R devel - May 2017 - stopifnot() does not stop at first non-TRUE argument

This is getting pretty convoluted.

The current behavior is consistent with the description at the top of
the help page -- it does not promise to stop evaluation once the first
non-TRUE is found.  That seems OK to me -- if you want sequencing you
can use

stopifnot(A)
stopifnot(B)

or

stopifnot(A && B)

I could see an argument for a change that in the multiple argumetn
case reports _all_ that fail; that would seem more useful to me than
twisting the code into knots.

Best,

luke

On Mon, 15 May 2017, Martin Maechler wrote:
>>>>>> Serguei Sokol <sokol at insa-toulouse.fr>
>>>>>>     on Mon, 15 May 2017 16:32:20 +0200 writes:
>
>    > Le 15/05/2017 ? 15:37, Martin Maechler a ?crit :
>    >>>>>>> Serguei Sokol <sokol at
insa-toulouse.fr>
>    >>>>>>> on Mon, 15 May 2017 13:14:34 +0200 writes:
>    >> > I see in the archives that the attachment cannot pass.
>    >> > So, here is the code:
>    >>
>    >> [....... MM: I needed to reformat etc to match closely to
>    >> the current source code which is in
>    >> https://svn.r-project.org/R/trunk/src/library/base/R/stop.R
>    >> or its corresponding github mirror
>    >>
https://github.com/wch/r-source/blob/trunk/src/library/base/R/stop.R
>    >> ]
>    >>
>    >> > Best,
>    >> > Serguei.
>    >>
>    >> Yes, something like that seems even simpler than Peter's
>    >> suggestion...
>    >>
>    >> It currently breaks 'make check' in the R sources,
>    >> specifically in tests/reg-tests-2.R (lines 6574 ff),
>    >> the new code now gives
>    >>
>    >> > ## error messages from (C-level) evalList
>    >> > tst <- function(y) { stopifnot(is.numeric(y)); y+ 1 }
>    >> > try(tst())
>    >> Error in eval(cl.i, pfr) : argument "y" is missing,
with no default
>    >>
>    >> whereas previously it gave
>    >>
>    >> Error in stopifnot(is.numeric(y)) :
>    >> argument "y" is missing, with no default
>    >>
>    >>
>    >> But I think that change (of call stack in such an error case)
is
>    >> unavoidable and not a big problem.
>
>    > It can be avoided but at price of customizing error() and warning()
calls with something like:
>    > wrn <- function(w) {w$call <- cl.i; warning(w)}
>    > err <- function(e) {e$call <- cl.i; stop(e)}
>    > ...
>    > tryCatch(r <- eval(cl.i, pfr), warning=wrn, error=err)
>
>    > Serguei.
>
> Well, a good idea, but the 'warning' case is more complicated
> (and the above incorrect): I do want the warning there, but
> _not_ return the warning, but rather, the result of eval() :
> So this needs even more sophistication, using  withCallingHandlers(.)
> and maybe that really get's too sophisticated and no
> more "readable" to 99.9% of the R users ... ?
>
> I now do append my current version -- in case some may want to
> comment or improve further.
>
> Martin
>
>
-- 
Luke Tierney
Ralph E. Wareham Professor of Mathematical Sciences
University of Iowa                  Phone:             319-335-3386
Department of Statistics and        Fax:               319-335-3017
    Actuarial Science
241 Schaeffer Hall                  email:   luke-tierney at uiowa.edu
Iowa City, IA 52242                 WWW:  http://www.stat.uiowa.edu

Hi,

On 05/15/2017 10:41 AM, luke-tierney at uiowa.edu wrote:
> This is getting pretty convoluted.
>
> The current behavior is consistent with the description at the top of
> the help page -- it does not promise to stop evaluation once the first
> non-TRUE is found.  That seems OK to me -- if you want sequencing you
> can use
>
> stopifnot(A)
> stopifnot(B)
>
> or
>
> stopifnot(A && B)
My main use case for using stopifnot() is argument checking. In that
context, I like the conciseness of

   stopifnot(
     A,
     B,
     ...
   )

