On Mon, Sep 27, 2021 at 3:08 PM Richard Biener via Gcc-patches
wrote:
>
> This avoids the last_vuse optimization hindering redundant store
> elimination by always also recording the original VUSE that was
> in effect on the load.
>
> Bootstrapped and tested on x86_64-unknown-linux-gnu.
>
> I'm still pondering on how to avoid the wastage of adding the ref
> twice and will at least record some statistics for this.
In stage3 gcc/*.o we have 3182752 times recorded a single
entry and 903409 times two entries (that's ~20% overhead).
With just recording a single entry the number of hashtable lookups
done when walking the vuse->vdef links to find an earlier access
is 28961618. When recording the second entry this makes us find
that earlier for donwnstream redundant accesses, reducing the number
of hashtable lookups to 25401052 (that's a ~10% reduction).
Overall I think it's a reasonable trade-off but as said, I'm pondering
a bit on how to reduce the overhead without too ugly hacks.
Richard.
> 2021-09-27 Richard Biener
>
> PR tree-optimization/100112
> * tree-ssa-sccvn.c (visit_reference_op_load): Record the
> referece into the hashtable twice in case last_vuse is
> different from the original vuse on the stmt.
>
> * gcc.dg/tree-ssa/ssa-fre-95.c: New testcase.
> ---
> gcc/testsuite/gcc.dg/tree-ssa/ssa-fre-95.c | 25 ++
> gcc/tree-ssa-sccvn.c | 17 +++
> 2 files changed, 38 insertions(+), 4 deletions(-)
> create mode 100644 gcc/testsuite/gcc.dg/tree-ssa/ssa-fre-95.c
>
> diff --git a/gcc/testsuite/gcc.dg/tree-ssa/ssa-fre-95.c
> b/gcc/testsuite/gcc.dg/tree-ssa/ssa-fre-95.c
> new file mode 100644
> index 000..b0936be5e77
> --- /dev/null
> +++ b/gcc/testsuite/gcc.dg/tree-ssa/ssa-fre-95.c
> @@ -0,0 +1,25 @@
> +/* PR100112 and dups. */
> +/* { dg-do compile } */
> +/* { dg-options "-O2 -fdump-tree-fre1-details -fdump-tree-optimized" } */
> +
> +int *c, *b;
> +void foo()
> +{
> + int *tem = b;
> + *tem = 0;
> + int *footem = c;
> + c = footem;
> +}
> +
> +void bar()
> +{
> + int *tem = b;
> + int *bartem = c;
> + *tem = 0;
> + c = bartem;
> +}
> +
> +/* We should elide the redundant store in foo, in bar it is not redundant
> since
> + the *tem = 0 store might alias. */
> +/* { dg-final { scan-tree-dump "Deleted redundant store c = footem" "fre1" }
> } */
> +/* { dg-final { scan-tree-dump "c = bartem" "optimized" } } */
> diff --git a/gcc/tree-ssa-sccvn.c b/gcc/tree-ssa-sccvn.c
> index e8b1c39184d..416a5252144 100644
> --- a/gcc/tree-ssa-sccvn.c
> +++ b/gcc/tree-ssa-sccvn.c
> @@ -5125,13 +5125,12 @@ static bool
> visit_reference_op_load (tree lhs, tree op, gimple *stmt)
> {
>bool changed = false;
> - tree last_vuse;
>tree result;
>vn_reference_t res;
>
> - last_vuse = gimple_vuse (stmt);
> - result = vn_reference_lookup (op, gimple_vuse (stmt),
> - default_vn_walk_kind, &res, true, &last_vuse);
> + tree vuse = gimple_vuse (stmt);
> + tree last_vuse = vuse;
> + result = vn_reference_lookup (op, vuse, default_vn_walk_kind, &res, true,
> &last_vuse);
>
>/* We handle type-punning through unions by value-numbering based
> on offset and size of the access. Be prepared to handle a
> @@ -5174,6 +5173,16 @@ visit_reference_op_load (tree lhs, tree op, gimple
> *stmt)
> {
>changed = set_ssa_val_to (lhs, lhs);
>vn_reference_insert (op, lhs, last_vuse, NULL_TREE);
> + if (vuse && SSA_VAL (last_vuse) != SSA_VAL (vuse))
> + {
> + if (dump_file && (dump_flags & TDF_DETAILS))
> + {
> + fprintf (dump_file, "Using extra use virtual operand ");
> + print_generic_expr (dump_file, last_vuse);
> + fprintf (dump_file, "\n");
> + }
> + vn_reference_insert (op, lhs, vuse, NULL_TREE);
> + }
> }
>
>return changed;
> --
> 2.31.1
