Fix issue https://github.com/official-stockfish/Stockfish/issues/2975

[mstembera]

This now compiles for me in windows using gcc 7.3, 8.1, and 10.1 and is actually a bit simpler.
After changing the code I could compile but was getting these errors
https://stackoverflow.com/questions/43152633/invalid-register-for-seh-savexmm-in-cygwin
with the older gcc versions but not 10.1.
The -fno-asynchronous-unwind-tables flag fixes those.
If there is a way to check gcc version in the Makefile we could skip this flag for gcc 10 and higher.
(I don't have gcc 9 to test)

Also @vondele asked if we need to adjust the avx512 flags.
We need -mavx512bw and -mavx512f. gcc silently includes -mavx512f when -mavx512bw is specified
but it's probably better to be explicit. (you can test this by adding -mno-avx512f to master and see it fail)
All avx512 implementations however also provide -mavx512cd -mavx512dq -mavx512vl instructions so
we can safely include them too.
https://en.wikichip.org/wiki/x86/avx-512
I measure no speed difference so either way is fine.

No functional change
NNUE: 3994357
bench: 4733874

[vondele]

I can confirm this builds on linux with gcc 8, 9, and 10.

I would not add the unneeded -mavx512cd -mavx512dq -mavx512vl, just what we need explicitly.

On Linux the flag -fno-asynchronous-unwind-tables is not needed, it is windows specific (see https://gcc.gnu.org/bugzilla/show_bug.cgi?id=65782). I don't think I would include that workaround, I don't think it is correct, even if it suppresses the error message (and comes from Agner Fog), it is also fixed on gcc 8, 9, 10 (but latest point releases of those versions).

[mstembera]

OK Thanks. PR updated.

[gvreuls]

@mstembera I bit the bullet and just finished upgrading my development system the next distro release which has GCC 10. (I was under the impression this wasn't going to get fixed, we're asking a lot of people to upgrade their compilers anyway, so why not to GCC 10.)

[mstembera]

@vondele If it's not too late could I please lobby to put the -fno-asynchronous-unwind-tables flag back in? I was working on some other AVX stuff and wanted to test older compilers. I spent a looong time trying to figure out why my new code was failing until I realized it was due to this -fno-asynchronous-unwind-tables flag missing again. I was indifferent earlier but very much an advocate now :)

[vondele]

Let's make it a separate PR, master is prepared for merging. I would like to understand better, and possibly make it conditional on a windows target? Do we know of side effects of the flag?

[mstembera]

NP Thanks. I will get back with a separate PR if it still seems to make sense after I look into it better.

[vondele]

Yes, my concern is that the code generation is actually wrong in unpatched versions of gcc, and the flag only suppresses the error message which happens to come from unwind-info generation: https://gcc.gnu.org/bugzilla/show_bug.cgi?id=65782#c8
But, I don't know these things well enough to judge.

Edit: the fact that it is backported to all open release branches, would support that, this is usually only done for wrong code generation and regressions.