Re: [PATCH] Common cpuid wrappers, use SYS_cpuid when available

[Joseph Myers <joseph at codesourcery dot com>]

On Wed, 9 Mar 2016, Piotr Henryk Dabrowski wrote:

> Sorry for the malformed patch in the first post.
> Here is a working one.
> 
> 	* config.h.in: Check for SYS_cpuid and define HAVE_SYS_CPUID
> 	* configure: Check for SYS_cpuid and define HAVE_SYS_CPUID
> 	* configure.ac: Check for SYS_cpuid and define HAVE_SYS_CPUID
> 	* misc/cpuid.h: Common cpuid wrappers, use SYS_cpuid when available

Non-sysdeps files should not have anything architecture-specific.

Now, *if* you need a configure test, you can't avoid changing the 
architecture-independent config.h.in (we don't yet have a way to split 
that by architecture).  But the rest can be avoided by using sysdeps 
configure fragments and headers.  And you don't need a configure test 
anyway - the code can use #ifdef __NR_cpuid to test if the syscall is 
available at compile time, and __ASSUME_CPUID to test if the syscall can 
be presumed to work (if the runtime kernel is known to be recent enough, 
since the kernel headers used to build glibc may be more recent than the 
kernel used by glibc at runtime).  Finally, misc/cpuid.h should go under a 
sysdeps directory, suitably named not to conflict with the compiler's 
<cpuid.h>.

There's a complication: sysdeps/x86 shouldn't contain anything 
Linux-specific either.  So what that suggests is that you have e.g. 
sysdeps/x86/x86-cpuid.h that uses just the cpuid instruction but that can 
be overridden by an OS-specific header that supports the syscall - and 
then have such an OS-specific version in sysdeps/unix/sysv/linux/x86 (you 
can also do more complicated schemes to avoid duplicating the code to use 
the instruction).

-- 
Joseph S. Myers
joseph@codesourcery.com