[llvm] [llvm][InstCombine] bitcast bfloat half castpair bug (PR #79832)

[Nikita Popov via llvm-commits]

================
@@ -3214,6 +3214,13 @@ unsigned CastInst::isEliminableCastPair(
         return secondOp;
       return 0;
     case 6:
+      // In cast pairs bfloat and half float shouldn't be treated as equivalent
+      // if the first operation is a bitcast i.e. if we have
+      // bitcast bfloat to half + fpext half to double we shouldn't reduce to
+      // fpext bfloat to double as this isn't equal to fpext half to double.
+      // This has been generalised for all float pairs that have the same width.
+      if(SrcTy->isFloatingPointTy() && MidTy->isFloatingPointTy())
+        return 0;
----------------
nikic wrote:

This implementation doesn't make much sense to me. Note that case 6 is the combination of bitcast + fptoui/fptosi/fpext/fptrunc, so you know that the MidTy is always a float. The below check also ensures that SrcTy is always a float. So this condition in the end is equivalent to "never fold". To implement that, you should replace the entries in the above table with `0` and remove this case entirely.

https://github.com/llvm/llvm-project/pull/79832