> -----Original Message-----
> From: Richard Biener <[email protected]>
> Sent: 07 August 2026 13:13
> To: [email protected]
> Cc: [email protected]; Tamar Christina <[email protected]>
> Subject: [PATCH 3/3] tree-optimization/126099 - SLP subgraph merging with
> low/highpart extracts
>
> The following implements merging of SLP subgraphs that overlap in
> their vectors low or highparts or in full vector copies using VEC_PERM
> SLP nodes which know how to perform these extractions, but not more
> in general.
>
> The operation is to build a reverse mapping of scalar stmt to SLP node
> for possible extraction starts - without knowing the extraction vector
> type that's even lanes of SLP nodes with an even number of lanes. And
> then for each SLP node lane zero checking whether it's fully contained
> in one of the candidates with the restrictions implied by
> vectorizable_slp_permutation.
>
> The motivating testcase is in PR126053 coming from 508.namd_r. The
> testcases gcc.dg/vect/bb-slp-pr126099-{4,5}.c are cases where schedule
> verification strips one of the graph entries as subgraph merging
> otherwise causes SSA verification failures.
>
> Bootstrapped and tested on x86_64-unknown-linux-gnu. I've also
> built SPEC CPU 2017 in various ways (which found quite some issues,
> all fixed now).
>
> Any comments?
>
> Thanks,
> Richard.
>
> PR tree-optimization/126099
> PR tree-optimization/126053
> * tree-vect-slp.cc (vect_cse_gather_part_starts): New function.
> (vect_cse_slp_node_parts): Likewise.
> (vect_optimize_slp): For BB SLP CSE to low/highparts of
> other nodes.
>
> * gcc.dg/vect/bb-slp-pr126099-1.c: New testcase.
> * gcc.dg/vect/bb-slp-pr126099-2.c: Likewise.
> * gcc.dg/vect/bb-slp-pr126099-3.c: Likewise.
> * gcc.dg/vect/bb-slp-pr126099-4.c: Likewise.
> * gcc.dg/vect/bb-slp-pr126099-5.c: Likewise.
> * gcc.dg/vect/costmodel/x86_64/costmodel-pr126053.c: Likewise.
> ---
> gcc/testsuite/gcc.dg/vect/bb-slp-pr126099-1.c | 24 +++
> gcc/testsuite/gcc.dg/vect/bb-slp-pr126099-2.c | 24 +++
> gcc/testsuite/gcc.dg/vect/bb-slp-pr126099-3.c | 24 +++
> gcc/testsuite/gcc.dg/vect/bb-slp-pr126099-4.c | 22 +++
> gcc/testsuite/gcc.dg/vect/bb-slp-pr126099-5.c | 23 +++
> .../costmodel/x86_64/costmodel-pr126053.c | 131 +++++++++++++++
> gcc/tree-vect-slp.cc | 154 ++++++++++++++++++
> 7 files changed, 402 insertions(+)
> create mode 100644 gcc/testsuite/gcc.dg/vect/bb-slp-pr126099-1.c
> create mode 100644 gcc/testsuite/gcc.dg/vect/bb-slp-pr126099-2.c
> create mode 100644 gcc/testsuite/gcc.dg/vect/bb-slp-pr126099-3.c
> create mode 100644 gcc/testsuite/gcc.dg/vect/bb-slp-pr126099-4.c
> create mode 100644 gcc/testsuite/gcc.dg/vect/bb-slp-pr126099-5.c
> create mode 100644
> gcc/testsuite/gcc.dg/vect/costmodel/x86_64/costmodel-pr126053.c
>
> diff --git a/gcc/testsuite/gcc.dg/vect/bb-slp-pr126099-1.c
> b/gcc/testsuite/gcc.dg/vect/bb-slp-pr126099-1.c
> new file mode 100644
> index 00000000000..7c3d08c2f56
> --- /dev/null
