jerry@violet.berkeley.edu ( Jerry Berkman ) (12/21/88)
How many records should be in the file created by:
do 10 i=1,10
write(12,8010) i
8010 format(' record ',i5, '.' )
10 continue
rewind 12
call skptoend( 12 )
write(12,8000)
8000 format(' new record at end')
end
subroutine skptoend( lunit )
100 continue
read ( lunit, '()', end=200 )
go to 100
200 backspace lunit
end
On our VAX UNIX, Sun UNIX, and IBM CMS systems, the file looks like:
record 1.
record 2.
record 3.
record 4.
record 5.
record 6.
record 7.
record 8.
record 9.
record 10.
new record at end
However, on our Cray UNICOS system and a VMS system I tested,
the record "record 10." is missing from the final output.
As I read the standard, section 12.10.4.1 (Backspace):
"... if the preceding record is an endfile record, the file
becomes positioned before the endfile record"
and section 12.6:
"[if] An endfile record is encountered during [reading]
the file is positioned after the endfile record."
Does anyone know DEC's or Cray's position on this?
- Jerry Berkman, U.C. Berkeley
jerry@violet.berkeley.edu, (415)642-4804hirchert@uxe.cso.uiuc.edu (12/28/88)
Jerry Berkman (jerry@violet.berkeley.edu) asks >How many records should be in the file created by: > > do 10 i=1,10 > write(12,8010) i >8010 format(' record ',i5, '.' ) >10 continue > rewind 12 > call skptoend( 12 ) > write(12,8000) >8000 format(' new record at end') > end > > subroutine skptoend( lunit ) >100 continue > read ( lunit, '()', end=200 ) > go to 100 > >200 backspace lunit > end Alas, this program is not standard conforming (in FORTRAN 77), so the answer is processor dependent. In FORTRAN 77, the END= clause is guaranteed to work only if the file contains an endfile record, and the file is guaranteed to contain an endfile record only if you execute an ENDFILE statement. Since you are trying to read past the last record you wrote (which is _not_ an endfile record), you are violating the rules in the standard and the results are processor dependent. In fact, every vendor I know of has extended their FORTRAN processor to support this case, but you get different results depending on how they did it: On most processors, there is no physical representation of the endfile record (it is simply the end of the file), and a flag in the library is used to keep track of whether the file is positioned before or after that endfile record. In effect, on such processors all files have endfile records. Thus in your example, the BACKSPACE command backspaces over the endfile record and you are positioned to add an 11th formatted record. On a few processors (and I believe Cray's is one of them), there is a physical end of file character that is the endfile record and END= is triggered by either an endfile record or by reaching the end of the file. On such processors, positioning with respect to the endfile record is handled exactly like the positioning for all other records. Thus in your example, the END= branch leaves you positioned at the end of file (immediately after the 10th record, since no endfile record was written), the BACKSPACE command backspaces over the 10th record, and you are positioned to replace the 10th formatted record. By the way, it might be interesting to rerun your example using unformatted I/O, as some processors that have no physical endfile record for formatted I/O do have a physical endfile record for unformatted I/O. Thus, you may see different results between formatted and unformatted I/O on the same processor. Since this behaviour is typically a function of the run-time library, not the hardware or compiler, you should be careful about reporting this. For example, were your "Cray" results on COS, CTSS using fortlib, CTSS using cftlib, UNICOS for a Cray 2, or UNICOS for a Cray 1, X-MP, or Y-MP. Each of these libraries is separately maintained and may produce different results. (Some of them aren't even maintained by Cray!) For what its worth, the Fortran 8x draft, in effect, requires all sequential files written by a Fortran program to end with an endfile record. (They are to be produced autotmatically in cases where the user does not execute ENDFILE explicitly.) Thus, your example program would be standard conforming on a Fortran 8x processor and would always produce 11 formatted records followed by an endfile record. Kurt W. Hirchert hirchert@ncsa.uiuc.edu National Center for Supercomputing Applications