AARCH64: 8139041: Redundant DMB instructions (CORRECTED )

[Andrew Haley]

Hi,

On 10/09/2015 03:59 PM, Roland Westrelin wrote:

>> There is a much simpler way: remove adjacent barriers in
>> MacroAssembler.  Thanks to the way that the AArch64 ISA is designed,
>> barriers can be merged simply by ORing them together.  Of course, this
>> technique works for C1 and C2, and it adds essentially nothing to the
>> compilation time.
>>
>> http://cr.openjdk.java.net/~aph/8139041/
>>
>> One thing which may be controversial is that I've added a field to
>> CodeBuffer to keep track of barriers and labels.  I had to do this
>> because when we're compiling there is (AFAICS) essentially nowhere
>> else to keep the state.
> 
> Isn’t your new field a bit like:
> 
>   address       insts_mark() const       { return _insts.mark();       }
>   void      set_insts_mark()             {        _insts.set_mark();   }
>   void    clear_insts_mark()             {        _insts.clear_mark(); }
> 
> which is used in very few locations AFAIK. Do you think you could reuse that one?

Yes, that's what it's based on.  I guess that is possible in theory,
but AbstractAssembler::InstructionMark() looks like this:

    InstructionMark(AbstractAssembler* assm) : _assm(assm) {
      assert(assm->inst_mark() == NULL, "overlapping instructions");
      _assm->set_inst_mark();
    }

so any instruction which leaves the mark set will trigger an assertion
failure the next time InstructionMark is used.  I suppose that in
extremis I could make every instruction which is not a memory barrier
clear the mark, but ewww.  :(

I suppose I could define an AArch64-specific version of
InstructionMark which does not have this assert, but I'm not sure I
like that either.

[An aside: we use InstructionMark unnecessarily in AArch64, and it's
on my list of things to remove, but that's for another day.]

Thanks for looking at this,

Andrew.