> +++ b/gcc/testsuite/gcc.dg/vect/bb-slp-pr126099-1.c
> @@ -0,0 +1,24 @@
> +/* { dg-do compile } */
> +/* { dg-require-effective-target vect_long } */
> +/* { dg-additional-options "-mavx2" { target avx2 } } */
> +
> +void foo (long *p, long *q, long *r)
> +{
> + long tem0 = r[0];
> + long tem1 = r[1];
> + long tem2 = r[2];
> + long tem3 = r[3];
> + tem0 = tem0 + 1;
> + tem1 = tem1 + 2;
> + tem2 = tem2 + 3;
> + tem3 = tem3 + 4;
> + p[0] = tem0;
> + p[1] = tem1;
> + q[0] = tem0;
> + q[1] = tem1;
> + q[2] = tem2;
> + q[3] = tem3;
> +}
> +
> +/* { dg-final { scan-tree-dump "CSEd node\[^\n\r\]*lowpart" "slp2" } } */
> +/* { dg-final { scan-tree-dump "BIT_FIELD_REF" "slp2" { target avx2 } } } */
> diff --git a/gcc/testsuite/gcc.dg/vect/bb-slp-pr126099-2.c
> b/gcc/testsuite/gcc.dg/vect/bb-slp-pr126099-2.c
> new file mode 100644
> index 00000000000..ce6cda61683
> --- /dev/null
> +++ b/gcc/testsuite/gcc.dg/vect/bb-slp-pr126099-2.c
> @@ -0,0 +1,24 @@
> +/* { dg-do compile } */
> +/* { dg-require-effective-target vect_long } */
> +/* { dg-additional-options "-mavx2" { target avx2 } } */
> +
> +void foo (long *p, long *q, long *r)
> +{
> + long tem0 = r[0];
> + long tem1 = r[1];
> + long tem2 = r[2];
> + long tem3 = r[3];
> + tem0 = tem0 + 1;
> + tem1 = tem1 + 2;
> + tem2 = tem2 + 3;
> + tem3 = tem3 + 4;
> + p[0] = tem2;
> + p[1] = tem3;
> + q[0] = tem0;
> + q[1] = tem1;
> + q[2] = tem2;
> + q[3] = tem3;
> +}
> +
> +/* { dg-final { scan-tree-dump "CSEd node\[^\n\r\]*highpart" "slp2" } } */
> +/* { dg-final { scan-tree-dump "BIT_FIELD_REF" "slp2" { target avx2 } } } */
> diff --git a/gcc/testsuite/gcc.dg/vect/bb-slp-pr126099-3.c
> b/gcc/testsuite/gcc.dg/vect/bb-slp-pr126099-3.c
> new file mode 100644
> index 00000000000..cfd58317ef4
> --- /dev/null
> +++ b/gcc/testsuite/gcc.dg/vect/bb-slp-pr126099-3.c
> @@ -0,0 +1,24 @@
> +/* { dg-do compile } */
> +/* { dg-require-effective-target vect_long } */
> +/* { dg-additional-options "-mavx2" { target avx2 } } */
> +
> +void foo (long *p, long *q, long *r, long *s)
> +{
> + long tem0 = r[0];
> + long tem1 = r[1];
> + long tem2 = r[2];
> + long tem3 = r[3];
> + tem0 = tem0 + s[0];
> + tem1 = tem1 - s[1];
> + tem2 = tem2 + s[2];
> + tem3 = tem3 - s[3];
> + p[0] = tem2;
> + p[1] = tem3;
> + q[0] = tem0;
> + q[1] = tem1;
> + q[2] = tem2;
> + q[3] = tem3;
> +}
> +
> +/* { dg-final { scan-tree-dump "CSEd node\[^\n\r\]*highpart" "slp2" } } */
> +/* { dg-final { scan-tree-dump "BIT_FIELD_REF" "slp2" { target avx2 } } } */
> diff --git a/gcc/testsuite/gcc.dg/vect/bb-slp-pr126099-4.c
> b/gcc/testsuite/gcc.dg/vect/bb-slp-pr126099-4.c
> new file mode 100644
> index 00000000000..0a2148eb832
> --- /dev/null
> +++ b/gcc/testsuite/gcc.dg/vect/bb-slp-pr126099-4.c
> @@ -0,0 +1,22 @@
> +/* { dg-do compile } */
> +
> +typedef unsigned v2si __attribute__((vector_size(8)));
> +typedef unsigned v4si __attribute__((vector_size(16)));
> +
> +unsigned bar(unsigned);
> +
> +v4si x;
> +v2si y;
> +
> +void foo(unsigned *p, unsigned i, unsigned j, unsigned k)
