_HAVE_STRING_ARCH_strcpy/strncpy and ppc64 macro expansion of strcpy/strncpy.

_HAVE_STRING_ARCH_strcpy/strncpy and ppc64 macro expansion of strcpy/strncpy.

[Aaron Sawdey]

At present, powerpc uses the generic macros for strcpy/strncpy
expansion, which includes an expansion into code that compares the
strings character by character if one string is a string constant and
is less than 4 chars. 

I'm currently working on builtin expansion of strcmp/strncmp in gcc7
for powerpc, similar to the memcmp builtin expansion I checked in a
couple weeks ago. This will potentialy generate much better code for
these short comparisons because gcc often knows the alignment which
allows it to generate word or doubleword comparisons reducing the
number of branches greatly.

However as things currently stand, comparisons to constant strings < 4
chars never see a strcmp/strncmp call because they get macro expanded
into comparison code. What would be helpful would be a change to have
powerpc specific code that doesn't invoke __strcmp_cg if compiled with
gcc7 or higher.

Thanks,
    Aaron

-- 
Aaron Sawdey, Ph.D.  acsawdey@linux.vnet.ibm.com
050-2/C113  (507) 253-7520 home: 507/263-0782
IBM Linux Technology Center - PPC Toolchain

Re: _HAVE_STRING_ARCH_strcpy/strncpy and ppc64 macro expansion of strcpy/strncpy.

[Adhemerval Zanella]

On 26/10/2016 13:04, Aaron Sawdey wrote:
> At present, powerpc uses the generic macros for strcpy/strncpy
> expansion, which includes an expansion into code that compares the
> strings character by character if one string is a string constant and
> is less than 4 chars. 
> 
> I'm currently working on builtin expansion of strcmp/strncmp in gcc7
> for powerpc, similar to the memcmp builtin expansion I checked in a
> couple weeks ago. This will potentialy generate much better code for
> these short comparisons because gcc often knows the alignment which
> allows it to generate word or doubleword comparisons reducing the
> number of branches greatly.
> 
> However as things currently stand, comparisons to constant strings < 4
> chars never see a strcmp/strncmp call because they get macro expanded
> into comparison code. What would be helpful would be a change to have
> powerpc specific code that doesn't invoke __strcmp_cg if compiled with
> gcc7 or higher.
> 
> Thanks,
>     Aaron
>

I think back when we lacked gcc support for string routing optimization
the string{2,3}.h header optimization made sense, but currently it add
more issue with multiple arch definition permutation and differences.

For your specific issue I think it would be feasible to create a
installed string.h with a arch-specific optimization (as for x86
and s390) with something like:

#define _HAVE_STRING_ARCH_strcmp
#ifndef _FORCE_INLINES
__STRING_INLINE int
strcmp (const char *__s1, const char *__s2)
{
  return __builtin_strcmp (__s1, __s2);
}
#endif

But I would rather prefer to evaluate if we *really* require the
default and convoluted strcmp macro optimization.  I think it would
be just better to just remove it and let it the compiler handle it.