<< Back to previous view

[CLJ-1325] Report warnings if *unchecked-math* and boxing happens Created: 16/Jan/14  Updated: 11/Nov/14  Resolved: 29/Aug/14

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

Type: Enhancement Priority: Major
Reporter: Alex Miller Assignee: Alex Miller
Resolution: Completed Votes: 1
Labels: errormsgs, math

Attachments: File boxed.diff     Text File boxedmath.txt     Text File clj-1325.patch     Text File clj-1325-v2.patch     Text File clj-1325-v3.patch    
Patch: Code and Test
Approval: Ok

 Description   

Currently, it is difficult to tell that the compiler is using boxed math unless you look at the generated bytecode. The proposed enhancement here is to emit new warnings if *unchecked-math* is on and boxed math is occurring.

Approach: In the compiler, when compiling a StaticMethodExpr, if *unchecked-math* is true and the class is clojure.lang.Numbers and one of the parameters of static method is of type java.lang.Object or java.lang.Number, then emit a warning at compile-time.

In addition, there is a new WarnBoxedMath Java annotation - a small number of methods on Numbers with Object parameters use this annotation to indicate that warning should not take place. The same annotation can be (but is not currently) used to mark methods on Numbers without Object/Number params that should warn. See boxedmath.txt for a list of methods and categories.

Patch: clj-1325-v3.patch

Screened by:



 Comments   
Comment by Alex Miller [ 14/Apr/14 10:56 PM ]

Moving to 1.7.

Comment by Alex Miller [ 15/Apr/14 10:17 AM ]

List of methods in Numbers and whether they should be considered "boxed math" or not, with some questions.

Comment by Alex Miller [ 14/May/14 2:34 PM ]

Ready for screening.

Comment by Alex Miller [ 16/May/14 11:19 AM ]

clj-1325-v2.patch is identical to last except for a cleaned up the commit message.

Comment by Alex Miller [ 16/May/14 11:51 AM ]

Added v3 patch that just reworks block/indentation style to match surrounding code better.

Comment by Stuart Sierra [ 16/May/14 1:15 PM ]

Screened. Comments:

1) There is no way to get both overflow checks and boxed-math warnings at the same time. Maybe this doesn't matter.

2) The error messages aren't ideal, because they refer to clojure.lang.Numbers, but we can assume that anyone savvy enough to be using *unboxed-math* will also be savvy enough to know what clojure.lang.Numbers is.

3) This doesn't protect me from autoboxing in arbitrary Java method calls, but normal reflection warnings should catch most real-world cases, since few Java APIs overload on primitive and Object.

Comment by Nicola Mometto [ 11/Nov/14 6:02 PM ]

With the new :warn-on-boxed, this code reports a warning:

user=> (defn f [x] (inc x))
Boxed math warning, NO_SOURCE_PATH:2:13 - call: public static java.lang.Number clojure.lang.Numbers.unchecked_in
#'user/f

but this does not:

user=> (defn f1 [^long x] (inc x))
#'user/f1

is this intentional?

Comment by Alex Miller [ 11/Nov/14 9:00 PM ]

The bytecode for those methods is:

f:
  public java.lang.Object invoke(java.lang.Object);
    Code:
       0: aload_1
       1: aconst_null
       2: astore_1
       3: invokestatic  #34                 // Method clojure/lang/Numbers.unchecked_inc:(Ljava/lang/Object;)Ljava/lang/Number;
       6: areturn

f1: 
  public final java.lang.Object invokePrim(long);
    Code:
       0: lload_1
       1: invokestatic  #36                 // Method clojure/lang/Numbers.unchecked_inc:(J)J
       4: invokestatic  #40                 // Method clojure/lang/Numbers.num:(J)Ljava/lang/Number;
       7: areturn

I assume your question is why the call to Numbers.num(long) at the end doesn't cause the warning due to the return type? I had those num() calls in my early list of questionables. This function is tricky because it's called from lots of other methods (many of which already trigger the warning), so it has the potential to cause multiple warnings on a single expression. But this does indeed seem like a common and important case to suggest a return type hint.

Any of these calls that take prims but return a Number or Object require a judgement call and an explicit annotation - there is certainly room for interpretation on some of them.

Adding the return type hint cleans things up pretty well:

public final long invokePrim(long);
    Code:
       0: lload_1
       1: lconst_1
       2: ladd
       3: lreturn
Comment by Alex Miller [ 11/Nov/14 9:06 PM ]

I created CLJ-1585 for this.

Generated at Sat Dec 20 11:27:41 CST 2014 using JIRA 4.4#649-r158309.