On Mon, Mar 23, 2020 at 10:53 AM Yangfei (Felix) <felix.y...@huawei.com> wrote:
>
> Hi,
>
> I created a bug for this issue:
> https://gcc.gnu.org/bugzilla/show_bug.cgi?id=94269
> Looks like widening_mul phase may move multiply instruction from outside
> the loop to inside the loop, merging with one add instruction inside the loop.
> This will increase the cost of the loop at least on aarch64 (4 cycles vs 1
> cycle). I think widening_mul should consider block frequency when doing such
> a combination.
> I mean something like:
As written in the PR I'd follow fma generation and restrict defs to the same BB.
Richard.
> diff --git a/gcc/tree-ssa-math-opts.c b/gcc/tree-ssa-math-opts.c
> index 54ba035..4439452 100644
> --- a/gcc/tree-ssa-math-opts.c
> +++ b/gcc/tree-ssa-math-opts.c
> @@ -2721,7 +2721,10 @@ convert_plusminus_to_widen (gimple_stmt_iterator *gsi,
> gimple *stmt,
> {
> if (!has_single_use (rhs1)
> || !is_widening_mult_p (rhs1_stmt, &type1, &mult_rhs1,
> - &type2, &mult_rhs2))
> + &type2, &mult_rhs2)
> + || (gimple_bb (rhs1_stmt) != gimple_bb (stmt)
> + && gimple_bb (rhs1_stmt)->count.to_frequency(cfun)
> + < gimple_bb (stmt)->count.to_frequency(cfun)))
> return false;
> add_rhs = rhs2;
> conv_stmt = conv1_stmt;
> @@ -2730,7 +2733,10 @@ convert_plusminus_to_widen (gimple_stmt_iterator *gsi,
> gimple *stmt,
> {
> if (!has_single_use (rhs2)
> || !is_widening_mult_p (rhs2_stmt, &type1, &mult_rhs1,
> - &type2, &mult_rhs2))
> + &type2, &mult_rhs2)
> + || (gimple_bb (rhs2_stmt) != gimple_bb (stmt)
> + && gimple_bb (rhs2_stmt)->count.to_frequency(cfun)
> + < gimple_bb (stmt)->count.to_frequency(cfun)))
> return false;
> add_rhs = rhs1;
> conv_stmt = conv2_stmt;
>
> Comments?
>
> Thanks,
> Felix
