Add support for SLP vectorization of OpenMP SIMD clone calls

This adds support for SLP vectorization of OpenMP SIMD clone calls.
There's a complication when vectorizing calls involving virtual
operands since this is now for the first time not only leafs (loads
or stores).  With SLP this runs into the issue that placement of
the vectorized stmts is not necessarily at one of the original
scalar stmts which leads to the magic updating virtual operands
in vect_finish_stmt_generation not working.  So we run into the
assert that updating virtual operands isn't necessary.  I've
papered over this similar to how we do for mismatched const/pure
attribution by setting vinfo->any_known_not_updated_vssa.

I've added two basic testcases with multi-lane SLP and verified
that with single-lane SLP enabled the rest of the existing testcases
pass.

	* tree-vect-slp.cc (mask_call_maps): New.
	(vect_get_operand_map): Handle IFN_MASK_CALL.
	(vect_build_slp_tree_1): Likewise.
	* tree-vect-stmts.cc (vectorizable_simd_clone_call): Handle
	SLP.

	* gcc.dg/vect/slp-simd-clone-1.c: New testcase.
	* gcc.dg/vect/slp-simd-clone-2.c: Likewise.

diff --git a/gcc/testsuite/gcc.dg/vect/slp-simd-clone-1.c b/gcc/testsuite/gcc.dg/vect/slp-simd-clone-1.c
new file mode 100644 (file)
index 0000000..6ccbb39
--- /dev/null
+++ b/gcc/testsuite/gcc.dg/vect/slp-simd-clone-1.c
@@ -0,0 +1,46 @@
+/* { dg-require-effective-target vect_simd_clones } */
+/* { dg-additional-options "-fopenmp-simd" } */
+
+#include "tree-vect.h"
+
+int x[1024];
+
+#pragma omp declare simd simdlen(4) notinbranch
+__attribute__((noinline)) int
+foo (int a, int b)
+{
+  return a + b;
+}
+
+void __attribute__((noipa))
+bar (void)
+{
+#pragma omp simd
+  for (int i = 0; i < 512; i++)
+    {
+      x[2*i+0] = foo (x[2*i+0], x[2*i+0]);
+      x[2*i+1] = foo (x[2*i+1], x[2*i+1]);
+    }
+}
+
+int
+main ()
+{
+  int i;
+  check_vect ();
+
+#pragma GCC novector
+  for (i = 0; i < 1024; i++)
+    x[i] = i;
+
+  bar ();
+
+#pragma GCC novector
+  for (i = 0; i < 1024; i++)
+    if (x[i] != i + i)
+      abort ();
+
+  return 0;
+}
+
+/* { dg-final { scan-tree-dump "vectorizing stmts using SLP" "vect" } } */

diff --git a/gcc/testsuite/gcc.dg/vect/slp-simd-clone-2.c b/gcc/testsuite/gcc.dg/vect/slp-simd-clone-2.c
new file mode 100644 (file)
index 0000000..98387c9
--- /dev/null
+++ b/gcc/testsuite/gcc.dg/vect/slp-simd-clone-2.c
@@ -0,0 +1,57 @@
+/* { dg-require-effective-target vect_simd_clones } */
+/* { dg-additional-options "-fopenmp-simd" } */
+/* { dg-additional-options "-mavx2" { target avx2_runtime } } */
+
+#include "tree-vect.h"
+
+int x[1024];
+
+#pragma omp declare simd simdlen(4) inbranch
+__attribute__((noinline)) int
+foo (int a, int b)
+{
+  return a + b;
+}
+
+void __attribute__((noipa))
+bar (void)
+{
+#pragma omp simd
+  for (int i = 0; i < 512; i++)
+    {
+      if (x[2*i+0] < 10)
+       x[2*i+0] = foo (x[2*i+0], x[2*i+0]);
+      if (x[2*i+1] < 20)
+       x[2*i+1] = foo (x[2*i+1], x[2*i+1]);
+    }
+}
+
+int
+main ()
+{
+  int i;
+  check_vect ();
+
+#pragma GCC novector
+  for (i = 0; i < 1024; i++)
+    x[i] = i;
+
+  bar ();
+
+#pragma GCC novector
+  for (i = 0; i < 1024; i++)
+    {
+      if (((i & 1) && i < 20)
+         || (!(i & 1) && i < 10))
+       {
+         if (x[i] != i + i)
+           abort ();
+       }
+      else if (x[i] != i)
+       abort ();
+    }
+
+  return 0;
+}
+
+/* { dg-final { scan-tree-dump "vectorizing stmts using SLP" "vect" { target avx2_runtime } } } */

