[PATCH v11 10/29] string: Improve generic stpcpy

[Adhemerval Zanella Netto]

On 01/02/23 14:29, Xi Ruoyao wrote:
> On Wed, 2023-02-01 at 14:03 -0300, Adhemerval Zanella wrote:
>> +static __always_inline char *
>> +stpcpy_unaligned_loop (op_t *restrict dst, const op_t *restrict src,
>> +                      uintptr_t ofs)
>> +{
>> +  op_t w2a = *src++;
>> +  uintptr_t sh_1 = ofs * CHAR_BIT;
>> +  uintptr_t sh_2 = OPSIZ * CHAR_BIT - sh_1;
> 
> Hmm, on 64-bit LoongArch if we "clone" the function 7 times to
> stpcpy_unaligned_loop_{1..7} and call them with a switch (ofs) { ... }
> construction, we'd be able to use bytepick.d instruction for MERGE,
> saving 2 instructions in the iteration.  But maybe this is going too
> far.  I'm not sure if this "optimization" applies for other
> architectures.

I think it should be feasible, I might get back to optimize the unaligned
loop with this strategy.  But I will need to check if compiler will indeed
exploit the fact that the shifts are now constants to optimize the merge.

It also increases the code size slight, on x86_64 text size went from 850
to 1993 and on loongarch from 864 to 2200 (so it might be something to
consider as well assuming that unaligned strings will have a equal probability
to happen, so icache pressure would be important).

> 
>> +  op_t w2 = MERGE (w2a, sh_1, (op_t)-1, sh_2);
>> +  if (!has_zero (w2))
>> +    {
>> +      op_t w2b;
>> +
>> +      /* Unaligned loop.  The invariant is that W2B, which is "ahead" of W1,
>> +        does not contain end-of-string.  Therefore it is safe (and necessary)
>> +        to read another word from each while we do not have a difference.  */
>> +      while (1)
>> +       {
>> +         w2b = *src++;
>> +         w2 = MERGE (w2a, sh_1, w2b, sh_2);
>> +         /* Check if there is zero on w2a.  */
>> +         if (has_zero (w2))
>> +           goto out;
>> +         *dst++ = w2;
>> +         if (has_zero (w2b))
>> +           break;
>> +         w2a = w2b;
>> +       }
>> +
>> +      /* Align the final partial of P2.  */
>> +      w2 = MERGE (w2b, sh_1, 0, sh_2);
>> +    }
>> +
>> +out:
>> +  return write_byte_from_word (dst, w2);
>> +}
>> +
>