If-conversion now applies rewrite_to_defined_overflow to the
address calculation in an IFN_MASK_LOAD. This means that we
end up with:
cast_base = (uintptr_t) base;
uncast_sum = cast_base + offset;
sum = (orig_type *) uncast_sum;
If the target supports IFN_MASK_GATHER_LOAD with pointer-sized
offsets for the given vectype, we wouldn't look through the sum
cast and so would needlessly vectorise the uncast_sum addition.
This showed up as several failures in gcc.target/aarch64/sve.
Regstrapped on aarch64-linux-gnu and x86_64-linux-gnu. OK to install?
Richard
gcc/
* tree-vect-data-refs.c (vect_check_gather_scatter): Continue
processing conversions if the current offset is a pointer.
---
gcc/tree-vect-data-refs.c | 1 +
1 file changed, 1 insertion(+)
diff --git a/gcc/tree-vect-data-refs.c b/gcc/tree-vect-data-refs.c
index f1d7f01a9ce..888ad72f3a9 100644
--- a/gcc/tree-vect-data-refs.c
+++ b/gcc/tree-vect-data-refs.c
@@ -4139,6 +4139,7 @@ vect_check_gather_scatter (stmt_vec_info stmt_info,
loop_vec_info loop_vinfo,
/* Don't include the conversion if the target is happy with
the current offset type. */
if (use_ifn_p
+ && !POINTER_TYPE_P (TREE_TYPE (off))
&& vect_gather_scatter_fn_p (loop_vinfo, DR_IS_READ (dr),
masked_p, vectype, memory_type,
TREE_TYPE (off), scale, &ifn,
--
2.25.1
