On Wed, May 25, 2022 at 03:13:09PM +0200, Martin Liška wrote:
> Btw. I noticed one discrepancy:
>
> ./libgomp/loop_ull.c:
>
> /* Cheap overflow protection. */
> if (__builtin_expect ((nthreads | ws->chunk_size_ull)
> < 1ULL << (sizeof (gomp_ull)
> * __CHAR_BIT__ / 2 - 1), 1))
> ws->mode = ws->end_ull < (__LONG_LONG_MAX__ * 2ULL + 1
> - (nthreads + 1) * ws->chunk_size_ull);
>
> while:
>
> ./libgomp/loop.c:
>
> /* Cheap overflow protection. */
> if (__builtin_expect ((nthreads | ws->chunk_size)
> >= 1UL << (sizeof (long)
> * __CHAR_BIT__ / 2 - 1), 0))
> ws->mode = 0;
>
> So once, it uses __builtin_expect(..., 1) and second time
> __builtin_expect(..., 0).
> Is it expected?
Yes. The comparison is different. The first case is
or_of_the_two_values < 0x80000000ULL which is predicted likely (usually both
number of threads and chunk size are small numbers).
While in the other case it is
or_of_the_two_values >= 0x80000000UL (or 0x8000UL for 32-bit arches)
and again, we predict both numbers to be small and so the comparison to be
false.
Jakub
