<< Back to previous view

[CLJ-1546] Widen vec to take Iterable/IReduce Created: 02/Oct/14  Updated: 11/Jan/15  Resolved: 10/Jan/15

Status: Closed
Project: Clojure
Component/s: None
Affects Version/s: Release 1.7
Fix Version/s: Release 1.7

Type: Enhancement Priority: Critical
Reporter: Alex Miller Assignee: Unassigned
Resolution: Completed Votes: 0
Labels: None

Attachments: Text File clj-1546-2.patch     Text File clj-1546-3.patch     Text File clj-1546-4.patch     Text File clj-1546-5.patch     Text File clj-1546-6.patch     Text File clj-1546.patch     PNG File numbers.png    
Patch: Code and Test
Approval: Ok


These examples should work but do not:

Something Iterable but not IReduce:

user> (def i (eduction (map inc) (range 100)))
user> (instance? java.util.Collection i)
user> (instance? Iterable i)
user> (vec i)
RuntimeException Unable to convert: class clojure.core.Iteration to Object[]

Something IReduceInit but not Iterable:

user=> (vec
  (reify clojure.lang.IReduceInit
    (reduce [_ f start]
      (reduce f start (range 10)))))
RuntimeException Unable to convert: class user$reify__15 to Object[]

Proposal: Add PersistentVector.create(Iterable) and PersistentVector.create(IReduceInit) to efficiently create PVs from those. See also the blog post http://insideclojure.org/2015/01/07/vec-perf/.

For performance, vec has several cases:
1) (vec) if vector?: return new vector w/o meta - this matches prior behavior but has a constant cost of a few ns, rather than linear cost. If not a vector, spill to LazilyPersistentVector.create(Object).

2) (LPV) instanceof IReduceInit: Anything reducible can reduce itself fastest. Right now this has a big benefit for PersistentList. on 1.7.0-alpha4 with list of size 1024, into=28 seconds, vec=18 seconds. After patch, vec=7 seconds. As more things become IReduce, they'll take this path as well. This is also the branch that handles the new Eduction and IReduceInit cases.

3) (LPV) instanceof ISeq: If the coll is a sequence already, best to just walk it rather than build an iterator or array from it. This calls into PersistentVector.create(ISeq). That implementation now contains an optimization to build into an array and construct the PersistentVector directly from the array for sequences <= 32 elements (which is common). Once that threshold is reached, it switches to building with transients. The benchmark shows that the patch makes vec substantially faster for all seqs and even faster than into in some cases.

4) (LPV) instanceof Iterable: For all non-Clojure collections (ArrayList) and current non-IReduce Clojure collections (PHM, PHS), this is fastest path. Iterators are preferred to seqs as they do not cache or hold onto the values as they go by. The PV.create() for Iterable uses transients. Due to slightly more overhead, small maps and sets are slightly slower but this is largely fixed by CLJ-1499 which adds direct iterators. ArrayLists with <= 32 are special-cased - we can toArray() and construct the PV with a seeded node in this case. This type and size is particularly common in real code. Even so, very small ArrayLists are a bit slower than they were due to increased number of conditional checks I think.

5) (LPV) otherwise RT.toArray(): catches Map, String, Object[], primitive array, etc. The important ones here are the arrays - they are slightly slower on small arrays due to overhead of checking more cases above, but big arrays are significantly faster than they were.

In addition, there was one hard-coded path in the Compiler into PersistentVector.create() and I re-routed that through LazilyPersistentVector instead as that code is now the place to choose the fastest path logic.

Patch: clj-1546-6.patch, see numbers.png for perf comparison

Screened by: Stu

Comment by Timothy Baldridge [ 02/Oct/14 9:44 AM ]

Is there a reason the final case for (vec something) can't just be a call to (into [] coll)? It seems a bit odd to do (to-array) on anything thats not a java collection or Iterable, when we have IReduce.

Comment by Rich Hickey [ 02/Oct/14 10:02 AM ]

re: Tim - yes, this needs to support IReduce (and thereby educe) as well

Comment by Alex Miller [ 14/Oct/14 9:56 AM ]

Added new patch that handles Iterable and IReduceInit in vec. It also makes calling with a vector much faster due to the first check. into is still faster for chunked seqs (due to special InternalReduce handling of chunking).

It would be possible to move more of the variant checking into LazilyPersistentVector or PersistentVector so it could be used in more contexts. I'm not sure how much to do with that.

It would also be possible to instead lean on reduce more from the Java side if there was a Java version of reduce (as defined in mikera's branch for http://dev.clojure.org/jira/browse/CLJ-1192 at https://github.com/mikera/clojure/compare/clj-1192-vec-performance. Something like that is the only way I can see of leveraging that same InternalReduce logic that makes into faster than vec.

Comment by Alex Miller [ 13/Nov/14 4:14 PM ]

Prior comments from Stu removed from description: "Open Question: Which branch should come first, Collection or IReduceInit? Collection reaches the fast path for small collections through LazilyPersistentVector, but IReduceInit should be faster for larger things. Related: Shouldn't the item count in LazilyPersistentVector be a bounded count?"

