Andarwinux wrote:
As a veclib user, I hope this can be more integrated with the LLVM/Clang
toolchain and coexist with -fveclib.
- mathvec should be part of compiler-rt builtins, distributed with the
toolchain, and enabled by default (via such as -fuse-mathvec-builtin?), so
users only need to add -fno-math-errno to vectorize all math functions. And
then could specify -fveclib to prefer to use other veclibs and with mathvec as
a fallback.
- mathvec shouldn't follow libmvec's symbol name mangling and calling
convention to pretend to be a clone of libmvec.
- As a portable vector math library based on llvm-libc, mathvec is supposed
to support multiple target arch - but what if there are target arches
unsupported by libmvec?
- mathvec is supposed to support multiple target OS other than Linux such as
Windows.
- It might be best to use some custom calling convention to improve
performance. For example, on Windows, __m256 and __m512 are passed via stack
unless use vectorcall.
https://github.com/llvm/llvm-project/pull/234621
_______________________________________________
cfe-commits mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
