[llvm-commits] [llvm] r154263 - in /llvm/trunk: lib/Analysis/ValueTracking.cpp test/Transforms/LoopRotate/dbgvalue.ll

[Chandler Carruth]

On Sat, Apr 7, 2012 at 9:24 PM, Eric Christopher <echristo at apple.com> wrote:

> I know some people did an audit of the passes a while ago to make sure
> that debug information didn't affect the output of the pass. Looks like
> they
> missed a few, but in general it shouldn't be too widespread.
>

Oh, I know that, that's not the problem.

Just disabling the one optimization (loop rotate) causes wildly different
behavior downstream. Block placement goes nuts, branch layout is crap,
loops don't get laid out correctly, extra branches in the output, etc.

And from what I can tell, this disables loop rotate for essentially every
natural for loop structure with -g...

I was trying to indicate how inaccurate profiles may have been if collected
with debug info because of this optimization not firing, not trying to
indicate that I think there are a bunch more problems like this waiting in
the wings.... This is the only one I've seen recently.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20120407/f669683c/attachment.html>