On Mon, 27 Jan 2025, Michael Matz wrote:
> Hello,
>
> On Thu, 23 Jan 2025, Richard Biener wrote:
>
> > When we get a zero distance vector we still have to check for the
> > situation of a common inner loop with zero distance. But we can
> > still allow a zero distance for the loop we distribute
> > (gcc.dg/tree-ssa/ldist-33.c is such a case). This is because
> > zero distances in non-outermost loops are a misrepresentation
> > of dependence by dependence analysis.
>
> I think as long as that is the case your proposed changes makes sense.
> But perhaps it's worth a comment to that effect, i.e. that because
> dependence analysis is wonky (sometime using zero as unknown) we do these
> "strange" tests. At least for the occasional reader it will be
> unintuitive why the obvious tests aren't working.
Like this?
diff --git a/gcc/tree-loop-distribution.cc b/gcc/tree-loop-distribution.cc
index 9d9d2ae592b..fc0cd3952d5 100644
--- a/gcc/tree-loop-distribution.cc
+++ b/gcc/tree-loop-distribution.cc
@@ -2193,7 +2193,9 @@ loop_distribution::pg_add_dependence_edges (struct
graph *rdg, int dir,
this_dir = -this_dir;
}
/* When then dependence distance of the innermost common
- loop of the DRs is zero we have a conflict. */
+ loop of the DRs is zero we have a conflict. This is
+ due to wonky dependence analysis which sometimes
+ ends up using a zero distance in place of unknown.
*/
auto l1 = gimple_bb (DR_STMT (dr1))->loop_father;
auto l2 = gimple_bb (DR_STMT (dr2))->loop_father;
int idx = index_in_loop_nest (find_common_loop (l1,
l2)->num,
