This is related to the discussion in D142071. It can solve the regression in D142071.
I tested llvm/tests/Transforms/CodeGenPrepare/X86/overflow-intrinsics.ll, no new regression.
Usually MathUsed is Add->hasNUsesOrMore(2), for optbranch_64, the add just has two uses.
MathUsed is true. If we could check these use must come from the same BB, perhaps MathUsed could be false.