cpluspluscheck vs ICU

Hi,

I was about to propose adding headerscheck / cpluspluscheck to the CI file so
that cfbot can catch future issues. Unfortunately running cpluspluscheck with
ICU enabled is, um, not fun: There's 30k lines of error output.

/home/andres/src/postgresql/src/tools/pginclude/cpluspluscheck /home/andres/src/postgresql /home/andres/build/postgres/dev-assert/vpath
In file included from /usr/include/c++/12/bits/stl_algobase.h:60,
from /usr/include/c++/12/memory:63,
from /usr/include/unicode/localpointer.h:45,
from /usr/include/unicode/unorm2.h:34,
from /usr/include/unicode/unorm.h:25,
from /usr/include/unicode/ucol.h:17,
from /home/andres/src/postgresql/src/include/utils/pg_locale.h:19,
from /home/andres/src/postgresql/src/include/tsearch/ts_locale.h:20,
from /tmp/cpluspluscheck.H59Y6V/test.cpp:3:
/usr/include/c++/12/bits/functexcept.h:101:3: error: conflicting declaration of C function ‘void std::__throw_ios_failure(const char*, int)’
101 | __throw_ios_failure(const char*, int) __attribute__((__noreturn__));
| ^~~~~~~~~~~~~~~~~~~
/usr/include/c++/12/bits/functexcept.h:98:3: note: previous declaration ‘void std::__throw_ios_failure(const char*)’
98 | __throw_ios_failure(const char*) __attribute__((__noreturn__));
| ^~~~~~~~~~~~~~~~~~~
In file included from /usr/include/c++/12/bits/stl_algobase.h:63:
/usr/include/c++/12/ext/numeric_traits.h:50:3: error: template with C linkage
50 | template<typename _Tp>
| ^~~~~~~~
/tmp/cpluspluscheck.H59Y6V/test.cpp:1:1: note: ‘extern "C"’ linkage started here
1 | extern "C" {
| ^~~~~~~~~~
...

with one warning for each declaration in numeric_traits.h, I think.

So, there's two questions:
1) How can we prevent this problem when ICU support is enabled?
2) Can we prevent such absurdly long error output?

For 2), perhaps we should just specify EXTRAFLAGS=-fmax-errors=10 in the
cpluspluscheck invocation, or add it in cpluspluscheck itself?

For 1), I don't immediately see a minimal solution other than ignoring it in
cpluspluscheck, similar to pg_trace.h/probes.h.

A different / complimentary approach could be to add -Wc++-compat to the
headerscheck invocation. Both gcc and clang understand that.

But neither of these really gets to the heart of the problem. There's still no
way for C++ code to include pg_locale.h correctly. And in contrast to
pg_trace.h/probes.h pg_locale.h is somewhat important.

This isn't a new problem, afaics.

Perhaps we should strive to remove the use of ICU headers from within our
headers? The members of pg_locale are just pointers and could thus be void *,
and HAVE_UCOL_STRCOLLUTF8 could be computed at configure time or such.

Greetings,

Andres Freund