Open romulasry opened 3 years ago
We have already been doing that for 6 or 7 years. Our GCC and LLVM are compiled to always produce code with a minimum of x86_64-v2 (a.k.a -march=westmere
) and we do v3 with autospec.
What about x86-64-v4? Using this may be more reliable to use via than your own custom method for the future anyways. Just a thought.
It may be extended past v4 and they may use Microarchitecture Feature Levels beyond just x86-64. If so, it may be easier to just use this then back port the changes to Clear Linux.
Edit: Nice to see something from Clear Linux now standardized in gcc 11 and llvm 12. Good job Intel.
We also do v4 via autospec (and have been doing for years).
Expect that any definition of what v5 is will come from us. I've been prodding our benchmark team to look into whether ICL or TGL has enough improvement in performance to warrant the next level.
https://www.phoronix.com/scan.php?page=news_item&px=GCC-11-x86-64-Feature-Levels https://www.phoronix.com/scan.php?page=news_item&px=LLVM-12-Clang-12-Feature-Over
How could this be implemented in clear linux?