GCC 12 Looking At Enabling Its Vectorizer For "-O2" Optimization Level

[phoronix]

Phoronix: GCC 12 Looking At Enabling Its Vectorizer For "-O2" Optimization Level

The GCC compiler when using the default "-O2" optimization level is likely to be slightly faster with next year's GCC 12 release as the developers are looking at enabling the vectorizer options by default...

https://www.phoronix.com/scan.php?page=news_item&px=GCC-O2-Vectorization-Talk

[jacob]

Which architectures will it support?

[perpetually high]

I'm still using gcc 9.3.0 (9.3.0-17ubuntu1~20.04) for my i5-4670K Haswell. Stupid move? Feels like maybe the best compiler for this chip but if anyone strongly disagrees, do share.

edit: someone asked why an old compiler would be best for any chip. My answer is that the optimizations are likely not coming in for Haswell anymore, and more code = longer execution time on stuff you don't need or gonna use. So I might as well use the last, greatest, stable version (9.3.0 in my case) instead of the bleeding edge 10/11/12. Put dumbly, of course. I'd love to be corrected.

[brucethemoose]

I'm still using gcc 9.3.0 (9.3.0-17ubuntu1~20.04) for my i5-4670K Haswell. Stupid move? Feels like maybe the best compiler for this chip but if anyone strongly disagrees, do share.

Thats an AVX2 chip. You would definitely benefit from a newer compiler, or better yet, the x86-64-V3 repo if you're on Arch.

[jabl]

I'm still using gcc 9.3.0 (9.3.0-17ubuntu1~20.04) for my i5-4670K Haswell. Stupid move? Feels like maybe the best compiler for this chip but if anyone strongly disagrees, do share.

edit: someone asked why an old compiler would be best for any chip. My answer is that the optimizations are likely not coming in for Haswell anymore, and more code = longer execution time on stuff you don't need or gonna use. So I might as well use the last, greatest, stable version (9.3.0 in my case) instead of the bleeding edge 10/11/12. Put dumbly, of course. I'd love to be corrected.

Most of the work that goes into the compiler is not tuning the pipeline descriptions for some particular CPU, so you're missing out on all that work.

That being said, in most cases the benefits are not huge, and the least-hassle way is to use the default version that comes with your distro (which is what you're using).

[perpetually high]

Most of the work that goes into the compiler is not tuning the pipeline descriptions for some particular CPU, so you're missing out on all that work.

That being said, in most cases the benefits are not huge, and the least-hassle way is to use the default version that comes with your distro (which is what you're using).

Yeah, I need to do benchmarks soon to really get to the bottom of this. Appreciate the reply.

[carewolf]

Which architectures will it support?

All of them?

It even have a generic mode that can operate without using SIMD instructions (2/4 16bit integers in 64bit word). Though that mode is naturally limited, and sometimes needs to be halved to deal with overflows.

[sdack]

Together with the pseudo x86-architectures for GCC's -march= switch that were added some time ago such as x86-64-v3 for AVX2 and x86-64-v4 for AVX512 do I hope to see distros picking up at least x86-64-v3 here and to provide more than the standard 64-bit binary packages.

By the time GCC 12 becomes stable would it be nice for distros like Debian to a add second x86-64 repository for AVX2-optimised binaries to make these gains available to all the Debian-based x86 distros. Perhaps the multimedia repository could start adding this.

[brad0]

Thats an AVX2 chip. You would definitely benefit from a newer compiler

For the most part that is very unlikely.

[pal666]

someone asked why an old compiler would be best for any chip. My answer is that the optimizations are likely not coming in for Haswell anymore

most optimizations are not chip-specific. new complier is always better and it's not limited to optimizations

and more code = longer execution time

not necessarily. sometimes it can be shorter execution time and more bugs found in your program at compile time.

the last, greatest, stable version (9.3.0 in my case)

last version in 9 branch is 9.4. i happen to use 9.3 in current project and i found deterministic ice in it(sometimes it ices but succeeds on rerun, but this one required disabling one optimization pass). i.e. 9.3 is definitely bad version