[llvm-testresults] smoosh-01.apple.com nightly tester results

[Daniel Dunbar]

Since we've been talking about inlining...

On Wed, Aug 26, 2009 at 10:01 PM, <daniel_dunbar at apple.com> wrote:
> LLC:
>  MultiSource/Applications/sgefa/sgefa: -6.32% (0.9500 => 0.8900)

This test consistently flip flops, see attached graphs (yes, we did
make it slower at some point).

It would be interesting to understand why; my personal guess is that
some thing is at the boundary of an inlining heuristic.

An interesting theory (obviously not true in practice) is that things
should never flip flop due to inlining -- if the heuristic is truly at
the boundary between inlining and not inlining, then if the heuristic
is correct inlining or not inlining should yield equivalent
performance. Since the test *is* flip flopping, this would mean some
heuristic is off base, if it is in fact due to an inlining policy
change.

Of course, it could be a million other things too. :)

 - Daniel

p.s. That graph is sampled approximately every 3 hours. Although there
is some truncation of results, this is a good comment on the amount of
true "noise" (i.e., sampling error) in nightly test results.
-------------- next part --------------
A non-text attachment was scrubbed...
Name: gXPerIW.png
Type: image/png
Size: 37845 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-testresults/attachments/20090826/66ec49e6/attachment.png>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: gXGEfd0.png
Type: image/png
Size: 41977 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-testresults/attachments/20090826/66ec49e6/attachment-0001.png>