My Browser Builds (Part 5)

[UCyborg]

It does run normally on capable CPUs natively. Intel SDE slows down even CPU-Z immensely. Not sure if there are any tricks to speed it up (not a developer with proper expertise to use the tool), but I can imagine the goal of making existing compiled Windows executable execute when it normally wouldn't being complicated.

[VistaLover]

7 hours ago, Leokids123 said:

i can't find bug 151066 anywhere, can you link it to me?

That would be Mozilla's Bugzilla bug no. 151066 (archaic , opened 23 years ago + closed 22 years ago), thus: 

https://bugzilla.mozilla.org/show_bug.cgi?id=151066

Edited December 29, 2024 by VistaLover

[Leokids123]

1 hour ago, UCyborg said:

It does run normally on capable CPUs natively. Intel SDE slows down even CPU-Z immensely. Not sure if there are any tricks to speed it up (not a developer with proper expertise to use the tool), but I can imagine the goal of making existing compiled Windows executable execute when it normally wouldn't being complicated.

Have you tried to bypass the check or try to run it manually on incapable CPUs? I tried it on a Pentium 3 system running Windows 7,named VM8 on CollabVM with Firefox 115esr and it seemingly worked well,except being obviously slow and the instability,It's more bearable,even with 2GB,cause it ain't even using higher ram than expected,only when using Youtube. Even Youtube seemingly is so smooth,the drop of frames is barely noticeable,only when looking at the stats for nerds that you can see it drop frames. I know of the consequences,the instability and the possibility of it crashing or memory leak. But so far that time,i didn't see anything unstable there.

[UCyborg]

There is no check since those are compile-time optimizations, so compiler uses AVX instructions wherever it sees fit when transforming C++ code into machine code which ends up in the final executable file. They still have SSE2 build here, built by one of the members of Pale Moon forum.

[Mathwiz]

5 hours ago, UCyborg said:

Running official 64-bit AVX optimized version of Pale Moon on an old processor without AVX instruction set through Intel Software Development Emulator version 8.59 (none of the current versions seem to emulate anything here):

It launches in over 3 minutes with extensions, loads web pages, but very slowly, practically unusable.

Thought it was slow before? Run it through Intel SDE! Now that's real slowness!

Not really surprising. Every time it hits an AVX instruction a software interrupt occurs; interrupt handler has to save everything, do what the AVX instruction would've done (if you actually had an AVX processor), restore everything and return to Pale Moon. Then a few nanoseconds later, it all happens again - over and over. Only reasonable solution is to use the build of Pale Moon without AVX instructions that you discovered above.

Not really clear why MCP did this (as opposed to why they say they did it); perhaps by limiting Pale Moon to AVX processors, they're effectively limiting it to newer processors, and thus (indirectly) to faster processors that can handle the Javledygook on modern Web sites without bogging down.

[Leokids123]

25 minutes ago, UCyborg said:

There is no check since those are compile-time optimizations, so compiler uses AVX instructions wherever it sees fit when transforming C++ code into machine code which ends up in the final executable file. They still have SSE2 build here, built by one of the members of Pale Moon forum.

Well,Firefox has a check which prevents you from installing it on non-sse2 PCs. If you try to bypass it,it'll will work but it'll be unstable and might crash. Why not Pale Moon?

[Leokids123]

45 minutes ago, Mathwiz said:

Not really surprising. Every time it hits an AVX instruction a software interrupt occurs; interrupt handler has to save everything, do what the AVX instruction would've done (if you actually had an AVX processor), restore everything and return to Pale Moon. Then a few nanoseconds later, it all happens again - over and over. Only reasonable solution is to use the build of Pale Moon without AVX instructions that you discovered above.

Not really clear why MCP did this (as opposed to why they say they did it); perhaps by limiting Pale Moon to AVX processors, they're effectively limiting it to newer processors, and thus (indirectly) to faster processors that can handle the Javledygook on modern Web sites without bogging down.

What they say they did it is the reason why they did it.

 

Modern processors these days support AVX,so that'll not be a problem. The problem is when you have a Celeron,Pentium and even Atom,that typically never supported AVX except in latest gens.

 

Heck,they even dropped support for "ancient" or "museum-grade" processors like Pentium 3,AMD Athlon XP and other processors not supporting SSE2,limiting them to SSE2 processors,for a similar reason. The same thing happened this year.

Edited December 29, 2024 by Leokids123
Museum-grade.....

[UCyborg]

AFAIK, Pale Moon installer does have a check and they made sure updater wouldn't update the old version to newer AVX version on non-AVX machines. I was trying out 7-zipped version.

Intel SDE is new to me. Maybe it was mentioned once on this forum, but I don't remember any discussion trying it out in practice. It sparked my interest after reading about it in some gaming circles. Intel doesn't keep majority of old versions around, no idea why new ones don't emulate anything here. Documentation doesn't mention any special must-have parameters. Too old Windows, maybe it doesn't like CPU for some reason, who knows... While technically interesting experiment, it's obviously not practical. No idea how it works, but however it does, it slows down even CPU-Z, which doesn't require any special instruction sets.

I'm aware that MCP doesn't want to target old hardware and I've been following recent developments, so also knew about contributed SSE2 builds right away. Pale Moon actually started out as a recompile of Firefox omitting support for pre-SSE2 CPUs. That was, when, 15 years ago? Now they're back at surpassing Firefox when it comes to CPU requirements.

