On Wed, Mar 18, 2026 at 04:21:49PM +0100, Tomasz Kaminski wrote:
> On Wed, Mar 18, 2026 at 4:16 PM Jakub Jelinek <[email protected]> wrote:
>
> > On Wed, Mar 18, 2026 at 03:56:17PM +0100, Tomasz Kaminski wrote:
> > > > Here is a new version of the plugin.
> > > > As requested, it doesn't diagnose any of the headers outside of
> > > > <bits/stdc++.h>, for -std=c++26 tests with also -freflection
> > -fcontracts
> > > > for now, contains a self test for the plugin to make sure it diagnoses
> > > > stuff it should diagnose and doesn't what it shouldn't, walks template
> > > > specializations, walks non-template functions inside of templates,
> > > > adds GC support to the plugin, whitelists submdspan_mapping manually
> > > > for C++26,
> > >
> > > Why exactly submdspan_mapping needs a special treatment, it should
> > > be no different from other ADL-only calls in the library, like begin/end?
> >
> > The plugin automatically whitelists non-uglified names of declarations
> > that it
> > sees in the std namespace and nested namespaces and in classes/class
> > templates
> > thereof except for declarations inside of non-inline _ prefixed namespaces,
> > trying to roughly reconstruct what is in
> > https://eel.is/c++draft/libraryindex .
> > The whitelisted ones are reported by the plugin too, but with a different
> > wording which doesn't make the testcase to fail.
> > And begin/end are whitelisted that way, submdspan_mapping is not.
> > E.g.
> > libstdc++-v3/include/bits/unicode.h:160:7: note: non-uglified whitelisted
> > name 'begin'
> > libstdc++-v3/include/bits/unicode.h:738:22: note: non-uglified whitelisted
> > name 'end'
> >
> We have a hidden friend declaration of submdspan_mapping both in the
> standard
> and in our implementation for each layout. For example in
> libstdc++/include/std/mdspan:2780.
Indeed, the plugin doesn't report anything on
// Test that uglification_plugin.cc works properly.
// { dg-options "-O0 -std=c++20" }
// Pretend this is in a libstdc++-v3 header.
#line 6 "gcc/libstdc++-v3/include/bits/universe.h"
namespace std {
template <int>
struct foo {
template <int>
friend int bar (int x) {
int y = x;
return y;
}
};
}
(or if either of those template lines are commented out but not
both).
I the plugin should walk even those, will need to find out where those
templated friends hang out when not instantiated yet.
Jakub
