Re: Optimizing stack access for a stack based VM

Jiri Svoboda wrote:


> Local variables should occur as natural operands of instructions (on
> the stack top) as much as possible, removing the need to fetch them
> from the middle of the stack (and encode displacement values).


IMO it doesn't matter whether you encode local variables as offsets
from the SP, or from another base pointer.




> Obviously anytime we want to execute an arithmetic operation it would be
> beneficial to have its operands on the stack top removing the need for
> stkld/stkst (and coding <offs>).
>
> While we get this for free for temporaries (= temporary results in
> arithmetic expressions) just by producing stack-based code from the
> source code, I'd like to extend it to work for variables, too.


Since access to local variables requires some selector (offset...)
value, you'll have to duplicate all operations, for using a local
variable as the second operand. Perhaps further sets are required, for
locals as first or both operands, and operations with non-local
variables would require further instructions. I'd stay with an RISC
instruction set, and only push/pop variables by dedicated
instructions.


Did you have a look at the i80x87 FPU instruction set? Or at the Java
VM, or at FORTH?


DoDi

Return to the comp.compilers page.
Search the comp.compilers archives again.