Hi Jim,

Attached a patch implementing your suggestion. Some notes:
I first tried to put the concatenation of comp_dir and dirname at the
top of dwarf2_start_subfile (like suggested), but this caused failures
in the testsuite: changing dirname at the beginning of the function
changes the semantics of the loop that tries to match dirname'/'filename
against existing subfiles. The subfile for the current compilation unit
wasn't matched correctly (DW_AT_name = $srcdir/file.c, lineinfo.dirname
= $srcdir, lineinfo.filename = file.c  where $srcdir is a relative
path). Putting the dirname ``absolutification'' after the loop solves
it, but it looks a bit more hackish.

OK to commit?

All this file matching seems quite fragile, and the current approach
will get it wrong sometimes. Shouldn't the loop in dwarf2_start_subfile
be killed in favor of a search using both xfullpath(dirname'/'filename)
at the start of buildsym.c::start_subfile?