> Am 24.02.2024 um 11:06 schrieb Richard Sandiford <richard.sandif...@arm.com>:
>
> During its forward pass, the SLP layout code tries to calculate
> the cost of a layout change on an incoming edge. This is taken
> as the minimum of two costs: one in which the source partition
> keeps its current layout (chosen earlier during the pass) and
> one in which the source partition switches to the new layout.
> The latter can sometimes be arranged by the backward pass.
>
> If only one of the costs is valid, the other cost was ignored.
> But the PR shows that this is not safe. If the source partition
> has layout 0 (the normal layout), we have to be prepared to handle
> the case in which that ends up being the only valid layout.
>
> Other code already accounts for this restriction, e.g. see
> the code starting with:
>
> /* Reject the layout if it would make layout 0 impossible
> for later partitions. This amounts to testing that the
> target supports reversing the layout change on edges
> to later partitions.
>
> Tested on aarch64-linux-gnu and x86_64-linux-gnu. OK to install?
Ok
Thanks,
Richard
> Richard
>
>
> gcc/
> PR tree-optimization/113205
> * tree-vect-slp.cc (vect_optimize_slp_pass::forward_cost): Reject
> the proposed layout if it does not allow a source partition with
> layout 2 to keep that layout.
>
> gcc/testsuite/
> PR tree-optimization/113205
> * gcc.dg/torture/pr113205.c: New test.
> ---
> gcc/testsuite/gcc.dg/torture/pr113205.c | 19 +++++++++++++++++++
> gcc/tree-vect-slp.cc | 4 ++++
> 2 files changed, 23 insertions(+)
> create mode 100644 gcc/testsuite/gcc.dg/torture/pr113205.c
>
> diff --git a/gcc/testsuite/gcc.dg/torture/pr113205.c
> b/gcc/testsuite/gcc.dg/torture/pr113205.c
> new file mode 100644
> index 00000000000..edfba7fcd0e
> --- /dev/null
> +++ b/gcc/testsuite/gcc.dg/torture/pr113205.c
> @@ -0,0 +1,19 @@
> +char a;
> +char *b, *c;
> +int d, e, f, g, h;
> +int *i;
> +
> +void
> +foo (void)
> +{
> + unsigned p;
> + d = i[0];
> + e = i[1];
> + f = i[2];
> + g = i[3];
> + p = d * b[0];
> + p += f * c[h];
> + p += e * b[h];
> + p += g * c[h];
> + a = (p + 8000) >> (__SIZEOF_INT__ * __CHAR_BIT__ / 2);
> +}
> diff --git a/gcc/tree-vect-slp.cc b/gcc/tree-vect-slp.cc
> index 7cf9504398c..895f4f7fb6b 100644
> --- a/gcc/tree-vect-slp.cc
> +++ b/gcc/tree-vect-slp.cc
> @@ -5034,6 +5034,10 @@ vect_optimize_slp_pass::forward_cost (graph_edge *ud,
> unsigned int from_node_i,
> cost.split (from_partition.out_degree);
> cost.add_serial_cost (edge_cost);
> }
> + else if (from_partition.layout == 0)
> + /* We must allow the source partition to have layout 0 as a fallback,
> + in case all other options turn out to be impossible. */
> + return cost;
>
> /* Take the minimum of that cost and the cost that applies if
> FROM_PARTITION instead switches to TO_LAYOUT_I. */
> --
> 2.25.1
>
