[llvm] [llubi] Initial support for floating-point numbers (PR #188453)

[Yingwei Zheng via llvm-commits]

================
@@ -187,6 +233,24 @@ class InstExecutor : public InstVisitor<InstExecutor, void> {
     });
   }
 
+  void visitFPUnOp(Instruction &I,
+                   function_ref<AnyValue(const APFloat &)> ScalarFn) {
+    FastMathFlags FMF = cast<FPMathOperator>(I).getFastMathFlags();
+
+    visitUnOp(I, [&](const AnyValue &Operand) -> AnyValue {
+      if (Operand.isPoison())
+        return AnyValue::poison();
+
+      // We don't flush denormals here since the only floating-point unary
+      // operation is fneg. And fneg is specified as a bitwise operation which
+      // only flips the sign bit of the input.
+
+      APFloat Result = ScalarFn(Operand.asFloat()).asFloat();
----------------
dtcxzyw wrote:

`ScalarFn` may return a poison. It would be better to change the return type to APFloat.

https://github.com/llvm/llvm-project/pull/188453