Re: [PATCH v7 2/5] x86/asm: add volatile, clobbers and zero-length check in inline memcmp

["H. Peter Anvin" <hpa@xxxxxxxxx>]

On 2026-07-24 16:35, H. Peter Anvin wrote:
> On 2026-07-24 14:53, H. Peter Anvin wrote:
>>
>> It's pretty bloaty; I don't know exactly how much the rest of the code 
>> bloats.
>>
>> I would be fine with having mem[n]eq() though as an explicit API. For the 
>> out-of-line version used for compiler-generated memcmp() calls it doesn't 
>> matter. 
>>
>> Here is a very compact way the out-of-line version can be written (showing 
>> the 64-bit version; formatting apologies, on a phone):
>>
>> xor %eax,%eax    /* Sets ZF */
>> repe cmpsb
>> jr z,1f
>> movzbl -1(%rdi),%edi
>> mov -1(%rsi),%al
>> sub %edi,%eax
>> 1: ret
>>
> With a C wrapper to work in any mode:
> 
> int memcmp(const void *a, const void *b, size_t size)
> {
>         int diff;
> 
>         asm volatile("xor %0,%0 ;"      /* Sets ZF for the size = 0 case */
>                      "repe cmpsb ;"
>                      "jz 1f ;"          /* When size = 0 loading is unsafe */
>                      "movb -1(%2),%b0 ;"
>                      "movzbl -1(%1),%k1 ;"
>                      "sub %k1,%0 ;"
>                      "1:"
>                      : "=&q" (diff), "+D" (a), "+S" (b), "+c" (size)
>                      : : "memory");
>         return diff;
> }
> 

And of course I got the comparison backwards (negative means a < b, thus we
want a - b not b - a). Here is a fixed version.

int memcmp(const void *a, const void *b, size_t size)
{
        int diff;

        asm volatile("xor %0,%0 ;"      /* Sets ZF for the size = 0 case */
                     "repe cmpsb ;"
                     "jz 1f ;"          /* When size = 0 loading is unsafe */
                     "movb -1(%1),%b0 ;"
                     "movzbl -1(%2),%k2 ;"
                     "sub %k2,%0 ;"
                     "1:"
                     : "=&q" (diff), "+D" (a), "+S" (b), "+c" (size)
                     : : "memory");
        return diff;
}

        -hpa