On 21.10.2024 16:50, LIU Hao wrote:
在 2024-10-21 19:11, LIU Hao 写道:
BTW, affected functions seem to be available in modern msvcrt
(msvcr120 and UCRT), so another solution for the bug would be to
move those function out of mingwex and use implementation provided
by OS when possible.
I can take a look later.
Attached is a patch for skipping the run-time check where SSE must
exist, for example, on Pentium 3 or on x86-64.
To elaborate on use of __SSE__ macro: it will be indeed mostly do the
right thing. On x86_64, it will always be present, so that's fine. On
32-bit x86, it will vary depending on compiler and command line
arguments. It may be explicitly enabled or disabled with CFLAG options
like -msse/-mno-sse. If it's enabled, the compiler is free to use it for
other parts of crt, so doing runtime checks is indeed not bringing any
compatibility benefits. As such, your solution looks fine to me.
Compilers seem to have different defaults on 32-bit targets: Clang
enables it by default, while GCC does not. While having different crt
behavior depending on the compiler is generally something we should
avoid, in this case it just reflects reality, so that seems okay.
Style-wise, I would perhaps prefer a solution that uses fewer #ifdefs.
We could for example change __mingw_has_sse to always return true when
it's defined. It would limit optimization opportunities for the
compiler, so perhaps we could have it in a header and make it a trivial
inline in case __SSE__ is defined? Anyway, the current patch matches how
similar things are done in mingw-w64, so I won't insist on that.
Even with that patch, the original patch caching cpuid result would
still be beneficial for cases when __SSE__ is not defined.
The Microsoft implementation can't be used, as explained in
e5d1fe451b81db3134d7ba70ff39d0a1b6c91b36.
I missed that, it's good to know.
Thanks,
Jacek
_______________________________________________
Mingw-w64-public mailing list
[email protected]
https://lists.sourceforge.net/lists/listinfo/mingw-w64-public