I think it's a common use case (and a pretty natural thing to do) to
order/organize the expressions in a way such that it only makes sense
to continue evaluating if all was OK so far e.g.

   stopifnot(
     is.numeric(x),
     length(x) == 1,
     is.na(x)
   )

At least that's how things are organized in the stopifnot() calls that
accumulated in my code over the years. That's because I was convinced
that evaluation would stop at the first non-true expression (as
suggested by the man page). Until recently when I got a warning issued
by an expression located *after* the first non-true expression. This
was pretty unexpected/confusing!

If I can't rely on this "sequencing" feature, I guess I can always
do

   stopifnot(A)
   stopifnot(B)
   ...

but I loose the conciseness of calling stopifnot() only once.
I could also use

   stopifnot(A && B && ...)

but then I loose the conciseness of the error message i.e. it's going
to be something like

   Error: A && B && ... is not TRUE

which can be pretty long/noisy compared to the message that reports
only the 1st error.

Conciseness/readability of the single call to stopifnot() and
conciseness of the error message are the features that made me
adopt stopifnot() in the 1st place. If stopifnot() cannot be revisited
to do "sequencing" then that means I will need to revisit all my calls
to stopifnot().
>
> I could see an argument for a change that in the multiple argumetn
> case reports _all_ that fail; that would seem more useful to me than
> twisting the code into knots.
Why not. Still better than the current situation. But only if that
semantic seems more useful to people. Would be sad if usefulness
of one semantic or the other was decided based on trickiness of
implementation.

Thanks,
H.
>
> Best,
>
> luke
>
> On Mon, 15 May 2017, Martin Maechler wrote:
>
>>>>>>> Serguei Sokol <sokol at insa-toulouse.fr>
>>>>>>>     on Mon, 15 May 2017 16:32:20 +0200 writes:
>>
>>    > Le 15/05/2017 ? 15:37, Martin Maechler a ?crit :
>>    >>>>>>> Serguei Sokol <sokol at
insa-toulouse.fr>
>>    >>>>>>> on Mon, 15 May 2017 13:14:34 +0200
writes:
>>    >> > I see in the archives that the attachment cannot pass.
>>    >> > So, here is the code:
>>    >>
>>    >> [....... MM: I needed to reformat etc to match closely to
>>    >> the current source code which is in
>>    >>
>>
https://urldefense.proofpoint.com/v2/url?u=https-3A__svn.r-2Dproject.org_R_trunk_src_library_base_R_stop.R&d=DwIFAw&c=eRAMFD45gAfqt84VtBcfhQ&r=BK7q3XeAvimeWdGbWY_wJYbW0WYiZvSXAJJKaaPhzWA&m=t9fJDOl9YG2zB-GF0wQXrXJTsW2jxTxMHE-qZfLGzHU&s=KGsvpXrXpHCFTdbLM9ci3sBNO9C3ocsgEqHMvZKvV9I&e>>
>> or its corresponding github mirror
>>    >>
>>
https://urldefense.proofpoint.com/v2/url?u=https-3A__github.com_wch_r-2Dsource_blob_trunk_src_library_base_R_stop.R&d=DwIFAw&c=eRAMFD45gAfqt84VtBcfhQ&r=BK7q3XeAvimeWdGbWY_wJYbW0WYiZvSXAJJKaaPhzWA&m=t9fJDOl9YG2zB-GF0wQXrXJTsW2jxTxMHE-qZfLGzHU&s=7Z5bPVWdGPpY2KLnXQP6c-_8s86CpKe0ZYkCfqjfxY0&e>>
>> ]
>>    >>
>>    >> > Best,
>>    >> > Serguei.
>>    >>
>>    >> Yes, something like that seems even simpler than
Peter's
>>    >> suggestion...
>>    >>
>>    >> It currently breaks 'make check' in the R sources,
>>    >> specifically in tests/reg-tests-2.R (lines 6574 ff),
>>    >> the new code now gives
>>    >>
>>    >> > ## error messages from (C-level) evalList
>>    >> > tst <- function(y) { stopifnot(is.numeric(y)); y+ 1
}
>>    >> > try(tst())
>>    >> Error in eval(cl.i, pfr) : argument "y" is
missing, with no default
>>    >>
>>    >> whereas previously it gave
>>    >>
>>    >> Error in stopifnot(is.numeric(y)) :
>>    >> argument "y" is missing, with no default
>>    >>
>>    >>
>>    >> But I think that change (of call stack in such an error
case) is
>>    >> unavoidable and not a big problem.
>>
>>    > It can be avoided but at price of customizing error() and
>> warning() calls with something like:
>>    > wrn <- function(w) {w$call <- cl.i; warning(w)}
>>    > err <- function(e) {e$call <- cl.i; stop(e)}
>>    > ...
>>    > tryCatch(r <- eval(cl.i, pfr), warning=wrn, error=err)
>>
>>    > Serguei.
>>
>> Well, a good idea, but the 'warning' case is more complicated
>> (and the above incorrect): I do want the warning there, but
>> _not_ return the warning, but rather, the result of eval() :
>> So this needs even more sophistication, using  withCallingHandlers(.)
>> and maybe that really get's too sophisticated and no
>> more "readable" to 99.9% of the R users ... ?
>>
>> I now do append my current version -- in case some may want to
>> comment or improve further.
>>
>> Martin
>>
>>
>
-- 
Herv? Pag?s

