It's rainy, but I'm beginning to see the light...
Things are more funny, but not really satisfactory...
Bad news first: still no notebook, so I'm working from my home desktop,
with just a dialup connection. This also means that my one year old son,
Michele, must generally sleep/play in my (and my wife's) bedroom, at least
while I am working.
Anyway, I managed to get something done. I rebuilt mono from the 20040502
CVS snapshot (had to get monolite for that, but never mind), and easily
applied the ABC removal patch to it. What is "not satisfactory" is that
the code works correctly (bounds checks are removed), but there are
no performance gains at all (actually performance gets worse!).
This completely contradicts previous tests, so I investigated a bit.
Machine code for this loop:
for (int i = 0; i <a.Length; i++)
{
a[i] = i;
}
Becomes like this:
2b: eb 0b jmp 38
2d: 8b c3 mov %ebx,%eax
2f: 8b cf mov %edi,%ecx
31: 8d 44 88 10 lea 0x10(%eax,%ecx,4),%eax
35: 89 38 mov %edi,(%eax)
37: 47 inc %edi
38: 8b c3 mov %ebx,%eax
3a: 8b 40 0c mov 0xc(%eax),%eax
3d: 3b f8 cmp %eax,%edi
3f: 7c ec jl 2d
And without bounds check removal like this:
2b: eb 14 jmp 41
2d: 8b c3 mov %ebx,%eax
2f: 8b cf mov %edi,%ecx
31: 39 48 0c cmp %ecx,0xc(%eax)
34: 0f 86 26 00 00 00 jbe 60
3a: 8d 44 88 10 lea 0x10(%eax,%ecx,4),%eax
3e: 89 38 mov %edi,(%eax)
40: 47 inc %edi
41: 8b c3 mov %ebx,%eax
43: 8b 40 0c mov 0xc(%eax),%eax
46: 3b f8 cmp %eax,%edi
48: 7c e3 jl 2d
Now, it is obvious that the bounds check has been removed. What is not so
obvious is why the code does not run faster (and yes, I know that JIT time
should be factored out)!
Maybe I'll have a look at the profiler, and see where the execution time is
actually spent. After all, during my next optimization tasks using a profiler
will be a must, so it's better starting immediately.
On other fronts, I have had a look at the existing implementations of
Array.Copy and Buffer.BlockCopy, understood most of how they work and
(even more important) understood how to provide and internal implementation
for methods (with "InternalCall", and adding the implementation in "icall.c").
At this point, the next thing to do is learn how to use the profiler well, so
that I can properly understand what's going on...
