On Mon, Jun 5, 2017 at 1:14 PM, Gaius Mulley
<gaius.mul...@southwales.ac.uk> wrote:
>
> Hi,
>
> here is a tiny patch which removes an assert which I believe is wrong.
> I think it is an anomaly as the only callee (determine_max_movement at
> gcc/tree-ssa-loop-im.c:749) tests the asserted result against NULL. (If
> the assert really were correct then the else statement is redundant
> (dead code) at line gcc/tree-ssa-loop-im.c:758).
That check checs for the first exit, !simple_mem_ref_in_stmt.
The assert is about having populated the hash properly when gathering
all simple_mem_refs.
> I think the else is correct and the assert wrong as the gnu modula-2
> front end can provoke the assert to fail (thus needing the else statement).
Then sth is wrong with either hashing, simple_mem_ref_in_stmt or the
IL generated by modula-2.
Richard.
> I've run the C regression tests and there are no changes in the results
> if the patch is applied.
>
> regards,
> Gaius
>
>
> --- gcc-versionno-orig/gcc/tree-ssa-loop-im.c 2017-06-01 17:06:50.228216946
> +0100
> +++ gcc-versionno/gcc/tree-ssa-loop-im.c 2017-06-01 21:34:55.623992245
> +0100
> @@ -602,7 +602,6 @@
> hash = iterative_hash_expr (*mem, 0);
> ref = memory_accesses.refs->find_with_hash (*mem, hash);
>
> - gcc_assert (ref != NULL);
> return ref;
> }
>
> @@ -2631,5 +2630,3 @@
> {
> return new pass_lim (ctxt);
> }
> -
> -
>
