#Imp vs Implicitly
I've made this repo in order to benchmark https://github.com/non/imp
that claims to implement a faster
I'm trying to check it's particular claim that
implicitly[ClassTag[Int]] is slower than
##Benchmarking framework used I do it by using the best benchmarking tool available for JVM: http://openjdk.java.net/projects/code-tools/jmh/
This is the state-of-the art tool avaliable for JVM. In this [thread] (https://groups.google.com/forum/#!topic/mechanical-sympathy/m4opvy4xq3U) you can see many problems that make most of benchmarking frameworks unreliable or simply wrong for that kind of benchmark.
I've checked. And what I can say is that on Oracle HotSpot version "1.8.0_72"(x86_64 ubuntu i7-4770) it's not. Actually, both of them get optimized by JVM to be a simple load of same constant.
[info] Benchmark Mode Cnt Score Error Units [info] ImplVsImplicitly.baseline avgt 30 2.507 ± 0.031 ns/op [info] ImplVsImplicitly.baseline:·asm avgt NaN --- [info] ImplVsImplicitly.measure avgt 30 2.491 ± 0.044 ns/op [info] ImplVsImplicitly.measure:·asm avgt NaN ---
Run it yourself
You can re-run those benchmarks by executing in sbt promt:
jmh:run -i 10 -wi 10 -f3 -t 1
Runtime low-level details
I've had a look on assembly generated for both methods. The generated assembly was the same. Hotspot was able to optimize both of them to a single read of the very same constant object.
baseline.assembly.txt contains the assembly for the
implicitly[ClassTag[Int]] where you can see that hotspot was able to get rid of all the overheads.
Most of those 2.5 ns are actually taken by benchmark framework itself. So, to be entirelly honest, the numbers that I showed earlier do not show anything. As those are simply benchmarking overheads of the jmh itself. What is a good indicator is the assembly printout iself: it shows that hotspot was able to remove static loads and virtual calls in this example.
In order to reproduce this output you would need to run on Linux with
-prof perfasm. You would also need a jvm that has dissassembler plugin installed. Setting it up is quite involved, and you'll likely need to build it from source of your version of JDK. Google for
hsdis for more details.
There is a small difference in the size of generated method:
0: getstatic #13 // Field scala/Predef$.MODULE$:Lscala/Predef$; 3: getstatic #18 // Field scala/reflect/ClassTag$.MODULE$:Lscala/reflect/ClassTag$; 6: invokevirtual #21 // Method scala/reflect/ClassTag$.Int:()Lscala/reflect/ClassTag; 9: invokevirtual #25 // Method scala/Predef$.implicitly:(Ljava/lang/Object;)Ljava/lang/Object; 12: checkcast #27 // class scala/reflect/ClassTag
0: getstatic #13 // Field scala/reflect/ClassTag$.MODULE$:Lscala/reflect/ClassTag$; 3: invokevirtual #16 // Method scala/reflect/ClassTag$.Int:()Lscala/reflect/ClassTag; 6: areturn
This may lead to methods containing many calls to
implicitly being less likely to be inlined, but it would argue this is very uncommon.
I'd be very glad to see a PR that would prove me wrong. I'll learn something this way :-).
@lloydmeta has contributed results from OSX (El Capitan), Oracle Java8 JDK 1.8.0_05:
[info] Benchmark Mode Cnt Score Error Units [info] ImplVsImplicitly.baseline avgt 30 3.268 ± 0.025 ns/op [info] ImplVsImplicitly.explicit avgt 30 3.292 ± 0.036 ns/op [info] ImplVsImplicitly.imply avgt 30 3.291 ± 0.039 ns/op
He's running a bit more of code by computing
1 + 1 in a generic way. I didin't get assembly printout for his modification, but I assume that HotSpot was able to constant-fold the whole expression in a similar way.