Andres Freund <andres@anarazel.de> writes:
> On 2018-12-01 17:28:54 -0500, Tom Lane wrote:
>> I guess I wasn't precise enough: I meant add -msse2 if on x86 and
>> compiler doesn't take -fexcess-precision=standard.
> Hm, I still don't like that: It'd silently bump the minimum required
> architecture. Like in the case of x86-32 freebsd, which doesn't require
> sse2 and uses clang, the maintainer wouldn't notice that they would have
> to switch to gcc to continue supporting their baseline.
Yeah, that's a fair point.
> It's not like there's that many people compiling for such platforms with
> insufficient compiler support, so forcing them to specify -msse2 if
> that's the desired escape hatch doesn't sound terrible to me.
I spent some time over the weekend investigating this, including going
so far as to build gcc 3.4.6 from source to try to reproduce the report
I had off-list. No luck; passed all regression tests just fine. In
fact, I was only able to produce any test failures at all with clang.
My conclusion is that it's really hard to get gcc to exhibit this type of
problem, even when using old versions that don't have -fexcess-precision.
I am now thinking that the right thing to do is nothing. If we insist
on having -msse2 or -fexcess-precision, the net effect is going to be
to break the build on more old platforms than we fix anything on. Yes,
updating to a newer compiler or moving your hardware baseline are both
recommendable things to do, but why should we force that on people who
aren't seeing a problem? Especially in stable back branches?
(In HEAD this should be moot anyway, given that C99 implies
-fexcess-precision=standard.)
regards, tom lane
