C# generated IL for ++ operator - when and why pre

问题:

Since this question is about the increment operator and speed differences with prefix/postfix notation, I will describe the question very carefully lest Eric Lippert discover it and flame me!

(further info and more detail on why I am asking can be found at http://www.codeproject.com/KB/cs/FastLessCSharpIteration.aspx?msg=3899456#xx3899456xx/)

I have four snippets of code as follows:-

(1) Separate, Prefix:

    for (var j = 0; j != jmax;) { total += intArray[j]; ++j; }

(2) Separate, Postfix:

    for (var j = 0; j != jmax;) { total += intArray[j]; j++; }

(3) Indexer, Postfix:

    for (var j = 0; j != jmax;) { total += intArray[j++]; }

(4) Indexer, Prefix:

    for (var j = -1; j != last;) { total += intArray[++j]; } // last = jmax - 1

What I was trying to do was prove/disprove whether there is a performance difference between prefix and postfix notation in this context (ie a local variable so not volatile, not changeable from another thread etc.) and if there was, why that would be.

Speed testing showed that:

• (1) and (2) run at the same speed as each other.

• (3) and (4) run at the same speed as each other.

• (3)/(4) are ~27% slower than (1)/(2).

Therefore I am concluding that there is no performance advantage of choosing prefix notation over postfix notation per se. However when the Result of the Operation is actually used, then this results in slower code than if it is simply thrown away.

I then had a look at the generated IL using Reflector and found the following:

• The number of IL bytes is identical in all cases.

• The .maxstack varied between 4 and 6 but I believe that is used only for verification purposes and so not relevant to performance.

• (1) and (2) generated exactly the same IL so its no surprise that the timing was identical. So we can ignore (1).

• (3) and (4) generated very similar code - the only relevant difference being the positioning of a dup opcode to account for the Result of the Operation. Again, no surprise about timing being identical.

So I then compared (2) and (3) to find out what could account for the difference in speed:

• (2) uses a ldloc.0 op twice (once as part of the indexer and then later as part of the increment).

• (3) used ldloc.0 followed immediately by a dup op.

So the relevant IL for the incrementing j for (1) (and (2)) is:

// ldloc.0 already used once for the indexer operation higher up
ldloc.0
ldc.i4.1
add
stloc.0

(3) looks like this:

ldloc.0
dup // j on the stack for the *Result of the Operation*
ldc.i4.1
add
stloc.0

(4) looks like this:

ldloc.0
ldc.i4.1
add
dup // j + 1 on the stack for the *Result of the Operation*
stloc.0

Now (finally!) to the question:

Is (2) faster because the JIT compiler recognises a pattern of ldloc.0/ldc.i4.1/add/stloc.0 as simply incrementing a local variable by 1 and optimize it? (and the presence of a dup in (3) and (4) break that pattern and so the optimization is missed)

And a supplementary: If this is true then, for (3) at least, wouldn't replacing the dup with another ldloc.0 reintroduce that pattern?

回答1:

OK after much research (sad I know!), I think have answered my own question:

The answer is Maybe. Apparently the JIT compilers do look for patterns (see http://blogs.msdn.com/b/clrcodegeneration/archive/2009/08/13/array-bounds-check-elimination-in-the-clr.aspx) to decide when and how array bounds checking can be optimized but whether it is the same pattern I was guessing at or not I don't know.

In this case, it is a moot point because the relative speed increase of (2) was due to something more than that. Turns out that the x64 JIT compiler is clever enough to work out whether an array length is constant (and seemingly also a multiple of the number of unrolls in a loop): So the code was only bounds checking at the end of each iteration and the each unroll became just:-

        total += intArray[j]; j++;
00000081 8B 44 0B 10          mov         eax,dword ptr [rbx+rcx+10h] 
00000085 03 F0                add         esi,eax 

I proved this by changing the app to let the array size be specified on the command line and seeing the different assembler output.

Other things discovered during this excercise:-

• For a standalone increment operation (ie the result is not used), there is no difference in speed between prefix/postfix.
• When an increment operation is used in an indexer, the assembler shows that prefix notation is slightly more efficient (and so close in the the original case that I assumed it was just a timing discrepency and called them equal - my mistake). The difference is more pronounced when compiled as x86.
• Loop unrolling does work. Compared to a standard loop with array bounds optimization, 4 rollups always gave an improvement of 10%-20% (and the x64/constant case 34%). Increasing the number of rollups gave varied timing with some very much slower in the case of a postfix in the indexer, so I'll stick with 4 if unrolling and only change that after extensive timing for a specific case.

回答2:

Interesting results. What I would do is:

• Rewrite the application to do the whole test twice.
• Put a message box between the two test runs.
• Compile for release, no optimizations, and so on.
• Start the executable outside of the debugger.
• When the message box comes up, attach the debugger
• Now inspect the code generated for the two different cases by the jitter.

And then you'll know whether the jitter is doing a better job with one than the other. The jitter might, for example, be realizing that in one case it can remove array bounds checks, but not realizing that in the other case. I don't know; I'm not an expert on the jitter.

The reason for all the rigamarole is because the jitter may generate different code when the debugger is attached. If you want to know what it does under normal circumstances then you have to make sure the code gets jitted under normal, non-debugger circumstances.

回答3:

I love performance testing and I love fast programs so I admire your question.

I tried to reproduce your findings and failed. On my Intel i7 x64 system running your code samples on .NET4 framework in the x86|Release configuration, all four test cases produced roughly the same timings.

To do the test I created a brand new console application project and used the QueryPerformanceCounter API call to get a high-resolution CPU-based timer. I tried two settings for jmax:

• jmax = 1000
• jmax = 1000000

because locality of the array can often make a big difference in how the performance behaves and the size of the of loop increases. However, both array sizes behaved the same in my tests.

I have done a lot of performance optimization and one of the things that I have learned is that you can very easily optimize an application so that it runs faster on one particular computer while inadvertently causing it to run slower on another computer.

I am not talking hypothetically here. I have tweaked inner loops and poured hours and days of work to make a program run faster, only to have my hopes dashed because I was optimizing it on my workstation and the target computer was a different model of Intel processor.

So the moral of this story is:

• Code snippet (2) runs faster than code snippet (3) on your computer but not on my computer

This is why some compilers have special optimization switches for different processors or some applications come in different versions even though one version could easily run on all supported hardware.

So if you are going to do testing like this, you have to do it same way that JIT compiler writers do: you have to perform your tests on a wide variety of hardware and then choose a blend, a happy-medium that gives the best performance on the most ubiquitous hardware.