Firefox Is Now Built With Clang+LTO Everywhere, Sizable Performance Wins For Linux

[Grinch]

Oh yikes, that was true at one point?

Yes, there was quite a ruckus when people realized that using windows Firefox under wine was faster than running native Linux Firefox binaries.

Turned out that the main reason was that the Windows binaries were built with PGO, which was not true for Firefox for Linux. Then PGO was fixed to compile with Linux, but I'm not sure what the case is these days. Also judging by the article, PGO is only working for Firefox Linux when compiling with Clang/LLVM at the moment.

It will be fixed undoubtably given that Mozilla has made it clear they will only support the Clang/LLVM toolchain for Firefox going forward.

[Grinch]

Arch uses clang for a long time

Ehh ? Do you mean that Arch has Clang/LLVM in their repositories and that it's used for projects which require it, that is true.

However the vast majority of packages are built using GCC as it's the default compiler toolchain for Arch. For example, Firefox which we are discussing here, is built using GCC on Arch.

[Grinch]

GCC has better optimizers than Clang. Only problem is that its LTO has some of bugs so some projects don't compile well with it.

That goes both ways, I've been trying to build Blender with Clang/LLVM using LTO but it won't build, meanwhile it works fine using GCC.

[pal666]

current clang with lto is faster than two year old gcc without lto. who would've thought

[pal666]

Clang enables cross-language inlining with Rust which is another huge performance step.

that's because rust is not a real language, but llvm frontend

[pal666]

GCC has better optimizers than Clang. Only problem is that its LTO has some of bugs so some projects don't compile well with it.

i think main problem here is that they are using gcc 6 instead of 8. lto support improved since then

[pal666]

also couldn't resist
mozilla: we are building a better internet. for windows users

[jacob]

Clang enables cross-language inlining with Rust which is another huge performance step.

I think that's the main thing. Rust itself is based on LLVM so as large parts of the Firefox code base move to it, having both Rust and C/C++ use the same compiler back-end is a huge advantage.

[hubicka]

Talos benchmarks GCC 7 builds using LTO+FDO from half a year ago comopared to today trunk (clang build I think) are here
https://treeherder.mozilla.org/perf....meRange=172800.

Comparing LTO+PGO between bth compilers seems to favour GCC.
I am building Firefox with LTO+PGO using GCC regularly. Usually linker errors are just missing attribute ((used)) on variables used by toplevel asm statements.

[Weasel]

Comparing LTO+PGO between bth compilers seems to favour GCC.
I am building Firefox with LTO+PGO using GCC regularly. Usually linker errors are just missing attribute ((used)) on variables used by toplevel asm statements.

Sounds like it's not a GCC LTO bug then, but a crappy project assuming stuff it shouldn't. I think it's very awesome that GCC removes such unused crap with LTO, you guys did a great job there for sure.

Please don't have GCC keep redundant garbage there just to deal with such crappy written projects (i.e. current behavior is fine!), but maybe some sort of other warnings would be easier to know the cause than just a linking error? Dunno if it's feasible. Been a while since I looked at GCC sources.