<< Back to previous view

[CLJ-1232] Functions with non-qualified return type hints force import of hinted classes when called from other namespace Created: 18/Jul/13  Updated: 03/Sep/15  Resolved: 03/Sep/15

Status: Closed
Project: Clojure
Component/s: None
Affects Version/s: Release 1.5, Release 1.6, Release 1.7
Fix Version/s: Release 1.8

Type: Defect Priority: Major
Reporter: Tassilo Horn Assignee: Unassigned
Resolution: Completed Votes: 8
Labels: compiler, typehints

Attachments: Text File 0001-auto-qualify-arglists-class-names.patch     Text File 0001-auto-qualify-arglists-class-names-v2.patch     Text File 0001-auto-qualify-arglists-class-names-v3.patch     Text File 0001-auto-qualify-arglists-class-names-v4.patch     Text File 0001-CLJ-1232-auto-qualify-arglists-class-names.patch     Text File 0001-throw-on-non-qualified-class-names-that-are-not-auto.patch    
Patch: Code and Test
Approval: Ok


You can add a type hint to function arglists to indicate the return type of a function like so.

user> (import '(java.util List))
user> (defn linkedlist ^List [] (java.util.LinkedList.))
user> (.size (linkedlist))

The problem is that now when I call `linkedlist` exactly as above from another namespace, I'll get an exception because java.util.List is not imported in there.

user> (in-ns 'user2)
#<Namespace user2>
user2> (refer 'user)
user2> (.size (linkedlist))
CompilerException java.lang.IllegalArgumentException: Unable to resolve classname: List, compiling:(NO_SOURCE_PATH:1:1)
user2> (import '(java.util List)) ;; Too bad, need to import List here, too.
user2> (.size (linkedlist))

There are two workarounds: You can import the hinted type also in the calling namespace, or you always use fully qualified class names for return type hints. Clearly, the latter is preferable.

Approach: Resolve the tags in the defn macro.

Patch: 0001-CLJ-1232-auto-qualify-arglists-class-names.patch

Screened by: Fogus

Alternate approach: Make the compiler resolve the return tags when necessary (tag is not a string, primitive tag (^long) or array tag (^longs)) and update the Var's :arglist appropriately. Patch: 0001-auto-qualify-arglists-class-names-v4.patch. Note that this patch had problems with Hystrix, which was using non-conformant arglists - Hystrix has since been patched.

Comment by Andy Fingerhut [ 16/Apr/14 3:47 PM ]

To make sure I understand, Nicola, in this ticket you are asking that the Clojure compiler change behavior so that the sample code works correctly with no exceptions, the same way as it would work correctly without exceptions if one of the workarounds were used?

Comment by Tassilo Horn [ 17/Apr/14 12:18 AM ]

Hi Andy. Tassilo here, not Nicola. But yes, the example should work as-is. When I'm allowed to use type hints with simple imported class names for arguments, then doing so for return values should work, too.

Comment by Rich Hickey [ 10/Jun/14 10:41 AM ]

Type hints on function params are only consumed by the function definition, i.e. in the same module as the import/alias. Type hints on returns are just metadata, they don't get 'compiled' and if the metadata is not useful to consumers in other namespaces, it's not a useful hint. So, if it's not a type in the auto-imported set (java.lang), it should be fully qualified.

Comment by Alex Miller [ 10/Jun/14 11:55 AM ]

Based on Rich's comment, this ticket should probably morph into an enhancement request on documentation, probably on http://clojure.org/java_interop#Java Interop-Type Hints.

Comment by Andy Fingerhut [ 10/Jun/14 3:13 PM ]

I would suggest something like the following for a documentation change, after this part of the text on the page Alex links in the previous comment:

For function return values, the type hint can be placed before the arguments vector:

(defn hinted
(^String [])
(^Integer [a])
(^java.util.List [a & args]))

-> #user/hinted

If the return value type hint is for a class that is outside of java.lang, which is the part auto-imported by Clojure, then it must be a fully qualified class name, e.g. java.util.List, not List.

Comment by Nicola Mometto [ 10/Jun/14 4:02 PM ]

I don't understand why we should enforce this complexity to the user.
Why can't we just make the Compiler (or even defn itself) update all the arglists tags with properly resolved ones? (that's what I'm doing in tools.analyzer.jvm)

Comment by Alexander Kiel [ 19/Jul/14 10:02 AM ]

I'm with Nicola here. I also think that defn should resolve the type hint according the imports of the namespace defn is used in.

