[llvm-commits] Fix for assertion in DAGCombiner

Dan Gohman gohman at apple.com
Mon May 3 17:05:05 PDT 2010

On Apr 29, 2010, at 12:26 PM, Jan Sjodin wrote:

>> Thinking about this a little more, I think this patch is 
>> fine, for its part. However, the current one doesn't apply cleanly;
>> at a glance, it looks like the whitespace isn't quite correct.
>> Could you re-send it, possibly as an attachment?
> I guess the argument could be made that this kind of lowering to
> get legal code should not be done in the DAGCombiner. Relying
> on a pure optimization to get correct code is not clean imo. However,
> it may already impossible to exclude the DAGCombiner to get correct
> code for all backends. In any case I updated and regenerated the patch
> Let me know if you have issues applying this. 

The patch is now applied, thanks.

In theory, it should be possible to disable the DAGCombiner passes
without breaking functionality. It does appear this VSETCC code is a
case where correctness depends on the DAGCombiner though. This is
unfortunate, but it's a consequence of LLVM's CodeGen not having
fully generalized vector comparison support yet.


More information about the llvm-commits mailing list