> +{
> + unsigned tem0 = p[0] ^ i;
> + unsigned tem1 = p[1] ^ j;
> + unsigned p2 = p[2];
> + unsigned p3 = p[3];
> + y = (v2si) { tem0, tem1 };
> + k = bar (k);
> + unsigned tem2 = p2 ^ k;
> + unsigned tem3 = p3 ^ k;
> + x = (v4si) { tem0, tem1, tem2, tem3 };
> +}
> diff --git a/gcc/testsuite/gcc.dg/vect/bb-slp-pr126099-5.c
> b/gcc/testsuite/gcc.dg/vect/bb-slp-pr126099-5.c
> new file mode 100644
> index 00000000000..1b1adf8c817
> --- /dev/null
> +++ b/gcc/testsuite/gcc.dg/vect/bb-slp-pr126099-5.c
> @@ -0,0 +1,23 @@
> +/* { dg-do compile } */
> +
> +unsigned bar(unsigned);
> +
> +unsigned x[4];
> +unsigned y[2];
> +
> +void foo(unsigned int *p, unsigned int i, unsigned int j, unsigned int k)
> +{
> + unsigned int tem0 = p[0] ^ i;
> + unsigned int tem1 = p[1] ^ j;
> + unsigned p2 = p[2];
> + unsigned p3 = p[3];
> + y[0] = tem0;
> + y[1] = tem1;
> + k = bar (k);
> + unsigned int tem2 = p2 ^ k;
> + unsigned int tem3 = p3 ^ k;
> + x[0] = tem0;
> + x[1] = tem1;
> + x[2] = tem2;
> + x[3] = tem3;
> +}
> diff --git a/gcc/testsuite/gcc.dg/vect/costmodel/x86_64/costmodel-
> pr126053.c b/gcc/testsuite/gcc.dg/vect/costmodel/x86_64/costmodel-
> pr126053.c
> new file mode 100644
> index 00000000000..f129a36b47b
> --- /dev/null
> +++ b/gcc/testsuite/gcc.dg/vect/costmodel/x86_64/costmodel-pr126053.c
> @@ -0,0 +1,131 @@
> +/* { dg-do compile } */
> +/* { dg-additional-options "-O3 -fno-signed-zeros -march=x86-64-v3 -fopt-
> info-vec" } */
> +
> +typedef double BigReal;
> +
> +struct Position {
> + BigReal x, y, z;
> +};
> +
> +struct CompAtom {
> + struct Position position;
> + float charge;
> + short vdwType;
> + unsigned char partition;
> + unsigned char nonbondedGroupSize;
> +};
> +
> +struct Force {
> + BigReal x, y, z;
> +};
> +
> +struct SimParams {
> + BigReal offset_x, offset_y, offset_z;
> +};
> +
> +enum { vXX, vXY, vXZ, vYY, vYZ, vZZ, fvXX, fvXY, fvXZ, fvYY, fvYZ, fvZZ };
> +void calc_pair_energy_fullelect(
> + const struct CompAtom *__restrict p_0,
> + const struct CompAtom *__restrict p_1,
> + const struct SimParams *__restrict params, const int *__restrict
> pairlist_n,
> + const int *__restrict pairlist_m, const int *__restrict npair_n_list,
> + const int *__restrict npair_m_list, int i_upper,
> + const BigReal *__restrict force_r_vals, struct Force *__restrict f_0,
> + struct Force *__restrict f_1, BigReal *__restrict reduction) {
> + BigReal virial_xx = 0, virial_xy = 0, virial_xz = 0;
> + BigReal virial_yy = 0, virial_yz = 0, virial_zz = 0;
> + BigReal fullElectVirial_xx = 0, fullElectVirial_xy = 0,
> + fullElectVirial_xz = 0;
> + BigReal fullElectVirial_yy = 0, fullElectVirial_yz = 0,
> + fullElectVirial_zz = 0;
> +
> + int pn = 0, pm = 0;
> + for (int i = 0; i < i_upper; ++i) {
> + const struct CompAtom *p_i = p_0 + i;
> + const BigReal p_i_x = params->offset_x + p_i->position.x;
> + const BigReal p_i_y = params->offset_y + p_i->position.y;
> + const BigReal p_i_z = params->offset_z + p_i->position.z;
