Description
As discussed in https://discourse.llvm.org/t/rfc-clarify-the-behavior-of-fp-operations-on-bit-strings-with-nsz-flag/85981, forbidding nsz flags in bitwise FP operations (select/phi/copysign/fabs/fneg) is infeasible as it blocks some common optimizations (e.g., (X < 0.0) ? -X : X -> fabs(X)). I am working on a better formal definition for the nsz flag (and other rewrite-based flags).
However, the nsz flag on fcmp is meaningless since fcmp treats both positive and negative zero as equal. It allows us to freely add nsz: https://alive2.llvm.org/ce/z/x2iNNo
Unfortunately, I notice that some existing optimizations in InstCombine rely on the nsz flag on fcmp. It is a strange behavior and causes some miscompilations (e.g., https://alive2.llvm.org/ce/z/J9ZBXX).
I have created a tool for metamorphic testing, to ensure that the nsz flag on fcmp doesn't affect the optimized IR.