[patch] Implementing -Wunused-local-typedefs

[Nico Weber]

(Forgot to say: This is also PR18265.)

On Sat, May 3, 2014 at 6:07 PM, Nico Weber <thakis at chromium.org> wrote:
> Hi,
>
> gcc has a warning -Wunused-local-typedefs that warns on unused local
> typedefs. Inspired by r207870, I thought it'd be nice if clang had
> this warning too. This warning requires the following three things:
>
>
> 1.) A bit to track if a typedef has been referenced.
>
> Decl already has isUsed and isReferenced, but they don't seem to be
> used for TypedefDecls, so I'm just using isReferenced on TypedefDecls
> for this.
>
>
> 2.) Setting that bit on typedefs that are used.
>
> The three strategies I can think of:
> a.) Do this in Sema::DiagnoseUseOfDecl(), this seems to be already
> called for decls all over the place, and an "if isa<TypedefDecl>(D)
> D->setReferenced()" seems to do the trick.
> b.) Do this in LookupName
> c.) Do this explicitly in the places where it's actually necessary.
>
> The attached patch goes with the last approach. The three places I
> found where it's needed are Sema::getTypeName(), Sema::ClassifyName(),
> and Sema::LookupInlineAsmField(). The first two are called by the
> parser for almost everything, while the third is called for references
> to structs from MS inline assembly. That last one is a bit scary as it
> means it's possible that there are other places this is needed that I
> missed.
>
>
> 3.) A way to check all typedefs in a scope when that scope ends.
>
> I added this to Sema::ActOnPopScope() which is where the
> unused-variable and unused-label warnings are created. This works
> well, but to catch the unused local typedef in
>
>   {
>     struct A {
>       struct B { typedef int SoDeep; };
>     };
>   }
>
> it means that that code now needs to recurse into all RecordDecl in a
> scope, which it didn't need to do previously.
>
>
> Let me know how badly I've chosen in each instance :-)
>
> Thanks,
> Nico
>
> ps: If this goes in, a follow-up question is if this should warn on
> unused C++11 type aliases too – it'd just require
> s/TypedefDecl/TypedefNameDecl/ in two places in the patch, so it
> should be an easy follow-up.