On 21 October 2011 14:52, Jakub Jelinek <ja...@redhat.com> wrote:
> On Fri, Oct 21, 2011 at 02:37:06PM +0200, Ira Rosen wrote:
>> > @@ -1620,7 +1615,13 @@ vectorizable_call (gimple stmt, gimple_s
>> >
>> > gcc_assert (!gimple_vuse (stmt));
>> >
>> > - if (modifier == NARROW)
>> > + if (slp_node || PURE_SLP_STMT (stmt_info))
>> > + {
>> > + if (modifier != NONE)
>> > + return false;
>> > + ncopies = 1;
>> > + }
>>
>> If you want to bail out if it's SLP and modifier != NONE, this check
>> is not enough. PURE_SLP means the stmt is not used outside the SLP
>> instance, so for hybrid SLP stmts (those that have uses outside SLP)
>> this check will not work. You need
>>
>> if (modifier != NONE && STMT_SLP_TYPE (stmt_info))
>> return false;
>
> I just blindly copied what vectorizable_operation does, without
> too much understanding what PURE_SLP_STMT or STMT_SLP_TYPE etc. mean.
> Didn't get that far.
> But modifier != NONE && something would sometimes allow modifier != NONE
> through, which at least the current code isn't prepared to handle.
> Did you mean || instead?
But it's OK to allow modifier != NONE if it's not SLP, so we need &&, no?
Something like:
if (modifier != NONE && STMT_SLP_TYPE (stmt_info))
return false;
if (slp_node || PURE_SLP_STMT (stmt_info))
ncopies = 1;
else if (modifier == NARROW)
ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_out;
else
ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_in;
>
>> But I wonder why not allow different type sizes? I see that we fail in
>> such cases in vectorizable_conversion too, but I think we should
>> support this as well.
>
> Merely because I don't know SLP enough, vectorizable_operation also
> handles just same size to same size, so I didn't have good examples
> on how to do it. For loops narrowing or widening operations are
> handled through ncopies != 1, but for SLP it seems it is always
> asserted it is 1...
There are vectorizable_type_promotion/demotion, and for the rest the
copies are "hidden" inside multiple vector operands that you get from
vect_get_vec_defs. But, of course, there is not need to handle
modifier == NARROW for SLP at the moment. I was just wondering out
loud.
Ira
>
> Jakub
>
