Glibc-HWCAPS To Help With AMD Zen Optimizations, Other Per-CPU Performance Bits

[phoronix]

Phoronix: Glibc-HWCAPS To Help With AMD Zen Optimizations, Other Per-CPU Performance Bits

Experimental patches under discussion for the GNU C Library (glibc) would make it easier to dynamically load optimized libraries (shared objects) on systems depending upon the CPU in use and its hardware capabilities. This glibc-hwcaps work stems from the desired work on being able to better leverage Linux performance optimizations on AMD Zen-based systems but the hardware capabilities patches themselves can help any CPU microarchitecture family in more easily shipping optimized support...

http://www.phoronix.com/scan.php?page=news_item&px=glibc-hwcaps-RFC

[pal666]

subj handles shared libraries. what is the plan for handling applications (/bin/*) ?

[milkylainen]

I don't get the difference between this and the already existing arch optimized implementations?
Per-cpu type function optimizations? Sounds nice but seems like a ridiculous mess...
Didn't glibc scratch all that shit when people tried to do better mem* optimizations for various CPUs within the same architecture?

I think the reasoning was more in the line of "unmaintainable mess".
Maybe my memory is failing me. Someone with more insight please?

[mlau]

This is about the dynamic loader (ld.so) loading libraries/object files from certain directories based on advertised hardware caps. it isn't restricted to glibc only,
you could build e.g. a libpython3.8m.so for plain x86_64, +avx, +avx2, +avx512 and so on, put them in the respective hwcaps subdirectory and
the loader would automatically load the one suited for the currently running platform. Probably very useful for most binary distros, but not Gentoo.

EDIT: as far as I understand it, glibc can do that already, but the directory naming is a mess. Mr Weimer's proposing to simplify it a bit.