[RFA 2/4] dwarf2_physname

[Keith Seitz]

On 12/08/2009 11:47 AM, Keith Seitz wrote:
> Yeah, this is *really* bad. Really, really bad. But this was just a an
> attempt at "making it work." Now comes the "make it faster" part. I have
> a few ideas about redundant calls and whatnot, but it will undoubtedly
> take some real effort and refactoring to really make the best of this.
>
> I'm working on some performance numbers now with OpenOffice.org Writer
> as Tom suggested. I'll follow up on this later in this thread when I
> have some more data collected.

Okay, so I have some performance numbers for my box using CVS HEAD 
patched and unpatched with my dwarf2_physname patches.

Yes, --readnow and things like "info func" and "info types" will take a 
big performance hit (really big right now).

So, without further ado...

I did three tests, all with OpenOffice.org Writer, as Tom suggested. So 
when you see $PID, you'll know what PID that's for.

1. time ./gdb -nx -q --batch-silent --pid $PID 2> /dev/null

HEAD:
real    1m14.906s  0m14.517s  0m12.229s  0m12.669s  0m12.638s
user    0m19.050s  0m11.985s  0m11.770s  0m11.738s  0m11.754s
sys     0m1.998s   0m0.484s   0m0.448s   0m0.437s   0m0.431s

dwarf2_physname:
real    1m6.761s   0m12.143s  0m12.070s  0m12.098s  0m12.128s
user    0m18.863s  0m11.710s  0m11.652s  0m11.695s  0m11.683s
sys     0m1.954s   0m0.420s   0m0.390s   0m0.390s   0m0.413s

2. time ./gdb -nx -q --batch-silent --pid $PID -ex "thread apply all bt 
full" 2> /dev/null

HEAD:
real    1m9.236s   0m12.739s  0m12.619s  0m12.613s  0m12.565s
user    0m19.059s  0m12.178s  0m12.170s  0m12.166s  0m12.102s
sys     0m1.883s   0m0.433s   0m0.440s   0m0.434s   0m0.456s

dwarf2_physname:
real    1m8.355s   0m13.276s  0m13.211s  0m13.208s  0m13.138s
user    0m19.863s  0m12.679s  0m12.750s  0m12.708s  0m12.708s
sys     0m0.978s   0m0.465s   0m0.436s   0m0.422s   0m0.425s

3. time ./gdb -nx -q --batch-silent --readnow --pid $PID 2> /dev/null

HEAD:
real    10m42.371s  6m57.499s  6m2.538s   7m50.482s  6m46.206s
user    1m43.927s   1m40.768s  1m40.768s  1m40.673s  1m41.129s
sys     0m9.349s    0m7.875s   0m7.038s   0m8.751s   0m7.552s

dwarf2_realname:
real    774m37.623s
user    3m48.010s
sys     0m14.745s

As you can see, --readnow (and by extrapolation "info func" and "info 
type") take an enormous performance hit. The dwarf2_physname patch as it 
currently exists, though, appears to have a pretty minimal impact on 
performance for more "normal" or "mundane" usage.

In other news, I've tested linux native using STABS, and there are two 
test files that show a total of five regressions, which I am digging 
into. [I've eliminated another five failures because they also exist as 
"failures" in the DWARF case (the library is still using DWARF!).]

As I'm sure we all know, STABS and C++ are such horrible bedfellows to 
begin with...

Keith