Since we didn't require SSE2 in 32-bit builds until this point, were floating-point results rounded unpredictably in those builds until now?
On Fri, May 13, 2016 at 1:42 PM, Benjamin Smedberg <benja...@smedbergs.us> wrote: > I am talking about requiring SSE2. That is a larger (but still quite small) > population, but the upside of being able to turn on SSE2 optimizations by > default is an important benefit. I've discussed and confirmed this with > Firefox product management. > > So yes, the plan of record is to require SSE2 starting in Firefox 49, and I > will update the tracking bugs to reflect that. > > --BDS > > On Fri, May 6, 2016 at 1:17 PM, Gregory Szorc <g...@mozilla.com> wrote: > > > On Fri, May 6, 2016 at 9:39 AM, Benjamin Smedberg <benja...@smedbergs.us > > > > wrote: > > > >> I agree that we should drop support for non-SSE2. It mattered 7 years > ago > >> (see https://bugzilla.mozilla.org/show_bug.cgi?id=500277) but it really > >> doesn't matter now. > >> > > > > Wait - are we talking about requiring SSE or SSE2? The thread up to this > > point was talking about requiring just SSE, not SSE2. I just want to make > > sure we're on the same page since according to mhoye's post the non-SSE2 > > population is ~25x larger than the non-SSE population... > > > > > >> > >> We do need to avoid updating these users to a build that will crash, and > >> do the same "unsupported" messaging we're doing for old versions of > MacOS. > >> Gregory, will you own that? You will probably need to add CPU feature > >> detection to the update URL/params for 47, or use some kind of system > addon > >> to shunt these users off the main update path. > >> > > > > Given that 47 is in Beta, is it too late/risky to make this change on > that > > channel? Should we revert to VS2013 on Aurora/48 and make the updater > > modifications on that channel? I think this will have minimal negative > > impact, as most of the impact to changing toolchains would be on central, > > as that is where most developers and automation live. > > > > > >> > >> On Fri, May 6, 2016 at 10:10 AM, Mike Hoye <mh...@mozilla.com> wrote: > >> > >>> On 2016-05-06 12:26 AM, Gregory Szorc wrote: > >>> > >>>> FWIW, the crashes we've seen so far are from incorrectly emitted movss > >>>> instructions. This instruction is part of the original SSE instruction > >>>> set, > >>>> which was initially unveiled by Intel on the Pentium 3 in 1999 and > >>>> later by > >>>> AMD on the Duron and Athlon XP in 2000-2001. I'm not sure why we still > >>>> need > >>>> Firefox to run on processors manufactured in the 90s. > >>>> > >>> Per an IRC conversation with chutten, Firefox users on CPUs that do not > >>> support SSE are 0.015% of our user base. (compared to 0.4% for > no-SSE2). A > >>> third of those are on otherwise-unsupported configurations (pre-SP3 XP, > >>> etc), this work provides continuity of support to 0.01% of our users. > >>> > >>> - mhoye > >>> > >>> > >>> 09:59 <chutten> So, to put it clearly and precisely, of the Firefox > >>> Population in release and beta who are reporting at least base > telemetry > >>> collection on machines running supported configurations, only 0.01% > cannot > >>> definitively say they have SSE. > >>> 10:00 <chutten> (according to a 1% random sample as stored in the > >>> longitudinal dataset) > >>> > >>> _______________________________________________ > >>> dev-platform mailing list > >>> dev-platform@lists.mozilla.org > >>> https://lists.mozilla.org/listinfo/dev-platform > >>> > >> > >> > > > _______________________________________________ > dev-platform mailing list > dev-platform@lists.mozilla.org > https://lists.mozilla.org/listinfo/dev-platform > _______________________________________________ dev-platform mailing list dev-platform@lists.mozilla.org https://lists.mozilla.org/listinfo/dev-platform
