at_end_of_stream on stdin in Prolog

I ask SICStus Prolog ?- at_end_of_stream. on the top-level and immediately I get: no.

GNU Prolog and Scryer Prolog do the same.

Traella Prolog and SWI-Prolog, however, choose to wait for input before answering, and rumor has it that behaving like this has been quite common among Prolog systems—especially in the past.

If I look at the ISO-Prolog standard, that behavior becomes spurious:

8.11.8.5 Bootstrapped built-in predicates
The built-in predicates at_end_of_stream/0 and at_the_of_stream/1 examine the single stream-property end_of_stream/1.

A goal at_end_of_stream is true iff the current input has a stream position end-of-stream or past-end-of-stream (7.10.2.9, 7.10.2.13).

A goal at_end_of_stream(S_or_a) is true iff the stream associated with stream or alias S_or_a has a stream position end-of-stream.
at_end_of_stream :-
  current_input(S),
  stream_property(S, end_of_stream(E)),
  !,
  (E = at ; E = past).
[The the sake of brevity, I omitted the code for at_end_of_stream/1.]

So it appears that the standard is quite clearly on the side of SICStus Prolog and GNU Prolog.

So my question boils down to this:

Is this "waiting" behavior simply non-conformance, a kind of anachronism justified on the basis of practicality / compatibility—or is there more to it?

Solution

Handling end of stream is hard, on the top-level it's even harder.

A very simple and basic top-level is like sh (dash). Prolog top-level is a bit similar, it differs by having auto-completion/choice (like Scryer Prolog, SWI-Prolog), not accepting input with ctrl-D but it could be like sh.

A query to test and better understand top-level is ?- get_char(C).%a (where C binds to %, source) and ?- get_char(C). (when submitting with enter, C binds to \n, if submitting with ctrl-D then waits or binds C to end_of_file).

Why those query? The top-level can be model as read_term(user_input, Goal, []), call(Goal):

Case get_char(C).%a submitted with enter: The query is parsed and the user_input is %a\n thus C binds to % and if get_char(C) is replaced by at_end_of_stream then at_end_of_stream/0 fails.
Case get_char(C). submitted with enter: The query is parsed and the user_input is \n thus C binds to \n and if get_char(C) is replaced by at_end_of_stream then at_end_of_stream/0 fails.
Case get_char(C). submitted with ctrl-D: The query is parsed and the user_input is empty. Now the stream position of user_input should be end-of-stream or past-end-of-stream if read_term/3 is executed as procedurally specified in 8.14.1.1:
- If user_input doesn't have an end then with the note of 7.10.2.9 the stream postion can't be end-of-stream or past-end-of-stream. Is the stream reset and has end_of_stream property not? What happens if the stream doesn't have the property eof_action(reset)? In any case get_char/1 waits and if get_char(C) is replaced by at_end_of_stream then at_end_of_stream/0 fails (there is an opportunity for at_end_of_stream/0 to wait here).
- Else user_input has a size but it's unknown (or variable but still unknown) then:
  - If stream position is end-of-stream then C binds to end_of_file and if get_char(C) is replaced by at_end_of_stream then at_end_of_stream/0 succeeds.
  - Else stream position is past-end-of-stream then get_char/1 waits and if get_char(C) is replaced by at_end_of_stream then at_end_of_stream/0 succeeds.

Sadly there is a bit of twisting at the end.

I might have missed/misread something if not then this undefinedness allows some optimization/simplification of the implementation, explains the difference between engine and also the actual top-level is more complicated that read_term(user_input, Goal, []), call(Goal) (even if it's the ideal one).

When it comes to at_end_of_stream/0, it doesn't seem like there is a reason to wait.

Digression

If at_end_of_stream/0 is implemented with peek_char/1 without executing eof_action then that might explain the wait.

On other operating system, it may be different. I use and test on Linux.

On GNU Prolog (version 1.5.0), there is the issue that a stream with property eof_action(reset) is never at end of stream. This explains why at_end_of_stream/0 fails.

$ echo -n | gprolog --init-goal "(at_end_of_stream, write(at), nl, halt ; peek_char(C), writeq(not(C)), nl, halt)"
not(end_of_file)
$ echo | gprolog --init-goal "(at_end_of_stream, writeq(at), nl, halt ; peek_char(C), writeq(not(C)), nl, halt)"
not('\n')

at_end_of_stream/0 fails even when the stdin is empty.

On Trealla Prolog (version v2.1.11):

$ echo -n | ./tpl -g "(at_end_of_stream, writeq(at), nl, halt ; peek_char(C), writeq(not(C)), nl, halt)"
at
$ echo | ./tpl -g "(at_end_of_stream, writeq(at), nl, halt ; peek_char(C), writeq(not(C)), nl, halt)"
not('\n')
$ echo -n | ./tpl -g "(stream_property(S, alias(user_input)), stream_property(S, end_of_stream(Eos)), at_end_of_stream(S), writeq(eos(Eos)), halt ; halt)"
eos(not)

The property end_of_stream doesn't agree with at_end_of_stream/1 (permuting stream_property/2 and at_end_of_stream/1 doesn't change the result).

On Scryer Prolog (on master (6b8e6204957bfc3136ea39ec659d30627775260d) or rebis-dev (c1945caf11c0d202f4121de446f1694854dcba47)):

$ echo -n | ./target/release/scryer-prolog -g "(at_end_of_stream, write(at), nl, halt ; peek_char(C), writeq(not(C)), nl, halt)"
not('\x0\')
$ echo | ./target/release/scryer-prolog -g "(at_end_of_stream, write(at), nl, halt ; peek_char(C), writeq(not(C)), nl, halt)"
not('\n')

at_end_of_stream/0 fails even when the stdin is empty.