[PATCH] D82940: [ASTReader][ASTWriter][PCH][Modules] Fix (de)serialization of SwitchCases

[Bruno Ricci via Phabricator via cfe-commits]

riccibruno added a comment.

In D82940#2125939 <https://reviews.llvm.org/D82940#2125939>, @martong wrote:

> In D82940#2125183 <https://reviews.llvm.org/D82940#2125183>, @riccibruno wrote:
>
> > > I'm not certain I understand, but I'm removing my LG while we figure it out. Can you expound a bit further?
> >
> > Take a look at the AST dump of a small modification of the reduced test case which doesn't trigger the assertion:
> >
> >   void should_not_crash_with_switch_in_lambda() {
> >     switch (1)
> >     default:;
> >     auto f = [] {
> >       switch (1)
> >       default:;
> >     };
> >   }
> >
> >
> > Before the serialization of `LambdaExpr` for the `LambdaExpr *E`, we have `E->getBody() == E->getCallOperator()->getBody()`.
> >  After serialization this is not true anymore because we are writing and reading the body directly when visiting the `LambdaExpr`
> >  (the captures have the same problem).
>
>
> The same is true in case of the simplest lambda example:
>
>   void foo() {   // line 11
>       []{};
>   }
>
>
> The AST that we got from the Parser has only one CompoundStmt node, but referenced in the CXXMethodDecl and added as a direct child to LambdaExpr:
>
>   `-LambdaExpr 0x5643235372c8 <line:12:5, col:8>
>     |-CXXRecordDecl 0x564323536c98 <col:5> col:5 implicit class definition
>     | |-CXXMethodDecl 0x564323536dd8 <col:6, col:8> col:5 constexpr operator() 'auto () const -> void' inline
>     | | `-CompoundStmt 0x564323536e88 <col:7, col:8>
>     `-CompoundStmt 0x564323536e88 <col:7, col:8>
>   
>
> After serialization and deserialization, the CompoundStmt is just copied, notice the different pointer values:
>
>   `-LambdaExpr 0x55a49f35f830 <line:12:5, col:8>
>     |-CXXRecordDecl 0x55a49f35f860 <col:5> col:5 imported implicit <undeserialized declarations> class definition
>     | |-CXXMethodDecl 0x55a49f35fac8 <col:6, col:8> col:5 imported constexpr operator() 'auto () const -> void' inline
>     | | `-CompoundStmt 0x55a49f360110 <col:7, col:8>
>     `-CompoundStmt 0x55a49f35f820 <col:7, col:8>
>   
>
> This suggests that we should prepare the ASTReader/ASTWriter to be able to handle references to Stmts similarly as it can already handle references to Decls (see `ASTWriter::DeclIDs`).
>  Does this sound as a good direction, should I move forward this way?


Yes, I think that my patch just traded-off a bug for another bug. What about the following instead:

Initially zero the `Stmt *` for the body, then when the body is needed (either with `getBody` or with `children`)
populate it if needed from the call operator. This is what was done before with `getBody`.

This preserve the benefit of having a lazily deserialized body while being a simple fix.

Thoughts? Am I missing something?

If not I can cook-up a patch either tonight or tomorrow.


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D82940/new/

https://reviews.llvm.org/D82940