RFR: 8217874: Shenandoah: AArch64: Clobbered register in ShenandoahBarrierSetAssembler::cmpxchg_oop()
adinn at redhat.com
Mon Feb 11 12:49:51 UTC 2019
On 11/02/2019 10:49, Roman Kennke wrote:
>> In AArch64, when called from C2, in
>> ShenandoahBarrierSetAssembler::cmpxchg_oop() the result register may
>> overlap with other input argument registers and thus fail the leading
>> assert, and lead to clobbered registers. In the body of the code block,
>> a temporary register should be used instead, and result should only get
>> filled in at the end.
That looks convincing except that the original code avoided using
rscratch1. Was there a reason for that?
It may be that way because the original cmpxchg code that this is
derived from allowed that callers might be holding something in
rscratch1 across the cmpxchg call. Will that ever happen for your usage
in the barrier set?
If you are sure that this won't happen and that all the gubbins you call
from the barrier cmpxchg_oop implementation doesn't use rscratch1 then I
think this is ok.
Of course, if the result/input overlap is only ever happening in calls
from C2 then you could fix this more simply in the C2 rules. Allocate a
temp register in each rule that encodes a call to cmpxchg_oop, pass that
temp register in as the 'res' argument and then plant a move to the
output register after the call to cmpxchg_oop.
Senior Principal Software Engineer
Red Hat UK Ltd
Registered in England and Wales under Company Registration No. 03798903
Directors: Michael Cunningham, Michael ("Mike") O'Neill, Eric Shander
More information about the hotspot-compiler-dev