RFR: 8291600: [vectorapi] vector cast op check is not always needed for vector mask cast [v5]
Jie Fu
jiefu at openjdk.org
Wed Aug 31 03:59:15 UTC 2022
On Wed, 31 Aug 2022 03:08:33 GMT, Xiaohong Gong <xgong at openjdk.org> wrote:
>> src/hotspot/share/opto/vectorIntrinsics.cpp line 2498:
>>
>>> 2496: }
>>> 2497: if (is_mask && is_floating_point_type(elem_bt_to)) {
>>> 2498: new_elem_bt_to = elem_bt_to == T_FLOAT ? T_INT : T_LONG;
>>
>> Why we need this change?
>
> Before `VectorCast` is used for vector mask casting, the src floating point type has been casted to an integral type that has the same element size. Please see the codes here: https://github.com/openjdk/jdk/blob/master/src/hotspot/share/opto/vectornode.cpp#L1741. We'd better to check the right type of the generated vector cast op.
OK, seems fine to me.
Thanks.
>> src/hotspot/share/opto/vectorIntrinsics.cpp line 2503:
>>
>>> 2501:
>>> 2502: // Make sure that vector cast is implemented to particular type/size combination.
>>> 2503: bool no_vec_cast_check = is_mask && (src_type->isa_vectmask() ||
>>
>> To make it safer, let's also check `dst_type->isa_vectmask()` like the following.
>>
>> if (is_mask) {
>> if ((dst_type->isa_vectmask() && src_type->isa_vectmask()) ||
>> (type2aelembytes(elem_bt_from) == type2aelembytes(elem_bt_to))) {
>> op = gvn().transform(new VectorMaskCastNode(op, dst_type));
>> } else {
>> op = VectorMaskCastNode::makeCastNode(&gvn(), op, dst_type);
>> }
>> }
>
> From the check https://github.com/openjdk/jdk/blob/master/src/hotspot/share/opto/vectorIntrinsics.cpp#L2485, we know the basic mask type of dst and src must be the same. So just checking `src_type` is ok here. Besides, this part of changes will be removed finally after I unify the vector mask cast operation.
But the mismatch seems strange to me.
Maybe, you can
1) Add `dst_type->isa_vectmask()` check here, or
2) Also delete `dst_type->isa_vectmask()` (https://github.com/openjdk/jdk/blob/master/src/hotspot/share/opto/vectorIntrinsics.cpp#L2551)
-------------
PR: https://git.openjdk.org/jdk/pull/9737
More information about the hotspot-compiler-dev
mailing list