veclower: improve selection of vector mode when lowering [PR 112787]

This patch addresses the issue reported in PR target/112787 by improving the
compute type selection.  We do this by not considering types with more elements
than the type we are lowering since we'd reject such types anyway.

gcc/ChangeLog:

	PR target/112787
	* tree-vect-generic.cc (type_for_widest_vector_mode): Change function to
	use original vector type and check widest vector mode has at most the
	same number of elements.
	(get_compute_type): Pass original vector type rather than the element
	type to type_for_widest_vector_mode and remove now obsolete check for
	the number of elements.

gcc/testsuite/ChangeLog:

	* gcc.target/aarch64/pr112787.c: New test.

diff --git a/gcc/testsuite/gcc.target/aarch64/pr112787.c b/gcc/testsuite/gcc.target/aarch64/pr112787.c
new file mode 100644 (file)
index 0000000..caca1bf
--- /dev/null
+++ b/gcc/testsuite/gcc.target/aarch64/pr112787.c
@@ -0,0 +1,11 @@
+/* { dg-do compile } */
+/* { dg-additional-options "-O2 -march=armv8-a+sve -mcpu=neoverse-n2" } */
+
+typedef int __attribute__((__vector_size__ (64))) vec;
+
+vec fn (vec a, vec b)
+{
+  return a + b;
+}
+
+/* { dg-final { scan-assembler-times {add\tv[0-9]+} 4 } } */

diff --git a/gcc/tree-vect-generic.cc b/gcc/tree-vect-generic.cc
index a7e6cb87a5e31e3dd2a893ea5652eeebf8d5d214..c906eb3521ea01fd2bdfc89c3476d02c555cf8cc 100644 (file)
--- a/gcc/tree-vect-generic.cc
+++ b/gcc/tree-vect-generic.cc
@@ -1343,12 +1343,16 @@ optimize_vector_constructor (gimple_stmt_iterator *gsi)
   gsi_replace (gsi, g, false);
 }
 \f
-/* Return a type for the widest vector mode whose components are of type
-   TYPE, or NULL_TREE if none is found.  */
+/* Return a type for the widest vector mode with the same element type as
+   type ORIGINAL_VECTOR_TYPE, with at most the same number of elements as type
+   ORIGINAL_VECTOR_TYPE and that is supported by the target for an operation
+   with optab OP, or return NULL_TREE if none is found.  */
 
 static tree
-type_for_widest_vector_mode (tree type, optab op)
+type_for_widest_vector_mode (tree original_vector_type, optab op)
 {
+  gcc_assert (VECTOR_TYPE_P (original_vector_type));
+  tree type = TREE_TYPE (original_vector_type);
   machine_mode inner_mode = TYPE_MODE (type);
   machine_mode best_mode = VOIDmode, mode;
   poly_int64 best_nunits = 0;
@@ -1371,7 +1375,9 @@ type_for_widest_vector_mode (tree type, optab op)
   FOR_EACH_MODE_FROM (mode, mode)
     if (GET_MODE_INNER (mode) == inner_mode
        && maybe_gt (GET_MODE_NUNITS (mode), best_nunits)
-       && optab_handler (op, mode) != CODE_FOR_nothing)
+       && optab_handler (op, mode) != CODE_FOR_nothing
+       && known_le (GET_MODE_NUNITS (mode),
+                    TYPE_VECTOR_SUBPARTS (original_vector_type)))
       best_mode = mode, best_nunits = GET_MODE_NUNITS (mode);
 
   if (best_mode == VOIDmode)
@@ -1702,9 +1708,8 @@ get_compute_type (enum tree_code code, optab op, tree type)
          || optab_handler (op, TYPE_MODE (type)) == CODE_FOR_nothing))
     {
       tree vector_compute_type
-       = type_for_widest_vector_mode (TREE_TYPE (type), op);
+       = type_for_widest_vector_mode (type, op);
       if (vector_compute_type != NULL_TREE
-         && subparts_gt (compute_type, vector_compute_type)
          && maybe_ne (TYPE_VECTOR_SUBPARTS (vector_compute_type), 1U)
          && (optab_handler (op, TYPE_MODE (vector_compute_type))
              != CODE_FOR_nothing))
@@ -1877,10 +1882,9 @@ expand_vector_conversion (gimple_stmt_iterator *gsi)
       /* Can't use get_compute_type here, as supportable_convert_operation
         doesn't necessarily use an optab and needs two arguments.  */
       tree vec_compute_type
-       = type_for_widest_vector_mode (TREE_TYPE (arg_type), mov_optab);
+       = type_for_widest_vector_mode (arg_type, mov_optab);
       if (vec_compute_type
-         && VECTOR_MODE_P (TYPE_MODE (vec_compute_type))
-         && subparts_gt (arg_type, vec_compute_type))
+         && VECTOR_MODE_P (TYPE_MODE (vec_compute_type)))
        {
          unsigned HOST_WIDE_INT nelts
            = constant_lower_bound (TYPE_VECTOR_SUBPARTS (vec_compute_type));