On Wed, Apr 05, 2023 at 04:10:25PM -0400, Andrew MacLeod via Gcc-patches wrote:
> When a statement is first processed, any SSA_NAMEs that are dependencies are
> cached for quick future access.
>
> if we ;later rewrite the statement (say propagate a constant into it), its
> possible the ssa-name in this cache is no longer active. Normally this is
> not a problem, but the changed to may_recompute_p forgot to take that into
> account, and was checking a dependency from the cache that was in the
> SSA_NAME_FREE_LIST. It thus had no SSA_NAME_DEF_STMT when we were expecting
> one.
>
> This patch simply rejects dependencies from consideration if they are in the
> free list.
>
> Bootstrapping on x86_64-pc-linux-gnu and presuming no regressio0ns, OK for
> trunk?
>
> Andrew
> commit ecd86e159e8499feb387bc4d99bd37a5fd6a0d68
> Author: Andrew MacLeod <amacl...@redhat.com>
> Date: Wed Apr 5 15:59:38 2023 -0400
>
> Check if dependency is valid before using in may_recompute_p.
>
> When the IL is rewritten after a statement has been processed and
> dependencies cached, its possible that an ssa-name in the dependency
> cache is no longer in the IL. Check this before trying to recompute.
>
> PR tree-optimization/109417
> gcc/
> * gimple-range-gori.cc (gori_compute::may_recompute_p): Check if
> dependency is in SSA_NAME_FREE_LIST.
>
> gcc/testsuite/
> * gcc.dg/pr109417.c: New.
Ok for trunk (mainly to unbreak the recent regression).
But please be ready to adjust if Richi disagrees next week.
> --- a/gcc/gimple-range-gori.cc
> +++ b/gcc/gimple-range-gori.cc
> @@ -1314,7 +1314,9 @@ gori_compute::may_recompute_p (tree name, basic_block
> bb, int depth)
> tree dep2 = depend2 (name);
>
> // If the first dependency is not set, there is no recomputation.
> - if (!dep1)
> + // Dependencies reflect original IL, not current state. Check if the
> + // SSA_NAME is still valid as well.
> + if (!dep1 || SSA_NAME_IN_FREE_LIST (dep1))
> return false;
>
> // Don't recalculate PHIs or statements with side_effects.
Jakub