I have attached a new patch that simplifies the impl to do it in LazilyPersistentVector instead of in vec, which was easier due to "and" not being able yet when vec is implemented to do the length check.

I have also done a considerable amount of analysis on the matrix of incoming collections and best path to follow and also collected some data on what collections are commonly passed into vec. The current patch reflects those findings. Some highlights:

  • vec is called with PersistentVector in all projects I tested. The instanceof check takes that case from typically 100s of nanos to ~5 ns. So I do think it is worth doing.
  • vec is overwhelmingly called with small collections - in most cases the incoming collection is <10 elements. In cases where the collection is not a sequence, the path of creating the Vector with an owning array is the fastest option, beating even IReduce and transient building (as that path has some checks involved).
  • PersistentList is the only IReduce likely to be encountered by vec right now and adding that branch is a significant performance boost from prior impl and vs into. If maps and sets were IReduce, they would gain this as well.
  • chunked seqs will be significantly faster with into than vec as into goes through CollReduce and can leverage many optimizations on reducing through chunks that are not available to vec.
  • seqs in general though are now faster with vec than they were due to leveraging transients.
  • eduction results support IReduce and are also faster with vec than into.
  • range is currently slower with vec, but when range is IReduce, it will probably be faster with vec

In summary, some new conventional wisdom (after this patch) on (into []) vs vec:

  • vec is faster if passed a vector, an IReduce, or an array
  • into is faster when working with seqs, but even vec is better than it used to be and may even be faster for things like range in the future
Comment by Michael Blume [ 13/Nov/14 7:24 PM ]

Latest patch won't build for me when applied to master

     [java] Exception in thread "main" java.lang.ExceptionInInitializerError
     [java] 	at clojure.lang.Compile.<clinit>(Compile.java:29)
     [java] Caused by: java.lang.NoSuchMethodError: clojure.lang.LazilyPersistentVector.create(Ljava/util/Collection;)Lclojure/lang/IPersistentVector;, compiling:(clojure/core.clj:14:23)
     [java] 	at clojure.lang.Compiler.load(Compiler.java:7206)
     [java] 	at clojure.lang.RT.loadResourceScript(RT.java:370)
     [java] 	at clojure.lang.RT.loadResourceScript(RT.java:361)
     [java] 	at clojure.lang.RT.load(RT.java:440)
     [java] 	at clojure.lang.RT.load(RT.java:411)
     [java] 	at clojure.lang.RT.doInit(RT.java:448)
     [java] 	at clojure.lang.RT.<clinit>(RT.java:329)
     [java] 	... 1 more
     [java] Caused by: java.lang.NoSuchMethodError: clojure.lang.LazilyPersistentVector.create(Ljava/util/Collection;)Lclojure/lang/IPersistentVector;
     [java] 	at clojure.lang.LispReader$VectorReader.invoke(LispReader.java:1073)
     [java] 	at clojure.lang.LispReader.readDelimitedList(LispReader.java:1138)
     [java] 	at clojure.lang.LispReader$ListReader.invoke(LispReader.java:972)
     [java] 	at clojure.lang.LispReader.read(LispReader.java:183)
     [java] 	at clojure.lang.LispReader$WrappingReader.invoke(LispReader.java:535)
     [java] 	at clojure.lang.LispReader.readDelimitedList(LispReader.java:1138)
     [java] 	at clojure.lang.LispReader$MapReader.invoke(LispReader.java:1081)
     [java] 	at clojure.lang.LispReader.read(LispReader.java:183)
     [java] 	at clojure.lang.LispReader$MetaReader.invoke(LispReader.java:716)
     [java] 	at clojure.lang.LispReader.readDelimitedList(LispReader.java:1138)
     [java] 	at clojure.lang.LispReader$ListReader.invoke(LispReader.java:972)
     [java] 	at clojure.lang.LispReader.read(LispReader.java:183)
     [java] 	at clojure.lang.Compiler.load(Compiler.java:7190)
     [java] 	... 7 more
Comment by Alex Miller [ 13/Nov/14 7:28 PM ]

Did you clean first? I replaced that static method call there with a wider version but if you are cleaning fresh it should be fine.

Comment by Michael Blume [ 13/Nov/14 7:31 PM ]

Apologies, maven just wasn't doing a good job of tracking changes, running mvn clean fixes the build.

Comment by Alex Miller [ 25/Nov/14 9:58 AM ]

Added benchmark.png showing times (in ns), tested with criterium, for into and vec on different types and sizes on 1.7.0-alpha4 and then vec again after the patch.

Comment by Jonas De Vuyst [ 11/Jan/15 6:47 AM ]

This patch breaks (vec (first {1 2}))
; ClassCastException clojure.lang.MapEntry cannot be cast to clojure.lang.IObj

Comment by Alex Miller [ 11/Jan/15 8:12 AM ]

Thanks for the report! I will get that fixed in the next alpha.

Generated at Thu Feb 22 03:11:05 CST 2018 using JIRA 4.4#649-r158309.