[m-dev.] --warn-non-tail-recursion and deep profiling

[Julien Fischer]

Hi Paul,

On Wed, 2 Mar 2016, Paul Bone wrote:

> On Wed, Mar 02, 2016 at 10:13:24AM +1100, Julien Fischer wrote:
>>
>> Hi,
>>
>> Was it intended that --warn-non-tail-recursion should emit warnings
>> in deep profiling grades?  Since those grades disable tail recursion
>> you obviously get a warning about every recursive predicate.
>
> Not specifically.
>
> Because this is disabled by default simply enabling deep profiling won't
> cause these warnings.  A user giving both options on the command line will
> very quickly figure out what is happening, but if --warn-non-tail-recursion
> is specified in a Mercury.options file, especially if it's setup by another
> developer on a team, then enabling --deep-profiling might cause some
> surprise.

That is the case I'm looking at.  Specifically, I'm looking at the
"failure" of the test case invalid/require_tailrec_2 in the grade
none.gc.profdeep.trseg.stseg.  We get additional warnings for this test
case because tail recursion is disabled.  There are various ways to
avoid this failure (e.g. extra expected output, don't run the test in
profdeep grades); I'm trying to determine which one to use here.

> I think that the question is: which surprise is better?  Lots of
> warnings

It's lots of warnings that are *not* present in non profdeep grades.
Also, in the profdeep grades, they are *useless* warnings since there's
nothing you can do about them.  (Presuambly, --warn-non-tail-recursion
respects --halt-at-warn, which is going to be a problem in profdeep
grades if the --halt-at-warn is also present in your Mercury.options
file.)

> or running out of stack space.

That won't come as a surprise since I've read the user's guide
and that's what it says will happen!

> Does the answer to this change if we take away the historical
> behaviour of breaking tail recursion in deep profiling grades?

How do you propose to do that?  We do have an undocumented option
for allowing (some) tail recursion in deep profiling grades, however
we currently say the following about it:

    % We do not currently enable (or publicly document) this option
    % because its use results in significant overheads. Also, it is not
    % compatible with coverage profiling, which is enabled by default. By
    % default, all deep profiling grades are also built with
    % --stack-segments in order to avoid problems caused by the lack of
    % tail recursion.

(I suspect it's actually broken due to bitrot in any case.)

Julien.