Re: [RFC] Vectorization of indexed elements

[Richard Biener <rguenther@suse.de>]

On Mon, 30 Sep 2013, Vidya Praveen wrote:

> On Mon, Sep 30, 2013 at 02:19:32PM +0100, Richard Biener wrote:
> > On Mon, 30 Sep 2013, Vidya Praveen wrote:
> > 
> > > On Fri, Sep 27, 2013 at 04:19:45PM +0100, Vidya Praveen wrote:
> > > > On Fri, Sep 27, 2013 at 03:50:08PM +0100, Vidya Praveen wrote:
> > > > [...]
> > > > > > > I can't really insist on the single lane load.. something like:
> > > > > > > 
> > > > > > > vc:V4SI[0] = c
> > > > > > > vt:V4SI = vec_duplicate:V4SI (vec_select:SI vc:V4SI 0)
> > > > > > > va:V4SI = vb:V4SI <op> vt:V4SI
> > > > > > > 
> > > > > > > Or is there any other way to do this?
> > > > > > 
> > > > > > Can you elaborate on "I can't really insist on the single lane load"?
> > > > > > What's the single lane load in your example? 
> > > > > 
> > > > > Loading just one lane of the vector like this:
> > > > > 
> > > > > vc:V4SI[0] = c // from the above scalar example
> > > > > 
> > > > > or 
> > > > > 
> > > > > vc:V4SI[0] = c[2] 
> > > > > 
> > > > > is what I meant by single lane load. In this example:
> > > > > 
> > > > > t = c[2] 
> > > > > ...
> > > > > vb:v4si = b[0:3] 
> > > > > vc:v4si = { t, t, t, t }
> > > > > va:v4si = vb:v4si <op> vc:v4si 
> > > > > 
> > > > > If we are expanding the CONSTRUCTOR as vec_duplicate at vec_init, I cannot
> > > > > insist 't' to be vector and t = c[2] to be vect_t[0] = c[2] (which could be 
> > > > > seen as vec_select:SI (vect_t 0) ). 
> > > > > 
> > > > > > I'd expect the instruction
> > > > > > pattern as quoted to just work (and I hope we expand an uniform
> > > > > > constructor { a, a, a, a } properly using vec_duplicate).
> > > > > 
> > > > > As much as I went through the code, this is only done using vect_init. It is
> > > > > not expanded as vec_duplicate from, for example, store_constructor() of expr.c
> > > > 
> > > > Do you see any issues if we expand such constructor as vec_duplicate directly 
> > > > instead of going through vect_init way? 
> > > 
> > > Sorry, that was a bad question.
> > > 
> > > But here's what I would like to propose as a first step. Please tell me if this
> > > is acceptable or if it makes sense:
> > > 
> > > - Introduce standard pattern names 
> > > 
> > > "vmulim4" - vector muliply with second operand as indexed operand
> > > 
> > > Example:
> > > 
> > > (define_insn "vmuliv4si4"
> > >    [set (match_operand:V4SI 0 "register_operand")
> > >         (mul:V4SI (match_operand:V4SI 1 "register_operand")
> > >                   (vec_duplicate:V4SI
> > >                     (vec_select:SI
> > >                       (match_operand:V4SI 2 "register_operand")
> > >                       (match_operand:V4SI 3 "immediate_operand)))))]
> > >  ...
> > > )
> > 
> > We could factor this with providing a standard pattern name for
> > 
> > (define_insn "vdupi<mode>"
> >   [set (match_operand:<mode> 0 "register_operand")
> >        (vec_duplicate:<mode>
> >           (vec_select:<scalarmode>
> >              (match_operand:<mode> 1 "register_operand")
> >              (match_operand:SI 2 "immediate_operand))))]
> 
> This is good. I did think about this but then I thought of avoiding the need
> for combiner patterns :-) 
> 
> But do you find the lane specific mov pattern I proposed, acceptable? 

The specific mul pattern?  As said, consider factoring to vdupi to
avoid an explosion in required special optabs.

> > (you use V4SI for the immediate?  
> 
> Sorry typo again!! It should've been SI.
> 
> > Ideally vdupi has another custom
> > mode for the vector index).
> > 
> > Note that this factored pattern is already available as vec_perm_const!
> > It is simply (vec_perm_const:V4SI <source> <source> <immediate-selector>).
> > 
> > Which means that on the GIMPLE level we should try to combine
> > 
> > el_4 = BIT_FIELD_REF <v_3, ...>;
> > v_5 = { el_4, el_4, ... };
> 
> I don't think we reach this state at all for the scenarios in discussion.
> what we generally have is:
> 
>  el_4 = MEM_REF < array + index*size >
>  v_5 = { el_4, ... }
> 
> Or am I missing something?

Well, but in that case I doubt it is profitable (or even valid!) to
turn this into a vector lane load from the array.  If it is profitable
to perform a vector read (because we're going to use the other elements
of the vector as well) then the vectorizer should produce a vector
load and materialize the uniform vector from one of its elements.

Maybe at this point you should show us a compilable C testcase
with a loop that should be vectorized using your instructions in
the end?

Thanks,
Richard.