On Thu, 9 Dec 2004 01:33:23 -0700 (MST), Gary Byers gb@clozure.com wrote:

OpenMCL dies doing the moral equivalent of:

(defun foo (s i) (declare (optimize (speed 3) (safety 0))) (schar s i))

(foo "" -1)

If I do:

? (let* ((policy ;; Be conservative about generating unsafe code, ;; regardless of OPTIMIZE declaration settings (new-compiler-policy :trust-declarations #'(lambda (env) (declare (ignore env)) nil) :inhibit-safety-checking #'(lambda (env) (declare (ignore env)) nil) :open-code-inline #'(lambda (env) (declare (ignore env)) nil)))) (set-current-compiler-policy policy) (set-current-file-compiler-policy policy))

and recompile cl-ppcre with those settings in effect, I get

? (ppcre::scan "\(.*\)$" "")

...

Error in process listener(1): Array index -1 out of bounds for "" . While executing: #<Anonymous Function #x648DD4E>

(The empty string that's being referenced is the value of CL-PPCRE::*STRING*, and -1 is the value of a local variable named CL-PPCRE::START-POS.)

Why this code was trying to call SCHAR on an empty string (with -1 for an index) isn't clear; it certainly -could- be a compiler bug or something similar, but it's also plausible to me that an unsafe (SCHAR "" -1) is a quieter error in other implementations.

Wherever the bug is, it might be easier to isolate if the compiler's operating under a policy that discourages generation of unsafe code.

Good catch, thanks. This didn't occur to me because it doesn't seem to happen with Lisps other than OpenMCL. I've just uploaded a new version (0.9.3) of CL-PPCRE which hopefully fixes this. Maybe someone using OpenMCL can try it.

http://weitz.de/files/cl-ppcre.tar.gz

Thanks again, Edi.