[clang] [analyzer] Model concrete floating-point values (PR #214098)

[Balázs Benics via cfe-commits]

================
@@ -288,8 +288,11 @@ void ExprEngine::VisitCast(const CastExpr *CastE, const Expr *Ex,
 
       if (const MemRegion *MR = State->getSVal(Ex, SF).getAsRegion()) {
         SVal OrigV = State->getSVal(MR);
-        CastedV = svalBuilder.evalCast(svalBuilder.simplifySVal(State, OrigV),
-                                       CastE->getType(), Ex->getType());
+        // evalCast converts the value, but we are doing a bitcast here, which
+        // is unmodeled for floats.
+        if (!OrigV.getAs<nonloc::ConcreteFloat>())
+          CastedV = svalBuilder.evalCast(svalBuilder.simplifySVal(State, OrigV),
+                                         CastE->getType(), Ex->getType());
----------------
steakhal wrote:

If `evalCast` does not model floats, then it should return `Unknown` - doesn't it? Which is the same as leaving `CastedV` as it was initialized (`Unknown`). So why do we have this guard here?

https://github.com/llvm/llvm-project/pull/214098