On Thu, Dec 03, 2015 at 07:26:20PM +0100, Martin Jambor wrote:
> this is a fix to the last "last" ICE of the hsa branch. THe problem
> turned out not to be in the gridification itself but, depending your
> point of view, in the gimple and tree walking infrastructure or in
> function copy_gimple_seq_and_replace_locals from tree-inline.c on
> which hsa gridification relies.
>
> The issue is that in between gimplification and omplow pass, there can
> be gimple sequences attached to OMP_CLAUSE trees that are attached to
> omp statements and that are neither copied by gimple_seq_copy nor
> walked by walk_gimple_seq.
>
> While the correct solution would probably be to extend tree and gimple
> walkers to handle them, that would be a big change. I have talked
> with Jakub about this yesterday on the IRC and he suggested that I
> enhance the internal walkers of copy_gimple_seq_and_replace_locals
> deal with this situation. Even though that leaves gimple_seq_copy,
> walk_gimple_seq and other to be technically incorrect, that is what I
> have done in the patch below, which fixes my last ICEs and which I
> have already committed to the branch.
The point is that those gimple_seqs are there only from gimplification
till omplower, and I believe nothing else for now cares about those.
> @@ -5200,6 +5231,18 @@ replace_locals_stmt (gimple_stmt_iterator *gsip,
> return NULL_TREE;
> }
>
> +/* Create a copy of SEQ and remap all decls in it. */
> +
> +static gimple_seq
> +duplicate_remap_omp_clause_seq (gimple_seq seq, struct walk_stmt_info *wi)
> +{
I would have expected an early if (seq == NULL) return NULL; either here,
or in the callers (not doing anything in the common case when it is NULL).
Jakub
