[vectorIntrinsics+mask] RFR: 8266287: Basic mask IR implementation for the Vector API masking feature support [v3]

[Xiaohong Gong]

On Tue, 22 Jun 2021 09:43:07 GMT, Jatin Bhateja <jbhateja at openjdk.org> wrote:

>> Hi @XiaohongGong ,
>> 
>> I made following changes in unbox and box creation routines and did not find any crashes as reported in any of VectorAPI jtreg over AVX-512 platform.
>> 
>> 
>> diff --git a/src/hotspot/share/opto/vectorIntrinsics.cpp b/src/hotspot/share/opto/vectorIntrinsics.cpp
>> index 21e238afb5b..76553c9f2c9 100644
>> --- a/src/hotspot/share/opto/vectorIntrinsics.cpp
>> +++ b/src/hotspot/share/opto/vectorIntrinsics.cpp
>> @@ -81,7 +81,7 @@ Node* GraphKit::box_vector(Node* vector, const TypeInstPtr* vbox_type, BasicType
>>    Node* ret = gvn().transform(new ProjNode(alloc, TypeFunc::Parms));
>> 
>>    assert(check_vbox(vbox_type), "");
>> -  const TypeVect* vt = TypeVect::make(elem_bt, num_elem, is_vector_mask(vbox_type->klass()));
>> +  const TypeVect* vt = TypeVect::make(elem_bt, num_elem);
>>    VectorBoxNode* vbox = new VectorBoxNode(C, ret, vector, vbox_type, vt);
>>    return gvn().transform(vbox);
>>  }
>> @@ -96,7 +96,7 @@ Node* GraphKit::unbox_vector(Node* v, const TypeInstPtr* vbox_type, BasicType el
>>      return NULL; // no nulls are allowed
>>    }
>>    assert(check_vbox(vbox_type), "");
>> -  const TypeVect* vt = TypeVect::make(elem_bt, num_elem, is_vector_mask(vbox_type->klass()));
>> +  const TypeVect* vt = TypeVect::make(elem_bt, num_elem);
>>    Node* unbox = gvn().transform(new VectorUnboxNode(C, vt, v, merged_memory(), shuffle_to_vector));
>>    return unbox;
>>  }
>> 
>> Can you kindly share a exact re-producer.
>
> Box/Unbox are macro nodes which encapsulates value, in this case the mask value is contained either in  a a vector or a predicate register. This should ideally not be disturbing the box's vector type() I think.

Did you implement the vector mask IRs with `TypeVectMask` type in backend? I made a workaround to make sure the tests can pass with this patch but without any backend implementation for AVX-512/SVE. Please see the change in `type.cpp`:

const TypeVect *TypeVect::makemask(const Type* elem, uint length) {
  if (Matcher::has_predicated_vectors() &&
      // TODO: remove this condition once the backend is supported.
      // Workround to make tests pass on AVX-512/SVE when predicate is not supported.
      // Could be removed once the backend is supported.
      Matcher::match_rule_supported_vector_masked(Op_StoreVectorMasked, MaxVectorSize, T_BOOLEAN)) {
    const TypeVect* mtype = Matcher::predicate_reg_type(elem, length);
    return (TypeVect*)(const_cast<TypeVect*>(mtype))->hashcons();
  } else {

The `TypeVectMask` can only be made successfully when the backend has implemented masked `Op_StoreVectorMasked`. This is just a workaround that would be removed once the AVX-512 add the backend implementation for all mask/masked IRs.

-------------

PR: https://git.openjdk.java.net/panama-vector/pull/78