On Wed, Mar 27, 2024 at 01:42:03PM +0100, Richard Biener wrote:
> The comment says it was added for (char)_Bool, probably
> gcc.dg/tree-ssa/vce-1.c will fail. But I'm not sure this optimization
> is important, it seems early SRA introduces a V_C_E here and then
> phiopt the conversion to unsigned char. But this is a V_C_E of
> _Bool -> unsigned char -> _Bool. We seem to apply both this
> pattern and then drop the V_C_E.
>
> Maybe we can reasonably optimize the case where type == TREE_TYPE (@1)
> and the intermediate precision is bigger or equal?
If type is INTEGRAL_TYPE_P and has TYPE_PRECISION <= that of TREE_TYPE (@1),
sure, it can be optimized whenever TREE_TYPE (@0) has precision >= that,
no matter what TYPE_SIGN TREE_TYPE (@1) has.
Anyway, will try to cook up something.
Jakub
