[Mono-devel-list] Performance of leave vs. br

[Ben Maurer]

Hello,

Today I noted that for the code:

> using System;
> 
> class T {
> 	static void Main ()
> 	{
> 		bool a = false, b = false, c = true;
> 		try {
> 			while (a)
> 				if (b) break;
> 			
> 			Console.WriteLine (c);
> 		} catch {
> 		}
> 	}
> }

That we emit:

>         .try { // 0
>           IL_0006:  br IL_0016
>  
>           IL_000b:  ldloc.2
>           IL_000c:  brfalse IL_0016
>  
>           IL_0011:  leave IL_001c
>  
>           IL_0016:  ldloc.1
>           IL_0017:  brtrue IL_000b
>  
>           IL_001c:  ldloc.0
>           IL_001d:  call void class [mscorlib]'System.Console'::'WriteLine'(bool)
>           IL_0022:  leave IL_002d
>  
>         } // end .try 0

In IL_0011, the op code could just as well be a `br' because it is not
leaving the try statement.

I was wondering if there is enough of a difference in the code we
generate to justify an optimization in MCS.
(Another interesting point is the brfalse. really, we could just write
for this loop:

>           IL_0006:  br IL_0016
>  
>           IL_000b:  ldloc.2
>           IL_000c:  brtrue IL_001
> 
>           IL_0016:  ldloc.1
>           IL_0017:  brtrue IL_000b
>  
>           IL_001c:  ldloc.0
>           IL_001d:  call void class [mscorlib]'System.Console'::'WriteLine'(bool)
>           IL_0022:  leave IL_002d

Can the JIT optimize this sort of case today? If not, should we
implement it in the JIT or the Compiler?

-- Ben