> On Jan 16, 2016, at 10:17 AM, Rasmus <rasmus@gmx.us> wrote:
> 
> Hi Anthony and Nicolas,
> 
> Anthony Cowley <acowley@gmail.com> writes:
> 
>> Thanks for taking a look, Rasmus! The bug is in Org, not Emacs. One
>> may perhaps assume the invariant that successive calls to
>> `current-time` will return non-decreasing values. One might also
>> assume the invariant that successive touches of different files will
>> result in file attributes with non-decreasing timestamps. It is mixing
>> the two that is the bug, and Emacs itself shouldn’t try to fix
>> that. To be clear, the limiting factor in this comparison is the
>> filesystem whose timestamp granularity is much coarser than the system
>> clock.
> 
> Thanks for carefully explaining exactly what is going on here.
> 
>> We could do something like generate an entirely separate temporary
>> file before compilation, and use its timestamp as the reference by
>> which compilation output is evaluated. However, the approach of
>> checking file freshness with a one-second granularity seems like a
>> pretty good compromise.
> 
> I think the approach is fine now.  See below.
> 
>>> Maybe you could just check the log directly for failures rather than
>>> checking the file attributes, if these are not reliable.
>>> 
>>> Also, what is the ‘take‘?  I don’t seem to have it in my Emacs…
>> 
>> Sorry about that. Believe it or not I had already rewritten those two
>> lines — as simple as they are — trying to not include extra
>> dependencies. I’ve attached an updated patch that uses subseq from
>> cl.el. I hope that is okay.
> 
> Actually, it's part of cl-lib now (in the broadest understanding of
> cl-lib).
> 
> Since this is a bugfix I guess you should use org-sublist (*shiver*) in
> maint and cl-subseq in master.  We can’t use cl-lib in maint, right
> Nicolas?
> 
> Nicolas Goaziou <mail@nicolasgoaziou.fr> writes:
> 
>> Anthony Cowley <acowley@gmail.com> writes:
>> 
>>> * lisp/ox-latex.el (org-latex-compile): Improve timestamp check on HFS+
>>>  filesystem by only considering 1-second clock resolution.
>>> 
>>> Previously, the call to (current-time) could return a timestamp with
>>> a non-zero microsecond or picosecond fields, while the file attribute
>>> always has zeros for these fields. The check that the generated file is
>>> newer than the reference timestamp only succeeded when the time to
>>> generate the file crossed a 1-second clock interval.
>>> 
>>> TINYCHANGE
>>> ---
>>> lisp/ox-latex.el | 3 ++-
>>> 1 file changed, 2 insertions(+), 1 deletion(-)
>>> 
>>> diff --git a/lisp/ox-latex.el b/lisp/ox-latex.el
>>> index a57677b..f803b7e 100644
>>> --- a/lisp/ox-latex.el
>>> +++ b/lisp/ox-latex.el
>>> @@ -3576,7 +3576,8 @@ Return PDF file name or an error if it couldn't be produced."
>>> 	;; Check for process failure.  Provide collected errors if
>>> 	;; possible.
>>> 	(if (or (not (file-exists-p pdffile))
>>> -		(time-less-p (nth 5 (file-attributes pdffile)) time))
>>> +		(time-less-p (subseq (nth 5 (file-attributes pdffile)) 0 2)
>>> +			     (subseq time 0 2)))
>> 
>> This sounds good. Thank you.
>> 
>> Although, I suggest to use `cl-subseq' instead of its alias. Also, it
>> may be worth commenting that trick right into the source.
> 
> Yes, it should.

I’ve attached two remixes of the original patch: the first uses cl-subseq, the second uses org-sublist.

Anthony