gcc/gcc/tree-ssa-loop-prefetch.c

7 under the terms of the GNU General Public License as published by the
121            the cache line accessed by (5) is the same with probability only
137       by machine description).  We prefetch as many prefetches as possible
171       estimated by taking the ratio between the prefetch count and the
181       -- make it behave sanely together with the prefetches given by user
187 /* Magic constants follow.  These should be replaced by machine specific
190 /* True if write can be prefetched by a read prefetch.  */
196 /* True if read can be prefetched by a write prefetch. */
202 /* The size of the block loaded by a single prefetch.  Usually, this is
279    than what is required by the prefetch.  This is to avoid redundant
370          by decreasing step.  */
735 /* Divides X by BY, rounding down.  */
738 ddown (HOST_WIDE_INT x, unsigned HOST_WIDE_INT by)
740   gcc_assert (by > 0);
743     return x / by;
745     return (x + by - 1) / by;
800 prune_ref_by_group_reuse (struct mem_ref *ref, struct mem_ref *by,
805   HOST_WIDE_INT delta_r = ref->delta, delta_b = by->delta;
842 	  != ddown (by->delta, PREFETCH_BLOCK))
931   /* The ref probably does not reuse by.  */
1047    time).  UNROLL_FACTOR is the factor by that the loop is going to be
1070   /* For now we just take memory references one by one and issue
1128    UNROLL_FACTOR is the factor by which LOOP was unrolled.  */
1151    is the factor by which LOOP was unrolled.  */
1209    factor by that LOOP was unrolled.  */
1347    this is the case, fill in DESC by the description of number of
1368 /* Determine the coefficient by that unroll LOOP, from the information
1386      from unrolling them enough to make exactly one cache line covered by each
1402      but bound the unrolling by UPPER_BOUND.  */
1440 	   the line divided by this number.  Otherwise, a cache line is
1472 /* Add the steps of ACCESS_FN multiplied by STRIDE to the array STRIDE
1625       /* Bound the volume by the L2 cache size, since above this bound,
1797   /* Prefetching improves performance by overlapping cache missing
1803      TODO: Implement a more precise computation by estimating the time
1837      of issued prefetch instructions is also affected by unrolling.  So,
1879   /* FIXME: the time should be weighted by the probabilities of the blocks in