> +
> + BigReal f_i_x = 0, f_i_y = 0, f_i_z = 0;
> +
> + {
> + const int npairi = npair_n_list[i];
> + const int *pli = pairlist_n + pn;
> + const BigReal *fr = force_r_vals + pn;
> + for (int k = 0; k < npairi; ++k) {
> + const int j = pli[k];
> + const struct CompAtom *p_j = p_1 + j;
> + struct Force *f_j = f_1 + j;
> + const BigReal p_ij_x = p_i_x - p_j->position.x;
> + const BigReal p_ij_y = p_i_y - p_j->position.y;
> + const BigReal p_ij_z = p_i_z - p_j->position.z;
> + const BigReal force_r = fr[k];
> + BigReal tmp_x = force_r * p_ij_x;
> + virial_xx += tmp_x * p_ij_x;
> + virial_xy += tmp_x * p_ij_y;
> + virial_xz += tmp_x * p_ij_z;
> + f_i_x += tmp_x;
> + f_j->x -= tmp_x; /* { dg-optimized "basic block part vectorized
> using 16
> byte vectors" } */
> + BigReal tmp_y = force_r * p_ij_y;
> + virial_yy += tmp_y * p_ij_y;
> + virial_yz += tmp_y * p_ij_z;
> + f_i_y += tmp_y;
> + f_j->y -= tmp_y;
> + BigReal tmp_z = force_r * p_ij_z;
> + virial_zz += tmp_z * p_ij_z;
> + f_i_z += tmp_z;
> + f_j->z -= tmp_z;
> + }
> + pn += npairi;
> + }
> + {
> + const int npairi = npair_m_list[i];
> + const int *pli = pairlist_m + pm;
> + const BigReal *fr = force_r_vals + pm;
> + for (int k = 0; k < npairi; ++k) {
> + const int j = pli[k];
> + const struct CompAtom *p_j = p_1 + j;
> + struct Force *f_j = f_1 + j;
> + const BigReal p_ij_x = p_i_x - p_j->position.x;
> + const BigReal p_ij_y = p_i_y - p_j->position.y;
> + const BigReal p_ij_z = p_i_z - p_j->position.z;
> + const BigReal force_r = fr[k];
> + BigReal tmp_x = force_r * p_ij_x;
> + virial_xx += tmp_x * p_ij_x;
> + virial_xy += tmp_x * p_ij_y;
> + virial_xz += tmp_x * p_ij_z;
> + f_i_x += tmp_x;
> + f_j->x -= tmp_x; /* { dg-optimized "basic block part vectorized
> using 16
> byte vectors" } */
> +
> + BigReal tmp_y = force_r * p_ij_y;
> + virial_yy += tmp_y * p_ij_y;
> + virial_yz += tmp_y * p_ij_z;
> + f_i_y += tmp_y;
> + f_j->y -= tmp_y;
> + BigReal tmp_z = force_r * p_ij_z;
> + virial_zz += tmp_z * p_ij_z;
> + f_i_z += tmp_z;
> + f_j->z -= tmp_z;
> + }
> + pm += npairi;
> + }
> +
> + f_0[i].x += f_i_x; /* { dg-optimized "basic block part vectorized using
> 16
> byte vectors" } */
> +
> + f_0[i].y += f_i_y;
> + f_0[i].z += f_i_z;
> + }
> +
> + reduction[vXX] += virial_xx; /* { dg-optimized "basic block part vectorized
> using 32 byte vectors" } */
> + reduction[vXY] += virial_xy;
> + reduction[vXZ] += virial_xz;
> + reduction[vYY] += virial_yy;
> + reduction[vYZ] += virial_yz; /* { dg-optimized "basic block part vectorized
> using 16 byte vectors" } */
> + reduction[vZZ] += virial_zz;
> + reduction[fvXX] += fullElectVirial_xx;
> + reduction[fvXY] += fullElectVirial_xy;
> + reduction[fvXZ] += fullElectVirial_xz;
> + reduction[fvYY] += fullElectVirial_yy;
> + reduction[fvYZ] += fullElectVirial_yz;
> + reduction[fvZZ] += fullElectVirial_zz;
> +}
> diff --git a/gcc/tree-vect-slp.cc b/gcc/tree-vect-slp.cc
> index 5ba19877f9e..944e36e490e 100644
> --- a/gcc/tree-vect-slp.cc
