optimizing acmp in L-World

John Rose john.r.rose at oracle.com
Thu Feb 8 07:03:59 UTC 2018


On Feb 7, 2018, at 5:27 PM, Tobias Hartmann <tobias.hartmann at oracle.com <mailto:tobias.hartmann at oracle.com>> wrote:
> 
> Knowing that ObjectAlignmentInBytes is guaranteed to be >= 2, we can do even better:
> 
>  __ testptr(rdx, rdx);
>  __ jcc(Assembler::zero, is_null);
>  __ movl(rbx, Address(rdx, oopDesc::klass_offset_in_bytes()));
>  __ andptr(rbx, Universe::oop_metadata_odd_mask());
>  __ orptr(rdx, rbx);

Yep.  And if metadata_odd_mask is some higher number like
0x10 or 0x20 (which I think it should be, in the long run*) then you
just need to add a right-shift to pull it down into to the mask 0x07.

In future systems if we use a color bit to mark the buffer pointer
of the value type, we can just clear that single bit, in one of the
operands, to force x != x.

— John

* P.S. So metadata_odd_mask is 0x01.  I missed that detail the
first time around.  If we align klass metadata pointers a little
more strongly we can make it a larger number like 0x10 or 0x20,
with the effect that we don't need an extra instruction to remove
that bit when using the klass pointer.  We could just align value
and object klasses a little differently, by allocating them (say)
even mod 0x20 and then (say) bump each value klass upward
by 0x10.  The fragmentation overhead would be small relative
to the (probable) speed improvement of omitting the detag
operation.



More information about the valhalla-dev mailing list