Fedora 29 Proposal "i686 Is For x86-64" Would Allow More Optimizations, Require SSE2

[grok]

You do that in 32-bit mode on a 64-bit processor? And you care about performance? Odd.

It would be interesting if x32-abi were better supported. This is x86-64 with 32-bit pointers. Should be quite useful on small systems (with perhaps 2G of RAM or less -- like the netbook on which I am typing).

x32 has to be a pure x32 system, so no proprietary software like Skype or drivers, no flash player (ok this is not important anymore but used to be even 5+ years ago), no Wine to play some silly Warcraft III. Biggest benefit should be lower CPU load from SSL/TLS.
And so, the niche for x32 was to be server VMs. But as we talk of dropping even i686, this should explain the lack of interest.

In other news Mint is fairly conservative in not dropping stuff, there are Mint 19 Mate edition and Mint 19 Xfce out ("BETA" but more a kind of single RC version) with 32bit i686.

[carewolf]

You do that in 32-bit mode on a 64-bit processor? And you care about performance? Odd.

What are you blathering about?

I specifically wondered why not set the new limit to SSE4 and also raise it for x64.

[Hugh]

Actually they are way more significant than SSE2. SSSE3 has the all important shuffle instruction and SSE4.1 as integer multiply, it more than 4 doubles the places a compiler can autovectorize in most codebases I have worked on.

Edit: Though that might be because I do more integer math than FP, it is not nearly as important for FP performance.

You do that in 32-bit mode on a 64-bit processor? And you care about performance? Odd.

It would be interesting if x32-abi were better supported. This is x86-64 with 32-bit pointers. Should be quite useful on small systems (with perhaps 2G of RAM or less -- like the netbook on which I am typing).

[carewolf]

The improvements in >SSE2 aren't as significant in general.

Actually they are way more significant than SSE2. SSSE3 has the all important shuffle instruction and SSE4.1 as integer multiply, it more than 4 doubles the places a compiler can autovectorize in most codebases I have worked on.

Edit: Though that might be because I do more integer math than FP, it is not nearly as important for FP performance.

[mczak]

Btw, why not set the new limit to SSSE3/SSE4 though? That would still cover all CPUs made the last 10 years

Not quite. While it's true that since the 2nd Gen Core 2 Duo (45nm Wolfdale/Penryn) all intel "big" cpus from a hw point of view can do sse41, pentiums/celerons had that disabled up to and including Clarkdale (Sandy Bridge introduced AVX, so from that point on celerons/pentiums have sse41 enabled and avx disabled), and Clarkdale is a 2010 cpu. Also, the small atom cores can only do it since Silvermont (late 2013), and while Bonnell/Saltwell definitely aren't stellar cpus there's no way you can drop support for them already. Although at least ssse3 is supported indeed with all intel cpus since 2006 or so (first gen Core 2, all atoms).
But it's worse for AMD. Later K8 models can do SSE3, but not SSSE3 (and no SSE41 support of course neither). K10 (that includes Phenom II) can't do it neither. You need Bulldozer for it (or Bobcat). So that would be not even 7 years (not to mention most people preferred K8/K10-based chips over BDs...).

[You-]

This could be awkward if intel release their next gen CPUs without SSE2 support.

AFAIK that was on the board to stop their designs performance being limited due to support for legacy instructions

[caligula]

Btw, why not set the new limit to SSSE3/SSE4 though? That would still cover all CPUs made the last 10 years, and up the x64 minimum at the same time.

The improvements in >SSE2 aren't as significant in general.

[carewolf]

Btw, why not set the new limit to SSSE3/SSE4 though? That would still cover all CPUs made the last 10 years, and up the x64 minimum at the same time.

[carewolf]

I am all for deprecating all CPUs before SS4. For 32-bit multilib call it i786

More like i868. i686 was Pentium Pro and Pentium 2. This is K8/P4 tech. so two generations later.

[gnufreex]

I am all for deprecating all CPUs before SS4. For 32-bit multilib call it i786