https://gcc.gnu.org/bugzilla/show_bug.cgi?id=98211
Richard Biener <rguenth at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
CC| |duanbo3 at huawei dot com,
| |rsandifo at gcc dot gnu.org
--- Comment #6 from Richard Biener <rguenth at gcc dot gnu.org> ---
OK, ICEs all over the place as expected.
Btw, on the GCC 10 branch we had
diff --git a/gcc/tree-vect-patterns.c b/gcc/tree-vect-patterns.c
index 6a5f9505792..52725f826c7 100644
--- a/gcc/tree-vect-patterns.c
+++ b/gcc/tree-vect-patterns.c
@@ -4452,16 +4452,7 @@ vect_recog_mask_conversion_pattern (vec_info *vinfo,
if (TREE_CODE (rhs1) != SSA_NAME)
{
tmp = vect_recog_temp_ssa_var (TREE_TYPE (rhs1), NULL);
- if (rhs1_op0_type
- && TYPE_PRECISION (rhs1_op0_type) != TYPE_PRECISION (rhs1_type))
- rhs1_op0 = build_mask_conversion (vinfo, rhs1_op0,
- vectype2, stmt_vinfo);
- if (rhs1_op1_type
- && TYPE_PRECISION (rhs1_op1_type) != TYPE_PRECISION (rhs1_type))
- rhs1_op1 = build_mask_conversion (vinfo, rhs1_op1,
- vectype2, stmt_vinfo);
- pattern_stmt = gimple_build_assign (tmp, TREE_CODE (rhs1),
- rhs1_op0, rhs1_op1);
+ pattern_stmt = gimple_build_assign (tmp, rhs1);
rhs1 = tmp;
append_pattern_def_seq (vinfo, stmt_vinfo, pattern_stmt, vectype2,
rhs1_type);
and going back fixes the issue as well. That was changed by g:91d80cf4bd28,
approved by Richard.
That said, using build_mask_conversion on sth not already a mask is definitely
wrong as can be seen here. The above conversions should, according to the
revs comment, only happen when rhs1_op0_type != rhs1_op1_type (those are
scalar types here - but revs comment says vector types mismatch).
That said, if we throw bool patterns ontop of the mask conversion one we end
up not vectorizing the thing because
x3.c:23:9: note: ==> examining statement: patt_22 = _28 ? 1 : 0;
x3.c:6:1: missed: not vectorized: relevant stmt not supported: patt_22 = _28
? 1 : 0;
which is because vect_is_simple_cond runs into
/* Mask case. */
if (TREE_CODE (cond) == SSA_NAME
&& VECT_SCALAR_BOOLEAN_TYPE_P (TREE_TYPE (cond)))
{
if (!vect_is_simple_use (vinfo, stmt_info, slp_node, 0, &cond,
&slp_op, &dts[0], comp_vectype)
|| !*comp_vectype
|| !VECTOR_BOOLEAN_TYPE_P (*comp_vectype))
return false;
so I guess that should end up as mask conversion pattern as well, recursively.
Meh.