diff --git a/gcc/tree-vect-slp.cc b/gcc/tree-vect-slp.cc
index fa098f9ff4eaebfaa2083124df11c435519573aa..af8f5031bd274778a7df6ef29f9efd511e66babe 100644 (file)
--- a/gcc/tree-vect-slp.cc
+++ b/gcc/tree-vect-slp.cc
@@ -505,6 +505,14 @@ static const int arg2_map[] = { 1, 2 };
 static const int arg1_arg4_map[] = { 2, 1, 4 };
 static const int arg3_arg2_map[] = { 2, 3, 2 };
 static const int op1_op0_map[] = { 2, 1, 0 };
+static const int mask_call_maps[6][7] = {
+  { 1, 1, },
+  { 2, 1, 2, },
+  { 3, 1, 2, 3, },
+  { 4, 1, 2, 3, 4, },
+  { 5, 1, 2, 3, 4, 5, },
+  { 6, 1, 2, 3, 4, 5, 6 },
+};
 
 /* For most SLP statements, there is a one-to-one mapping between
    gimple arguments and child nodes.  If that is not true for STMT,
@@ -547,6 +555,15 @@ vect_get_operand_map (const gimple *stmt, unsigned char swap = 0)
          case IFN_MASK_STORE:
            return arg3_arg2_map;
 
+         case IFN_MASK_CALL:
+           {
+             unsigned nargs = gimple_call_num_args (call);
+             if (nargs >= 2 && nargs <= 7)
+               return mask_call_maps[nargs-2];
+             else
+               return nullptr;
+           }
+
          default:
            break;
          }
@@ -1070,7 +1087,7 @@ vect_build_slp_tree_1 (vec_info *vinfo, unsigned char *swap,
       if (call_stmt)
        {
          combined_fn cfn = gimple_call_combined_fn (call_stmt);
-         if (cfn != CFN_LAST)
+         if (cfn != CFN_LAST && cfn != CFN_MASK_CALL)
            rhs_code = cfn;
          else
            rhs_code = CALL_EXPR;
@@ -1085,6 +1102,7 @@ vect_build_slp_tree_1 (vec_info *vinfo, unsigned char *swap,
              rhs_code = CFN_MASK_STORE;
            }
          else if ((cfn != CFN_LAST
+                   && cfn != CFN_MASK_CALL
                    && internal_fn_p (cfn)
                    && !vectorizable_internal_fn_p (as_internal_fn (cfn)))
                   || gimple_call_tail_p (call_stmt)

diff --git a/gcc/tree-vect-stmts.cc b/gcc/tree-vect-stmts.cc
index ce925cc1d53d73a789883159ad5c37b17600e28b..33b557c2a4991d6d0e8b0e76557485eb0e4114f1 100644 (file)
--- a/gcc/tree-vect-stmts.cc
+++ b/gcc/tree-vect-stmts.cc
@@ -4208,10 +4208,6 @@ vectorizable_simd_clone_call (vec_info *vinfo, stmt_vec_info stmt_info,
   if (loop_vinfo && nested_in_vect_loop_p (loop, stmt_info))
     return false;
 
-  /* FORNOW */
-  if (slp_node)
-    return false;
-
   /* Process function arguments.  */
   nargs = gimple_call_num_args (stmt) - arg_offset;
 
@@ -4220,6 +4216,8 @@ vectorizable_simd_clone_call (vec_info *vinfo, stmt_vec_info stmt_info,
     return false;
 
   arginfo.reserve (nargs, true);
+  auto_vec<slp_tree> slp_op;
+  slp_op.safe_grow_cleared (nargs);
 
   for (i = 0; i < nargs; i++)
     {
@@ -4231,9 +4229,12 @@ vectorizable_simd_clone_call (vec_info *vinfo, stmt_vec_info stmt_info,
       thisarginfo.op = NULL_TREE;
       thisarginfo.simd_lane_linear = false;
 
-      op = gimple_call_arg (stmt, i + arg_offset);
-      if (!vect_is_simple_use (op, vinfo, &thisarginfo.dt,
-                              &thisarginfo.vectype)
+      int op_no = i + arg_offset;
+      if (slp_node)
+       op_no = vect_slp_child_index_for_operand (stmt, op_no);
+      if (!vect_is_simple_use (vinfo, stmt_info, slp_node,
+                              op_no, &op, &slp_op[i],
+                              &thisarginfo.dt, &thisarginfo.vectype)
          || thisarginfo.dt == vect_uninitialized_def)
        {
          if (dump_enabled_p ())
@@ -4244,7 +4245,13 @@ vectorizable_simd_clone_call (vec_info *vinfo, stmt_vec_info stmt_info,
 
       if (thisarginfo.dt == vect_constant_def
          || thisarginfo.dt == vect_external_def)
-       gcc_assert (thisarginfo.vectype == NULL_TREE);
+       {
+         gcc_assert (vec_stmt || thisarginfo.vectype == NULL_TREE);
+         if (!vec_stmt)
+           thisarginfo.vectype = get_vectype_for_scalar_type (vinfo,
+                                                              TREE_TYPE (op),
+                                                              slp_node);
+       }
       else
        gcc_assert (thisarginfo.vectype != NULL_TREE);
 
