On 13/08/2025 18:33, Jeremy Drake via Cygwin-patches wrote:
On Tue, 5 Aug 2025, Thirumalai Nagalingam wrote:
Hi all,
This patch adds support for the `fsqrt` instruction on AArch64 platforms
in the `__FLT_ABI(sqrt)` implementation.
This looks OK as far as it goes, but I have a few thoughts.
From the comments, it appears this code originally came from mingw-w64.
Their current version of this code has aarch64 implementations. The
difference with this one is they have a version for float as well as
double. The versions here seem to only be used for long double (which on
aarch64 is the same as double).
Given that long double is the same as double on aarch64, might it make
sense to redirect/alias the long double names to the double
implementations in the def file (cygwin.din) on aarch64, rather than
providing two different implementations (one in newlib for double and one
in this cygwin/math directory for long double)? It seems like that's
asking for subtle discrepancies between the implementations. I'm not
seeing any obvious preprocesor-like operations in gendef (mingw-w64 uses
cpp to preprocess .def.in => .def files for arch-specific #ifdefs) so
maybe this would be more complicated.
Sorry about the long delay looking into this.
So, I was about to apply Thiru's v2 patch, since that all seems
reasonable to me. But now I'm not so sure...
I think that a good goal is to keep this file aligned with the mingw-w64
version, if possible.
If I'm understanding correctly, if we do that, this problem goes away,
but at the cost that fsqrtd and fsqrtl are potentially different
(although surely since it all boils down to a single instruction, that's
never going to happen :) )?
