On Fri, 7 Feb 2025, Jacek Caban wrote:
On 7.02.2025 10:22, Martin Storsjö wrote:
On Fri, 7 Feb 2025, LIU Hao wrote:
在 2025-02-06 22:23, Jacek Caban 写道:
FWIW, that's the approach we take in Wine for entire CRT modules (not
just static parts like mingw- w64-crt's counterpart), and it works well.
Most optimizations lost this way can be addressed by writing efficient C
code from the start, avoiding the need to second-guess how the compiler
will handle it. I'd be in favor of applying the same approach to
mingw-w64.
I agree.
Maybe it's just me, but committing compiler-generated assembly to the
tree feels like a sign that something isn't quite right...
As for this specific case, handwritten assembly is hardly any better than
this generated stuff.
Yes it's not that we could do better - it's just unnecessary code to have
standalone assembly for something where the compiler could do the same.
There are of course a number of cases where we _do_ need assembly, but
ideally we should limit that to the cases where we really do need it.
Yes, I meant those specific cases. In other situations, assembly is the
better choice, or even necessary.
(For arm64ec support, there will probably be a need to touch a lot of the
assembly files we have.)
Yes, making ARM64 assembly work in most cases is straightforward, it mainly
involves mangling symbol names. However, if we want to mix x86 and EC code,
additional thunks are needed, and those aren't fun to write manually.
Leveraging the C compiler as much as possible makes things much easier.
If there's no opposition to building the whole mingw-w64-crt with
-fno-builtin, I think that would be a good direction; this is quite special
code and we want to make sure that we call exactly the function we intend
to, not letting the compiler rewrite things back into what we're trying to
implement.
It seems we have a consensus on the path forward. Thanks!
I think it's still probably best to apply this patchset as is, and then
apply -fno-builtin, and we can convert these files, the corresponding
existing arm ones, plus other similar functions back to C code.
Sure, sounds good to me.
I pushed these patches now; I'll try to look into moving towards
-fno-builtin sometime soon.
// Martin
_______________________________________________
Mingw-w64-public mailing list
Mingw-w64-public@lists.sourceforge.net
https://lists.sourceforge.net/lists/listinfo/mingw-w64-public
