[llvm-commits] [llvm] r131781 - in /llvm/trunk: lib/Analysis/BasicAliasAnalysis.cpp test/Analysis/BasicAA/2010-09-15-GEP-SignedArithmetic.ll test/Analysis/TypeBasedAliasAnalysis/dynamic-indices.ll

[Dan Gohman]

On May 22, 2011, at 5:04 AM, Duncan Sands wrote:

>>> Author: djg
>>> Date: Fri May 20 20:05:08 2011
>>> New Revision: 131781
>>> 
>>> URL: http://llvm.org/viewvc/llvm-project?rev=131781&view=rev
>>> Log:
>>> When BasicAA can determine that two pointers have the same base but
>>> differ by a dynamic offset, return PartialAlias instead of MayAlias.
>>> See the comment in the code for details. This fixes PR9971.
>> 
>> I speculatively reverted this commit to see if doing so fixes this breakage:
>> http://google1.osuosl.org:8011/builders/dragonegg-x86_64-linux/builds/1494
> 
> It did fix the buildbot.  At the point of failure, gcc and dragonegg (but not
> yet LLVM) had been built by dragonegg.  Assuming that dragonegg miscompiled
> itself (which seems likely given that the failure was an APInt assertion) it
> may be feasible to reduce a testcase.  Would you like me to try?


Yes, if you could.

In theory, this change should have made the optimizer strictly less
aggressive.

Dan