In case anyone's wondering about the performance hit of boxing the struct in a Nullable object (to avoid the double type check from is
and the cast), there is a non-negligible overhead.
tl;dr: Use is
& cast in this scenario.
struct Foo : IEquatable<Foo>
{
public int a, b;
public Foo(int a, int b)
{
this.a = a;
this.b = b;
}
public override bool Equals(object obj)
{
#if BOXING
var obj_ = obj as Foo?;
return obj_ != null && Equals(obj_.Value);
#elif DOUBLECHECK
return obj is Foo && Equals((Foo)obj);
#elif MAGIC
?
#endif
}
public bool Equals(Foo other)
{
return a == other.a && b == other.b;
}
}
class Program
{
static void Main(string[] args)
{
RunBenchmark(new Foo(42, 43), new Foo(42, 43));
RunBenchmark(new Foo(42, 43), new Foo(43, 44));
}
static void RunBenchmark(object x, object y)
{
var sw = Stopwatch.StartNew();
for (var i = 0; i < 100000000; i++) x.Equals(y);
sw.Stop();
Console.WriteLine(sw.ElapsedMilliseconds);
}
}
Results:
BOXING
EQ 8012 7973 7981 8000
NEQ 7929 7715 7906 7888
DOUBLECHECK
EQ 3654 3650 3638 3605
NEQ 3310 3301 3319 3297
Warning: This test might be flawed in many ways, though I did verify that the benchmark code itself wasn't optimized in an odd fashion.
Looking at the IL, the double-check method compiles a little cleaner.
Boxing IL:
.method public hidebysig virtual
instance bool Equals (
object obj
) cil managed
{
// Method begins at RVA 0x2060
// Code size 37 (0x25)
.maxstack 2
.locals init (
[0] valuetype [mscorlib]System.Nullable`1<valuetype StructIEqualsImpl.Foo> obj_
)
IL_0000: ldarg.1
IL_0001: isinst valuetype [mscorlib]System.Nullable`1<valuetype StructIEqualsImpl.Foo>
IL_0006: unbox.any valuetype [mscorlib]System.Nullable`1<valuetype StructIEqualsImpl.Foo>
IL_000b: stloc.0
IL_000c: ldloca.s obj_
IL_000e: call instance bool valuetype [mscorlib]System.Nullable`1<valuetype StructIEqualsImpl.Foo>::get_HasValue()
IL_0013: brfalse.s IL_0023
IL_0015: ldarg.0
IL_0016: ldloca.s obj_
IL_0018: call instance !0 valuetype [mscorlib]System.Nullable`1<valuetype StructIEqualsImpl.Foo>::get_Value()
IL_001d: call instance bool StructIEqualsImpl.Foo::Equals(valuetype StructIEqualsImpl.Foo)
IL_0022: ret
IL_0023: ldc.i4.0
IL_0024: ret
} // end of method Foo::Equals
Double-check IL:
.method public hidebysig virtual
instance bool Equals (
object obj
) cil managed
{
// Method begins at RVA 0x2060
// Code size 23 (0x17)
.maxstack 8
IL_0000: ldarg.1
IL_0001: isinst StructIEqualsImpl.Foo
IL_0006: brfalse.s IL_0015
IL_0008: ldarg.0
IL_0009: ldarg.1
IL_000a: unbox.any StructIEqualsImpl.Foo
IL_000f: call instance bool StructIEqualsImpl.Foo::Equals(valuetype StructIEqualsImpl.Foo)
IL_0014: ret
IL_0015: ldc.i4.0
IL_0016: ret
} // end of method Foo::Equals
Props to Roman Reiner for spotting a mistake that really wasn't making me look good.