Comment by Max Penet [ 22/Jul/14 7:06 AM ]

Same here, I was bit by this in the past. The current behavior is clearly counterintuitive.

Comment by Nicola Mometto [ 28/Aug/14 12:58 PM ]

Attached two patches implementing two different solutions:

  • 0001-auto-qualify-arglists-class-names.patch makes the compiler automatically qualify all the tags in the :arglists
  • 0001-throw-on-non-qualified-class-names-that-are-not-auto.patch makes the compiler throw an exception for all public defs whose return tag is a symbol representing a non-qualified class that is not in the auto-import list (approach proposed in IRC by Alex Miller)
Comment by Tassilo Horn [ 29/Aug/14 1:49 AM ]

For what it's worth, I'd prefer the first patch because the second doesn't help in situations where the caller lives in a namespace where the called function's return type hinted class is `ns-unmap`-ed. And there a good reasons for doing that. For example, Process is a java.lang class and Process is a pretty generic name. So in some namespace, I want to define my own Process deftype or defrecord. Without unmapping 'Process first to get rid of the java.lang.Process auto-import, I'd get an exception:

user> (deftype Process [])
IllegalStateException Process already refers to: class java.lang.Process in namespace: user  clojure.lang.Namespace.referenceClass (Namespace.java:140)

Now when I call some function from some library that has a `^Process` return type hint (meaning java.lang.Process there), I get the same exception as in my original report.

I can even get into troubles when only using standard Clojure functions because those have `^String` and `^Class` type hints. IMO, Class is also a pretty generic name I should be able to name my custom deftype/defrecord. And I might also want to have a custom String type/record in my astrophysics system.

Comment by Andy Fingerhut [ 30/Sep/14 4:39 PM ]

Not sure whether the root cause of this behavior is the same as the example in the description or not, but seems a little weird that even for fully qualified Java class names hinting the arg vector, it makes a difference whether it is done with defn or def:

Clojure 1.6.0
user=> (set! *warn-on-reflection* true)
user=> (defn f1 ^java.util.LinkedList [coll] (java.util.LinkedList. coll))
user=> (def f2 (fn ^java.util.LinkedList [coll] (java.util.LinkedList. coll)))
user=> (.size (f1 [2 3 4]))
user=> (.size (f2 [2 3 4]))
Reflection warning, NO_SOURCE_PATH:5:1 - reference to field size can't be resolved.
Comment by Alex Miller [ 30/Sep/14 6:21 PM ]

Andy, can you file that as a separate ticket?

Comment by Andy Fingerhut [ 30/Sep/14 9:08 PM ]

Created ticket CLJ-1543 for the issue raised in my comment earlier on 30 Sep 2014.

Comment by Andy Fingerhut [ 01/Oct/14 12:38 PM ]

Tassilo (or anyone), is there a reason to prefer putting the tag on the argument vector in your example? It seems that putting it on the Var name instead avoids this issue:

user=> (clojure-version)
user=> (set! *warn-on-reflection* true)
user=> (import '(java.util List))
user=> (defn ^List linkedlist [] (java.util.LinkedList.))
user=> (.size (linkedlist))
user=> (ns user2)
user2=> (refer 'user)
user2=> (.size (linkedlist))

I suppose that only allows a single type tag, rather than an independent one for each arity.

Comment by Tassilo Horn [ 02/Oct/14 3:16 AM ]

I wasn't aware of the fact that you can put it on the var's name. That's not documented at http://clojure.org/java_interop#Java Interop-Type Hints. But IMHO the documented version with putting the tag on the argument vector is more general since it supports different return type hints for the different arity version. In any case, if both forms are permitted then they should be equivalent in the case the function has only one arity.

Comment by Rich Hickey [ 16/Mar/15 12:02 PM ]

Please work on the simplest patch that resolves the names

Comment by Alex Miller [ 16/Mar/15 4:34 PM ]

Nicola, in this:

if (tag != null &&
                        !(tag instanceof String) &&
                        primClass((Symbol)tag) == null &&
                        !tagClass((Symbol) tag).getName().startsWith("["))
                            argvec = (IPersistentVector)((IObj)argvec).withMeta(RT.map(RT.TAG_KEY, Symbol.intern(tagClass((Symbol)tag).getName())));

doesn't tagClass already handle most of these cases properly already? Can this be simplified? Is there an optimization case in avoiding lookup for a dotted name?

