improper insistence on fixnum type

[xrme]

From #ccl on freenode:

(defun bug ()
  (let ((indentation 0))	
    (macrolet ((with-indent (inc &body body)
		 `(prog2 (incf indentation ,inc)
		      (progn ,@body)
		    (decf indentation ,inc))))
      (labels ((foo () (1+ indentation)))
	(with-indent 2.5 "")))))

Expected result is "". Actual result is an error: "The value 2.5 is not of the expected type FIXNUM".

The labels form seems to be implicated somehow.

---

Want to back this issue? Post a bounty on it! We accept bounties via Bountysource.

[informatimago]

I don't reproduce it on "Version 1.12-dev (v1.12-dev.3-27-g7a42fe0c) Darwinx8664" either with compile-file or load, evaluating the defun in the slime repl, or evaluating (bug) in the three cases.

[kpoeck]

Still happening to me on macosx, latest everything

[xrme]

Gilbert Baumann wrote on openmcl-devel:

I looked into before-mentioned issue on github. The smallest test case exposing this bug that I could find is like the following:

(DEFUN BUG ()
  (DECLARE (OPTIMIZE (SPEED 1) (SAFETY 3) (DEBUG 1) (COMPILATION-SPEED 0)))
  (LET ((X 0))
    (LABELS ((FOO () (1+ X)))
      (SETQ X PI)
      (1+ X))))

I get

(bug)
=> #<BOGUS object @ #x302001E0E255>

Which is bad. It goes away with (COMPILATION-SPEED 3). Remove the LABELS and the bug disappears.

I nailed that down to the following change: https://github.com/Clozure/ccl/commit/bed5cf811ae157298903036276c079e336ae1cf9

   diff --git a/compiler/nx0.lisp b/compiler/nx0.lisp
   index 110c6c41..15101941 100644
   --- a/compiler/nx0.lisp
   +++ b/compiler/nx0.lisp
   @@ -1599,7 +1599,7 @@ Or something. Right? ~s ~s" var varbits))
               (with-program-error-handler (lambda (c) (runtime-program-error-form c))
                 (parse-body (%cddr lambda-form) *nx-lexical-environment* t))
              (setf (afunc-acode p) (nx1-lambda (%cadr lambda-form) body decls)))))
   -
   +    (rewrite-acode-form (afunc-acode p))
        (nx1-transitively-punt-bindings *nx-punted-vars*)
        (setf (afunc-blocks p) *nx-blocks*)
        (setf (afunc-tags p) *nx-tags*)

I suggest until we find the real issue with REWRITE-ACODE-FORM, we disable it, if we can.

[xrme]

A more targeted change would seem to be edit acode-rewrite.lisp and change the definition of acode-rewrite-labels-flet so that it doesn't do any rewriting there.

In other words, do the inverse of

+
 (def-acode-rewrite acode-rewrite-labels-flet (labels flet)  asserted-type (vars funcs body p2decls)
-  (declare (ignore vars funcs))
+  (declare (ignore vars))
+  (dolist (func funcs)
+    (let* ((*nx-current-function* func))
+      (rewrite-acode-form (afunc-acode func))))
   (with-acode-declarations p2decls (rewrite-acode-form body asserted-type)))

(which was changed in 3117330ad50003ec2c5923bcb62d84ef4f55acd6)

[svspire]

Source of the problem is this line: (LET ((X 0)) if this is changed to (LET ((X 0.0d0)) then everything works fine.

The optimizer is looking at what it thinks the type of X is, which is a fixnum because of that 0, and then compiling a fixnum add/2 instruction, which when handed 1 and PI, creates a bogus object. It's not recognizing that it guessed wrong about the type of X.

When you set (COMPILATION-SPEED 3), this optimization is not done, so everything works.

The bug is somewhere in def-acode-rewrite acode-rewrite-add2.