This is basically a copy of my post from https://groups.google.com/forum/#!topic/clojure/Ozt5HQyM36I on the mailing list. Based on the replies there I'm not sure whether this should be logged as an enhancement or a defect. Please change the designation to whatever is appropriate.
I have found what appears to be a bug in AOT-compiled Clojure when ns-unmap is used; the root cause of this probably impacts other code that redefines Vars as well. I have the following reduced case:
(ns unmap-test.core)
(def a :test-1)
(ns-unmap 'unmap-test.core 'a)
(def a :test-2)
It turns out that a is not resolvable when this namespace is loaded. When I looked at the compiled bytecode,
it appears that the following operations occur:
- A call to RT.var withe 'unmap-test.core and 'a returns a Var, which is bound to a constant.
This var is added to the namespaces's mapping during this call.
- Same as 1.
- The var from 1 is bound to :test-1.
- ns-unmap is called.
- The var from 2 is bound to :test-2.
Disclaimer: This is the first time I have had occasion to look directly at bytecode and I could be missing something.
The basic problem here is that the var is accessible from the load method, but when step 5 executes the var is no longer
accessible from the namespace mappings. Thus, the root of the Var is set to :test-2, but that Var is not mapped from the namespace.
This works when there is no AOT compilation, as well as when I use
(ns unmap-test.core)
(def a :test-1)
(ns-unmap 'unmap-test.core 'a)
(intern 'unmap-test.core 'a :test-2)
I realize that creating defs, unmapping them, and then recreating them is generally poor practice in Clojure.
We have an odd case in that we need to have an interface and a Var of the same name in the same namespace. Simply
doing definterface and then def causes a compilation failure:
user=> (definterface abc)
user.abc
user=> (def abc 1)
CompilerException java.lang.RuntimeException: Expecting var, but abc is mapped to interface user.abc, compiling:(/private/var/folders/3m/tvc28b5d7p50v5_8q5ntj0pmflbdh9/T/form-init4734176956271823921.clj:1:1)
Without going into too much detail, this is basically a hack to allow us to refactor our internal framework code
without immediately changing a very large amount of downstream consumer code. We get rid of the usage of the interface during macroexpansion,
but it still needs to exist on the classpath so it can be imported by downstream namespaces.
There are a number of other ways to accomplish this, so it isn't a particularly big problem for us, but I thought the issue was worth raising.
This was just the first thing I tried and I was surprised when it didn't work.
Note that I used the 1.8.0 RC4 version of Clojure in my sample project, but I had the same behavior on 1.7.0.
Relevant links:
- Bytecode for the load method of the init class: https://gist.github.com/WilliamParker/d8ef4c0555a30135f35a
- Bytecode for the init0 method: https://gist.github.com/WilliamParker/dc606ad086670915efd9
- Decompiled Java code for the init class. Note that this does not completely line up with the bytecode as far as I can tell,
but it is a quicker way to get a general idea of what is happening than the bytecode.
https://gist.github.com/WilliamParker/4cc47939f613d4595d94
- A simple project containing the code above: https://github.com/WilliamParker/unmap-test
Note that if you try it without AOT compilation the target folder with any previously compiled classes should be removed.
