The testcase in the PR shows that when we have a reduction chain
with a wrapped conversion we fail to properly fall back to a
regular reduction, resulting in wrong-code. The following fixes
this by failing discovery. The testcase has other issues, so
I'm not including it here.
Bootstrapped and tested on x86_64-unknown-linux-gnu, pushed.
PR tree-optimization/121592
* tree-vect-slp.cc (vect_analyze_slp): When SLP reduction chain
discovery fails, fail overall when the tail of the chain
isn't also the entry for the non-SLP reduction.
---
gcc/tree-vect-slp.cc | 5 +++++
1 file changed, 5 insertions(+)
diff --git a/gcc/tree-vect-slp.cc b/gcc/tree-vect-slp.cc
index 628d8a02ba9..fd793a36da8 100644
--- a/gcc/tree-vect-slp.cc
+++ b/gcc/tree-vect-slp.cc
@@ -5004,6 +5004,11 @@ vect_analyze_slp (vec_info *vinfo, unsigned
max_tree_size,
vinfo = next;
}
STMT_VINFO_DEF_TYPE (first_element) = vect_internal_def;
+ /* ??? When there's a conversion around the reduction
+ chain 'last' isn't the entry of the reduction. */
+ if (STMT_VINFO_DEF_TYPE (last) != vect_reduction_def)
+ return opt_result::failure_at (vect_location,
+ "SLP build failed.\n");
/* It can be still vectorized as part of an SLP reduction. */
loop_vinfo->reductions.safe_push (last);
}
--
2.43.0
