<< Back to previous view

[CLJ-1214] Compiler runs out of memory on a small snippet of code Created: 31/May/13  Updated: 18/Apr/14  Resolved: 18/Apr/14

Status: Closed
Project: Clojure
Component/s: None
Affects Version/s: Release 1.5
Fix Version/s: None

Type: Defect Priority: Major
Reporter: Praki Prakash Assignee: Unassigned
Resolution: Declined Votes: 0
Labels: compiler

Linux 3.2.0-39-generic

Attachments: File fubar.clj    


Clojure compiler runs out of memory when loading the attached file. Transcript below.

$ java -cp ~/.m2/repository/org/clojure/clojure/1.5.1/clojure-1.5.1.jar:. clojure.main
Clojure 1.5.1
user=> (load "fubar")
OutOfMemoryError GC overhead limit exceeded  [trace missing]

The file contents are:

  (ns fu.bar)

  (defn foo[l] (concat (drop-last l) (repeat (last l))))

  (def ^:const bar (foo [#(print "") #(println ";")]))


If I remove the metadata on bar, it works. Removal of namespace seems to fix it as well. Pretty strange.

Although I realize this code is not quite kosher, it would be nice to have the compiler deal with it.

Comment by Andy Fingerhut [ 01/Jun/13 7:40 PM ]

If you really do have 'bar' on a line by itself at the end of file fubar.clj, it seems you are asking it to evaluate the value of bar, which by the definitions is an infinite list, and will of course exhaust all available memory if you attempt to evaluate it.

It seems to me the more odd thing is not that it runs out of memory as shown, but that it does not run out of memory when you remove the metadata on bar.

What is the purpose of having 'bar' on a line by itself at the end of the file?

If I try this but remove 'bar' as the last line of the file, loading the file causes no errors regardless of whether there is metadata on bar's definition or not. It is strange that doing (load "fubar") followed by (first fu.bar/bar) seems to go into an infinite loop if the :const is there on bar, but quickly returns the correct answer if the metadata is removed.

Comment by Praki Prakash [ 01/Jun/13 8:40 PM ]

This code snippet is a minimal test case to show that the compiler runs out of memory. What I meant by "it works" was that the compiler doesn't run out of memory and successfully loads the file (or in my real code base, the namespace is compiled).

In my code, I use bar (or whatever the real thing is) as source of sequence of functions. The sole reference to bar is needed to trigger this issue. I believe that bar is not being fully evaluated here and thus no infinite loop. If I try to print it, yes, it will ultimately fail.

Comment by Praki Prakash [ 01/Jun/13 9:04 PM ]

Having thought about this a bit more, it seems to me that when bar is annotated with const, the compiler is trying to evaluate the associated expression which exhausts the heap? I have never looked at the compiler source and thus not sure if this is what is happening. If it is, then it seems like one should be really careful when adding metadata.

That still leaves the other question about the namespace requirement to cause memory exhaustion. I quite distinctly recall having to add the namespace when trying to come up with minimal test case to reproduce the bug.

If you think this is really user error, I would accept it

Comment by Andy Fingerhut [ 02/Jun/13 4:56 AM ]

It is not just any old metadata that causes this issue, only :const metadata. I tried testing with :const replaced with :dynamic and :private, and there was no problem.

This might shed some light on the issue: https://github.com/clojure/clojure/blob/master/changes.md#215-const-defs

It appears that ^:const is causing the compiler to evaluate the value at compile time. The value in your example is unbounded, so that can never complete.

Comment by Kevin Downey [ 17/Apr/14 10:39 PM ]

the problem is for complex objects :const results in the pr-str of the object being embedded in the bytecode, a pr-str of an infinite seq is going to blow the heap. I think the limits on the usage of :const are not well defined and instead of falling back on pr-str it should throw a compilation error for constants that are not jvm primitves or strings

Comment by Alex Miller [ 17/Apr/14 10:46 PM ]

There is support for handling a variety of other useful cases (core Clojure collections for example) as constant objects. And it is useful to allow code that is not constant but evaluates to a constant result (for creating data tables, etc). In the face of that, I'm not sure it's possible to bound this usefully without solving the halting problem. I'm tempted to just put this issue in the category of "don't do that".

Comment by Alex Miller [ 18/Apr/14 7:08 AM ]

After sleeping on it, I don't think this is worth working on (if there is anything that could even be done).

Generated at Sun Jan 21 05:10:41 CST 2018 using JIRA 4.4#649-r158309.