[PATCH] D51431: [WIP][IPSCCP] Add lattice value for != constant and propagate nonnull.

[Florian Hahn via Phabricator via llvm-commits]

fhahn added inline comments.


================
Comment at: lib/Transforms/Scalar/SCCP.cpp:844
+    if (IV.isNotConstant())
+      return (void)markOverdefined(&PN);
+
----------------
efriedma wrote:
> I guess you could try to resolve a PHI to NotConstant...?  Not sure how important that is.
Yep, there should be a TODO here (and in other places)


================
Comment at: lib/Transforms/Scalar/SCCP.cpp:1785
+      IPNumNonNullAdded++;
+      CS.addParamAttr(ArgNo, Attribute::NonNull);
+    }
----------------
efriedma wrote:
> I'm not sure this is right: nonnull is UB if the value is null, but SCCP can merge an "Undef" into a NotConstant.
I think I am missing something. tryToAddNonNull is used after we finished solving and a lattice value of notconstant(Null) should mean "we proved that the value is != null".  If we  merge an undef into a NotConstant, isn't that similar to when we merge an undef into a Constant, where it is safe to use the constant after solving?


https://reviews.llvm.org/D51431