[llvm-dev] Default hashing function for integers (DenseMapInfo.h)

[mats petersson via llvm-dev]

On 10 January 2017 at 15:35, Francois Fayard via llvm-dev <
llvm-dev at lists.llvm.org> wrote:

>
> > It is not clear what the test-case is (what source, what compiler
> options). My suspicion is that your differences are in the noise, and most
> of the time is spent doing other things than hashing. Did you profile the a
> run, and check how much of the total time is spent in the hash-function
> [you may need to tweak the code a bit to not inline the actual hash
> function]. Also publishing the RANGE for each set of tests, since the
> variation between "best and worst" may be more important than the overall
> best time.
>
> I agree that I did not present a clear test case, which was just a random
> .cpp file that benchmarks different hash tables from LLVM/Google/STL
> compiled with -O3 option. I’ll try to make it better tomorrow, but I can
> assure you that it was not the noise. I have disabled turboboost and
> hyper-threading on my machine and ran the benchmark 10 times for every test
> and the difference was way larger than the observed standard deviation. But
> anyway, for the time being, the winner for hashing pointers is the current
> one from of LLVM.
>

Yes, it seems to be "good". But it's always good to understand WHY.


>
> > Counting the number of collisions or some such would probably also help.
>
> Yes, I could do that. If I want to report all the collision, I might do
> that in the destructor of DenseMap. Sorry for my ignorance, but is
> clang/llvm multithreaded? Is it safe to write to a file in a destructor for
> logging purposes?
>

I don't think you can rely on that ALWAYS, but clang and LLVM doesn't spawn
threads - someone may use two threads simultaneously to call into llvm and
clang functionality as a library tho'. In a limited addition to the code to
report something like `std::cout << "Number colliisions: " << NumCollisions
<< std::endl;` should be safe enough when using clang as the test [and
opening a file in append mode should also be reasonably safe]. If you
wanted to submit such a change as a patch, it would probably require a fair
bit more work to guarantee threadsafety... [I have certainly done similar
quick "print something" additions to Clang and LLVM, including in
destructors].

>
> François Fayard
>
> _______________________________________________
> LLVM Developers mailing list
> llvm-dev at lists.llvm.org
> http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20170110/2bc8bf06/attachment.html>