Re: [PATCH v12 18/31] arm: Add string-fza.h

[Adhemerval Zanella Netto <adhemerval.zanella@linaro.org>]

On 20/02/23 10:45, Szabolcs Nagy wrote:
> The 02/20/2023 13:24, Szabolcs Nagy via Libc-alpha wrote:
>> The 02/02/2023 15:11, Adhemerval Zanella via Libc-alpha wrote:
>>> +static __always_inline find_t
>>> +find_zero_all (op_t x)
>>> +{
>>> +  /* Use unsigned saturated subtraction from 1 in each byte.
>>> +     That leaves 1 for every byte that was zero.  */
>>> +  op_t ones = repeat_bytes (0x01);
>>> +  return __builtin_arm_uqsub8 (ones, x);
>>> +}
>>
>> __builtin_arm_uqsub8 is only available >=gcc-10
>>
>> so now the build fails with gcc-9
>>
>> ../sysdeps/arm/armv6t2/string-fza.h:36:10: error: implicit declaration of function ‘__builtin_arm_uqsub8’; did you mean ‘__builtin_arm_stc’? [-Werror=implicit-function-declaration]
>>    36 |   return __builtin_arm_uqsub8 (ones, x);
>>
>> so this code should be conditional on gcc version.
>>
> 
> i think
> 
>  asm ("uqsub8  %0, %0, %1" : "+r" (ones) : "r" (x));
> 
> should be a good fallback (untested).

This is what we have on v7:

static __always_inline op_t
find_zero_all (op_t x)
{
  /* Use unsigned saturated subtraction from 1 in each byte.
     That leaves 1 for every byte that was zero.  */
  op_t ret, ones = repeat_bytes (0x01);
  asm ("uqsub8 %0,%1,%2" : "=r"(ret) : "r"(ones), "r"(x));
  return ret;
}

Maybe extend with:

static __always_inline op_t
find_zero_all (op_t x)
{
  op_t ones = repeat_bytes (0x01);
#if __GNUC_PREREQ (10, 0)
  return __builtin_arm_uqsub8 (ones, x);
#else
  op_t ret;
  asm ("uqsub8 %0,%1,%2" : "=r"(ret) : "r"(ones), "r"(x));
  return ret;
#endif
}