The vectorizer handles a _Float16 to __bf16 conversion through
vectorizable_assignment, thinking it's a noop. The following
fixes this by requiring the same vector component mode when
checking for CONVERT_EXPR_CODE_P, being stricter than for
VIEW_CONVERT_EXPR.
This variant splits the check for VIEW_CONVERT_EXPR and
CONVERT_EXPR_CODE_P since it otherwise regresses gcc.dg/vect/vect-120.c.
PR tree-optimization/114921
* tree-vect-stmts.cc (vectorizable_assignment): Require
same vector component modes for input and output for
CONVERT_EXPR_CODE_P.
---
gcc/tree-vect-stmts.cc | 19 +++++++++++--------
1 file changed, 11 insertions(+), 8 deletions(-)
diff --git a/gcc/tree-vect-stmts.cc b/gcc/tree-vect-stmts.cc
index f8d8636b139..7e571968a59 100644
--- a/gcc/tree-vect-stmts.cc
+++ b/gcc/tree-vect-stmts.cc
@@ -5955,14 +5955,17 @@ vectorizable_assignment (vec_info *vinfo,
if (!vectype_in)
vectype_in = get_vectype_for_scalar_type (vinfo, TREE_TYPE (op), slp_node);
- /* We can handle NOP_EXPR conversions that do not change the number
- of elements or the vector size. */
- if ((CONVERT_EXPR_CODE_P (code)
- || code == VIEW_CONVERT_EXPR)
- && (!vectype_in
- || maybe_ne (TYPE_VECTOR_SUBPARTS (vectype_in), nunits)
- || maybe_ne (GET_MODE_SIZE (TYPE_MODE (vectype)),
- GET_MODE_SIZE (TYPE_MODE (vectype_in)))))
+ /* We can handle VIEW_CONVERT conversions that do not change the number
+ of elements or the vector size or other conversions when the component
+ mode keeps the same. */
+ if (!vectype_in
+ || maybe_ne (TYPE_VECTOR_SUBPARTS (vectype_in), nunits)
+ || (code == VIEW_CONVERT_EXPR
+ && maybe_ne (GET_MODE_SIZE (TYPE_MODE (vectype)),
+ GET_MODE_SIZE (TYPE_MODE (vectype_in))))
+ || (CONVERT_EXPR_CODE_P (code)
+ && (TYPE_MODE (TREE_TYPE (vectype))
+ != TYPE_MODE (TREE_TYPE (vectype_in)))))
return false;
if (VECTOR_BOOLEAN_TYPE_P (vectype) != VECTOR_BOOLEAN_TYPE_P (vectype_in))
--
2.35.3
