I wrote:
> That's confusing because it implies that -fno-common is the default,
> which it evidently is not. But anyway, my diagnosis is that you're
> breaking something about the linker's behavior with that switch.
Oh! Looking closer, the core dump happens here:
const printTextFormat pg_utf8format;
printTextFormat *popt = (printTextFormat *) &pg_utf8format;
--> popt->name = "unicode";
So apparently, the relevant property of "-fno-common" is that it
causes "const" variables to actually get placed in read-only data.
I think this code is new in 9.5, which'd explain why you didn't see
the failure with older PG versions. It's surely busted though.
I shall get rid of the const-ness, as well as the lame casting away
of it, and I think I will also go make buildfarm member longfin use
"-fno-common". It is truly sad that we apparently have no test
machine that enforces that const means const ...
regards, tom lane