On 2018-09-05 14:08:11 -0400, Tom Lane wrote:
> Andres Freund <andres@anarazel.de> writes:
> > On 2018-09-05 10:05:26 -0400, Tom Lane wrote:
> >> One thought is that maybe we should provide a way to override this,
> >> in case somebody really can't or doesn't want to use -msse2, and
> >> is willing to put up with platform-dependent float behavior.
>
> > IDK, people that are capable of making that decision can just hack
> > configure.
>
> "Just hacking configure" is a pretty high bar for most folk.
Sure. But being able to accurately judge whether you're ok that math
behaviour doesn't work as documented seems like a high bar too. And
this'd only be relevant if they insist on using clang rather than gcc.
> If you wanted to argue that the set of people who still want to run PG
> on pre-SSE2 hardware is the empty set, that'd be an easier sell
> really.
I think it's an empty set, yea, especially with clang.
> But what I'm really concerned about here, given that we're apparently
> talking about an ABI change, is that someone might want to run on a
> platform where the libraries insist on the x87 way.
> I'm actually a bit surprised to hear that you can just randomly add
> -msse2 on BSDen. Do they ship separate copies of libc et al to make
> that work?
I don't think we're talking about an ABI change - with -msse2 the
calling conventions are unchanged, the stuff just gets moved out of the
x87 registers into SSE ones / xmm for the actual math. Which in turn
solves our "80bit register" problem. There really shouldn't be any need
for libc6 itself to care.
Greetings,
Andres Freund