Hi,
the attached simple patch slightly reduces the amount of syscalls
gfortran does. The main issue addressed is that the Fortran standard
forbids opening the same file on multiple units. Hence when opening an
existing file, we need to go through every open file and compare the
device and inode numbers. The patch caches these values when a file is
opened, so it's no longer necessary to call fstat().
Secondly, it's not necessary to call lseek() to test whether a file is
seekable, as that information can already be found in the stat buffer.
Third, when deciding whether to use buffering or not, only buffer
regular files and block devices. While this won't solve the particular
FIFO deadlock issue that was recently reported on the mailing list, it
should at least reduce the chance of similar issues when using special
files. Also, the isatty() test is moved last, doing the cheaper tests
first.
Finally, there is a token effort to support block devices, in case
anyone is crazy enough to attempt accessing such things with gfortran.
This is not tested, and I suspect that stuff will go wrong when trying
to write past EOF.
Tested on x86_64-unknown-linux-gnu, Ok for trunk?