Richard Henderson <r...@redhat.com> writes:
> +/* Recognize broadcast patterns for the Loongson. */
> +
> +static bool
> +mips_expand_vpc_loongson_bcast (struct expand_vec_perm_d *d)
> +{
> + unsigned i, elt;
> + rtx t0, t1;
> +
> + if (!(TARGET_HARD_FLOAT && TARGET_LOONGSON_VECTORS))
> + return false;
> + /* Note that we've already matched V2SI via punpck and V4HI via pshufh. */
> + if (d->vmode != V8QImode)
> + return false;
> + if (!d->one_vector_p)
> + return false;
> +
> + elt = d->perm[0];
> + for (i = 1; i < 8; ++i)
> + if (d->perm[i] != elt)
> + return false;
> +
> + if (d->testing_p)
> + return true;
> +
> + /* With one interleave we put two of the desired element adjacent. */
> + t0 = gen_reg_rtx (V8QImode);
> + if (elt < 4)
> + emit_insn (gen_loongson_punpcklbh (t0, d->op0, d->op0));
> + else
> + emit_insn (gen_loongson_punpckhbh (t0, d->op0, d->op0));
> +
> + /* Shuffle that one HImode element into all locations. */
> + elt &= 3;
> + elt *= 0x55;
> + t1 = gen_reg_rtx (V4HImode);
> + emit_insn (gen_loongson_pshufh (t1, gen_lowpart (V4HImode, t0),
> + force_reg (SImode, GEN_INT (elt))));
> +
> + emit_move_insn (d->target, gen_lowpart (V8QImode, t1));
> + return true;
Probably one of those where you had to stop following the rathole,
but could this be generalised to handle pairs in which perm[1]
== perm[0] + 8? Something like:
> + if (!(TARGET_HARD_FLOAT && TARGET_LOONGSON_VECTORS))
> + return false;
> + /* Note that we've already matched V2SI via punpck and V4HI via pshufh. */
> + if (d->vmode != V8QImode)
> + return false;
> + if (d->perm[1] != (d->one_vector_p ? : d->perm[0] : d->perm[0] + 8))
> + return false;
> +
> + for (i = 2; i < 8; ++i)
> + if (d->perm[i] != d->perm[i & 1])
> + return false;
then use both d->op0 and d->op1 in the pack?
> + memset (&d, 0, sizeof(d));
missing space
Richard
