Thread

Re: Broken ./configure checks for __cpuid() and __cpuidex()

Lukas Fittl <lukas@fittl.com> — 2025-07-29T03:53:43Z

On Mon, Jul 28, 2025 at 8:40 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:

>
> > How isn't that a bug in MinGW itself?  I'm puzzled my the macro
> > definition of __cpuid() that reports a conflict.  __cpuidex() and
> > __cpuid() are both detected by ./configure, the PG use of __cpuid() in
> > pg_crc32c_sse42_choose.c causes a failure.
>
> Why is the configure probe succeeding?  Maybe pg_crc32c_sse42_choose.c
> is including something the configure check isn't?
>

Could it be that the problem only happens when including both cpuid.h and
intrin.h, because they both define __cpuid? (the configure check only
includes intrin.h)

My theory when I worked on the patch that Michael referenced in the
original email was that intrin.h is only for MSVC (for GCC at least,
__cpuidex is defined in cpuid.h).

I'm not sure how to get CI to run MinGW (it appears paused for me?), so I
can't test this myself easily.

But the relevant change would be to change "defined(HAVE__CPUIDEX)" to
"(defined(HAVE__CPUIDEX) && defined(_MSC_VER))" for the guard on both
intrin.h includes.

Thanks,
Lukas

-- 
Lukas Fittl