On Fri, Feb 04, 2022 at 10:53:22AM +0100, Eric Botcazou wrote:
> > Well, yes, we have to fix it.
>
> Here's the fix we agreed upon in the audit trail, OK for the mainline?
>
> PR tree-optimization/104356
> * match.pd (X / bool_range_Y is X): Add guard.
> (X / X is one): Likewise.
> (X / abs (X) is X < 0 ? -1 : 1): Likewise.
> (X / -X is -1): Likewise.
> (1 / X -> X == 1): Likewise.
Looks mostly good, just a few nits.
> --- a/gcc/match.pd
> +++ b/gcc/match.pd
> @@ -401,27 +401,35 @@ DEFINE_INT_AND_FLOAT_ROUND_FN (RINT)
> /* X / bool_range_Y is X. */
> (simplify
> (div @0 SSA_NAME@1)
> - (if (INTEGRAL_TYPE_P (type) && ssa_name_has_boolean_range (@1))
> + (if (INTEGRAL_TYPE_P (type)
> + && ssa_name_has_boolean_range (@1)
> + && !flag_non_call_exceptions)
ssa_name_has_boolean_range call is certainly more expensive than
!flag_non_call_exceptions check, can you swap those two?
> @@ -444,6 +452,7 @@ DEFINE_INT_AND_FLOAT_ROUND_FN (RINT)
> (trunc_div integer_onep@0 @1)
> (if (INTEGRAL_TYPE_P (type)
> && !integer_zerop (@1)
> + && (!flag_non_call_exceptions || tree_expr_nonzero_p (@1))
> && TYPE_PRECISION (type) > 1)
And similarly, TYPE_PRECISION (type) > 1 check is very cheap, can
it be done before the && !integer_zerop (@1) line?
I admit this one is already preexisting, but tree_expr_nonzero_p
can be quite expensive.
Otherwise LGTM.
Jakub
