Hello All, I just realized that M0 is stack based. M0 is dedicated to use the least op to do the work, it's extreme makes it being stack based, not register based. If the latter is more advanced, then it's a retrogression. Here is some compares:
jvm-ish: push I0, 123 push I1, 123 add I2, I1, I0 CPU-ish: mov %eax, 123 add %eax, 356 or add %eax, %ebx or add %eax, [0] # [0] is memory unit. or add [0], %eax # [0] is memory unit.
In M0, we can't do this: add_i I2, 123, 356
which should be right way of register based VM, and which parrot currently does. I don't think parrot current ops implementation is bad, we should reuse it to implement M0 ops. And I don't want M0 is stack based for easier JIT too. Any thought or viewpoint?
> I just realized that M0 is stack based. M0 is dedicated to use the least op to do the work, it's extreme makes it being stack based, not register based. If the latter is more advanced, then it's a retrogression. Here is some compares:
> M0: > set_imm I0, 0, 123 > set_imm I1, 0, 356 > add_i I2, I1, I0 > CPU-ish: > mov %eax, 123 > add %eax, 356 > or > add %eax, %ebx > or > add %eax, [0] # [0] is memory unit. > or > add [0], %eax # [0] is memory unit.
> In M0, we can't do this: > add_i I2, 123, 356
I have thought it would be nice if M0 had some concept of addressing modes. Indirect, indexed, and immediate are things that jump to mind quickly. They could be handled by the core runloop before dispatching off to the ops for simplicity.
On Sun, Apr 1, 2012, at 23:51, Brian Gernhardt wrote:
> On Apr 1, 2012, at 11:17 PM, Jimmy Zhuo wrote:
> > I just realized that M0 is stack based. M0 is dedicated to use the > > least op to do the work, it's extreme makes it being stack based, > > not register based. If the latter is more advanced, then it's a > > retrogression. Here is some compares:
> > CPU-ish: > > mov %eax, 123 > > add %eax, 356 > > or > > add %eax, %ebx > > or > > add %eax, [0] # [0] is memory unit. > > or > > add [0], %eax # [0] is memory unit.
> > In M0, we can't do this: > > add_i I2, 123, 356
> I have thought it would be nice if M0 had some concept of addressing > modes. Indirect, indexed, and immediate are things that jump to mind > quickly. They could be handled by the core runloop before dispatching > off to the ops for simplicity.
> ~~ Benabik
The big goal of M0 is stupid simplicity, but I also don't want meaningfully efficient execution to require an optimizing assembler. Adding addressing modes isn't out of the question.
I'm busier than I should be at the moment, but I'd love to see an experiment with addressing modes. If either of you have the time, please feel free to fork the m0 branch and hack out something on the Perl interpreter. This kind of experimentation is exactly what it's for.
