RFR: 8183546: [MVT] cleanup/improve vreturn convention implementation

Frederic Parain frederic.parain at oracle.com
Wed Jul 5 19:11:31 UTC 2017


Thank you very much for improving the vreturn convention implementation.

Just a minor nit:

valueKlass.cpp:442: Missing message in assert?


> On Jul 5, 2017, at 03:33, Roland Westrelin <rwestrel at redhat.com> wrote:
> http://cr.openjdk.java.net/~roland/valhalla/vreturn-improv/webrev.00/
> This patch:
> - reworks the verification logic in
>  SharedRuntime::store_value_type_fields_to_buf() to address the crash
>  that Paul encountered
> - changes the vreturn convention so if we return the klass pointer +
>  fields, the klass pointer is tagged by setting its least significant
>  bit. This way it's cheaper for the caller to detect if it's getting an
>  oop, a buffered value or a list of fields back from the callee. (this
>  is based on a discussion with Fred)
> - stores precomputed extended signature and vreturn registers in the
>  ValueKlass to not have to compute them every time they are
>  needed. This is also something Fred suggested. I considered
>  initializaing those fields on first use but came to the conclusion it
>  would make things too complicated. For instance the vreturn registers
>  are sometimes use at a point where the thread can't safepoint making
>  it impossible to use a metadata Array.
> Roland.

More information about the valhalla-dev mailing list