Why does throwing 2 exceptions in a row not generate an unreachable code warning?

Solution 1:

It is clearly a compiler bug, and it was introduced in C# 3.0 -- right around the time that I heavily refactored the reachability checker. This is probably my bad, sorry.

The bug is completely benign; basically, we just forgot a case in the warning reporter. We generate the reachability information correctly; as others have noted, we correctly trim out the unreachable code before codegen.

The bug is nothing more than a missing case in the warning generator. We have some tricky code in there that ensures that we do not report a zillion warnings when you make some large section of code unreachable. The compiler has code for specifically reporting warnings on unconditional gotos ("goto", "break", "continue"), conditional gotos ("if", "while" and so on), try-catch-finally (which includes forms equivalent to try-catch-finally, like lock and using), blocks, returns (yield return and regular return), local declarations, labelled statements, switches, and expression statements.

Do you see "throw statements" on that list? Me neither. That's because we forgot it.

Apologies for the inconvenience. I'll send a note to QA and we'll get a fix for this into a future version of the language.

Thanks for bringing this to my attention.

Solution 2:

It could give a compiler warning / error but sadly it does not. But if you look at IL code only first exception is regarded. You can log in to connect.microsoft.com and raise this as something you would like to see.

if you ILDasm the code below

static void Main(string[] args)
        {
            Console.Write("Line 1");
           throw new Exception(); 
           throw new Exception();
           Console.Write("Line 4");
        }

You will get this

.method private hidebysig static void  Main(string[] args) cil managed
{
  .entrypoint
  // Code size       18 (0x12)
  .maxstack  8
  IL_0000:  nop
  IL_0001:  ldstr      "Line 1"
  IL_0006:  call       void [mscorlib]System.Console::Write(string)
  IL_000b:  nop
  IL_000c:  newobj     instance void [mscorlib]System.Exception::.ctor()
  IL_0011:  throw
} // end of method Program::Main

After the first Exception object nothing else is converted to IL.

Solution 3:

This bug (as Lippert calls it above) has some strange consequences. Of course code like this also gives no compile-time warnings:

static int Main()
{
  return 0;

  throw new Exception("Can you reach me?");
}

If you're creative, you can still make the throw statement induce (unrelated) warnings. In this curious example, the code generates a warning only because "green" is unreachable:

static int Main()
{
  return 0;

  throw new Exception(((Func<string>)(() => { if (2 == 2) { return "yellow"; } return "green"; }))());
}

(code just creates a delegate instance from a lambda and invokes the delegate).

But this example is simpler and seems worse:

static int Main()
{
  int neverAssigned;

  return 0;

  throw new Exception(neverAssigned.ToString());
}

This last code sample also compiles with no warning! There's no problem in "using" neverAssigned because the "usage" is unreachable. But you also get no warning about a local variable never assigned to (and never "really" read). So to repeat, no warning at all, which seems very wrong.

I wonder if this behavior will be changed in future versions of Visual C#? Changing it will give people warnings they didn't have before (which in my opinion they deserve).

Addition: This behavior appears to be unchanged with the Roslyn-based C# 6.0 compiler of Visual Studio 2015.