Re: [Jikesrvm-researchers] how to add a new IR instruction?

[David F. B. <df...@wa...>]

What do you plan to use this for?  We already emit CAS for certain things, like
locking primitives and for pointer updates in the concurrent collector.  See
VM_Magic.prepare() and VM_Magic.attempt().  However, there may be a better
general-purpose solution.

david

David Hovemeyer wrote:

> Hi everyone,
>
> I want to modify Jikes RVM's optimizing compiler to emit code for
> atomic compare-and-swap (CAS) in certain circumstances.  Since I don't
> know much about the internals of RVM, I have some beginner questions.
>
> Should CAS should be an HIR or LIR instruction?
>
> CAS needs at least three operands: memory location, expected value,
> updated value.  Will I need to define a new instruction format for
> this?
>
> Does it also need a guard operand, for the result of the null check on
> the memory location?  (What exactly are the guard operands used for?)
>
> What is the procedure for adding a new IR instruction?  (Which files
> need to be changed.)
>
> How/where is the machine code actually emitted for IR instructions?
>
> What configuration should I use if I don't want the optimizing compiler
> compiled into the boot image?  I'm assuming that it would be better to
> load it at runtime, to avoid having to make a new boot image each time
> I modify the opt compiler.  Currently I'm using OptOptSemispace.
>
> Initially, I would like to get this working with PowerPC Linux, using
> the lwarx and stwcx instructions.  Later I might want to implement it
> on x86 using the locked cmpxchg instruction.
>
> Any help would be greatly appreciated.
>
> Thanks,
> Dave
> _______________________________________________
> Jikesrvm-researchers mailing list
> Jik...@ww...
> http://www-124.ibm.com/developerworks/oss/mailman/listinfo/jikesrvm-researchers