With new clang based intel compilers (ifx, icx, icpx) we should use -march=native

[Micket]

easybuild-framework/easybuild/toolchains/compiler/inteliccifort.py

Line 62 in b37f707

'optarch': 'xHost',

We should set this to -march=native but only when the new ifx, icx, icpx are used.

This will be better for AMD cpus. I think @bartoldeman said that from 2025 onwards the old compilers are gone, so, we shouldn't use that old bad -x flags that are only supported on Intel cpus.

[Thyre]

With 2024.0.0, icc and icpc were removed. With 2025.0.0, ifort is gone, leaving only the "new" ones. I think switching to -march=native beginning with 2025.0.0 is therefore a good idea, since a toolchain using 2024.x.x might still use ifort.

[boegel]

Is there any reference material on this?

I'm not sure we should trust the Intel (oneAPI) compilers to "do the right thing" on AMD-based systems when using -march=native...

[Thyre]

Is there any reference material on this?

I'm not sure we should trust the Intel (oneAPI) compilers to "do the right thing" on AMD-based systems when using -march=native...

Intel doesn't mention what happens for -march=native unfortunately, see: https://www.intel.com/content/www/us/en/docs/dpcpp-cpp-compiler/developer-guide-reference/2025-0/march.html

[Micket]

at most they say

Many of the Clang settings for option -m are supported. For more information on Clang settings for option -m, see the Clang documentation.

but for me it was @bartoldeman telling me to just use it, and from my testing i can confirm that yes, that works.

[Thyre]

If one looks at the downstream source code, the behavior is basically the same from what I've found for LLVM:

• If native, call llvm::sys::getHostCPUFeatures() (clang/lib/Driver/ToolChains/Arch/X86.cpp#L131)
• Figure out supported features (llvm/lib/TargetParser/Host.cpp#L1753)

This seems completely independent of the CPU vendor. Who knows though if they are using the same code in their oneAPI packages...