Re: ISA checks for factories

[Andrew Cagney <cagney at redhat dot com>]

Phil,

unfortunate timing, we crossed e-mails during the night,

Phil Muldoon wrote:
> I've written several factories to decide what class to instantiate 
> depending on the ISA. These usually end up as hacks, probably because 
> I do not understand something quite right, or there is no real decent  
> "What ISA does this task/proc belong" discovery.
>
> I can instance check an ISA with:
>
> Isa isa = proc.getIsa();
There is now also:
   ISA isa = task.getISA()
which returns a frysk.isa.ISA, which looks smells and behaves like an 
ENUM.  This new ISA just the following attributes:
  - family (i386, x8664, PPC)
  - word size (4, 8)
  - byte order (little, big)
leading to the combinations: IA32, X8664, PPC32LE, PPC32BE, PPC64LE, 
PPC64BE.

> if (isa instanceof IsaIA32) ....

It can then be used as either:

   if (isa == ISA.IA32)
     ..
   else if (isa = ISA.X8664)
     ..

or better?:

  ISAMap map = new ISAMap("why am I here")
      .put(ISA.IA32, new My(IA32))
      .put(ISA.X8664, new My(X8664))
      ;
  ...
      // Throws a RuntimeException if ISA missing
      My o = (My) map.get(task.getISA);

which could easily be wrapped up in a factory.

Andrew


> and so on, but that will not work for 32 on 64 ISAs. For example, on a 
> 32 bit process on a 64 bit system, the instance returned would be
>
> LinuxIa32On64
>
> And those cannot be instance checked as they are package private, 
> while the "native" ISAs are public.
>
> So what is the solution here? Open the scope of the xxOnxx ISAs? I get 
> the feeling that instanceof checks tell me that we need to work on a 
> better solution anyway. Reverting to reflection screams hack to me.
>
> Regards
>
> Phil