Int64 indices are silently truncated when used to access MD array elements

[mikedn]

The following code should produce an exception (IndexOutOfRangeException but OverflowException would do too I guess) but it does not:

        static void Main() => Test(new int[42, 42], 41, 41);
        [MethodImpl(MethodImplOptions.NoInlining)]
        static void Test(int[,] a, long x, long y) => a[x | (0x12345678L << 32), y] = 42;

The generated IL code is:

    ldarg.0
    ldarg.1
    ldc.i8     0x1234567800000000
    or
    conv.ovf.i
    ldarg.2
    conv.ovf.i
    ldc.i4.s   42
    call instance void int32[0..., 0...]::Set(int32,int32,int32)

The spec appears to be silent about this case. From the generated code it looks like the intention was to handle this case by using a checked cast, the same is done with SZ arrays.

There must have been some confusion with ldelem. That one accepts both int32 and native int but MD array accessors support only int32 indices and passing a native int argument in this case would result in silent truncation. That conv.ovf.i should really be conv.ovf.i4.

Reproduced with VS 15.8.7 but I'd guess that the problem is much older.

[mikedn]

It seems that UInt32 indices also cause problems. The C# spec sort of rejects the idea of negative array indices, there's no way to create such an array in C#. But you can create one using Array.CreateInstance and then cast it to a C# array type. Then large unsigned indices get treated as valid negative indices:

var a = Array.CreateInstance(typeof(int), new int[] { 10, 10 }, new int[] { -10, -10 });
var md = (int[,])a;
md[-1, -1] = 42;
Console.WriteLine("Yay {0}!", md[uint.MaxValue, uint.MaxValue]); // Happily prints Yay 42!

It would seem that we need to insert a conv.ovf.i4 for UInt32 indices. Not particularly thrilled about that, sometimes it's useful to use uint instead of int for perf reasons. It would be a pity to end up with overflow checks when doing that.