[PATCH] D24399: [RS4GC] Avoid usage of unrelocated values while rematerializing geps and casts

[Philip Reames via llvm-commits]

reames added inline comments.

================
Comment at: lib/Transforms/Scalar/RewriteStatepointsForGC.cpp:1859
@@ -1858,1 +1858,3 @@
       }
+      // The rematerialized gep/cast instructions should use the relocated base
+      // value (note that the current phi will never be relocated). So, all its
----------------
I'm not really following why this is needed.  How can there be uses of the original phi which don't get updated?  Isn't that a missed relocation bug?  Looking at the added test case, I don't see how that's related to the change?

If you could expand a bit more on the scenario which causes this and how we reach the problematic situation, I'd appreciate it.


https://reviews.llvm.org/D24399