tree-optimization/114921 - _Float16 -> __bf16 isn't noop

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.

	PR tree-optimization/114921
	* tree-vect-stmts.cc (vectorizable_assignment): Require
	same vector component modes for input and output for
	CONVERT_EXPR_CODE_P.

(cherry picked from commit 87e35da16d)
This commit is contained in:
Richard Biener
2024-05-02 13:55:15 +02:00
parent f86f1970e0
commit fa7e05d90b

View File

@@ -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))