[Bug gas/15295] gas rebuffer_line() routine (listing.c) issues excessive read syscalls

[jtison at us dot ibm.com]

http://sourceware.org/bugzilla/show_bug.cgi?id=15295

--- Comment #2 from Jim Tison <jtison at us dot ibm.com> 2013-03-25 16:07:13 
UTC ---
Looks like a good fread() at line 626 is going to return 1, not 99, at least
according to the manual page. When I put the patch under gdb, I see '1'
consistently. I get 6,473 assembler warnings stating "Warning: unable to
rebuffer file".

I tried making that change to fread(), and now 'as' returns these for times and
strace summary:

time as-2.23-1 -o offldr.o -alshdg=offldr.lst offldr.s
real    1m6.492s
user    0m55.526s
sys    0m10.656s

strace -c as-2.23-1 -o offldr.o -alshdg=offldr.lst offldr.s
% time     seconds  usecs/call     calls    errors syscall
------ ----------- ----------- --------- --------- ----------------
 88.57  113.230591           3  43456033           lseek
 10.91   13.946411           4   3160831           read
  0.15    0.190372          14     13462       208 open
  0.14    0.184666          14     13248           munmap
  0.08    0.096489           7     13254           close
  0.06    0.082295          55      1501           write
  0.05    0.061030           5     13272           mmap
  0.03    0.040859           3     13255           fstat
  0.00    0.001632           4       401           brk
  0.00    0.001362           7       192       189 stat
  0.00    0.000863         863         1           unlink
  0.00    0.000238         238         1           execve
  0.00    0.000053          13         4           fadvise64
  0.00    0.000039           7         6           mprotect
  0.00    0.000007           7         1         1 access
  0.00    0.000006           3         2           fcntl
  0.00    0.000003           3         1           getrusage
  0.00    0.000003           3         1           lstat
------ ----------- ----------- --------- --------- ----------------
100.00  127.836919              46685466       398 total

Improved, but still a bit inefficient (? what say you ?) with respect to
lseek() syscalls, I'd think. I have this nagging feeling I've missed something
else. There are a lot of read() calls, too, but in this case, something has
sped up the wall clock execution by a factor of 4, which I'll take if that's
the best we can do. The quality of the listing also seems perfect: the source
code matches up with the .loc counters from start to finish.

Re "no listings", I've been arguing that case for at least 10 years and losing
every time. We have a shop full of people who can read & write assembler but
don't trust themselves in C. Someday I (or my successor) will win this
argument.

Thank you, Nick. I look forward to your next patch (at least that fread() needs
to be fixed). Appreciate the effort & attention.

-- 
Configure bugmail: http://sourceware.org/bugzilla/userprefs.cgi?tab=email
------- You are receiving this mail because: -------
You are on the CC list for the bug.