_______________________________________________
cfe-commits mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
