[Patch, libgfortran, 4.5] PR25561, 37754: New low level I/O library

[Dominique Dhumieres]

Hi Janne,

I have applied your patch in 
http://gcc.gnu.org/ml/fortran/2009-01/msg00075.html on a clean tree of
i686-apple-darwin9. First the god news: it regtested without regression 
and my NIST testsuite passed for both -m32 and -m64; I see the speed 
improvement for unformatted write:

unpatched gfc -m64 -O3 IO_perf.f90

Unformatted sequential write performance test
Record size                 MB/s
================================
	  4   5.5993388198961620     
	  8   7.4793002075283219     
	 16   10.866773532402469     
	 32   17.165706271034924     
	 64   28.460467625254942     
	128   45.496996344593001     
	256   72.012566925508892     
	512   96.633305309305015     
       1024   110.66870232125038     
       2048   111.34043526016019     
       4096   105.74522966710947     
       8192   156.75644642594767     
      16384   246.79700599827905     
      32768   343.65024852758239     
      65536   457.39986636369247     
     131072   540.66200958727222     
     262144   601.36693859508853     
     524288   626.05532938624162     
4.881u 6.710s 0:38.69 29.9%	0+0k 40+80io 31pf+0w

patched gfc -m64 -O3 IO_perf.f90

Unformatted sequential write performance test
Record size                 MB/s
================================
	  4   10.813002095241160     
	  8   14.092293335580440     
	 16   20.070869469471486     
	 32   30.740473125399852     
	 64   48.170370698139074     
	128   71.158336484321254     
	256   100.45156268770435     
	512   125.81038021510702     
       1024   138.18200039555481     
       2048   145.64013285668358     
       4096   133.40498332295562     
       8192   166.92994066261576     
      16384   240.92061911529214     
      32768   348.36583935169870     
      65536   454.29682572143855     
     131072   537.49385060699592     
     262144   576.82790874873228     
     524288   602.51514953193987     
3.369u 4.569s 0:36.34 21.7%	0+0k 33+41io 0pf+0w

Note that, as far as I understand it, this is testing the MB/s relative to
the CPU time and not the actual bandwith of the disk subsystem (I'ld be
glad to have such a throughput on my macbook!-).  The picture is less clear
for formatted IOs: although I did not see any timing regressions, the
improvement ranges from 0 to 10%, still twice slower than executables from
g77.

Now the bad news: I see two regressions on my own tests:

(1) the following test aborts:

character(len=20) :: b
! write something no advance
open(10,FILE="fort.10",POSITION="REWIND")
write(10, '(a,t1,a)',advance='no') 'XXXXXX', 'ABC'
close(10)
! append some data
open(10,FILE="fort.10",POSITION="APPEND")
write(10, '(a)') 'DEF'
close(10)
! check what is in the first record
open(10,FILE="fort.10",POSITION="REWIND")
read(10,'(a)') b
IF (b.NE."ABCXXX") CALL ABORT()
END

(2) countlines.f does not work correctly on file containing the ascii 
character 255, as shown by the file

if (iachar ("ÿ")/= 255) call abort
end program main

for which the executable returns 0 instead of 2 (note that it is likely 
that "ÿ" will be messed by the mailing systems).

Thanks for the work,

Dominique