Program in Computational Biology
Division of Public Health Sciences
Fred Hutchinson Cancer Research Center
1100 Fairview Ave. N, M1-B514
P.O. Box 19024
Seattle, WA 98109-1024

E-mail: hpages at fredhutch.org
Phone:  (206) 667-5791
Fax:    (206) 667-1319

>>>>> Herv? Pag?s <hpages at fredhutch.org>
>>>>>     on Mon, 15 May 2017 16:54:46 -0700 writes:
    > Hi,
    > On 05/15/2017 10:41 AM, luke-tierney at uiowa.edu wrote:
    >> This is getting pretty convoluted.
    >> 
    >> The current behavior is consistent with the description at the top
of
    >> the help page -- it does not promise to stop evaluation once the
first
    >> non-TRUE is found.  That seems OK to me -- if you want sequencing
you
    >> can use
    >> 
    >> stopifnot(A)
    >> stopifnot(B)
    >> 
    >> or
    >> 
    >> stopifnot(A && B)

    > My main use case for using stopifnot() is argument checking. In that
    > context, I like the conciseness of

    > stopifnot(
    > A,
    > B,
    > ...
    > )

    > I think it's a common use case (and a pretty natural thing to do)
to
    > order/organize the expressions in a way such that it only makes sense
    > to continue evaluating if all was OK so far e.g.

    > stopifnot(
    > is.numeric(x),
    > length(x) == 1,
    > is.na(x)
    > )

I agree.  And that's how I have used stopifnot() in many cases
myself, sometimes even more "extremely" than the above example,
using assertions that only make sense if previous assertions
were fulfilled, such as

    stopifnot(is.numeric(n), length(n) == 1, n == round(n), n >= 0)

or in the Matrix package, first checking some class properties
and then things that only make sense for objects with those properties.


    > At least that's how things are organized in the stopifnot() calls
that
    > accumulated in my code over the years. That's because I was
convinced
    > that evaluation would stop at the first non-true expression (as
    > suggested by the man page). Until recently when I got a warning issued
    > by an expression located *after* the first non-true expression. This
    > was pretty unexpected/confusing!

    > If I can't rely on this "sequencing" feature, I guess I
can always
    > do

    > stopifnot(A)
    > stopifnot(B)
    > ...

