Forked from
cmucl / cmucl
Source project has a limited visibility.
-
pw authored
foreign functions taking floating-point arguments after integer arguments but before the fourth argument (more or less; mips has unusual C calling conventions), and making cmucl work on IRIX 6.2 on the Onyx (where the system page size has changed). I've included the patches below, along with some discussion of the changes and a couple of problems with cmucl's interaction with dynamic loading. Also, there's a misc change to make ioctls work (I'm not sure for what version the released code worked, but it doesn't work on anything I have). The patches do not fix the R10000; cmucl continues to not work on it (O2 and Octane). The fixes make cmucl on the sgi use shared libraries so it can use dlopen et al in load-foreign; this is based on the solaris implementation. The problem with this is that the addresses of foreign variables are included directly in the dumped lisp image (without relocation on loading), which might result in a problem if cmucl is run with newer shared libraries than those it was compiled and dumped with. As far as I can tell, for the SGI at least, this does not affect function calls because the addresses used (and included in the dumped image) are those of stub routines that jump to the actual routines, thus are located in the executable proper rather than in the shared libraries, and so shouldn't be changed by a new version of the shared library. Unfortunately, this is not the case for global variables. Fortunately, there appear to only be two foreign variables used by cmucl in any of its system code winding up in the dumped image: errno and tzname. Apparently, Solaris relocates data references in dsos to match the program, while IRIX relocates data references in the program to match the dso. Thus, on Solaris, you don't need to do anything special for cmucl to work with a newer runtime libc, but on IRIX it gets the wrong value of errno and tzname. Unfortunately, the address of errno is different between irix 6.2, 6.3, and 6.4 (all of which we have to have because SGI doesn't have a single cross-platform OS version yet), which would require a different lisp.core for each machine (you just need to run mk-lisp, not recompile) if it weren't for the hack below. But fortunately again, a lisp compiled for one version seems to be able to run on another version well enough to generate the new lisp.core (presumably the worst that would happen is that error messages could be confusing). It would be really nice to avoid this problem. Perhaps some kind of relocation table could be added to the dumped image format? I don't know enough about the system to know how possible or ridiculous that suggestion is, unfortunately. Is the Solaris behavior guaranteed to be the same in future releases? If not, or if any other platforms have the irix behavior, this might become a more general problem. In the meantime, we use a hack: we *define* errno in undefineds.c rather than just referencing it; this results in warnings from the linker, but it resolves the conflict in favor of the definition in the program, relocating libc to match, thus fixing this problem for errno. The other global variables, tzname et al, can't be fixed this way as easily because they're listed as weak symbols, essentially aliases for the real symbol, so redefining them wouldn't cause libc to reference the new version. Thus, as long as you don't use tzname, there's no problem right now (although there may be if any other library global variables are used in cmucl in the future). The definition of tzname is in a progn preceeded by a comment which says not to use it, so hopefully it's not in fact used and therefore not a problem. These patches have been tested on an R5000 O2 running irix 6.3 and a dual R4400 Onyx running irix 6.2. A previous version of these patches (for 17f) was tested on the onyx when it was running 5.3 (I think, maybe 5.2). The patches here are from the main branch of the CVS tar as of August 13. Incidentally, it was quite painful figuring out how to compile and bootstrap the current sources under the distributed irix 18a binary, even without trying :new-random or :signed-array (which I haven't tried to do yet); it might be nice to put a newer binary up...
pw authoredforeign functions taking floating-point arguments after integer arguments but before the fourth argument (more or less; mips has unusual C calling conventions), and making cmucl work on IRIX 6.2 on the Onyx (where the system page size has changed). I've included the patches below, along with some discussion of the changes and a couple of problems with cmucl's interaction with dynamic loading. Also, there's a misc change to make ioctls work (I'm not sure for what version the released code worked, but it doesn't work on anything I have). The patches do not fix the R10000; cmucl continues to not work on it (O2 and Octane). The fixes make cmucl on the sgi use shared libraries so it can use dlopen et al in load-foreign; this is based on the solaris implementation. The problem with this is that the addresses of foreign variables are included directly in the dumped lisp image (without relocation on loading), which might result in a problem if cmucl is run with newer shared libraries than those it was compiled and dumped with. As far as I can tell, for the SGI at least, this does not affect function calls because the addresses used (and included in the dumped image) are those of stub routines that jump to the actual routines, thus are located in the executable proper rather than in the shared libraries, and so shouldn't be changed by a new version of the shared library. Unfortunately, this is not the case for global variables. Fortunately, there appear to only be two foreign variables used by cmucl in any of its system code winding up in the dumped image: errno and tzname. Apparently, Solaris relocates data references in dsos to match the program, while IRIX relocates data references in the program to match the dso. Thus, on Solaris, you don't need to do anything special for cmucl to work with a newer runtime libc, but on IRIX it gets the wrong value of errno and tzname. Unfortunately, the address of errno is different between irix 6.2, 6.3, and 6.4 (all of which we have to have because SGI doesn't have a single cross-platform OS version yet), which would require a different lisp.core for each machine (you just need to run mk-lisp, not recompile) if it weren't for the hack below. But fortunately again, a lisp compiled for one version seems to be able to run on another version well enough to generate the new lisp.core (presumably the worst that would happen is that error messages could be confusing). It would be really nice to avoid this problem. Perhaps some kind of relocation table could be added to the dumped image format? I don't know enough about the system to know how possible or ridiculous that suggestion is, unfortunately. Is the Solaris behavior guaranteed to be the same in future releases? If not, or if any other platforms have the irix behavior, this might become a more general problem. In the meantime, we use a hack: we *define* errno in undefineds.c rather than just referencing it; this results in warnings from the linker, but it resolves the conflict in favor of the definition in the program, relocating libc to match, thus fixing this problem for errno. The other global variables, tzname et al, can't be fixed this way as easily because they're listed as weak symbols, essentially aliases for the real symbol, so redefining them wouldn't cause libc to reference the new version. Thus, as long as you don't use tzname, there's no problem right now (although there may be if any other library global variables are used in cmucl in the future). The definition of tzname is in a progn preceeded by a comment which says not to use it, so hopefully it's not in fact used and therefore not a problem. These patches have been tested on an R5000 O2 running irix 6.3 and a dual R4400 Onyx running irix 6.2. A previous version of these patches (for 17f) was tested on the onyx when it was running 5.3 (I think, maybe 5.2). The patches here are from the main branch of the CVS tar as of August 13. Incidentally, it was quite painful figuring out how to compile and bootstrap the current sources under the distributed irix 18a binary, even without trying :new-random or :signed-array (which I haven't tried to do yet); it might be nice to put a newer binary up...