Shouldn't modulus with a readonly integer not be reduced?

[Genbox]

I've always assumed that the compiler did the good ol' trick of reducing modulus (to an and instruction) when the divisor is 2^n.
Turns out to not be true for readonly integers.

Benchmark:

private readonly uint _mod = 1024;
private const uint _modConst = 1024;

[Benchmark] public long Var() => 3_000_000_000 % _mod;
[Benchmark] public long VarConst() => 3_000_000_000 % _modConst;
[Benchmark] public long VarShift() => 3_000_000_000 & 1023;

Results:

[Host] : .NET 9.0.5 (9.0.525.21509), X64 RyuJIT AVX2

| Method   | Mean      | Error     | StdDev    | Median    |
|--------- |----------:|----------:|----------:|----------:|
| Var      | 0.1045 ns | 0.0387 ns | 0.0256 ns | 0.0894 ns |
| VarConst | 0.0023 ns | 0.0090 ns | 0.0047 ns | 0.0000 ns |
| VarShift | 0.0093 ns | 0.0209 ns | 0.0093 ns | 0.0084 ns |

There are times where const cannot be used and readonly is the only choice. I couldn't find an issue on this, so maybe I'm missing something obvious?

[EgorBo]

There are times where const cannot be used and readonly is the only choice. I couldn't find an issue on this, so maybe I'm missing something obvious?

readonly instance fields cannot be constant folded, you probably meant static readonly.

[Genbox]

I do mean instance fields. If there are no writes to the field, which is enforced with readonly, then at least outside the ctor, it can be assumed the value is constant, no?

[EgorBo]

it can be assumed the value is constant, no?

No, it cannot, it may be changed with reflection and if you can't use const it is implied you have a more complex expression there. There is not much JIT can legally do in this case, only static readonly is allowed to be constant-folded.

[Genbox]

Well, that at least explains it from an ECMA-standard perspective (the semantics breaks with unsafe code).