Subject: publishing
chambers@cs.washington.edu
Date: Tue May 30 2000 - 11:58:23 PDT
I'm publishing a couple of small things, and also a new vortex
executable (in /projects/cecil/bin/solaris) and vortex.nCCP profile
(in /projects/cecil/bin/pic-files).
*) There appears to be a g++ linker problem with our huge vortex.pic
executables, where some strings (in particular, the strings that are
printed out with the PIC profiling output) aren't being linked
properly, but are left as garbage addresses. ??!!?!?!?!? This causes
the profiling output gathered by make-profile to be ill-formed at
times, and causes the pic-filter script to produce vortex.nCCP data
that isn't syntactically correct, not to mention semantically
incorrect. Since over the course of a week of debugging I've
convinced myself that it's a linker bug, not a Vortex bug, I modified
the pic.c runtime file to comment-out PIC output information that
appears to be bogus. This makes downstream processing work correctly
again (we can run make-profile on vortex.pic to produce new
vortex.nCCP files again; this has been broken for a year at least it
seems), at the cost of losing some PIC entries. Only PIC entries
corresponding to in-line type tests and static links are lost; regular
dynamically dispatched PICs all seem to be output correctly. (I still
have a hard time believing that the linker has a bug, since it only
seems to affect string literals involved in certain kinds of PIC
output. At least that's the only time I've detected a string literal
problem.)
*) I modified the profile reading code in Vortex (the implementation
of load_profile) to be more robust in the face of syntactically
malformed profile files. I.e., syntax errors are generated to the
user, rather than generating internal errors within the compiler.
*) I extended the pic-filter script to allow comments in the
vortex.pic output file.
*) I revised the commentary in the make-profile script.
*) I extended the Vortex makefiles to allow regular debug versions of
.pic executables (in addition to optimized and full_debug versions).
I also fixed a bug that caused non-pic versions of the
full_debug_pic_files option to be built in addition to the pic version.
*) I regenerated the standard RTL files for the Java benchmarks into
/projects/cecil/RTL-src/Java, using the updated Java front-end. I
also updated the RTL code for the example program in
$VORTEX_HOME/Java/src/example/rtl.
-- Craig
This archive was generated by hypermail 2b25 : Tue Oct 03 2000 - 15:21:35 PDT