tree 115a39824dd012e2a9aab81738d09bbada8c4826
parent 731386735808a18351eeb480b78b8c1686f815c3
author bjacob <benoitjacob@google.com> 1694634807 -0400
committer GitHub <noreply@github.com> 1694634807 -0400
gpgsig -----BEGIN PGP SIGNATURE-----
 
 wsBcBAABCAAQBQJlAhM3CRBK7hj4Ov3rIwAAUoAIAKLH+iwdC8BhO+eHt+4M53Lt
 h21M/S4H9637ae2R4veJ2F2zo4thckpKuLixM250w6UxXlaX5ofwae82FadptH/0
 SCtw/dR0xj+1yehuFial0IMPHuPL46ByYM5YfoL3H17SWupTnm5HQzn9sWpoIfud
 9jtu+Hef/S2dyR/dRzzo2tS1MzuVSJGqdcUG8e3F7TetqL7JBL+bB4ugXDrBlkiO
 TTxZungQj7ccyzvE/22Gg/CPHRAnVzTQ4v1SLgcj9vaYlVINJTPMPAnJcYZI/P2p
 Zx0mr0vrrru8+EU/CxR0lAtK5BdxbEfSrnDR0Xuk0Fo4eltT0sd+8onp5Ffu5dg=
 =ZcBg
 -----END PGP SIGNATURE-----
 

Tentatively work around MSVC compiler bug in CPUID (#14968)

The `noinline` is a tentative work-around for what might be a MSVC
miscompile.

The symptom is that MSVC builds incorrectly report some CPU features as
supported. This only happens for CPU feature bits in the EDX output
register, which would make sense if the compiler omitted to save that
output register after CPUID instructions.

The other things done in this PR are:

1. Use `memcpy` for proper bit-casting, don't just cast `uint32_t` to
`int` and back: that is UB when not in `int` range, which happens
concretely here with compile-time constant values (hence the compiler
can actually see that's UB): `eax=0x80000000u`).
2. Drop some futile `inline` keywords.