So, progress on SBCL PPC#

Sun, 16 Dec 2001 00:22:52 +0000

So, progress on SBCL PPC. Turns out to be an assembler routine which was jumping into the middle of EQL instead of starting at the beginning. EQL expects to use 16 bytes of number stack; by joining half way through we were omitting to allocate it, but freeing it as normal.

Why we were jumping into the middle of EQL, I don't know. The assembler routines were calculating their offsets in a way that would have been correct if static-function-offset did what its docstring says it should have done. However it doesn't, and various other stuff (basically, the generation of cold-sbcl.map) seems to depend on it working as implemented

-  (inst j code-tn
-       (- ( function-code-offset word-bytes) function-pointer-type))
+  (inst j code-tn 0)
+;;     (- ( function-code-offset word-bytes) function-pointer-type))

Now we have some problem with EQUALP instead, which will probably turn out to be related. From my cvs log entry: "I'm always reluctant to claim that code found in cmucl could not possibly have worked. Still not sure how this did, though. "