[foreign-memaccess+abi] RFR: 8275646: Implement optimized upcall stubs on AArch64 [v2]

[Jorn Vernee]

On Mon, 8 Nov 2021 09:13:43 GMT, Nick Gasson <ngasson at openjdk.org> wrote:

>> This is a fairly direct port of the X86 code.  The changes to
>> frame_aarch64 and foreign_globals_aarch64 are identical: perhaps
>> ForeignGlobals::parse_call_regs_impl() could be moved to shared code?
>> 
>> The X86 version has a call to reinit_heapbase() before the return from
>> the stub.  I think this is redundant because the heap base register will
>> be clobbered immediately by the native code?
>> 
>> I hit a really weird bug in release builds where the first few
>> instructions in the code buffer were overwritten by the fields of
>> OptimizedEntryBlob.  I think we need to pass sizeof(OptimizedEntryBlob)
>> instead of sizeof(BufferBlob) as the header_size argument to the
>> RuntimeBlob constructor (none of the other subclasses of BufferBlob have
>> extra fields).  I added a header_size argument to BufferBlob's
>> constructor to thread this through.
>> 
>> I removed the calls to change the W^X state in on_entry/on_exit calls:
>> in the on_entry case the stub must already be executable because we
>> called into the VM from there, and for on_exit we need the code to be
>> executable not writable otherwise we'll get a SIGBUS as soon as we
>> return to the stub.  The newly added stack tests in TestUpcall hit
>> JDK-8275584 on MacOS/AArch64 so I've problem-listed that for now.
>> 
>> JMH results from org.openjdk.bench.jdk.incubator.foreign.Upcalls before:
>> 
>> 
>> Benchmark                Mode  Cnt     Score    Error  Units
>> Upcalls.jni_args10       avgt   30   450.417 ?  4.755  ns/op
>> Upcalls.jni_args5        avgt   30   245.898 ?  3.171  ns/op
>> Upcalls.jni_blank        avgt   30   195.606 ?  5.459  ns/op
>> Upcalls.jni_identity     avgt   30   369.788 ? 15.165  ns/op
>> Upcalls.panama_args10    avgt   30  1253.189 ? 62.261  ns/op
>> Upcalls.panama_args5     avgt   30   927.101 ? 35.369  ns/op
>> Upcalls.panama_blank     avgt   30   637.708 ? 11.353  ns/op
>> Upcalls.panama_identity  avgt   30   697.109 ?  9.971  ns/op
>> 
>> 
>> After:
>> 
>> 
>> Benchmark                Mode  Cnt    Score    Error  Units
>> Upcalls.jni_args10       avgt   30  455.304 ? 21.838  ns/op
>> Upcalls.jni_args5        avgt   30  247.279 ?  2.513  ns/op
>> Upcalls.jni_blank        avgt   30  194.113 ?  4.317  ns/op
>> Upcalls.jni_identity     avgt   30  366.145 ?  4.912  ns/op
>> Upcalls.panama_args10    avgt   30  236.337 ? 11.072  ns/op
>> Upcalls.panama_args5     avgt   30  223.858 ? 12.345  ns/op
>> Upcalls.panama_blank     avgt   30  203.631 ?  8.840  ns/op
>> Upcalls.panama_identity  avgt   30  208.783 ?  9.914  ns/op
>> 
>> 
>> Tested tier1 and jdk_foreign on Linux/AArch64 and MacOS/AArch64.
>
> Nick Gasson has updated the pull request incrementally with one additional commit since the last revision:
> 
>   Review comments

LGTM. Feel free to integrate like this.

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

Marked as reviewed by jvernee (Committer).

PR: https://git.openjdk.java.net/panama-foreign/pull/610