[m-dev.] status of --deep-profile-tail-recursion
Julien Fischer
jfischer at opturion.com
Tue Mar 31 19:00:34 AEDT 2015
On Tue, 31 Mar 2015, Zoltan Somogyi wrote:
> On Tue, 31 Mar 2015 16:39:31 +1100 (AEDT), Julien Fischer <jfischer at opturion.com> wrote:
>> The current situation where --deep-profile-tail-recursion is enabled by
>> default and --deep-profile-tail-recursion is not enabled by default is
>> not ideal, and also contrary to what the user's guide currently says,
>> namely:
>>
>> With deep profiling, there are other modifications as well, the most
>> important impact of which is the loss of tail-recursion for groups of
>> mutually tail-recursive predicates (self-tail-recursive predicates stay
>> tail-recursive).
>>
>> Assuming that --deep-profile-tail-recursion *is* working (I will look
>> into this as I happen to be profiling stuff at the moment anyway), are
>> there any objections to switching the default values of these two
>> options?
>
> The thing is, deep profiling tail recursion is not as useful as it may appear,
> which of course I found out the hard way. It has significant overheads,
> and enabling it did not result in meaningful speedups. It does reduce stack
> usage, but only for self-recursive predicates. Since it does not reduce
> stack usage for mutually recursive predicates, you can still run out of
> stack in a deep prof grade, and if I recall correctly, that used to happen
> regularly to me when I forgot to set MERCURY_OPTIONS to avoid this.
> (Several compiler tasks that operate on large data structures are
> implemented using mutually recursive predicates.)
>
> I think a better fix is to tie the grades that break tail call optimization
> (which include the debug grades as well as the deep prof grades) to
> the use of stack segments, and to fix the above quote in the user guide.
The debug grades already use stack segments by default anyway. I will
adjust the configure script to make the profdeep grades do so as well.
(And update the user's guide and the comment in compiler/options.m
accordingly.)
Cheers,
Julien.
More information about the developers
mailing list