tabs ↹ over ␣ ␣ ␣ spaces

by Jiří {x2} Činčura

Is it better to "!= 0” or ”== mask” when working with enums (C#, RyuJIT)

20 May 2019 .NET Core, JIT, RyuJIT

First thing first, I know there’s an Enum.HasFlag method, yet I’m not interested in it (or optimizations around it). What I like to know is whether one of these two ways – (EnumValue & mask) != 0 and (EnumValue & mask) == mask – to check for a (single value) flag in an enum is better.

I’m writing better, because I don’t want to say faster, as other aspects of such primitive operation might affect the overall result (like code size, register usage, etc.).

Hypothesis

My guess is that the "!= 0” is going be the winner, because computers are good doing stuff around value zero (like jnz instruction).

Code

In all the code following, I’ll be using this enum I aptly named MyEnum. Nothing special.

[Flags]
public enum MyEnum
{
    None = 0,
    A = 0x01,
    B = 0x02,
    C = 0x04,
    D = 0x08,
}

And two simple methods (with x64/AMD64 assembly right below each).

public bool A(MyEnum myEnum)
{
    return (myEnum & MyEnum.B) == MyEnum.B;
}
test dl, 0x2
setnz al
movzx eax, al
ret
public bool B(MyEnum myEnum)
{
    return (myEnum & MyEnum.B) != 0;
}
test dl, 0x2
setnz al
movzx eax, al
ret

Surprisingly both result in exactly the same assembly produced by 64-bit RyuJIT (on .NET Core 2.2.4). Although, it kind of makes sense, both are doing the same. I also learned about setnz and movzx instructions (the movzx strictly speaking is the for returning the value, not for the compare) along the way, as I wasn’t familiar with these.

I could stop here and say that both are the same and use whatever I feel like using given the circumstances. But I like to look under.

Where’s the magic happening in RyuJIT

I’m nowhere familiar of internals of RyuJIT, thus I had no idea where is this happening. Time for little thinking. I’m working with enums, but enums are just numbers, hence maybe the same behavior works for numbers too. That way I know whether to keep grepping for something around enums or look for something more general.

Sadly doing (value & 0x02) == 0x02 and (value & 0x02) != 0 (where value is plain int) shows same behavior. I’ll spare you my two hours of semi-methodically (read randomly 😃) walking through the RyuJIT’s code and I’ll jump right into the action.

The first magic happens in Lowering::OptimizeConstCompare, where the “compare and” is transformed into “test” and even “inventing” the 0 comparison if possible. That covers transforming ”== mask” into "!= 0”. The other part of the puzzle can be found in the code generation itself. The test reg, reg instruction is shorter than doing cmp reg, 0. And although the saving is not huge (1 byte), if you’re writing JIT, every byte and CPU cycle more than counts.

Moreover, as I was researching, I learned about macro-fusing instructions in certain architectures and test can (i.e. Intel Core 2, Intel Nehalem, …) macro-fuse with more conditional jump instructions, compared to cmp.

Real world usage

Above I said, that both are the same. But carefully looking into Lowering::OptimizeConstCompare at this line, one can see the 0 comparison is “invented” only if the second operand is integral single bit mask constant. That means, if you have a code, where the “mask” is i.e. a variable, it’s better to use "!= 0” if you can, else you’re out of luck with RyuJIT’s smartness.

public bool A(MyEnum myEnum, MyEnum value)
{
    return (myEnum & value) == value;
}
and edx, r8d
cmp edx, r8d
setz al
movzx eax, al
ret
public bool B(MyEnum myEnum, MyEnum value)
{
    return (myEnum & value) != 0;
}
test r8d, edx
setnz al
movzx eax, al
ret

Summary

Although my hypothesis with "!= 0” was correct, I learned the direct “compare” with 0 wasn’t really the reason. Quite the opposite. And finally, I spent time learning something new and it was fun.

In case you’d ask; the Enum.HasFlag is optimized and doing the same at the end.

Profile Picture Jiří Činčura is an independent developer focusing on data and business layers, language constructs, parallelism and databases. Specifically Entity Framework, asynchronous and parallel programming, cloud and Azure. He's Microsoft Most Valuable Professional and you can read his articles, guides, tips and tricks at www.tabsoverspaces.com.