    > but I loose the conciseness of calling stopifnot() only once.
    > I could also use

    > stopifnot(A && B && ...)

    > but then I loose the conciseness of the error message i.e. it's
going
    > to be something like

    > Error: A && B && ... is not TRUE

    > which can be pretty long/noisy compared to the message that reports
    > only the 1st error.


    > Conciseness/readability of the single call to stopifnot() and
    > conciseness of the error message are the features that made me
    > adopt stopifnot() in the 1st place. 

Yes, and that had been my design goal when I created it.

I do tend agree with  Herv? and Serguei here.

    > If stopifnot() cannot be revisited
    > to do "sequencing" then that means I will need to revisit all
my calls
    > to stopifnot().

    >> 
    >> I could see an argument for a change that in the multiple argumetn
    >> case reports _all_ that fail; that would seem more useful to me
than
    >> twisting the code into knots.

Interesting... but really differing from the current documentation,

    > Why not. Still better than the current situation. But only if that
    > semantic seems more useful to people. Would be sad if usefulness
    > of one semantic or the other was decided based on trickiness of
    > implementation.

Well, the trickiness  should definitely play a role.
Apart from functionality and semantics, long term maintenance
and code readibility, even elegance have shown to be very
important aspects of good code in ca 30 years of S and R programming.

OTOH, as mentioned above, the creation of good error messages
has been an important design goal of  stopifnot()  and hence I'm
willing to accept the extra complexity of "patching up" the call
used in the error / warning messages.

Also, as a change to what I posted yesterday, I now plan to follow
Peter Dalgaard's suggestion of using
             eval( ..<n> ) 
instead of   eval(cl[[i]], envir = <parent.frame(.)>)
as there may be cases where the former behaves better in lazy
evaluation situations.
(Other opinions on that ?)

Martin

    > Thanks,
    > H.

    >> 
    >> Best,
    >> 
    >> luke
    >> 
    >> On Mon, 15 May 2017, Martin Maechler wrote:
    >> 
    >>>>>>>> Serguei Sokol <sokol at
insa-toulouse.fr>
    >>>>>>>> on Mon, 15 May 2017 16:32:20 +0200 writes:
    >>> 
    >>> > Le 15/05/2017 ? 15:37, Martin Maechler a ?crit :
    >>> >>>>>>> Serguei Sokol <sokol at
insa-toulouse.fr>
    >>> >>>>>>> on Mon, 15 May 2017 13:14:34 +0200
writes:
    >>> >> > I see in the archives that the attachment cannot
pass.
    >>> >> > So, here is the code:
    >>> >>
    >>> >> [....... MM: I needed to reformat etc to match closely
to
    >>> >> the current source code which is in
    >>> >>
    >>>
https://urldefense.proofpoint.com/v2/url?u=https-3A__svn.r-2Dproject.org_R_trunk_src_library_base_R_stop.R&d=DwIFAw&c=eRAMFD45gAfqt84VtBcfhQ&r=BK7q3XeAvimeWdGbWY_wJYbW0WYiZvSXAJJKaaPhzWA&m=t9fJDOl9YG2zB-GF0wQXrXJTsW2jxTxMHE-qZfLGzHU&s=KGsvpXrXpHCFTdbLM9ci3sBNO9C3ocsgEqHMvZKvV9I&e
>>> >> or its corresponding github mirror
    >>> >>
    >>>
https://urldefense.proofpoint.com/v2/url?u=https-3A__github.com_wch_r-2Dsource_blob_trunk_src_library_base_R_stop.R&d=DwIFAw&c=eRAMFD45gAfqt84VtBcfhQ&r=BK7q3XeAvimeWdGbWY_wJYbW0WYiZvSXAJJKaaPhzWA&m=t9fJDOl9YG2zB-GF0wQXrXJTsW2jxTxMHE-qZfLGzHU&s=7Z5bPVWdGPpY2KLnXQP6c-_8s86CpKe0ZYkCfqjfxY0&e
>>> >> ]
    >>> >>
    >>> >> > Best,
    >>> >> > Serguei.
    >>> >>
    >>> >> Yes, something like that seems even simpler than
Peter's
    >>> >> suggestion...
    >>> >>
    >>> >> It currently breaks 'make check' in the R
sources,
    >>> >> specifically in tests/reg-tests-2.R (lines 6574 ff),
    >>> >> the new code now gives
    >>> >>
    >>> >> > ## error messages from (C-level) evalList
    >>> >> > tst <- function(y) { stopifnot(is.numeric(y));
y+ 1 }
    >>> >> > try(tst())
    >>> >> Error in eval(cl.i, pfr) : argument "y" is
missing, with no default
    >>> >>
    >>> >> whereas previously it gave
    >>> >>
    >>> >> Error in stopifnot(is.numeric(y)) :
    >>> >> argument "y" is missing, with no default
    >>> >>
    >>> >>
    >>> >> But I think that change (of call stack in such an
error case) is
    >>> >> unavoidable and not a big problem.
    >>> 
    >>> > It can be avoided but at price of customizing error() and
    >>> warning() calls with something like:
    >>> > wrn <- function(w) {w$call <- cl.i; warning(w)}
    >>> > err <- function(e) {e$call <- cl.i; stop(e)}
    >>> > ...
    >>> > tryCatch(r <- eval(cl.i, pfr), warning=wrn, error=err)
    >>> 
    >>> > Serguei.
    >>> 
    >>> Well, a good idea, but the 'warning' case is more
complicated
    >>> (and the above incorrect): I do want the warning there, but
    >>> _not_ return the warning, but rather, the result of eval() :
    >>> So this needs even more sophistication, using 
withCallingHandlers(.)
    >>> and maybe that really get's too sophisticated and no
    >>> more "readable" to 99.9% of the R users ... ?
    >>> 
    >>> I now do append my current version -- in case some may want to
    >>> comment or improve further.
    >>> 
    >>> Martin

