db07f91 Assert that what might look like a possible division by zero is actually unreachable (practicalswift)
Tree-SHA512: f1652eb37196a5b72f356503a1fbb44fb98aa8a94954ad1765f86d81ebf41a2337d4eb58c4f19937fda3752f5d2d642756e44afdbd438015b87ac20801246bff
The initialization order of global data structures in different
implementation units is undefined. Making use of this is essentially
gambling on what the linker does, the so-called [Static initialization
order fiasco](https://isocpp.org/wiki/faq/ctors#static-init-order).
In this case it apparently worked on Linux but failed on OpenBSD and
FreeBSD.
To create it on first use, make the registration structure local to
a function.
Fixes#8910.
Make sure that the count is a zero modulo the new mask before
scaling, otherwise the next time until a measure triggers
will take only 1/2 as long as accounted for. This caused
the 'min time' to be potentially off by as much as 100%.
Previously the benchmark code used an integer division (%) with
a non-constant in the inner-loop. This is quite slow on many
processors, especially ones like ARM that lack a hardware divide.
Even on fairly recent x86_64 like haswell an integer division can
take something like 100 cycles-- making it comparable to the
runtime of siphash.
This change avoids the division by using bitmasking instead. This
was especially easy since the count was only increased by doubling.
This change also restarts the timing when the execution time was
very low this avoids mintimes of zero in cases where one execution
ends up below the timer resolution. It also reduces the impact of
the overhead on the final result.
The formatting of the prints is changed to not use scientific
notation make it more machine readable (in particular, gnuplot
croaks on the non-fixedpoint, and it doesn't sort correctly).
This also hoists out all the floating point divisions out of the
semi-hot path because it was easy to do so.
It might be prudent to break out the critical test into a macro
just to guarantee that it gets inlined. It might also make sense
to just save out the intermediate counts and times and get the
floating point completely out of the timing loop (because e.g.
on hardware without a fast hardware FPU like some ARM it will
still be slow enough to distort the results). I haven't done
either of these in this commit.
Avoid calling gettimeofday every time through the benchmarking loop, by keeping
track of how long each loop takes and doubling the number of iterations done
between time checks when they take less than 1/16'th of the total elapsed time.
Benchmarking framework, loosely based on google's micro-benchmarking
library (https://github.com/google/benchmark)
Wny not use the Google Benchmark framework? Because adding Even More Dependencies
isn't worth it. If we get a dozen or three benchmarks and need nanosecond-accurate
timings of threaded code then switching to the full-blown Google Benchmark library
should be considered.
The benchmark framework is hard-coded to run each benchmark for one wall-clock second,
and then spits out .csv-format timing information to stdout. It is left as an
exercise for later (or maybe never) to add command-line arguments to specify which
benchmark(s) to run, how long to run them for, how to format results, etc etc etc.
Again, see the Google Benchmark framework for where that might end up.
See src/bench/MilliSleep.cpp for a sanity-test benchmark that just benchmarks
'sleep 100 milliseconds.'
To compile and run benchmarks:
cd src; make bench
Sample output:
Benchmark,count,min,max,average
Sleep100ms,10,0.101854,0.105059,0.103881
