MVT-based vectors: first stab
Vladimir Ivanov
vladimir.x.ivanov at oracle.com
Wed Jun 28 16:03:11 UTC 2017
> thanks a lot for the nice summary! I think it would makes sense to file enhancements/bugs for the individual issues.
FYI here's the list of issues:
RFEs
https://bugs.openjdk.java.net/browse/JDK-8183130
core-libs/j.l.i [MVT] Implement intrinsics for Q-typed LambdaForms
https://bugs.openjdk.java.net/browse/JDK-8183137
hotspot/compiler [MVT] C2 doesn't eliminate redundant value
type allocations inside loops
Bugs
https://bugs.openjdk.java.net/browse/JDK-8183129
hotspot/compiler [MVT] Type::get_typeflow_type() doesn't
distinguish between TypeValueTypePtr & TypeValueType
https://bugs.openjdk.java.net/browse/JDK-8183131
core-libs/j.l.i [MVT] 2-slot type (long & double) support is
broken in LambdaFormBuilder
https://bugs.openjdk.java.net/browse/JDK-8183132
core-libs/j.l.i [MVT] MHI.unboxResultHandle misses unboxing
step for Q-types
https://bugs.openjdk.java.net/browse/JDK-8183133
core-libs/j.l.i [MVT] Arraylength method handle on Q-typed
array shouldn't use ArrayAccessor.lengthL(Object[])
https://bugs.openjdk.java.net/browse/JDK-8183135
core-libs/j.l.i [MVT] Code builder doesn't fully support
constant pool patching
https://bugs.openjdk.java.net/browse/JDK-8183138
core-libs/j.l.i Code Builder: Problems with handling complex
control flow
Best regards,
Vladimir Ivanov
> On 28.06.2017 15:17, Vladimir Ivanov wrote:
>> I hit some bugs along the way
>
> I had a closer look and (as you already figured out) one problem is that typeflow analysis incorrectly sets the type of a PhiNode to TypeValueType instead of TypeValueTypePtr for __Value. When inlining a method handle intrinsic in CallGenerator::for_method_handle_inline() we should cast all arguments to their actual types but we miss the PhiNode because it's already of TypeValueType:
>
> t: <ciValueKlass name=valhalla/vector/Long2$Value loader=0x00000006d8776320 loaded=true initialized=true finalized=false subklass=false size=32 flags=public,final,super super=java/lang/__Value ident=1108 address=0x00007f218008e270>
> arg_type: valuetype[0]:{empty}
> sig_type: valuetype* valhalla/vector/Long2$Value:NotNull
> arg: 659 Phi === 652 _ 626 [[ 616 662 678 680 684 730 760 785 790 793 837 867 874 889 891 894 ]] #valuetype[0]:{empty} !jvms: 1310540333::invoke @ bci:170 858242339::invokeExact_MT @ bci:19 VectorTest::acc @ bci:4 VectorTest::main @ bci:69
>
> We then fail in Parse::do_vunbox because the input node is a Phi (and not a ValueTypeNode).
>
> I think in addition to the fix in type.cpp, we should add this assert to catch such problems in the future:
>
> --- old/src/share/vm/opto/callGenerator.cpp 2017-06-28 15:25:40.812039870 +0200
> +++ new/src/share/vm/opto/callGenerator.cpp 2017-06-28 15:25:40.720039875 +0200
> @@ -896,10 +896,11 @@
> if (t->is_valuetype()) {
> assert(!(arg_type->isa_valuetype() && t == kit.C->env()->___Value_klass()), "need a pointer to the value type");
> if (arg_type->isa_valuetypeptr() && t != kit.C->env()->___Value_klass()) {
> - const Type* sig_type = TypeOopPtr::make_from_klass(t->as_klass());
> Node* cast = gvn.transform(new CheckCastPPNode(kit.control(), arg, sig_type));
> Node* vt = ValueTypeNode::make(gvn, kit.merged_memory(), cast);
> kit.set_argument(arg_nb, vt);
> + } else {
> + assert(t == kit.C->env()->___Value_klass() || arg->is_ValueType(), "inconsistent argument");
> }
> } else {
> if (arg_type->isa_oopptr() && !arg_type->higher_equal(sig_type)) {
>
> I'll take the chance to look at the method handle implementation in a bit more detail and try to come up with a regression test for this to extend our ValueTypesTestBench.
>
> Best regards,
> Tobias
>
More information about the valhalla-dev
mailing list