Labels as values and threaded-code interpretation

[Dmitry Olshansky]

02-Jun-2013 22:25, Alex Rønne Petersen пишет:
> On 02-06-2013 19:58, Dmitry Olshansky wrote:
>>
>> One pointer to a struct MyInterpreterState. Think of it as 'this'
>> pointer :)
>
> That's one load and store for every single step in the interpreter.
> Sounds harmless, but every cycle matters when every goto start(...)
> amounts to a single instruction in the program you're running.
>

If the instruction set is RISC-y then you are in trouble anyway (writing 
interpreter not in ASM). One way out is fusing multiple simple 
instructions into macro-instructions and interpreting those.

> Either way, evidence suggests that this style of interpreter has
> performance advantages in practice.
>

Agreed. Though this day I'd go for what I call "bastardized JIT". In 
essense just produce a chunk of "call XYZ" instructions, comnditionals 
would have to have a test + jmp. And then there is 'ret' at the end of 
this chunk ~4 instructions to have to deal with. What you get is branch 
prediction done on the hardware level and no instruction counter to 
worry about.

-- 
Dmitry Olshansky