Re: cpluspluscheck vs ICU again

John Naylor <johncnaylorls(at)gmail(dot)com> writes:
> On Mon, Jul 7, 2025 at 11:06 PM Tom Lane <tgl(at)sss(dot)pgh(dot)pa(dot)us> wrote:
>> John Naylor <johncnaylorls(at)gmail(dot)com> writes:
>>> I see that now. If extensions follow the practice of including system
>>> headers before Postgres headers, it should be fine. I've attached v2
>>> which removes the useless #undef and drafts an explanatory commit
>>> message.

>> Works for me.

> Pushed.

Just when you thought it was safe to go back in the water ...
I tried cpluspluscheck with late-model libicu (76.1 on Fedora 42)
and darned if it didn't blow up in exactly the same way.
Investigation reveals that they've split the "U_SHOW_CPLUSPLUS_API"
symbol into two, and now if you really really don't want any C++
stuff you need to also set "U_SHOW_CPLUSPLUS_HEADER_API" to zero.

I've confirmed that this re-silences the failures:

diff --git a/src/include/utils/pg_locale.h b/src/include/utils/pg_locale.h
index 931f5b3b880..2b072cafb4d 100644
--- a/src/include/utils/pg_locale.h
+++ b/src/include/utils/pg_locale.h
@@ -18,6 +18,8 @@
/* only include the C APIs, to avoid errors in cpluspluscheck */
#undef U_SHOW_CPLUSPLUS_API
#define U_SHOW_CPLUSPLUS_API 0
+#undef U_SHOW_CPLUSPLUS_HEADER_API
+#define U_SHOW_CPLUSPLUS_HEADER_API 0
#include <unicode/ucol.h>
#endif

This shouldn't complicate extensions' lives any further than
before; the rule still is "include ICU headers first
if you want their C++ symbols".

BTW, I see that you applied ed26c4e25 only to master, but don't
we want to back-patch? cpluspluscheck is not just an exercise in a
vacuum, it's to ensure that C++-coded extensions don't have trouble
with our headers.

regards, tom lane