Constants are leaked in case of a reentrant eval

Question

Compiling a function that references a non loaded (or uninitialized) class triggers its init static. When the init static loads clojure code, some constants (source code I think) are leaked into the constants pool of the function under compilation.

It prevented CCW from working in some environments (Rational) because the static init of the resulting function was over 64K.

Steps to reproduce:

Load the leak.main ns and run the code in comments: the first function has 15 extra fields despite being identical to the second one.

`
(ns leak.main)

(defn first-to-load []
leak.Klass/foo)

(defn second-to-load []
leak.Klass/foo)

(comment
=> (map (comp count #(.getFields %) class) [first-to-load second-to-load])
(16 1)
)
`

`
package leak;

import clojure.lang.IFn;
import clojure.lang.RT;
import clojure.lang.Symbol;

public class Klass {
static {

RT.var("clojure.core", "require").invoke(Symbol.intern("leak.leaky"));

}
public static IFn foo = RT.var("leak.leaky", "foo");
}
`

`
(ns leak.leaky)

(defn foo
"Some doc"
[]
"hello")

(def unrelated 42)
`

https://gist.github.com/cgrand/5dcb6fe5b269aecc6a5b#file-main-clj-L10

Patch: clj-1620-v5.patch

Answer 1

Comment made by: alexmiller

Rich has ok'ed screening this one for 1.7 but I do not feel that I can mark it screened without understanding it much better than I do. The description, code, and cause information here is not sufficient for me to understand what the problem actually is or why the fix is the right one. The fix seems to address the symptom but I worry that it is just a symptom and that a better understanding of the actual cause would lead to a different or better fix.

The evolution of the patches was driven by bugs in CLJ-1544 (a patch which has been pulled out for being suspect for other reasons). Starting fresh, were those modifications necessary and correct?

Why does this set of vars need to push clean impls into the bindings? Why not some of the other vars (like those pushed in load())? The set chosen here seems to match that from the ReifyParser - why? Why should they only be pushed if they are bound (that is, why is "not bound" not the same as "bound but empty")? Are we affecting performance?

Popping all the way out, is the thing being done by CCW even a thing that should be doable? The description says "Compiling a function that references a non loaded (or uninitialized) class triggers its init static" - should this load even happen? Can we get an example that actually demonstrates what CCW was doing originally?

Answer 2

Comment made by: laurentpetit

Alex, the question of "should what CCW is doing be doable" can be answered if you answer it on the given example, I think.

The question "should the initialization of the class occur when it could just be loaded" is a good one. Several reports have been made on the Clojure list about this problem, and I guess there is at least one CLJ issue about changing some more classForName into classForNameNonLoading here and there in Clojure.
For instance, it prevents referencing java classes which have code in their static initializers as soon as the code does some supposition about the runtime it is initialized in. This is a problem with Eclipse / SWT, this a problem with Cursive as I remember Colin mentioning a similar issue. And will probably is a problem that can appear each time one tries to AOT compile clojure code interoperating with java classes who happen to have, somewhere within static initializers triggered by the compilation (and this is transitive), assumptions that they are initialized in the proper target runtime environment.

What I don't know is if preventing the initialization to occur in the first place would be sufficient to get rid of the class of problems this bug and the proposed patch tried to solve. I do not claim to totally what is happening either (Christophe and Nicolas were of great help to analyze the issue and create the patch), but as I understand it, it's a kind of "Inception-the-movie-like" bug. Compiling a fn which triggers compiling another fn (here through the loading of clojure namespaces via a java initializer).

If preventing the initialization of class static methods when they are referenced (through interop calls - constructor, field, method, static field, static method-) is the last remaining bit that could cause such "compilation during compilation" scenario, then yes, protecting the compilation process like Nicolas tried to do may not be necessary, and just fixing the undesired loading may be enough.

Answer 3

Reference: https://clojure.atlassian.net/browse/CLJ-1620 (reported by cgrand)