On Tue, Jun 10, 2025 at 02:52:24PM +0000, Varghese, Vipin wrote: > [Public] > > Snipped > > > > > > > In above log I get `2 instances of march`; logs `-march=native -mrtm > > > -DALLOW_EXPERIMENTAL_API -DALLOW_INTERNAL_API -Wno-format- > > truncation -Wno-address-of-packed-member - > > DRTE_LOG_DEFAULT_LOGTYPE=pmd.net.i40e -DCC_AVX512_SUPPORT - > > march=skylake-avx512`. > > > > > > Question-1: I think this is not expected right? The `-march=native` is > > > populated > > from `cflags` and `-march= skylake-avx512` is populated from ` > > cc_avx512_flags`. > > > > The above command is correct. So long as the compiler supports AVX-512 we > > will > > always compile the AVX-512 code paths for runtime selection. In practice, > > all > > supported compilers have AVX-512 support, so in reality we have the two > > scenarios > > you tested: > > > > * The target architecture e.g. znver3 in your case, doesn't support avx512, > > so the meson.build file adds on the necessary flags to add this support, > > i.e. that file is compiled with -march=skylake=avx512, which is the > > minimum ISA that gives you the necessary support. > > * The target architecture, e.g. znver4, does support AVX-512, then no > > additional flags are added and the files are compiled "as normal" > > > > In both these cases, whether the target architecture is specified as > > "native" or > > explicitly makes no difference. > > So, for files which needs avx512 support like acl_avx512, fib_tire_avx512 and > others, we then only pass `cc_avx512_flags`. > All other cases this is ignored. Thank you that makes sense. > > > > > > Question-2: if the target is meet minimal ISA why not we use > > > `-march=x86-64- > > v4`? > > > > > > > Good point, that would indeed be better. I'm just not sure whether it is > > supported > > widely enough on our compilers. Do you know what gcc and clang versions > > support > > that target? > > As I recollect we have been using `x86-64-v4` this from gcc 12.3 and clang-14. > I am not sure if the older versions supports `avx512 that is x86-64-v4` >
Checking with godbolt.org, to see what versions give errors or not. * Clang accepts the argument from v12 onwards * GCC accepts the argument from v11 onwards. Unfortunately, that means that we need to at least keep some form of backward compatibility for older compilers until such time as we have those versions as minimum. If you are ok with the patch in general, I'll see if I can respin a version where we check for support for -march=x86-64-v4 and use that when we can, otherwise falling back to skylake-avx512 as here. Does that seem a reasonable approach? /Bruce
