[PATCH] D30638: [SLP] Fixed non-determenistic behavior in Loop Vectorizer.

[Michael Kuperstein via Phabricator via llvm-commits]

mkuper accepted this revision.
mkuper added a comment.
This revision is now accepted and ready to land.

LGTM, modulo the minor issues inline.



================
Comment at: lib/Analysis/LoopAccessAnalysis.cpp:825
           if ((IsWrite || IsReadOnlyPtr) && SetHasWrite) {
-            CheckDeps.insert(Access);
+            CheckDeps.push_back(Access);
             IsRTCheckAnalysisNeeded = true;
----------------
aaboud wrote:
> mkuper wrote:
> > We never have duplicates here anyway, right? So it doesn't matter if it's a set or a list.
> I do not know this code Michael, but even if we end up having duplication, we are going to process them once due to the "Visited" set.
> 
Yes, but if there's a lot of duplication, then the loop now has many more iterations. I think this is fine, though, since a pointer can't be in more than one AliasSet.


================
Comment at: lib/Analysis/LoopAccessAnalysis.cpp:1611
   }
+  CheckDeps.clear();
 
----------------
aaboud wrote:
> mkuper wrote:
> > Is this necessary?
> > I mean, it used to happen as a side-effect of the way the loop was structured, but is CheckDeps every actually used after this?
> Again, I am not familiar with the code, from my check, this is why I preferred to keep same behavior as before.
> 
> I do not mind removing this line.
I don't know this code well enough either to positively say whether it's safe to remove. That's why I'm asking you to check. :-)


https://reviews.llvm.org/D30638