Signed overflow optimization hazards in the kernel [LWN.net]

Signed overflow optimization hazards in the kernel

Posted Aug 20, 2012 16:45 UTC (Mon) by daglwn (guest, #65432)
In reply to: Signed overflow optimization hazards in the kernel by cmccabe
Parent article: Signed overflow optimization hazards in the kernel

You don't believe vectorization can significantly speed up code?

to post comments

Signed overflow optimization hazards in the kernel

Posted Aug 21, 2012 7:58 UTC (Tue) by jezuch (subscriber, #52988) [Link] (2 responses)

As someone who insists on rebuilding some of Debian's packages for my own machine (for fun and not for profit), I can tell that vectorization very rarely has any significant impact. Yes, it can provide a great boost in some synthetic benchmarks or maybe in HPC code (but then, HPC code relies more on optimization by hand and leaves little to the compiler, as I understand it), but very few loops in real-world code are well-formed enough to be suitable for auto-vectorization. I won't make your browser visibly faster :)

Signed overflow optimization hazards in the kernel

Posted Aug 21, 2012 14:30 UTC (Tue) by daglwn (guest, #65432) [Link] (1 responses)

> but then, HPC code relies more on optimization by hand and leaves little
> to the compiler, as I understand it)

Not true. The Intel compiler, for example, will vectorize its brains out automatically.

A code using lots of non-restrict pointers will certainly be difficult to vectorize. Such code can sometimes be auto-parallelized but I don't believe gcc has that capability.

gcc's vectorization is also pretty weak, though it is getting better. That may be part of what you're seeing.

Signed overflow optimization hazards in the kernel

Posted Aug 22, 2012 10:39 UTC (Wed) by jezuch (subscriber, #52988) [Link]

> A code using lots of non-restrict pointers will certainly be difficult to vectorize. Such code can sometimes be auto-parallelized but I don't believe gcc has that capability.

It sorta-kinda does (-ftree-parallelize-loops) but I haven't tested it much. But I expect it to be even weaker than vectorization.

> gcc's vectorization is also pretty weak, though it is getting better. That may be part of what you're seeing.

It may be that. Or it may be that most of the code in non-HPC world does not lend itself to vectorization easily. I actually don't know, I'm just an amateur and hobbyist :)