On Wednesday 20 April 2011 19:11:07, Roman Leshchinskiy wrote:
Daniel Fischer wrote:
Further investigation of the sorting code in vector-algorithms revealed no bugs there, and if the runtime was forced to keep a keen eye on the indices, by replacing unsafeRead/Write/Swap with their bounds-checked counterparts or by 'trace'ing enough of their uses, the NaNs did not appear.
Did you replace them in vector-algorithms or in vector itself?
vector-algorithms only.
So, is it possible that some change in ghc-7.0.3 vs. the previous versions caused a bad interaction between ghc-optimisations and vector fusion resulting in bad vector reads/writes?
Am I right in assuming that this happens in code which uses only mutable vectors?
Yes, the sorting uses mutable vectors, in this case unboxed Double vectors.
Fusion only works for immutable ones so it shouldn't really affect things here.
Ah, didn't know that. Another suspect gone.
Have you tried playing around with code generation flags like -msse2?
No, not yet. So far only -O2 (with -fspec-constr-count=5 in the presence of many trace calls) and -O0.
In any case, I would try to take a look at this if you tell me how to reproduce.
I'll prepare a bundle, I'm afraid it won't be small, though. And it might be architecture dependent, so I can't guarantee that you will be able to reproduce it. But Bryan said on IRC yesterday that others have reported similar issues with criterion output, so it may well be cross-platform reproducible. Cheers, Daniel