Windows: use __declspec(thread) TLS implementation, it is MUCH faster…

[Araq]

… than _Thread_local [backport]

[github-actions]

Thanks for your hard work on this PR!
The lines below are statistics of the Nim compiler built from ec3bca8

Hint: mm: orc; opt: speed; options: -d:release
167104 lines; 9.132s; 613.176MiB peakmem

[RSDuck]

Why did you change it to use __thread on mingw/GCC?

For me __thread didn't yield any improvement (with --threads:on more than halving the performance). But then I tried using __declspec(thread) and performance is now comparable to Linux, making --threads:on finally useable for me. Apparently this attribute is just ignored with a warning by GCC, so this is not valid.

EDIT 2: apparently gcc does not support native TLS on Windows at all (https://gcc.gnu.org/bugzilla/show_bug.cgi?id=80881), though atleast clang does, so I guess I have to stick with it.

[Araq]

I tried clang but even "hello world" crashes with clang on Windows for me. Which clang do you use? I'm willing to change the default C compiler that we ship with Nim if this finally solved this madness.

[RSDuck]

I'm using the one from msys2. While I like its integrated model with its package manager and closed environments, it's probably not for everyone.

One thing of note is that for TLS to work it's mingw clang environment has to be used, as in the default mingw x64 environment all the libraries (and clang) are compiled with gcc, so TLS emulation is used everywhere.