@@ -4301,15 +4308,14 @@ vectorizable_simd_clone_call (vec_info *vinfo, stmt_vec_info stmt_info,
          && thisarginfo.dt != vect_constant_def
          && thisarginfo.dt != vect_external_def
          && loop_vinfo
-         && !slp_node
          && TREE_CODE (op) == SSA_NAME)
        vect_simd_lane_linear (op, loop, &thisarginfo);
 
       arginfo.quick_push (thisarginfo);
     }
 
-  poly_uint64 vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo);
-  if (!vf.is_constant ())
+  if (loop_vinfo
+      && !LOOP_VINFO_VECT_FACTOR (loop_vinfo).is_constant ())
     {
       if (dump_enabled_p ())
        dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
@@ -4318,6 +4324,8 @@ vectorizable_simd_clone_call (vec_info *vinfo, stmt_vec_info stmt_info,
       return false;
     }
 
+  poly_uint64 vf = loop_vinfo ? LOOP_VINFO_VECT_FACTOR (loop_vinfo) : 1;
+  unsigned group_size = slp_node ? SLP_TREE_LANES (slp_node) : 1;
   unsigned int badness = 0;
   struct cgraph_node *bestn = NULL;
   if (STMT_VINFO_SIMD_CLONE_INFO (stmt_info).exists ())
@@ -4328,7 +4336,8 @@ vectorizable_simd_clone_call (vec_info *vinfo, stmt_vec_info stmt_info,
       {
        unsigned int this_badness = 0;
        unsigned int num_calls;
-       if (!constant_multiple_p (vf, n->simdclone->simdlen, &num_calls)
+       if (!constant_multiple_p (vf * group_size,
+                                 n->simdclone->simdlen, &num_calls)
            || n->simdclone->nargs != nargs)
          continue;
        if (num_calls != 1)
@@ -4454,7 +4463,10 @@ vectorizable_simd_clone_call (vec_info *vinfo, stmt_vec_info stmt_info,
 
   fndecl = bestn->decl;
   nunits = bestn->simdclone->simdlen;
-  ncopies = vector_unroll_factor (vf, nunits);
+  if (slp_node)
+    ncopies = vector_unroll_factor (vf * group_size, nunits);
+  else
+    ncopies = vector_unroll_factor (vf, nunits);
 
   /* If the function isn't const, only allow it in simd loops where user
      has asserted that at least nunits consecutive iterations can be
@@ -4469,6 +4481,15 @@ vectorizable_simd_clone_call (vec_info *vinfo, stmt_vec_info stmt_info,
 
   if (!vec_stmt) /* transformation not required.  */
     {
+      if (slp_node)
+       for (unsigned i = 0; i < nargs; ++i)
+         if (!vect_maybe_update_slp_op_vectype (slp_op[i], arginfo[i].vectype))
+           {
+             if (dump_enabled_p ())
+               dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
+                                "incompatible vector types for invariants\n");
+             return false;
+           }
       /* When the original call is pure or const but the SIMD ABI dictates
         an aggregate return we will have to use a virtual definition and
         in a loop eventually even need to add a virtual PHI.  That's
@@ -4477,6 +4498,11 @@ vectorizable_simd_clone_call (vec_info *vinfo, stmt_vec_info stmt_info,
          && !gimple_vdef (stmt)
          && TREE_CODE (TREE_TYPE (TREE_TYPE (bestn->decl))) == ARRAY_TYPE)
        vinfo->any_known_not_updated_vssa = true;
+      /* ???  For SLP code-gen we end up inserting after the last
+        vector argument def rather than at the original call position
+        so automagic virtual operand updating doesn't work.  */
+      if (gimple_vuse (stmt) && slp_node)
+       vinfo->any_known_not_updated_vssa = true;
       STMT_VINFO_SIMD_CLONE_INFO (stmt_info).safe_push (bestn->decl);
       for (i = 0; i < nargs; i++)
        if ((bestn->simdclone->args[i].arg_type
@@ -4526,8 +4552,14 @@ vectorizable_simd_clone_call (vec_info *vinfo, stmt_vec_info stmt_info,
 
