On Fri, Feb 03, 2017 at 06:07:56PM -0600, Segher Boessenkool wrote:
> On Fri, Feb 03, 2017 at 04:25:00PM -0500, Michael Meissner wrote:
> > +;; Return 1 if operand is either a vector constant of all 0 bits of a
> > vector
> > +;; constant of all 1 bits.
> > +(define_predicate "vector_int_same_bit"
> > + (match_code "const_vector")
> > +{
> > + if (GET_MODE_CLASS (mode) != MODE_VECTOR_INT)
> > + return 0;
> > +
> > + else
> > + return op == CONST0_RTX (mode) || op == CONSTM1_RTX (mode);
> > +})
>
> This predicate is unused as far as I see?
Right. It was used in my first attempt when I had a peephole2 to
eliminate the extra loads. Since I moved the processing to when we create the
conditional vector assignment and deleted the peephole2, I missed deleting the
predicate. Thanks for catching it.
> > + /* Optimize vec1 == vec2, to know the mask generates -1/0. */
> > + if (GET_MODE_CLASS (dest_mode) == MODE_VECTOR_INT)
> > {
> > - tmp = op_true;
> > - op_true = op_false;
> > - op_false = tmp;
> > + if (op_true == constant_m1 && op_false == constant_0)
> > + {
> > + emit_move_insn (dest, mask);
> > + return 1;
> > + }
> > +
> > + else if (op_true == constant_0 && op_false == constant_m1)
> > + {
> > + emit_insn (gen_rtx_SET (dest, gen_rtx_NOT (dest_mode, mask)));
> > + return 1;
> > + }
> > }
>
> Do you need to test for dest_mode == mask_mode here, like below?
Yes, because there is support for vcondv4siv4sf and vcondv4sfv4si where the
mask is one of V4SI or V4SF and the destination is the other. The dest_mode ==
mask_mode checks for that, and ignore that case.
I.e. something like:
static float af[1024], bf[1024], cf[1024];
static int di[1024], ei[1024];
// ..
for (i = 0; i < 1024; i++)
af[i] = (di[i] == ei[i]) ? bf[i] : cf[i];
>
> > + if (op_true == constant_m1 && dest_mode == mask_mode)
> > + op_true = mask;
> > + else if (!REG_P (op_true) && !SUBREG_P (op_true))
> > + op_true = force_reg (dest_mode, op_true);
> > +
> > + if (op_false == constant_0 && dest_mode == mask_mode)
> > + op_false = mask;
> > + else if (!REG_P (op_false) && !SUBREG_P (op_false))
> > + op_false = force_reg (dest_mode, op_false);
>
> Another thing you could try is, if either op_true or op_false is 0
> or -1, let the result be
> (mask & op_true) | (~mask & op_false)
>
> and let the rest of the optimisers sort it out (it's a single vor/vand
> or vorc/vandc, or a vnot, or nothing). A later improvement perhaps.
> Or does it already handle all cases now :-)
I don't know if it would handle it.
>
> Okay for trunk with the unused predicate removed, and the dest_mode ==
> mask_mode thing looked at. Thanks!
--
Michael Meissner, IBM
IBM, M/S 2506R, 550 King Street, Littleton, MA 01460-6245, USA
email: meiss...@linux.vnet.ibm.com, phone: +1 (978) 899-4797
