PR68577: Handle narrowing for vector popcount, etc.

[Richard Sandiford <richard.sandiford@arm.com>]

This patch adds support for simple cases where the a vector internal
function returns wider results than the scalar equivalent.  It punts
on other cases.

Tested on powerpc64-linux-gnu and x86_64-linux-gnu.  OK to install?

Thanks,
Richard


gcc/
	PR tree-optimization/68577
	* tree-vect-stmts.c (simple_integer_narrowing): New function.
	(vectorizable_call): Restrict internal function handling
	to NONE and NARROW cases, using simple_integer_narrowing
	to test for the latter.  Add cost of narrowing operation
	and insert it where necessary.

gcc/testsuite/
	PR tree-optimization/68577
	* gcc.dg/vect/pr68577.c: New test.

diff --git a/gcc/testsuite/gcc.dg/vect/pr68577.c b/gcc/testsuite/gcc.dg/vect/pr68577.c
new file mode 100644
index 0000000..999c1c8
--- /dev/null
+++ b/gcc/testsuite/gcc.dg/vect/pr68577.c
@@ -0,0 +1,25 @@
+/* { dg-do compile } */
+
+int a, b;
+
+void
+__sched_cpucount (void)
+{
+  while (b)
+    {
+      long l = b++;
+      a += __builtin_popcountl(l);
+    }
+}
+
+void
+slp_test (int *x, long *y)
+{
+  for (int i = 0; i < 512; i += 4)
+    {
+      x[i] = __builtin_popcountl(y[i]);
+      x[i + 1] = __builtin_popcountl(y[i + 1]);
+      x[i + 2] = __builtin_popcountl(y[i + 2]);
+      x[i + 3] = __builtin_popcountl(y[i + 3]);
+    }
+}
diff --git a/gcc/tree-vect-stmts.c b/gcc/tree-vect-stmts.c
index 3b078da..af86bce 100644
--- a/gcc/tree-vect-stmts.c
+++ b/gcc/tree-vect-stmts.c
@@ -2122,6 +2122,40 @@ vectorizable_mask_load_store (gimple *stmt, gimple_stmt_iterator *gsi,
   return true;
 }
 
+/* Return true if vector type VECTYPE_OUT has integer elements and
+   if we can narrow two integer vectors with the same shape as
+   VECTYPE_IN to VECTYPE_OUT in a single step.  On success,
+   return the binary pack code in *CONVERT_CODE and the types
+   of the input vectors in *CONVERT_FROM.  */
+
+static bool
+simple_integer_narrowing (tree vectype_out, tree vectype_in,
+			  tree_code *convert_code, tree *convert_from)
+{
+  if (!INTEGRAL_TYPE_P (TREE_TYPE (vectype_out)))
+    return false;
+
+  if (!INTEGRAL_TYPE_P (TREE_TYPE (vectype_in)))
+    {
+      unsigned int bits
+	= GET_MODE_BITSIZE (TYPE_MODE (TREE_TYPE (vectype_in)));
+      tree scalar_type = build_nonstandard_integer_type (bits, 0);
+      vectype_in = get_same_sized_vectype (scalar_type, vectype_in);
+    }
+
+  tree_code code;
+  int multi_step_cvt = 0;
+  auto_vec <tree, 8> interm_types;
+  if (!supportable_narrowing_operation (NOP_EXPR, vectype_out, vectype_in,
+					&code, &multi_step_cvt,
+					&interm_types)
+      || multi_step_cvt)
+    return false;
+
+  *convert_code = code;
+  *convert_from = vectype_in;
+  return true;
+}
 
 /* Function vectorizable_call.
 
@@ -2288,7 +2322,13 @@ vectorizable_call (gimple *gs, gimple_stmt_iterator *gsi, gimple **vec_stmt,
   tree callee = gimple_call_fndecl (stmt);
 
   /* First try using an internal function.  */
-  if (cfn != CFN_LAST)
+  tree_code convert_code = ERROR_MARK;
+  tree convert_from = NULL_TREE;
+  if (cfn != CFN_LAST
+      && (modifier == NONE
+	  || (modifier == NARROW
+	      && simple_integer_narrowing (vectype_out, vectype_in,
+					   &convert_code, &convert_from))))
     ifn = vectorizable_internal_function (cfn, callee, vectype_out,
 					  vectype_in);
 
@@ -2328,7 +2368,7 @@ vectorizable_call (gimple *gs, gimple_stmt_iterator *gsi, gimple **vec_stmt,
 
   if (slp_node || PURE_SLP_STMT (stmt_info))
     ncopies = 1;
-  else if (modifier == NARROW)
+  else if (modifier == NARROW && ifn == IFN_LAST)
     ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_out;
   else
     ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_in;
@@ -2344,6 +2384,10 @@ vectorizable_call (gimple *gs, gimple_stmt_iterator *gsi, gimple **vec_stmt,
         dump_printf_loc (MSG_NOTE, vect_location, "=== vectorizable_call ==="
                          "\n");
       vect_model_simple_cost (stmt_info, ncopies, dt, NULL, NULL);
+      if (ifn != IFN_LAST && modifier == NARROW && !slp_node)
+	add_stmt_cost (stmt_info->vinfo->target_cost_data, ncopies / 2,
+		       vec_promote_demote, stmt_info, 0, vect_body);
+
       return true;
     }
 