Comment by Nicola Mometto [ 16/Mar/15 5:10 PM ]

Patch 0001-auto-qualify-arglists-class-names-v2.patch avoids doing unnecessary lookups (dotted names, special tags (primitive tags, array tags)) and adds a testcase

Comment by Michael Blume [ 20/May/15 1:13 PM ]

I'm seeing an odd failure with this patch and hystrix defcommands, will post a small reproduction shortly

Comment by Michael Blume [ 20/May/15 1:20 PM ]


passes lein check with 1.7 beta3, fails with v3 patch applied

Comment by Nicola Mometto [ 20/May/15 1:40 PM ]

During analysis the compiler understands only arglists in the form of (quote ([..]*)) (see https://github.com/clojure/clojure/blob/master/src/jvm/clojure/lang/Compiler.java#L557-L558), hystrix emits arglists in the form of (list (quote [..])*).

Not sure what to do about this.

Comment by Michael Blume [ 20/May/15 1:51 PM ]

Possibly just ask Hystrix not to do that?

Comment by Alex Miller [ 20/May/15 2:30 PM ]

test.generative uses non-standard arglists too. I haven't looked at the patch, but if it's sensitive to that, it's probably not good enough.

Comment by Nicola Mometto [ 20/May/15 6:51 PM ]

test.generative uses non-standard :tag, not :arglists

Comment by Alex Miller [ 20/May/15 10:22 PM ]

ah, yes. sorry.

Comment by Stuart Halloway [ 10/Jul/15 12:15 PM ]

Opened https://github.com/Netflix/Hystrix/issues/831 to see what is up with hystrix.

Comment by Stuart Halloway [ 31/Jul/15 9:16 AM ]

Can somebody with context update this patch to apply to latest master?

Comment by Nicola Mometto [ 31/Jul/15 9:37 AM ]

patch updated

Comment by Matt Jacobs [ 06/Aug/15 12:59 PM ]

I've released hystrix-clj 1.4.14, with a fix for https://github.com/Netflix/Hystrix/issues/831: https://github.com/Netflix/Hystrix/releases/tag/v1.4.14. If anything else is needed on the Hystrix side, please open another issue.

Comment by Fogus [ 07/Aug/15 11:03 AM ]

Applied the 0001-CLJ-1232-auto-qualify-arglists-class-names.patch to master (1.8) and ran through the example scenario to check the veracity of the change. Additionally, I ran a modified snippet using def and verified that it too worked. Finally, after checking the code it seems reasonable in implementation and scope.

Comment by Rich Hickey [ 08/Aug/15 9:56 AM ]

I'd rather not change the compiler and it seems hystrix was broken. Also please make it clear what the single strategy and patch is at the top of the ticket.

Comment by Alex Miller [ 11/Aug/15 5:01 PM ]

Made preferred approach clear in the description and moved back to vetted. I believe the other ticket was screened by Fogus so this needs to be re-screened with the preferred approach ticket.

Comment by Alex Miller [ 11/Aug/15 5:08 PM ]

Nicola and Fogus, would appreciate an eye as to whether I just made the proper alterations in the description.

Comment by Nicola Mometto [ 12/Aug/15 6:17 AM ]

Alex Miller Reading Fogus' last comment, it seems to me that he actually screened the current patch (the one who resolves tags in defn rather than in the compiler) "Applied the 0001-CLJ-1232-auto-qualify-arglists-class-names.patch to master [..]"

Comment by Fogus [ 14/Aug/15 9:18 AM ]

It would be better if the description said the word "preferred," but based on my reading resolving the tags in defn is the winner (0001-CLJ-1232-auto-qualify-arglists-class-names.patch). Thankfully that's the one that I screened as I agree that non-trivial changes to the compiler are to be avoided. That said, of course that patch does modify the compiler but since it's a change that pulls out some code into a useful method I let it through. To my mind that was a trivial (and justifiable change). Of course, if we want to avoid any compiler changes then this patch would have to be reworked.

Comment by Alex Miller [ 18/Aug/15 5:03 PM ]

Moving back to screened as fogus screened the preferred (by Rich) version.

Generated at Wed Oct 07 05:10:39 CDT 2015 using JIRA 4.4#649-r158309.