replace memcpy with memmove

[vtjnash]

Uses of memcpy in Base may need to be replaced by memmove, except where we can guarantee that there is no pointer aliasing. (similarly, uses of LAPACK copy should be avoided, unless we first check for pointer aliasing)

[stevengj]

Some useful benchmarks here on whether it is even worth figuring out whether memcpy is safe. As a first pass, it looks like we wouldn't lose much if we did s/memcpy/memmove/ on everything.

[timholy]

I read those benchmarks as saying "just use loops." I know the author doesn't say that, but check out the curves: for all compilers except pgcc, the loops-with-indices result is right on top of memcpy.

That confirms something I had noticed recently: on my current machine, with a current julia, I also couldn't see a difference. I found this interesting because some time ago I went to a certain amount of effort (#190) to incorporate memcpy into getindex/setindex. At the time it was a major improvement (something like 5-fold, depending on array size etc). I don't know whether it's because I've since changed platforms (x86->x86_64), because of advances in LLVM, or because we're turning on more optimization passes, but when I benchmarked this a week or two ago I couldn't see the difference between them.

If other people can't see any differences between for loops and ccall(:memcpy,...) either, I propose we yank out calls to memcpy and go back to for loops. It's not often we get to shrink base...

[pao]

I don't know whether it's because I've since changed platforms (x86->x86_64), because of advances in LLVM, or because we're turning on more optimization passes, but when I benchmarked this a week or two ago I couldn't see the difference between them.

I'd be wary of this if we don't know which change is the source of the speedup. I've seen (recently) order-of-magnitude differences in the speed of memcpy vs. loop copying of arrays (gcc 4.6.3, -O2, for what it's worth.)

[simonster]

There are also LLVM intrinsics for memcpy/memmove, although I don't know whether there's any advantage to using them.

[vtjnash]

that trivial loop has the same bug I am advising against here. the use of loops in gcc is basically hoping that it's pattern matching algorithm will turn it into a call to memcpy (as shown in the blog).

[timholy]

Fair enough. I agree it looks like the performance cost is negligible.

[vtjnash]

@dcjones was that the only memcpy usage that needed to be switched?

[dcjones]

I looked at the memcpy ccalls in Base and this seemed like the only risky one.