wanted: DEFINE-THREAD-LOCAL-VARIABLE

[Nikodemus Siivola]

A way to initialize thread local variables.

Discussed a number of times on sbcl-devel.

Collect links to those discussions /here/.

This thread on sbcl-devel starts with a proposed patch and ends with another:

http://thread.gmane.org/gmane.lisp.steel-bank.devel/7351/focus=8794

Dealing with already running threads is not implemented.

[Tobias C. Rittweiler (tcr)]

Another way would be to introduce a special form THREAD-LOCAL-VALUE that is
inspired by LOAD-TIME-VALUE.

    (THREAD-LOCAL-VALUE expr) reserves space in thread-local memory, initialized
                  by the result of evaluating EXPR at load-time (or at execution time, but
                  guaranteed to be evaluated only once.)

This way DEFINE-THREAD-LOCAL-VARIABLE could be implemented as

  (DEFGLOBAL FOO (THREAD-LOCAL-VALUE expr))

or similiarly. The advantage is that this would also allow C-style static variables:

  (defun quux (x)
     (let ((count (thread-local-value 0)))
       (incf count)
       (list :x :count count)))

[Paul Khuong (pvk)]

I already have code hanging around to set the value of a special binding, adding a thread-local one if necessary. THREAD-LOCAL-VALUE could easily be implemented on top of that. See <http://paste.lisp.org/display/63257>.

[Tobias C. Rittweiler (tcr)]

I was about to say that the THREAD-LOCAL-VALUE special form may be
a succinct way to express what currently *zap-array-data-temp* and
*merge-sort-temp-vector* are for. However, there also come interrupts
into play -- and THREAD-LOCAL-VALUE would enable useres to write
non-reentrant functions very easily, wouldn't it?

I was thinking of THREAD-LOCAL-VALUE as "static variables but done
right" -- I guess that belief was mistakenly naive?

[Gábor Melis (melisgl)]

I thought of what might be a better way to deal with the thread local problem. Implementation is here:
https://github.com/melisgl/sbcl/commits/thread-local

There is no way to inherit from the parent thread, but it handles existing threads gracefully. See the documentation diff for a tutorialish treatment.