Suboptimal code generation for == operator for nullable primitive types

[zlatanov]

Version Used: 16.4.2

If we write the following code:

static bool M(int? x, int? y) => x == y;

The compiler generates IL that introduces two locals.

.maxstack 3
.locals init (
	[0] valuetype [System.Private.CoreLib]System.Nullable`1<int32>,
	[1] valuetype [System.Private.CoreLib]System.Nullable`1<int32>
)

IL_0000: ldarg.0
IL_0001: stloc.0
IL_0002: ldarg.1
IL_0003: stloc.1
IL_0004: ldloca.s 0
IL_0006: call instance !0 valuetype [System.Private.CoreLib]System.Nullable`1<int32>::GetValueOrDefault()
IL_000b: ldloca.s 1
IL_000d: call instance !0 valuetype [System.Private.CoreLib]System.Nullable`1<int32>::GetValueOrDefault()
IL_0012: ceq
IL_0014: ldloca.s 0
IL_0016: call instance bool valuetype [System.Private.CoreLib]System.Nullable`1<int32>::get_HasValue()
IL_001b: ldloca.s 1
IL_001d: call instance bool valuetype [System.Private.CoreLib]System.Nullable`1<int32>::get_HasValue()
IL_0022: ceq
IL_0024: and
IL_0025: ret

If we write the code as:

static bool M(int? x, int? y) => x.HasValue == y.HasValue 
    && x.GetValueOrDefault() == y.GetValueOrDefault();

.maxstack 8

IL_0000: ldarga.s x
IL_0002: call instance bool valuetype [System.Private.CoreLib]System.Nullable`1<int32>::get_HasValue()
IL_0007: ldarga.s y
IL_0009: call instance bool valuetype [System.Private.CoreLib]System.Nullable`1<int32>::get_HasValue()
IL_000e: bne.un.s IL_0021

IL_0010: ldarga.s x
IL_0012: call instance !0 valuetype [System.Private.CoreLib]System.Nullable`1<int32>::GetValueOrDefault()
IL_0017: ldarga.s y
IL_0019: call instance !0 valuetype [System.Private.CoreLib]System.Nullable`1<int32>::GetValueOrDefault()
IL_001e: ceq
IL_0020: ret

IL_0021: ldc.i4.0
IL_0022: ret

then the compiler generates code that has no locals and has 3 IL instructions less. The JITed code difference is bigger - the first version is 18 instructions and the second 13.

Here is a repro in sharplab.

Even microbenchmarking the code shows that the second version is better:

BenchmarkDotNet=v0.12.0, OS=Windows 10.0.18363
Intel Core i7-4790K CPU 4.00GHz (Haswell), 1 CPU, 8 logical and 4 physical cores
.NET Core SDK=3.1.100
  [Host]     : .NET Core 3.1.0 (CoreCLR 4.700.19.56402, CoreFX 4.700.19.56404), X64 RyuJIT
  DefaultJob : .NET Core 3.1.0 (CoreCLR 4.700.19.56402, CoreFX 4.700.19.56404), X64 RyuJIT

Method	x	y	Mean	Error	StdDev	Ratio
Default	1	2	2.758 ns	0.0218 ns	0.0193 ns	1.00
Custom	1	2	2.486 ns	0.0281 ns	0.0263 ns	0.90

EDIT:

This is what the code looks like when the values being compared are not arguments: here.

    public static bool M()
    {
        var x = Random();
        var y = Random();
        
        return x == y;
    }
    
    
    [MethodImpl( MethodImplOptions.NoInlining )]
    static int? Random() => 42;

.maxstack 3
.locals init (
	[0] valuetype [System.Private.CoreLib]System.Nullable`1<int32>,
	[1] valuetype [System.Private.CoreLib]System.Nullable`1<int32>,
	[2] valuetype [System.Private.CoreLib]System.Nullable`1<int32>
)

IL_0000: call valuetype [System.Private.CoreLib]System.Nullable`1<int32> C::Random()
IL_0005: stloc.0
IL_0006: call valuetype [System.Private.CoreLib]System.Nullable`1<int32> C::Random()
IL_000b: ldloc.0
IL_000c: stloc.1
IL_000d: stloc.2
IL_000e: ldloca.s 1
IL_0010: call instance !0 valuetype [System.Private.CoreLib]System.Nullable`1<int32>::GetValueOrDefault()
IL_0015: ldloca.s 2
IL_0017: call instance !0 valuetype [System.Private.CoreLib]System.Nullable`1<int32>::GetValueOrDefault()
IL_001c: ceq
IL_001e: ldloca.s 1
IL_0020: call instance bool valuetype [System.Private.CoreLib]System.Nullable`1<int32>::get_HasValue()
IL_0025: ldloca.s 2
IL_0027: call instance bool valuetype [System.Private.CoreLib]System.Nullable`1<int32>::get_HasValue()
IL_002c: ceq
IL_002e: and
IL_002f: ret

This time the code generated is almost the same and the performance is the same. One thing I don't understand is why in this case the compiler generates 3 locals? Both GetValueOrDefault and HasValue are readonly methods and will not mutate the objects?

[tannergooding]

GetValueOrDefault and HasValue were only recently (6 days ago: dotnet/runtime#1727) made readonly, so this might be a point in time issue and it would be worth checking against the latest dotnet/runtime nightly build.