8:42 p.m.
Hi,
this is a bug that is probably triggered only in unusual circumstances,
but it happened to me in a kind-of real world function development scenario:
<hansl>
clear
function void serc (const series s) # the 'const' here is important
if exists(s)
print "to be"
else
eval typename(s)
print "or not to be"
endif
end function
open denmark
print "call 1"
serc(LRM)
print "call 2"
serc(const) # "const" is a series ID 0 in all datasets
# -> produces "null" and "or not to be", as if "s" inside
the function
didn't exist
</hansl>
Notice two things: No gretl errors are thrown here, so it's not that
gretl doesn't recognize "const" as a valid series argument. Secondly,
when the function signature doesn't have the "const" arg modifier, then
it works fine.
This is with snapshot Jan 27th.
thanks
sven
12:42 a.m.
On Thu, 1 Feb 2024, Sven Schreiber wrote:
this is a bug that is probably triggered only in unusual
circumstances, but
it happened to me in a kind-of real world function development scenario:
<hansl>
function void serc (const series s) # the 'const' here is important
if exists(s)
print "to be"
else
eval typename(s)
print "or not to be"
endif
end function
open denmark
print "call 1"
serc(LRM)
print "call 2"
serc(const) # "const" is a series ID 0 in all datasets
# -> produces "null" and "or not to be", as if "s"
inside
# the function didn't exist
</hansl>
This behavior has been there "forever" (or more precisely, ever
since the "const" modifier for hansl function arguments has been
supported).
It's quite an interesting problem. When a series is given as a
function argument without the "const" modifier, gretl gives the
function a copy of the incoming series, under the name given by the
function parameter. But when the "const" flag is given, the series
in question is shifted from the function-execution level of the
caller to that of the callee, and is renamed using the parameter
name -- and then these changes are reverted when the function exits.
However, that treatment of the "const" modifier cannot work when the
series passed is the one named "const" (with ID number 0). That
series is supposed to be automatically available at all levels of
function execution, and renaming it would make in unavailable under
its proper name inside the function -- a real bug.
This issue hasn't been reported before, which is not surprising
since passing series 0 as an argument is in general pointless, since
it's already available everywhere and is of no interest other than
as the intercept in a regression. I think I see a way of making it
harmless but I need to do some more testing before I'm sure.
Allin
11:46 a.m.
Am 02.02.2024 um 01:42 schrieb Allin Cottrell:
However, that treatment of the "const" modifier cannot work when the
series passed is the one named "const" (with ID number 0). That series
is supposed to be automatically available at all levels of function
execution,
Well, but only if the dataset context is passed on to the called
function, right? Otherwise no series is available inside, not even
"const" (ID 0). This is the background of what I'm trying to achieve,
with a little bit of borderline hacking. I'm working with MPI and
(sometimes, not always in my general meta function) need to pass on the
dataset context. And I thought the safest way is to use the series that
always exists, namely "const".
This issue hasn't been reported before, which is not surprising since
passing series 0 as an argument is in general pointless, since it's
already available everywhere and is of no interest other than as the
intercept in a regression. I think I see a way of making it harmless
but I need to do some more testing before I'm sure.
Thanks. Again, the interest here is not the "const" series per se, but
the idea is to "piggyback" the dataset context on an arbitrarily chosen
series. But I think I can work around this particular problem; for
example, two hackish pieces of syntax that work instead of the passed
arg "const" are:
1) "(dataset)[1]" - assumes at least one other series, which is OK; not
that "dataset[1]" didn't work
2) "deflist(const)[1]" - uglier, but also works
(More background explanation: I first tried to use a gretl list that is
of more direct interest, but the problem was that this doesn't work in
the also covered case that no dataset is present. That is, I can use the
function arg specification 'series s[null]' and the null default also
works without an active dataset, but 'list L[null]' yields an error
then. I know this is because of the special behavior of lists, which
always exist even when "null-ed" or omitted. But that's causing a
problem here, hence my workaround with "const".)
If it's too dangerous to accommodate the passing of just "const" (and
since usually there isn't much point in it), perhaps it's better to just
ban it? Or perhaps ban the combination of passing "const" to a function
that uses the const modifier for its arg?
thanks
sven
2:12 p.m.
On Fri, 2 Feb 2024, Sven Schreiber wrote:
Am 02.02.2024 um 01:42 schrieb Allin Cottrell:
>
> However, that treatment of the "const" modifier cannot work when
> the series passed is the one named "const" (with ID number 0).
> That series is supposed to be automatically available at all
> levels of function execution,
Well, but only if the dataset context is passed on to the called
function, right?
Right, but that happens automatically within any given gretl
process, regardless of whether any series, or list of series, is
passed to the called function. As a trivial illustration:
<hansl>
function void f (void)
print const
smpl
end function
nulldata 8
setobs 4 2016:1
f()
</hansl>
In the MPI context the dataset (if present) is not automatically
passed between distinct gretl processes, but this is ensured by the
--send-data option for a mpi block. Another trivial illustration:
<hansl>
function void f (void)
printf "process %d\n", $mpirank
print const
smpl
end function
nulldata 8
setobs 4 2016:1
mpi --send-functions --send-data
f()
end mpi --np=2
</hansl>
There's one gotcha to note here. If you don't need to pass the
entire dataset in the MPI context you can use --send-data=L, where L
is a list of series, BUT this won't work if L contains just const.
That's because inter-process transfer of data is done using a gdt or
gdtb file, and "const" is not explicitly represented in such a file.
Trying to save a gdt[b] file containing no series other than const
will fail with the message "No data to save!". It's fine if L
contains just a single series, but it can't be series 0. You could
insert "list L = index" before the mpi block above, and make the
option --send-data=L.
Anyway, the behavior you found with exists() is clearly wrong, and
I'm satisfied that a proper fix can be engineered: we'll give the
called function a copy of const under the parameter name rather then
doing the "shift and rename" thing. That should be in git quite
soon.
Allin
4:39 p.m.
Am 02.02.2024 um 15:12 schrieb Allin Cottrell:
Right, but that happens automatically within any given gretl process,
regardless of whether any series, or list of series, is passed to the
called function. As a trivial illustration:
....
In the MPI context the dataset (if present) is not automatically
passed between distinct gretl processes, but this is ensured by the
--send-data option for a mpi block. Another trivial illustration:
Aha! That's very interesting, I wasn't aware of it, because I thought it
would go against the idea of function encapsulation. But very useful
here, so I don't need any workaround, it seems.
There's one gotcha to note here. If you don't need to pass
the entire
dataset in the MPI context you can use --send-data=L, where L is a
list of series, BUT this won't work if L contains just const. That's
because inter-process transfer of data is done using a gdt or gdtb
file, and "const" is not explicitly represented in such a file. Trying
to save a gdt[b] file containing no series other than const will fail
with the message "No data to save!". It's fine if L contains just a
single series, but it can't be series 0.
That's OK.
You could insert "list L = index" before the mpi block
above, and make
the option --send-data=L.
I guess a "genr index" would also be needed,
because not every existing
dataset has the index series. But it's not really needed anyway.
Anyway, the behavior you found with exists() is clearly wrong, and I'm
satisfied that a proper fix can be engineered: we'll give the called
function a copy of const under the parameter name rather then doing
the "shift and rename" thing. That should be in git quite soon.
OK, good.
sven
637
days inactive
638
days old
4 comments
2 participants
participants (2)
-
Allin Cottrell -
Sven Schreiber