[vectorIntrinsics] RFR: Refactor JVM Interface

[Viswanathan, Sandhya]

Hi Vladimir,

>> The patch also fixes one long standing vbox related issue that we were seeing on one of our platforms.

>Can you elaborate here a bit, please? Is it related to missing safepoint polls?

The failure used to be reported as follows for VectorConversionTests.java on KNL platform. With the refactor now the test passes.
 
# To suppress the following error report, specify this argument
# after -XX: or in .hotspotrc:  SuppressErrorAt=/compile.cpp:2760
#
# A fatal error has been detected by the Java Runtime Environment:
#
#  Internal Error ( vectorIntrinsics/src/hotspot/share/opto/compile.cpp:2760), pid=268802, tid=268817
#  Error: assert(!vbox->is_Phi()) failed
...
# Problematic frame:
# V  [libjvm.so+0x932921]  Compile::expand_vbox_node_helper(Node*, Node*, TypeInstPtr const*, TypeVect const*)+0x81
#

From hs_err file:

Current CompileTask:
C2: 639330 2033 %           VectorConversionTests::conversion_kernel @ 84 (300 bytes)

Stack: [0x00007ff4680ef000,0x00007ff4681f0000],  sp=0x00007ff4681eb450,  free space=1009k
Native frames: (J=compiled Java code, A=aot compiled Java code, j=interpreted, Vv=VM code, C=native code)
V  [libjvm.so+0x932921]  Compile::expand_vbox_node_helper(Node*, Node*, TypeInstPtr const*, TypeVect const*)+0x81
V  [libjvm.so+0x932d84]  Compile::expand_vbox_node(VectorBoxNode*)+0x84
V  [libjvm.so+0x942cc8]  Compile::expand_vbox_nodes()+0x128
V  [libjvm.so+0x9437e8]  Compile::Optimize()+0x648
V  [libjvm.so+0x945612]  Compile::Compile(ciEnv*, C2Compiler*, ciMethod*, int, bool, bool, bool, DirectiveSet*)+0x1172
V  [libjvm.so+0x7b55d5]  C2Compiler::compile_method(ciEnv*, ciMethod*, int, DirectiveSet*)+0x305
V  [libjvm.so+0x9523af]  CompileBroker::invoke_compiler_on_method(CompileTask*)+0x38f
V  [libjvm.so+0x953263]  CompileBroker::compiler_thread_loop()+0x393
V  [libjvm.so+0x17ce56a]  JavaThread::thread_main_inner()+0x17a
V  [libjvm.so+0x17d3bc6]  Thread::call_run()+0xf6
V  [libjvm.so+0x13972b5]  thread_native_entry(Thread*)+0x125

Best Regards,
Sandhya

-----Original Message-----
From: Vladimir Ivanov <vladimir.x.ivanov at oracle.com> 
Sent: Wednesday, March 25, 2020 3:22 AM
To: Viswanathan, Sandhya <sandhya.viswanathan at intel.com>; panama-dev <panama-dev at openjdk.java.net>
Subject: Re: [vectorIntrinsics] RFR: Refactor JVM Interface

Hi Sandhya,

> The vector api tests pass on all Intel platforms with this patch.

Thank you!

> Only one minor build issue was there due to a last remaining UseVectorApiIntrinsics use in stubGenerator_x86_64.cpp.

Good catch. Fixed.

> The patch also fixes one long standing vbox related issue that we were seeing on one of our platforms.

Can you elaborate here a bit, please? Is it related to missing safepoint polls?

> Please go ahead and check in.

Pushed.

Best regards,
Vladimir Ivanov

> -----Original Message-----
> From: panama-dev <panama-dev-bounces at openjdk.java.net> On Behalf Of 
> Vladimir Ivanov
> Sent: Tuesday, March 24, 2020 2:07 PM
> To: panama-dev <panama-dev at openjdk.java.net>
> Subject: [vectorIntrinsics] RFR: Refactor JVM Interface
> 
> http://cr.openjdk.java.net/~vlivanov/panama/vector/new_jvm_interface/w
> ebrev.00/
> 
> (First cleanup pass over the JVM implementation.
> The focus is on JVM-JDK interface.)
> 
> New JVM interface is minimal and contains intrinsics, well-known 
> classes, and opcode declarations. It resides in java.base
> (jdk.internal.vm.vector.VectorSupport) and exports the interface to jdk.incubator.vector module.
> 
> JVM doesn't need to know about typed vectors anymore: VectorPayload, 
> VectorMask, and VectorShuffle are enough to represent all interesting 
> cases and provide enough information to the JVM how to handle vector
> classes:
> 
>     - VectorPayload class represents vector values which are amenable to aggressive box elimination transformation
>        * vector on-heap representation remains primitive array-backed;
>        * classes which extend VectorPayload should declare 2 static 
> fields (VLENGTH and ETYPE) which describe vector on-heap 
> representation to the JVM (array length and its element type);
> 
>     - VectorShuffle and VectorMask mark vector shuffles and masks for 
> the JVM;
> 
>     - VectorSpecies and Vector are there to enhance type checking of 
> JVM intrinsic usages;
> 
> 
> Also, additional refactorings/fixes:
> 
>     * merged reinterpret and cast intrinsics into one (convert)
> 
>     * refactored rematerialization support
> 
>     * don't remove unused VectorBoxAllocate nodes, but replace them 
> with safepoints
> 
>     * VM vector support is turned off by default, but automatically 
> enabled when user adds jdk.incubator.vector incubator module to the 
> graph
> 
>     * multiple minor fixes and cleanups along the way
> 
> Testing: jdk/incubator/vector tests
> 
> Best regards,
> Vladimir Ivanov
>