On November 11, 2014 6:49:34 PM CET, Jakub Jelinek <ja...@redhat.com> wrote:
>On Tue, Nov 11, 2014 at 06:40:25PM +0100, Marek Polacek wrote:
>> --- gcc/fold-const.c
>> +++ gcc/fold-const.c
>> @@ -7862,9 +7862,15 @@ fold_unary_loc (location_t loc, enum tree_code
>code, tree type, tree op0)
>> return fold_view_convert_expr (type, op0);
>>
>> case NEGATE_EXPR:
>> - tem = fold_negate_expr (loc, arg0);
>> - if (tem)
>> - return fold_convert_loc (loc, type, tem);
>> + if (TREE_CODE (arg0) == INTEGER_CST
>> + || TREE_CODE (arg0) == REAL_CST
>> + || TYPE_OVERFLOW_WRAPS (type)
>> + || (flag_sanitize & SANITIZE_SI_OVERFLOW) == 0)
>> + {
>> + tem = fold_negate_expr (loc, arg0);
>> + if (tem)
>> + return fold_convert_loc (loc, type, tem);
>> + }
>
>a) if arg0 is INTEGER_CST, but e.g. INT_MIN, should we
> really fold it for -fsanitize=signed-integer-overflow
> (I'd say in that case we should fold only if it is not
> the minimum value)
Right. I think you should fix VRP instead.
>b) if the argument is not INTEGRAL_TYPE_P (type), shouldn't
> we fold no matter what flag_sanitize says? Is the REAL_CST
> case needed in that case?
Yes and no I guess.
Richard.
> Jakub
