[r6rs-discuss] [OT] Should compilers preserve termination?

From: Paul Schlie <schlie>
Date: Fri Mar 2 01:00:02 2007

> Thomas Lord wrote:
>> Joe Marshall wrote:
>>> Paul Schlie <schlie_at_comcast.net> wrote:
>>>
>>> Thereby although creating a logical loop testing for a termination
>>> condition prior to returning a value and thereby enabling control
>>> flow to proceed, may not be the best way to temporally synchronize
>>> program execution to external events, it's seemingly valid; and
>>> thereby seemingly only valid to optimize away evaluation if it can
>>> be proven to terminate exclusively based on synthetic state values
>>> within the context of the logical program, and not values imported
>>> from the outside world through I/O ports or other similar means for
>>> example.
>>
>> This seems weird. If I were to write a loop that just ran for timing
>> purposes, I'd be insane to run it through an optimizing compiler.
>>
>> As it stands, collatz conjecture hasn't been proven, but no divergent
>> value has ever been discovered. Would it be ok to optimize out the code
>> for the cases for which it is known to converge?

- personally, only if the compilation of collatz determined that its
  temporal behavior and value termination state is not dependent on
  external values/events. Which in practice seems fairly easy in the
  general case, as the presence of any conditional termination condition
  based on an externally controllable value/event, such as being dependent
  on the return of a value through a port, would preclude collatz from
  being optimized away (as otherwise one could arguably optimize away
  elements of a REPL, which presumably would not be desirable)

>> proof were found, would it be ok to optimize out the call in that case?

- yes, although I doubt that one could ever prove sequential temporal
  insensitivity without claiming omnipotence.

>> Would it be ok even if someone wrote timing code that depended on such
>> a proof not being found?

- personally no, as otherwise the optimizer is for a different language.

> Joe, I think he is saying that all low-level port-events should be
> defined to be well-ordered points on a hypothetical world-line of
> a single program run and that the semantics of the order-of-execution
> partially determine that ordering. Because low-level port events are
> externally observable, they can't be optimized away and, in the view
> he's expressing, have to occur in one of the expected orders.

- yes

> I tried for a while to argue against that but came to the conclusion
> that Paul is right. Even an optimizing compiler should be sensitive
> to the observable side effects implied by the semantics.
>
> But I don't disagree that the whole topic is "weird", at least
> to the extent that that is roughly synonymous with "interesting"
> or "fascinating".

- not really so weird, although arguably not a great coding practice;
  but in the absents of standardized language alternatives, it should
  be honored.

> You could still build an optimizing compiler that didn't honor
> this restriction and it could still be very useful -- but it should,
> in this view, come with a warning that it may not compile some
> portable programs faithfully.

- personally, an optimizing compiler should honor all language semantics
Received on Fri Mar 02 2007 - 00:59:24 UTC