@@ -2357,9 +2401,9 @@ vectorizable_call (gimple *gs, gimple_stmt_iterator *gsi, gimple **vec_stmt,
   vec_dest = vect_create_destination_var (scalar_dest, vectype_out);
 
   prev_stmt_info = NULL;
-  switch (modifier)
+  if (modifier == NONE || ifn != IFN_LAST)
     {
-    case NONE:
+      tree prev_res = NULL_TREE;
       for (j = 0; j < ncopies; ++j)
 	{
 	  /* Build argument list for the vectorized call.  */
@@ -2387,12 +2431,30 @@ vectorizable_call (gimple *gs, gimple_stmt_iterator *gsi, gimple **vec_stmt,
 		      vec<tree> vec_oprndsk = vec_defs[k];
 		      vargs[k] = vec_oprndsk[i];
 		    }
-		  if (ifn != IFN_LAST)
-		    new_stmt = gimple_build_call_internal_vec (ifn, vargs);
+		  if (modifier == NARROW)
+		    {
+		      tree half_res = make_ssa_name (convert_from);
+		      new_stmt = gimple_build_call_internal_vec (ifn, vargs);
+		      gimple_call_set_lhs (new_stmt, half_res);
+		      vect_finish_stmt_generation (stmt, new_stmt, gsi);
+		      if ((i & 1) == 0)
+			{
+			  prev_res = half_res;
+			  continue;
+			}
+		      new_temp = make_ssa_name (vec_dest);
+		      new_stmt = gimple_build_assign (new_temp, convert_code,
+						      prev_res, half_res);
+		    }
 		  else
-		    new_stmt = gimple_build_call_vec (fndecl, vargs);
-		  new_temp = make_ssa_name (vec_dest, new_stmt);
-		  gimple_call_set_lhs (new_stmt, new_temp);
+		    {
+		      if (ifn != IFN_LAST)
+			new_stmt = gimple_build_call_internal_vec (ifn, vargs);
+		      else
+			new_stmt = gimple_build_call_vec (fndecl, vargs);
+		      new_temp = make_ssa_name (vec_dest, new_stmt);
+		      gimple_call_set_lhs (new_stmt, new_temp);
+		    }
 		  vect_finish_stmt_generation (stmt, new_stmt, gsi);
 		  SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt);
 		}
@@ -2436,6 +2498,21 @@ vectorizable_call (gimple *gs, gimple_stmt_iterator *gsi, gimple **vec_stmt,
 	      new_temp = make_ssa_name (vec_dest);
 	      new_stmt = gimple_build_assign (new_temp, new_var);
 	    }
+	  else if (modifier == NARROW)
+	    {
+	      tree half_res = make_ssa_name (convert_from);
+	      new_stmt = gimple_build_call_internal_vec (ifn, vargs);
+	      gimple_call_set_lhs (new_stmt, half_res);
+	      vect_finish_stmt_generation (stmt, new_stmt, gsi);
+	      if ((j & 1) == 0)
+		{
+		  prev_res = half_res;
+		  continue;
+		}
+	      new_temp = make_ssa_name (vec_dest);
+	      new_stmt = gimple_build_assign (new_temp, convert_code,
+					      prev_res, half_res);
+	    }
 	  else
 	    {
 	      if (ifn != IFN_LAST)
@@ -2447,17 +2524,16 @@ vectorizable_call (gimple *gs, gimple_stmt_iterator *gsi, gimple **vec_stmt,
 	    }
 	  vect_finish_stmt_generation (stmt, new_stmt, gsi);
 
-	  if (j == 0)
+	  if (j == (modifier == NARROW ? 1 : 0))
 	    STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt;
 	  else
 	    STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt;
 
 	  prev_stmt_info = vinfo_for_stmt (new_stmt);
 	}
-
-      break;
-
-    case NARROW:
+    }
+  else if (modifier == NARROW)
+    {
       for (j = 0; j < ncopies; ++j)
 	{
 	  /* Build argument list for the vectorized call.  */
@@ -2528,10 +2604,7 @@ vectorizable_call (gimple *gs, gimple_stmt_iterator *gsi, gimple **vec_stmt,
 	      vargs.quick_push (vec_oprnd1);
 	    }
 
-	  if (ifn != IFN_LAST)
-	    new_stmt = gimple_build_call_internal_vec (ifn, vargs);
-	  else
-	    new_stmt = gimple_build_call_vec (fndecl, vargs);
+	  new_stmt = gimple_build_call_vec (fndecl, vargs);
 	  new_temp = make_ssa_name (vec_dest, new_stmt);
 	  gimple_call_set_lhs (new_stmt, new_temp);
 	  vect_finish_stmt_generation (stmt, new_stmt, gsi);
@@ -2545,13 +2618,10 @@ vectorizable_call (gimple *gs, gimple_stmt_iterator *gsi, gimple **vec_stmt,
 	}
 
       *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info);
-
-      break;
-
-    case WIDEN:
-      /* No current target implements this case.  */
-      return false;
     }
+  else
+    /* No current target implements this case.  */
+    return false;
 
   vargs.release ();