   auto_vec<vec<tree> > vec_oprnds;
   auto_vec<unsigned> vec_oprnds_i;
-  vec_oprnds.safe_grow_cleared (nargs, true);
   vec_oprnds_i.safe_grow_cleared (nargs, true);
+  if (slp_node)
+    {
+      vec_oprnds.reserve_exact (nargs);
+      vect_get_slp_defs (vinfo, slp_node, &vec_oprnds);
+    }
+  else
+    vec_oprnds.safe_grow_cleared (nargs, true);
   for (j = 0; j < ncopies; ++j)
     {
       /* Build argument list for the vectorized call.  */
@@ -4558,9 +4590,10 @@ vectorizable_simd_clone_call (vec_info *vinfo, stmt_vec_info stmt_info,
                      gcc_assert ((k & (k - 1)) == 0);
                      if (m == 0)
                        {
-                         vect_get_vec_defs_for_operand (vinfo, stmt_info,
-                                                        ncopies * o / k, op,
-                                                        &vec_oprnds[i]);
+                         if (!slp_node)
+                           vect_get_vec_defs_for_operand (vinfo, stmt_info,
+                                                          ncopies * o / k, op,
+                                                          &vec_oprnds[i]);
                          vec_oprnds_i[i] = 0;
                          vec_oprnd0 = vec_oprnds[i][vec_oprnds_i[i]++];
                        }
@@ -4596,10 +4629,11 @@ vectorizable_simd_clone_call (vec_info *vinfo, stmt_vec_info stmt_info,
                        {
                          if (m == 0 && l == 0)
                            {
-                             vect_get_vec_defs_for_operand (vinfo, stmt_info,
-                                                            k * o * ncopies,
-                                                            op,
-                                                            &vec_oprnds[i]);
+                             if (!slp_node)
+                               vect_get_vec_defs_for_operand (vinfo, stmt_info,
+                                                              k * o * ncopies,
+                                                              op,
+                                                              &vec_oprnds[i]);
                              vec_oprnds_i[i] = 0;
                              vec_oprnd0 = vec_oprnds[i][vec_oprnds_i[i]++];
                            }
@@ -4670,10 +4704,11 @@ vectorizable_simd_clone_call (vec_info *vinfo, stmt_vec_info stmt_info,
                             elements as the current function.  */
                          if (m == 0)
                            {
-                             vect_get_vec_defs_for_operand (vinfo, stmt_info,
-                                                            o * ncopies,
-                                                            op,
-                                                            &vec_oprnds[i]);
+                             if (!slp_node)
+                               vect_get_vec_defs_for_operand (vinfo, stmt_info,
+                                                              o * ncopies,
+                                                              op,
+                                                              &vec_oprnds[i]);
                              vec_oprnds_i[i] = 0;
                            }
                          vec_oprnd0 = vec_oprnds[i][vec_oprnds_i[i]++];
@@ -4817,7 +4852,11 @@ vectorizable_simd_clone_call (vec_info *vinfo, stmt_vec_info stmt_info,
 
                  if (j == 0 && l == 0)
                    *vec_stmt = new_stmt;
-                 STMT_VINFO_VEC_STMTS (stmt_info).safe_push (new_stmt);
+                 if (slp_node)
+                   SLP_TREE_VEC_DEFS (slp_node)
+                     .quick_push (gimple_assign_lhs (new_stmt));
+                 else
+                   STMT_VINFO_VEC_STMTS (stmt_info).safe_push (new_stmt);
                }
 
              if (ratype)
@@ -4860,7 +4899,11 @@ vectorizable_simd_clone_call (vec_info *vinfo, stmt_vec_info stmt_info,
 
              if ((unsigned) j == k - 1)
                *vec_stmt = new_stmt;
-             STMT_VINFO_VEC_STMTS (stmt_info).safe_push (new_stmt);
+             if (slp_node)
+               SLP_TREE_VEC_DEFS (slp_node)
+                 .quick_push (gimple_assign_lhs (new_stmt));
+             else
+               STMT_VINFO_VEC_STMTS (stmt_info).safe_push (new_stmt);
              continue;
            }
          else if (ratype)
@@ -4883,7 +4926,10 @@ vectorizable_simd_clone_call (vec_info *vinfo, stmt_vec_info stmt_info,
 
       if (j == 0)
        *vec_stmt = new_stmt;
-      STMT_VINFO_VEC_STMTS (stmt_info).safe_push (new_stmt);
+      if (slp_node)
+       SLP_TREE_VEC_DEFS (slp_node).quick_push (gimple_get_lhs (new_stmt));
+      else
+       STMT_VINFO_VEC_STMTS (stmt_info).safe_push (new_stmt);
     }
 
   for (i = 0; i < nargs; ++i)