Implement ListBuffer.isEmpty / nonEmpty / head efficiently #5802
Conversation
| // Override with efficient implementations using the extra size information available to ListBuffer. | ||
| override def isEmpty: Boolean = len == 0 | ||
| override def nonEmpty: Boolean = len > 0 | ||
| override def head: A = if (len > 0) start.asInstanceOf[::[A]].head else Nil.head |
There was a problem hiding this comment.
This one doesn't look very cute.. start.head would be a virtual call with two cases, so C2 should be able to inline it. But before C2 it's probably slower than the monomorphic calls.
There was a problem hiding this comment.
Yes, you are probably right about head. I removed that one.
Uses the extra length information to provide more efficient implementations. Evaluating these methods turns up with about 5-6% of akka-http message parsing.
43f194d to
40f0514
Compare
|
I basically just saw these turning up in play / akka-http benchmarks. It seems one reason that I saw For the |
|
I searched a bit before but couldn't find it |
|
Was it this one? It's a JDK patch by Jason, but it's about forwarders not https://gist.github.com/retronym/d27090a68570485c3e329a52db0c7b25 |
|
Interesting information. FWIW, in this case, methods from |
|
My JDK patch was intended to make the extra indirection of the trait encoding "invisible" to the inliner heuristics. I've been meaning to get back to it, but I haven't been able to get my dev environment setup for OpenJDK builds recently :( |
|
I'm seeing a 6% improvement to the performance of our compiler benchmark suite with Related discussion on SO: Why does the JVM has a max inline depth? It will be interesting to see how much of that improvement I could get under the default limit with my JDK patch applied. |
|
@lrytz - Yes, this patch looks sensible to me too. |
Uses the extra length information in comparison to list to provide efficient implementations.
Evaluating these three methods turns up with about 6-7% of akka-http message parsing.
From the profiles it looks the likely reason are polymorphic invokeinterface calls against List (i.e. just forwarding to
start.isEmptyetc wouldn't help).