[llvm-commits] [patch] Use the new edge dominance functions in GVN

[Rafael Espíndola]

> In any case,  an updated patch is attached. I made it slightly less
> aggressive but faster by keeping isOnlyReachableViaThisEdge as a fast
> dominance approximation. I understand that it is taking a technical
> debt, but I think I should be able to repay it :-)

Looking at the code of the new GVN pass I see that it implements a
similar feature by doing

      AU.addRequiredID(BreakCriticalEdgesID)

While this is a possibility, applying this to the current GVN
implementation causes some problems. First, we have to mark
BreakCriticalEdges as preserving AliasAnalysis and
MemoryDependenceAnalysis, which I guess is fine.

The addition of extra basic blocks seems to prevent some optimizations
and causes two tests to fail:

    LLVM :: CodeGen/Thumb2/2009-12-01-LoopIVUsers.ll
    LLVM :: Transforms/GVN/2011-06-01-NonLocalMemdepMiscompile.l

I have not debugged exactly why. Finally, it seems that when looking
at the time of a full opt -O2 run, at least with the current
algorithm, the impact of splitting the critical edge is larger than
the impact of directly reasoning about a use being dependent on an
edge.

In any case. I have attached both the previous patch and a variation
that uses critical edge splitting instead. If there is a preference
for the edge splitting one I can try to figure out the two remaining
test failures.

Cheers,
Rafael
-------------- next part --------------
A non-text attachment was scrubbed...
Name: pr13307.patch
Type: application/octet-stream
Size: 13096 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20120813/7e9ee788/attachment.obj>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: pr13307-split.patch
Type: application/octet-stream
Size: 5346 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20120813/7e9ee788/attachment-0001.obj>