Re: Languages don't matter. A mathematical refutation

["Randy Brukardt" <randy@rrsoftware.com>]

"Paul Rubin" <no.email@nospam.invalid> wrote in message 
news:87lhi5ayuv.fsf@jester.gateway.sonic.net...
> "Dmitry A. Kazakov" <mailbox@dmitry-kazakov.de> writes:
>> I don't know anything about jet engine, but typical controllers run 
>> 1-10ms
>> cycles. Things requiring shorter times (e.g. frequency measurements) are
>> performed by dedicated microcontrollers.
>
> I assumed that program was on a dedicated microcontroller.
>
>> No, the benefit of arena is unbeatable speed of allocation and zero
>> deallocation time.
>
> You get that from gc with region inference too, and you sort of get it
> with semispace gc.  Unless the compiler is spending a lot of its time in
> gc, arenas by definition can't be that big a win.

Unless the compiler was designed by trained apes, GC isn't going to be much 
of a win, either. Janus/Ada (the compiler I've been working on for the last 
34+ years [gosh!]) does almost no deallocation of anything. Back in the old 
days, we used to deallocate a few things (needed to for compiling on 
MS-DOS), but even that proved to be too expensive and thus the compile never 
deallocates anything at this point unless there is an extreme memory 
shortage (and I've never seen a case where that even worked). GC overhead 
(and the corresponding allocation overhead) would be a lot more than the 
combination of free chaining and static allocation that we ended up with. 
(And yes, we verified that with profiling; I would never have believed it 
otherwise.) Moreover, the vast majority of memory used in the compiler is 
never freed anyway (it's in global data like the symbol table), until the 
end of the compiler pass. Any overhead to manage that is completely wasted.

                                     Randy.