> +++ b/gcc/tree-vect-slp.cc
> @@ -8494,6 +8494,127 @@ vect_cse_slp_nodes
> (scalar_stmts_to_slp_tree_map_t *bst_map, slp_tree& node)
> *bst_map->get (SLP_TREE_SCALAR_STMTS (node)) = node;
> }
>
> +/* Associate stmts with possible starts of a subset of lanes of NODE
> + in PART_STARTS. */
> +
> +static void
> +vect_cse_gather_part_starts (hash_set<slp_tree> &visited,
> + vec<vec<slp_tree>> part_starts, slp_tree node)
> +{
> + /* CSEing external nodes complicates scheduling since we materialize
> + those at the latest position, so avoid that. */
> + if (SLP_TREE_DEF_TYPE (node) != vect_internal_def
> + || visited.add (node))
> + return;
> +
> + /* Besides some VEC_PERM_EXPR, two-operator nodes also lack scalar
> stmts
> + and thus CSE doesn't work. For now gather two-lane aligned starts
> + of nodes with a multiple of two number of lanes. */
> + if (!SLP_TREE_SCALAR_STMTS (node).is_empty ()
> + && SLP_TREE_LANES (node) > 2
> + && (SLP_TREE_LANES (node) & 1) == 0)
> + {
> + auto_vec<unsigned, 8> uids;
> + for (unsigned i = 0; i < SLP_TREE_LANES (node); i += 2)
> + {
> + stmt_vec_info s = SLP_TREE_SCALAR_STMTS (node)[i];
> + if (!s)
> + continue;
> + unsigned uid = gimple_uid (s->stmt);
> + if (!uids.contains (uid))
> + {
> + uids.safe_push (uid);
> + part_starts[uid].safe_push (node);
> + }
> + }
> + }
> +
> + for (slp_tree &child : SLP_TREE_CHILDREN (node))
> + if (child)
> + vect_cse_gather_part_starts (visited, part_starts, child);
> +}
> +
> +/* Apply CSE to NODE and its children using lowparts of nodes in BST_MAP.
> */
> +
> +static void
> +vect_cse_slp_node_parts (hash_set<slp_tree> &visited,
> + const vec<vec<slp_tree>> part_starts,
> + vec<slp_tree> &drops, slp_tree node)
> +{
> + if (SLP_TREE_DEF_TYPE (node) != vect_internal_def
> + || visited.add (node))
> + return;
> +
> + /* Besides some VEC_PERM_EXPR, two-operator nodes also
> + lack scalar stmts and thus CSE doesn't work. */
> + unsigned HOST_WIDE_INT c;
> + if (!SLP_TREE_SCALAR_STMTS (node).is_empty ()
> + && SLP_TREE_SCALAR_STMTS (node)[0]
> + /* Avoid touching loads which need care with load permutations
> + and specialities like load-lane representations. */
> + && !STMT_VINFO_DATA_REF (SLP_TREE_REPRESENTATIVE (node)))
> + for (slp_tree cand
> + : part_starts[gimple_uid (SLP_TREE_SCALAR_STMTS (node)[0]-
> >stmt)])
> + /* ??? There is a possible ordering/optimality problem in that
> + the CSE then can keep a wider feeding live even though it itself
> + becomes dead by means of CSE. Which might be solvable by doing
> + the CSE in a wide-to-narrow order. */
> + if (SLP_TREE_LANES (cand) > SLP_TREE_LANES (node)
> + /* We can do high/lo extracts and full vector copies. */
I wonder if the restriction isn't too strict here? especially for low extract?
extracting e.g. a V4QI from a V16QI from the lowpart should be equally
cheap no? just a BIT_FIELD_REF? It looks like the code below it handles it
fine?
Aside from this question this looks good to me.
