https://gcc.gnu.org/bugzilla/show_bug.cgi?id=124068
Bug ID: 124068
Summary: vect_recog_over_widening_pattern fails to perform
target check
Product: gcc
Version: 16.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: tree-optimization
Assignee: unassigned at gcc dot gnu.org
Reporter: rguenth at gcc dot gnu.org
Target Milestone: ---
vect_recog_over_widening_pattern happily turns a vector int << int operation
into a vector short << short even when that is not supported. This can be
seen when testing gcc.dg/vect/pr121949_1.c with -mavx2 which supports
variable shift with SImode but not HImode (AVX512BW is required for this).
Of course pattern detection has no idea whether we'll be doing a vector by
scalar or a vector by vector shift. Vector by scalar shift is supported
with SSE already.
The following adds such check, but it will regress the vector-by-scalar case
(heuristically a check for external/constant def would be possible, not
considering SLP)
diff --git a/gcc/tree-vect-patterns.cc b/gcc/tree-vect-patterns.cc
index 97130206a21..34c6bc00450 100644
--- a/gcc/tree-vect-patterns.cc
+++ b/gcc/tree-vect-patterns.cc
@@ -3167,7 +3167,10 @@ vect_recog_over_widening_pattern (vec_info *vinfo,
where beneficial. */
tree new_vectype = get_vectype_for_scalar_type (vinfo, new_type);
tree op_vectype = get_vectype_for_scalar_type (vinfo, op_type);
- if (!new_vectype || !op_vectype)
+ optab optab1;
+ if (!new_vectype || !op_vectype
+ || !(optab1 = optab_for_tree_code (code, op_vectype, optab_vector))
+ || !can_implement_p (optab1, TYPE_MODE (op_vectype)))
return NULL;
if (dump_enabled_p ())
