Re: codiff misses changes if inline -> not inlined?

[Arnaldo Carvalho de Melo <acme-H+wXaHxf7aLQT0dZR+AlfA@public.gmane.org>]

Em Fri, Jan 04, 2008 at 01:24:02AM +0200, Ilpo Järvinen escreveu:
> On Thu, 3 Jan 2008, Arnaldo Carvalho de Melo wrote:
> 
> > Em Thu, Jan 03, 2008 at 03:40:16PM +0200, Ilpo Järvinen escreveu:
> > > Hi,
> > > 
> > > I've had a problem with codiff when playing around with inlines. It seems 
> > > to miss completely the resizement of inlined function from zero to 
> > > something. Here's one example (relevant patch attached):
> > > 
> > > $ codiff tcp_input.o.old tcp_input.o
> > > net/ipv4/tcp_input.c:
> > >   tcp_dsack_extend |  -73
> > >   tcp_data_queue   |  -17
> > >  2 functions changed, 90 bytes removed
> > > 
> > > $ pfunct -s tcp_input.o.old | grep "tcp_sack_extend"
> > > $ pfunct -s tcp_input.o | grep "tcp_sack_extend"
> > > tcp_sack_extend: 66
> > > 
> > > Isn't that tcp_sack_extend | +66 missing from codiff's output???
> > > 
> > > 
> > > There's nothing special in the tcp_sack_extend(), any inline (at least in 
> > > .c files) will do.
> > 
> > From codiff's point of view its a "new" function... I'm checking this right now, just a sec :)
> 
> More very interesting cases found... :-/ The tool removed inline from 
> ip_ufo_append_data and got this result:
> 
> $ codiff -V net/ipv4/ip_output.o.old net/ipv4/ip_output.o
> net/ipv4/ip_output.c:
>   dst_output        |  +11 (uninlined)
>   ip_send_check     |  +69 (uninlined)
>   ip_finish_output2 | +594 (uninlined)
>  3 functions changed, 674 bytes added, diff: +674
> 
> ...It may well be right due to some strange inlining heurestics 
> interaction but I'm a bit suspicious because of the large numbers and 
> pfunct -i tells a different story for at least ip_finish_output2 and 
> ip_send_check (in both cases the sizes are constant 0 in -i output!?!).

Before manually removing inline from the function:

[acme@doppio net-2.6.25]$ pfunct -f ip_ufo_append_data /tmp/ip_output.o.old
inline int ip_ufo_append_data(struct sock * sk, int (*getfrag)(void *, char *, int, int, int, struct sk_buff *), void * from, int length, int hh_len, int fragheaderlen, int transhdrlen, int mtu, unsigned int flags);
[acme@doppio net-2.6.25]$

And the compiler is not, at least in my kernel configuration, trying to
do (un)inlining differently from what is in the source code:

[acme@doppio net-2.6.25]$ pfunct --cc_uninlined /tmp/ip_output.o.old
[acme@doppio net-2.6.25]$ pfunct --cc_inlined /tmp/ip_output.o.old
[acme@doppio net-2.6.25]$

[acme@doppio net-2.6.25]$ grep INLIN ../build/net-2.6.25/doppio/.config
# CONFIG_FORCED_INLINING is not set

[acme@doppio net-2.6.25]$ pfunct -T -f ip_append_data /tmp/ip_output.o.old

Shows that ip_ufo_append_data is being inlined in ip_append_data, taking
336 bytes in my config (x86_64).

Then I remove the inline and...

[acme@doppio net-2.6.25]$ codiff /tmp/ip_output.o.old
/tmp/ip_output.o.new
/home/acme/git/net-2.6.25/net/ipv4/ip_output.c:
  dst_output        |  +14
  ip_send_check     |  +65
  ip_finish_output2 | +499
 3 functions changed, 578 bytes added, diff: +578
[acme@doppio net-2.6.25]$

Which should have shown ip_ufo_append_data as being uninlined... ah!

[acme@doppio pahole]$ pfunct --cc_inlined /tmp/ip_output.o.new
ip_ufo_append_data

The compiler decided that even if we remove the 'inline' keyword the
function is only called from ip_append_data, so it inlines it anyway.

But why we end up using 578 bytes more? Lemme see...

It is a bug:

[acme@doppio pahole]$ size /tmp/ip_output.o.old /tmp/ip_output.o.new
   text    data     bss     dec     hex filename
  12262       4       0   12266    2fea /tmp/ip_output.o.old
  12262       4       0   12266    2fea /tmp/ip_output.o.new
[acme@doppio pahole]$

Investigating...

- Arnaldo
-
To unsubscribe from this list: send the line "unsubscribe dwarves" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html