If you look around, some software started requiring AVX earlier, eg. this query on Super User dated almost 8 years ago. Discussions around this particular instruction set seemed to have become more frequent earlier this decade, it has presumably started to become more common requirement since.

[j7n]

Is there any appreciable benefit from AVX? People say AVX was created a long time ago now. But back then there was no gain from paying for an "i3" when a "Pentium" was exactly the same without AVX and hyperthreading that I always saw a bit of a gimmick.

[UCyborg]

There are certainly folks that appreciated AVX builds since they were published in early 2022.

Those old Phenoms like I have don't even have SSE 4.1/4.2. I don't remember anyone pointing out back then instruction sets might mean something in the future. Today, it's a difference in being able / not being able to run certain software. Still, 15+ years is stretching it and an old CPU will lag behind new ones regardless of extra instruction sets.

So long term, it may still not make that much difference if you were to eventually replace the old one anyway.

Edited December 29, 2024 by UCyborg

[Leokids123]

4 minutes ago, UCyborg said:

There are certainly folks that appreciated AVX builds since they were published in early 2022.

Those old Phenoms like I have don't even have SSE 4.1/4.2. I don't remember anyone pointing out back then instruction sets might mean something in the future. Today, it's a difference in being able / not being able to run certain software. Still, 15+ years is stretching it and an old CPU will lag behind new ones regardless of extra instruction sets.

Celeron Sandy Bridge got sse4.1/4.2 and....allows you to run pcsx2....

[Mathwiz]

AVX instructions would be of some benefit on CPU-intensive applications like AI, but probably not much on a typical Web browser, and certainly not on an email client.

PM might be an exception because its old FF 52 Javascript engine takes a lot more CPU than the corresponding engines in more modern browsers. Back in 2017, the JS engine could afford to be inefficient, but Web sites are much more bloated with JS in 2024. But if MCP is merely using AVX as a proxy for "faster CPU," then it seems silly to me. Instead of AVX, the PM installer could just run a simple loop to test CPU speed, and pop up a warning if the CPU is found to be "too slow."

It'd be interesting to run some side-by-side tests of AVX and SSE2 Pale Moon builds on the same PC (with an AVX processor), so we could see how much improvement actually comes from AVX vs. the improvement just from a faster CPU. I suspect it's mostly the latter, but either way, it seems to me there would be a much greater ROI from backporting a more modern JS engine to UXP than from just building PM with AVX instructions to try to brute-force their way out of an old, slow JS engine.

(Their stubborn insistence on a single process doesn't help either. Most modern CPUs have at least eight cores, but PM will basically use only one of them.)

[Leokids123]

33 minutes ago, Mathwiz said:

AVX instructions would be of some benefit on CPU-intensive applications like AI, but probably not much on a typical Web browser, and certainly not on an email client.

PM might be an exception because its old FF 52 Javascript engine takes a lot more CPU than the corresponding engines in more modern browsers. Back in 2017, the JS engine could afford to be inefficient, but Web sites are much more bloated with JS in 2024. But if MCP is merely using AVX as a proxy for "faster CPU," then it seems silly to me. Instead of AVX, the PM installer could just run a simple loop to test CPU speed, and pop up a warning if the CPU is found to be "too slow."

It'd be interesting to run some side-by-side tests of AVX and SSE2 Pale Moon builds on the same PC (with an AVX processor), so we could see how much improvement actually comes from AVX vs. the improvement just from a faster CPU. I suspect it's mostly the latter, but either way, it seems to me there would be a much greater ROI from backporting a more modern JS engine to UXP than from just building PM with AVX instructions to try to brute-force their way out of an old, slow JS engine.

(Their stubborn insistence on a single process doesn't help either. Most modern CPUs have at least eight cores, but PM will basically use only one of them.)

They insist modern browsers are like fully fledged 3d games.... that's why they got avx....

[UCyborg]

2 hours ago, Mathwiz said:

It'd be interesting to run some side-by-side tests of AVX and SSE2 Pale Moon builds on the same PC (with an AVX processor), so we could see how much improvement actually comes from AVX vs. the improvement just from a faster CPU.

Faster CPU alone makes the most practical difference from my own experience. When they were voting on the forum what optimizations should be used in mainline build, they were quite divided between SSE2 and AVX. The difference can be noticed between the two, probably depends on web sites and with what kind of extensions you're bogging the browser down with. Slow down of GUI is quite noticeable with CuteButtons extension due to CSS used, AVX doesn't help with that.

1 hour ago, Leokids123 said:

They insist modern browsers are like fully fledged 3d games....

WebRender was actually documented to work like 3D game engine.

[UCyborg]

Holy s***, this site actually wants SSE 4.1.

https://donoharm.report/

Looks like .NET (Microsoft) is in the background. I doubt UXP does WASM SIMD at all. Edit: Nope, there's an open issue here.

4 hours ago, Leokids123 said:

Celeron Sandy Bridge got sse4.1/4.2 and....allows you to run pcsx2....

I'll probably play with PCSX2 again whenever I choose to buy a newer PC, the old Phenom is hopeless for these emulators last time I checked (unless you're happy with smooth intro cinematics alone).

Edited December 29, 2024 by UCyborg