Bug#930935: webkit2gtk: Baseline violation on i386

[Adrian Bunk]

On Mon, Jun 24, 2019 at 10:29:56AM +0200, Alberto Garcia wrote:
> On Mon, Jun 24, 2019 at 12:34:26AM +0300, Adrian Bunk wrote:
> 
> > > Building with SSE2 is not a decision made by the Debian
> > > maintainers.  Upstream WebKit dropped support for non-SSE2
> > > processors and the code assumes that SSE2 is always available:
> > > 
> > >    https://bugs.webkit.org/show_bug.cgi?id=194853
> > 
> > "as part of Igalia's work on 32bit support"
> > "there were no negative comments"
> > 
> > Someone should have told upstream at Igalia that non-SSE is
> > supported by the Debian i386 port.
> 
> This was actually discussed, IIRC the concusion was that this was
> a reasonable choice because (a) SSE2 is almost 20 years old,

Hardware without SSE2 was still pretty common 15 years ago
(e.g. Athlon XP), and surprisingly many people still have
such ancient hardware running somewhere.

The non-SSE AMD Geode was quite popular in embedded devices,
and is still in production.

There were also many people unhappy when stretch dropped i586 support,
since this dropped support for hardware they were still using.

> (b) supporting non-SSE2 CPUs was increasing the maintaince burden for
> upstream, and (c) other browsers had already followed the same path
> for a few yeas.
> 
> > > And so did Firefox:
> > > 
> > >    https://support.mozilla.org/en-US/kb/your-hardware-no-longer-supported
> > 
> > The Debian maintainers have patched both Rust and Firefox to make
> > them work on non-SSE i386 despite upstream no longer supporting it.
> > 
> > There were plenty of bug reports when an SSE-using Firefox recently
> > slipped through - and in general based on the numbers of bugs
> > submitted non-SSE i386 has more users than half our release
> > architectures combined.
> 
> Ok, that's interesting information.
> 
> Let's try to move forward then. I see three possibilities:
> 
> 1) Detect SSE2 at runtime. This doesn't seem to be possible at this
>    point, or at least not trivially so. Although we still have the
>    JavaScriptCoreUseJIT variable, it's unclear whether this would be
>    enough. I need to test it.
> 
> 2) Build with SSE2 completely disabled (using WTF_CPU_UNKNOWN, or
>    somethig else, I'm still discussing this with the team). This would
>    probably solve the problem but would make all i386 builds much
>    slower (in already slow machines) so I suppose many users would
>    also complain.

Users on such hardware are usually aware that their hardware is slow.

The slowest i386 hardware is the one without SSE, and people are running 
Firefox on such hardware and submit bugs when this breaks.

> 3) Make two builds for i386: one with SSE2 and one without, and have
>    two separate packages. Maybe it's too late for this?

I'm adding debian-release to Cc for that.

>    In any case
>    I think this would only affect the libjavascriptcore package, the
>    libwebkit2gtk build should be valid for all CPUs. So we'd have
>    libjavascriptcoregtk-4.0-18 and libjavascriptcoregtk-4.0-18-sse2 or
>    something like that.

The old hardware capabilities (see ld.so(8)) could be used for providing 
these in one package, with runtime detection done by the dynamic linker
for optimized libraries placed in /usr/lib/sse2/.

> Berto

cu
Adrian

-- 

       "Is there not promise of rain?" Ling Tan asked suddenly out
        of the darkness. There had been need of rain for many days.
       "Only a promise," Lao Er said.
                                       Pearl S. Buck - Dragon Seed