Le 15/05/2017 ? 19:41, luke-tierney at uiowa.edu a ?crit
:
> This is getting pretty convoluted.
>
> The current behavior is consistent with the description at the top of
> the help page -- it does not promise to stop evaluation once the first
> non-TRUE is found.
Hm... we can read in the man page :
?stopifnot(A, B)? is conceptually equivalent to

       { if(any(is.na(A)) || !all(A)) stop(...);
         if(any(is.na(B)) || !all(B)) stop(...) }
and this behavior does promise to stop at first non-TRUE value
without evaluation of the rest of conditions.

Sergue?.

On Tue, 16 May 2017, Serguei Sokol wrote:
> Le 15/05/2017 ? 19:41, luke-tierney at uiowa.edu a ?crit :
>> This is getting pretty convoluted.
>> 
>> The current behavior is consistent with the description at the top of
>> the help page -- it does not promise to stop evaluation once the first
>> non-TRUE is found.
> Hm... we can read in the man page :
> ?stopifnot(A, B)? is conceptually equivalent to
>
>      { if(any(is.na(A)) || !all(A)) stop(...);
>        if(any(is.na(B)) || !all(B)) stop(...) }
> and this behavior does promise to stop at first non-TRUE value
> without evaluation of the rest of conditions.
Yes: that is why I explicitly referenced the description at the top of
the page.

Changing the 'conceptually equivalent' bit to reflect what is
happening is easy.  The changes being discussed, and their long term
maintenance, ar not.

Best,

luke

>
> Sergue?.
>
-- 
Luke Tierney
Ralph E. Wareham Professor of Mathematical Sciences
University of Iowa                  Phone:             319-335-3386
Department of Statistics and        Fax:               319-335-3017
    Actuarial Science
241 Schaeffer Hall                  email:   luke-tierney at uiowa.edu
Iowa City, IA 52242                 WWW:  http://www.stat.uiowa.edu