On Fri, Oct 30, 2015 at 11:26 AM, Marc Glisse <marc.gli...@inria.fr> wrote:
> On Fri, 30 Oct 2015, Richard Biener wrote:
>
>> +/* Convert (A + A) * C -> A * 2 * C. */
>> +(simplify
>> + (mult:c (convert? (plus @0 @0)) (convert? @1))
>> + (if (tree_nop_conversion_p (TREE_TYPE (@0), type))
>> + (convert (mult @0 (mult { build_int_cst (TREE_TYPE (@1), 2); } @1)))))
>> +(simplify
>> + (mult:c (convert? (plus @0 @0)) INTEGER_CST@1)
>> + (if (tree_nop_conversion_p (TREE_TYPE (@0), type))
>> + (convert (mult @0 (mult { build_int_cst (TREE_TYPE (@1), 2); } @1)))))
>>
>> fold-const.c only handles constant C, so we only need to 2nd pattern.
>> Also the :c on the mult in that is not needed due to canonicalization
>> rules.
>> Please build the result of the inner multiplication directly.
>> I think the fold-const.c code has an overflow issue when the outer
>> multiplication
>> is signed and the inner addition unsigned. (signed)((unsigned)INT_MAX
>> + (unsigned)INT_MAX)*2
>> is valid but INT_MAX * 4 is not as it overflows. So I think we should
>> _not_ allow
>> nop conversions here (it's fine if all ops are signed or unsigned).
>
>
> Is there a reason why the simple transformation A+A->2*A is not
> generally a desirable canonicalization? We currently restrict it to
> SCALAR_FLOAT_TYPE_P.
No special reason I know of.
>
> --
> Marc Glisse
