On Mon, Aug 19, 2024 at 01:55:38PM +0000, Tamar Christina wrote:
> So would this not be the simplest fix:
>
> diff --git a/gcc/tree-vect-patterns.cc b/gcc/tree-vect-patterns.cc
> index 87b3dc413b8..fcbc83a49f0 100644
> --- a/gcc/tree-vect-patterns.cc
> +++ b/gcc/tree-vect-patterns.cc
> @@ -4558,6 +4558,9 @@ vect_recog_sat_add_pattern (vec_info *vinfo,
> stmt_vec_info stmt_vinfo,
>
> if (gimple_unsigned_integer_sat_add (lhs, ops, NULL))
But then you call gimple_unsigned_integer_sat_add with mismatching types,
not sure if that is ok.
> {
> + if (TREE_CODE (ops[1]) == INTEGER_CST)
> + ops[1] = fold_convert (TREE_TYPE (ops[0]), ops[1]);
> +
This would be only ok if the conversion doesn't change the value
of the constant.
.ADD_OVERFLOW etc. could have e.g. int and unsigned arguments, you don't
want to change the latter to the former if the value has the most
significant bit set.
Similarly, .ADD_OVERFLOW could have e.g. unsigned and unsigned __int128
arguments, you don't want to truncate the constant.
So, you could e.g. do the fold_convert and then verify if
wi::to_widest on the old and new tree are equal, or you could check for
TREE_OVERFLOW if fold_convert honors that.
As I said, for INTEGER_CST operands of .ADD/SUB/MUL_OVERFLOW, the infinite
precision value (aka wi::to_widest) is all that matters.
Jakub
