Re: [PATCH i386] Enable -freorder-blocks-and-partition

[Teresa Johnson <tejohnson at google dot com>]

On Thu, Dec 12, 2013 at 5:13 PM, Jan Hubicka <hubicka@ucw.cz> wrote:
>> On Wed, Dec 11, 2013 at 1:21 AM, Martin LiÅka <marxin.liska@gmail.com> wrote:
>> > Hello,
>> >    I prepared a collection of systemtap graphs for GIMP.
>> >
>> > 1) just my profile-based function reordering: 550 pages
>> > 2) just -freorder-blocks-and-partitions: 646 pages
>> > 3) just -fno-reorder-blocks-and-partitions: 638 pages
>> >
>> > Please see attached data.
>>
>> Thanks for the data. A few observations/questions:
>>
>> With both 1) (your (time-based?) reordering) and 2)
>> (-freorder-blocks-and-partitions) there are a fair amount of accesses
>> out of the cold section. I'm not seeing so many accesses out of the
>> cold section in the apps I am looking at with splitting enabled. In
>
> I see you already comitted the patch, so perhaps Martin's measurement assume
> the pass is off by default?
>
> I rebuilded GCC with profiledboostrap and with the linkerscript unmapping
> text.unlikely.  I get ICE in:
> (gdb) bt
> #0  diagnostic_set_caret_max_width(diagnostic_context*, int) () at ../../gcc/diagnostic.c:108
> #1  0x0000000000f68457 in diagnostic_initialize (context=0x18ae000 <global_diagnostic_context>, n_opts=n_opts@entry=1290) at ../../gcc/diagnostic.c:135
> #2  0x000000000100050e in general_init (argv0=<optimized out>) at ../../gcc/toplev.c:1110
> #3  toplev_main(int, char**) () at ../../gcc/toplev.c:1922
> #4  0x00007ffff774cbe5 in __libc_start_main () from /lib64/libc.so.6
> #5  0x0000000000f7898d in _start () at ../sysdeps/x86_64/start.S:122
>
> That is relatively early in startup process. The function seems inlined and
> it fails only on second invocation, did not have time to investigate further,
> yet while without -fprofile-use it starts...

I'll see if I can reproduce this and investigate, although at this
point that might have to wait until after my holiday vacation.

>
> On our periodic testers I see off-noise improvement in crafty 2200->2300
> and regression on Vortex, 2900->2800, plus code size increase.

I had only run cpu2006, but not cpu2000. I'll see if I can reproduce
this as well.

I have been investigating a few places where I saw accesses in the
cold split regions in internal benchmarks. Here are a couple, and how
I have addressed them so far:

1) loop unswitching

In this case, loop unswitching hoisted a branch from within the loop
to outside the loop, and in doing so it was effectively speculated
above several other branches. In it's original location it always went
to only one of the successors (biased 0/100%). But when it was hoisted
it sometimes took the previously 0% path. This led to executing out of
the cold region, since we didn't update the branch probability when
hoisting. I worked around this by assigning a small non-zero
probability after hoisting with the following change:

Index: tree-ssa-loop-unswitch.c
===================================================================
--- tree-ssa-loop-unswitch.c    (revision 205590)
+++ tree-ssa-loop-unswitch.c    (working copy)
@@ -384,6 +384,8 @@ tree_unswitch_loop (struct loop *loop,

   extract_true_false_edges_from_block (unswitch_on, &edge_true, &edge_false);
   prob_true = edge_true->probability;
+  if (!prob_true)
+    prob_true = REG_BR_PROB_BASE/10;
   return loop_version (loop, unshare_expr (cond),
                       NULL, prob_true, prob_true,
                       REG_BR_PROB_BASE - prob_true, false);

This should probably be refined (if prob_true is 100% we want to
assign a small non-zero probability to the false path), and 10% may be
too high (maybe give it 1%?).

2) More COMDAT issues

My earlier patch handled the case where the comdat had 0 counts since
the linker kept the copy in a different module. In that case we
prevent the guessed frequencies from being dropped by counts_to_freq,
and then later mark any reached via non-zero callgraph edges as
guessed. Finally, when one such 0-count comdat is inlined the call
count is propagated to the callee blocks using the guessed
probabilities.

However, in this case, there was a comdat that had a very small
non-zero count, that was being inlined to a much hotter callsite. I
believe this could happen when there was a copy that was ipa-inlined
in the profile gen compile, so the copy in that module gets some
non-zero counts from the ipa inlined instance, but when the out of
line copy was eliminated by the linker (selected from a different
module). In this case the inliner was scaling the bb counts up quite a
lot when inlining. The problem is that you most likely can't trust
that the 0 count bbs in such a case are really not executed by the
callsite it is being into, since the counts are very small and
correspond to a different callsite.

The problem is how to address this. We can't simply suppress
counts_to_freq from overwriting the guessed frequencies in this case,
since the profile counts are non-zero and would not match the guessed
probabilities. But we can't figure out which are called by much hotter
callsites (compared to their entry count) until later when the
callgraph is built, which is when we would know that we want to ignore
the profile counts and use the guessed probabilities instead. The
solution I came up with is to allow the profile counts to overwrite
the guessed probabilites in counts_to_freq. But then when we inline we
re-estimate the probabilities in the callee when the callsite count is
much hotter than the entry count, and then follow the same procedure
we were doing in the 0-count case (propagate the call count into the
callee bb counts via the guessed probabilities). Is there a better
solution?

Thanks,
Teresa


>
> Honza



-- 
Teresa Johnson | Software Engineer | tejohnson@google.com | 408-460-2413