Proper use of x86/x86_64 CPUID instruction with extended assembler

[Avi Kivity]

On 08/19/2015 09:28 AM, Jeffrey Walton wrote:
> Forgive my ignorance....
>
> The following retrieves information from a x86/x86_64 processor. Its
> used in an inline assembly block under Linux. The function is also
> inlined in a header. The other detail is, its System V ABI, so the
> caller owns RBP, RBX and R12-R15 under x86_64 (I'm not sure about the
> i386 variant).
>
>      struct CPUIDinfo
>      {
>          word32 EAX;
>          word32 EBX;
>          word32 ECX;
>          word32 EDX;
>      };
>      ....
>
>      static inline void CpuId(unsigned int func, unsigned int subfunc,
> CPUIDinfo& info)
>      {
>          __asm__ __volatile__ (
>              "cpuid"
>              : "=a"(info.EAX), "=b"(info.EBX), "=c"(info.ECX), "=d"(info.EDX)
>              : "a"(func), "c"(subfunc)
>              : "b"

This "b" clobber is unneeded, since "b" is an output register.

>          );
>      }
>
> The CPUID instruction always uses EAX/EBX/ECX/EDX, even on 64-bit
> platforms. On 64-bit platforms, the CPUID instruction sets the high
> 32-bit words of RAX/RBX/RCX.RDX equal to 0.
>
> My question is, will the extended assembly preserve the registers? Is
> it sufficient to list "b" as a clobber? Or do I need to do something
> special?

It will not preserve the registers, except those that are required to by 
the ABI ("b").  The clobber tells gcc which registers are destroyed by 
the instruction, not which registers to preserve; it already knows the 
latter.

>
> If I need to do something special, then what needs to be done?

Nothing much; just drop that "b" clobber.

>
> (I also looked at
> https://gcc.gnu.org/ml/gcc-patches/2007-09/msg00324.html, which is a
> GCC patch for cpuid.h. But its not clear to me if the above is correct
> because the operands are 32-bit in size. Naively, if I use "a" and "b"
> with a 32-bit operand, then I would expect code for EAX and EBX; and
> not RAX and RBX).

It is correct.

> Thanks in advance.