Re: expected baseline changes for forky => proposal no more FPU and 80bits float/double on i386

Stephen Kitt <[email protected]> Sat, 15 Aug 2026 15:43:24 +0200

Newsgroups	gmane.linux.debian.devel.general
Message-ID	<[email protected]>

Hi Adrian,

On Sat, 15 Aug 2026 15:41:05 +0300, Adrian Bunk <[email protected]> wrote:
> On Sat, Aug 15, 2026 at 02:22:24PM +0200, Stephen Kitt wrote:
> > On Sat, 15 Aug 2026 13:20:20 +0300, Adrian Bunk <[email protected]> wrote:  
> > > On Fri, Aug 14, 2026 at 09:23:58PM +0900, Simon Richter wrote:  
> > > > The floating point results change, though, so the behaviour is
> > > > different from what the software was tested with, especially the
> > > > rounding behaviour is a lot more aggressive.
> > > >...    
> > > 
> > > you can compile software relying on the old behaviour with
> > > -mfpmath=387.  
> > 
> > The typical use-case here though is that you have an existing i386 binary,
> > which you can’t recompile, running with distribution-provided i386
> > libraries (or a mixture of i386 libraries provided alongside the binary
> > and distribution-provided libraries), and it’s the behaviour of the
> > libraries that may have changed.  
> 
> this whole line of arguing comes across as asking to continue to do a 
> lot of work with a mostly theoretical argument.

My point is only that if we end up in a situation where i386 software people
care about (which, as you point out elsewhere, isn’t part of the
distribution) is broken, recompiling would most likely not be an option.

If i386 ends up only being useful for running such software (which is not yet
the case), and that ends up broken, or untestable in any reasonable sense, I
would say that argues for dropping i386 entirely.

> glibc already uses SSE when available, and there aren't that many other 
> relevant libraries without any ABI break in decades.
> 
> If there are a handful relevant libraries, continuing to use x87 for 
> these would be easily possible.

+1

> Does anyone have an example of actual breakage?

I don’t. In any case most dynamically-linked, i386-only games that might be
broken are already broken for other reasons and need old libraries to run.

Regards,

Stephen

signature.asc (application/pgp-signature, 833 B)