Eric Marsden <[EMAIL PROTECTED]> writes: > >>>>> "gm" == Gerd Moellmann <[EMAIL PROTECTED]> writes: > > gm> There is good and bad news: Your test case fails with CVS CMUCL > gm> (basically 18e) as well as with 18d, but it seems to work with > gm> my PCL, to be installed post-18e. > gm> > gm> What I've seen in the backtrace of the failing cases brings back > gm> vague memories of fixing something similar a longer while ago in > gm> my PCL, but I want to drop dead if I remember what the fix > gm> exactly was. > > the same problem (well, same symptoms) occurs while compiling CL-XML > with CVS CMUCL, and IIRC with LISA (an inference engine); in both > cases metaclasses are being used. These are important applications; it > would be great to find a fix.
It looks like the MAKE-INSTANCE optimization in CVS PCL, fast-init.lisp, is the culprit. In my PCL, I've replaced fast-init with something faster and apparently more bug-free, a version of which is used in SBCL. That's why the test case works in both CMUCL/my PCL and recent SBCL. I'm afraid that makes a fix in 18e pretty unlikely. Even if someone invested the huge effort of porting back the new MAKE-INSTANCE optimization from my PCL to CVS PCL (unlikely to happen :), it still wouldn't have been tested much, which would be pretty suboptimal for a release.