Thanks,
Tamar
> + && constant_multiple_p
> + (TYPE_VECTOR_SUBPARTS (SLP_TREE_VECTYPE (cand)),
> + TYPE_VECTOR_SUBPARTS (SLP_TREE_VECTYPE (node)), &c)
> + && c <= 2)
> + {
> + unsigned HOST_WIDE_INT s;
> + bool const_p
> + = TYPE_VECTOR_SUBPARTS (SLP_TREE_VECTYPE
> (node)).is_constant (&s);
> + unsigned i;
> + for (i = 0; i <= SLP_TREE_LANES (cand) - SLP_TREE_LANES (node);)
> + {
> + unsigned j;
> + for (j = 0; j < SLP_TREE_LANES (node); ++j)
> + if (!SLP_TREE_SCALAR_STMTS (node)[j]
> + || (SLP_TREE_SCALAR_STMTS (cand)[i+j]
> + != SLP_TREE_SCALAR_STMTS (node)[j]))
> + break;
> + if (j == SLP_TREE_LANES (node))
> + break;
> + if (!const_p)
> + {
> + i = SLP_TREE_LANES (cand);
> + break;
> + }
> + /* We can extract only aligned on node vector type boundary. */
> + i += s;
> + }
> + if (i > SLP_TREE_LANES (cand) - SLP_TREE_LANES (node))
> + continue;
> + /* Found node within cand at i. Put a permute in place
> + of it, selecting the subset from cand. */
> + if (dump_enabled_p ())
> + dump_printf (MSG_NOTE, "CSEd node %p as %spart of node %p\n",
> + (void *)node, i == 0 ? "low" : "high", (void *)cand);
> + for (slp_tree child : SLP_TREE_CHILDREN (node))
> + /* Delay SLP tree release since we might still reference a node
> + from the part_starts map. */
> + drops.safe_push (child);
> + SLP_TREE_CHILDREN (node).truncate (1);
> + SLP_TREE_REF_COUNT (cand)++;
> + SLP_TREE_CHILDREN (node)[0] = cand;
> + SLP_TREE_CODE (node) = VEC_PERM_EXPR;
> + SLP_TREE_REPRESENTATIVE (node) = NULL;
> + SLP_TREE_LANE_PERMUTATION (node).create (SLP_TREE_LANES
> (node));
> + for (unsigned j = i; j < i + SLP_TREE_LANES (node); ++j)
> + SLP_TREE_LANE_PERMUTATION (node).quick_push (std::make_pair
> (0, j));
> + return;
> + }
> +
> + for (slp_tree &child : SLP_TREE_CHILDREN (node))
> + if (child)
> + vect_cse_slp_node_parts (visited, part_starts, drops, child);
> +}
> +
> /* Optimize the SLP graph of VINFO. */
>
> void
> @@ -8511,6 +8632,39 @@ vect_optimize_slp (vec_info *vinfo)
> vect_cse_slp_nodes (bst_map, SLP_INSTANCE_TREE (inst));
>
> release_scalar_stmts_to_slp_tree_map (bst_map);
> +
> + if (!is_a <bb_vec_info> (vinfo))
> + return;
> +
> + /* Attempt to merge SLP sub-graphs that intersect in low or highparts of
> + each other. Build the reverse mapping from stmt to SLP node for
> + lanes starting at the low or high part.
> + ??? In the future we can extend this to do a two-step permute
> + and extract or extract and permute to put the high/low part in
> + place on the original vector or permute the hogh/low part to
> + match up the target lane order. */
> + hash_set<slp_tree> visited;
> + vec<vec<slp_tree>> start_for_part;
> + start_for_part.create (vinfo->stmt_vec_infos.length () + 1);
> + start_for_part.quick_grow_cleared (vinfo->stmt_vec_infos.length () + 1);
> + for (auto inst : vinfo->slp_instances)
> + vect_cse_gather_part_starts (visited,
> + start_for_part, SLP_INSTANCE_TREE (inst));
> +
> + /* Now replace low/highpart copies with extracting permutes. */
> + auto_vec<slp_tree> drops;
> + visited.empty ();
> + for (auto inst : vinfo->slp_instances)
> + vect_cse_slp_node_parts (visited, start_for_part, drops,
> + SLP_INSTANCE_TREE (inst));
> +
> + /* Now perform delayed releases of nodes. */
> + for (slp_tree node : drops)
> + vect_free_slp_tree (node);
> +
> + for (auto v : start_for_part)
> + v.release ();
> + start_for_part.release ();
> }
>
> /* Gather loads reachable from the individual SLP graph entries. */
> --
> 2.51.0
