Target Environment - Cross-compilation
doug.simon at oracle.com
Fri Jan 20 09:42:15 UTC 2017
> On 20 Jan 2017, at 10:31, Andrew Haley <aph at redhat.com> wrote:
> On 19/01/17 23:55, David Clarke wrote:
>> So on x86 we want to optimise away a full fence that immediately
>> follows a weakCompareAndSet and on ARM, we don’t. This could be
>> delayed until the code generation phase, but I would like to keep
>> the optimisations at the IR or LIR level so that we retain the
>> greatest generality for as long as possible and avoid repeating
>> similar code in several distinct code generation phases.
> There's a related problem on ARMv8. At the present time we emit
> fences very early, which means that we don't generate the STLR/LDAR
> instructions we really want for Java volailes. At the very least this
> is inefficient.
>> So my questions are “How can the code within a Node find out the
>> target environment?" and “Can we fiddle this so that the target
>> environment is different from the actual host environment?”
> I think that the real answer is to handle all of this properly in the
> Graal IR by lowering the nodes in a way that is controlled by the back
> end. This will mean that the compiler will have to be taught about
> how to handle the nodes appropriately.
Yes, you will need something like the AddressLowering/AddressLoweringPhase mechanism Graal currently has.
More information about the graal-dev