On Mon, 1 Nov 2021, Tamar Christina wrote:
> Hi All,
>
> The complex FMS detection is missing a check on if the nodes of the VEC_PERM
> has the amount of children we expect before it recurses.
>
> This check is there on MUL and FMA but was missing for FMS, due to this the
> compiler goes on further than it should and hits an assert.
>
> Bootstrapped Regtested on aarch64-none-linux-gnu, x86_64-pc-linux-gnu
> and no issues.
>
> Ok for master?
OK.
Thanks,
Richard.
> Thanks,
> Tamar
>
> gcc/ChangeLog:
>
> PR tree-optimizations/103007
> * tree-vect-slp-patterns.c (complex_fms_pattern::matches): Add elem
> check.
>
> gcc/testsuite/ChangeLog:
>
> PR tree-optimizations/103007
> * g++.dg/pr103007.C: New test.
>
> --- inline copy of patch --
> diff --git a/gcc/testsuite/g++.dg/pr103007.C b/gcc/testsuite/g++.dg/pr103007.C
> new file mode 100644
> index
> 0000000000000000000000000000000000000000..1631a85080039f29b83c97d2f62c66be9eac109f
> --- /dev/null
> +++ b/gcc/testsuite/g++.dg/pr103007.C
> @@ -0,0 +1,19 @@
> +/* { dg-do compile } */
> +/* { dg-additional-options "-O3" } */
> +
> +typedef float MushMeshVector[4];
> +struct MushMeshQuaternionPair {
> + void VectorRotate(MushMeshVector &);
> + MushMeshVector m_first;
> + MushMeshVector m_second;
> +};
> +void
> +MushMeshQuaternionPair::
> +VectorRotate(MushMeshVector &ioVec) {
> + ioVec[2] = (2 - m_first[1] + m_first[3] * 0);
> + ioVec[3] = (m_first[3] + m_first[1] - m_first[2] * 0);
> + float c = ioVec[2], d = ioVec[3];
> + ioVec[2] = (0 - d * m_second[1]);
> + ioVec[3] = (2 - c * m_second[1]);
> +}
> +
> diff --git a/gcc/tree-vect-slp-patterns.c b/gcc/tree-vect-slp-patterns.c
> index
> 6b37e9bac6f3f86a51d1a532a4c570a37af76eac..5e64a9bc417ab6b855e8791fd482dba23287f467
> 100644
> --- a/gcc/tree-vect-slp-patterns.c
> +++ b/gcc/tree-vect-slp-patterns.c
> @@ -1250,13 +1250,17 @@ complex_fms_pattern::matches (complex_operation_t op,
>
> auto childs = SLP_TREE_CHILDREN (nodes[0]);
> auto l0node = SLP_TREE_CHILDREN (childs[0]);
> - auto l1node = SLP_TREE_CHILDREN (childs[1]);
>
> /* Now operand2+4 may lead to another expression. */
> auto_vec<slp_tree> left_op, right_op;
> left_op.safe_splice (SLP_TREE_CHILDREN (l0node[1]));
> right_op.safe_splice (SLP_TREE_CHILDREN (nodes[1]));
>
> + /* If these nodes don't have any children then they're
> + not ones we're interested in. */
> + if (left_op.length () != 2 || right_op.length () != 2)
> + return IFN_LAST;
> +
> bool is_neg = vect_normalize_conj_loc (left_op);
>
> bool conj_first_operand = false;
>
>
>
--
Richard Biener <rguent...@suse.de>
SUSE Software Solutions Germany GmbH, Maxfeldstrasse 5, 90409 Nuernberg,
Germany; GF: Ivo Totev; HRB 36809 (AG Nuernberg)
