Subject: publishing
From: Craig Chambers (chambers@cs.washington.edu)
Date: Sun Apr 14 2002 - 18:25:02 PDT
I'm publishing some significant improvements to Whirlwind. I've
worked on the global passes for constructing the class hierarchy and
generating dispatcher data structures (e.g. vtbls), and made them work
reasonably in the face of recompilation. They now use dependencies,
and trigger invalidations of compiled files when their information
changes. You can touch a source file, and things are precise enough
that only that file will end up being recompiled. Interrupting
compilation in the middle (e.g. typing ^C) and recovering to the
Whirlwind prompt and recompiling also works. Some details could still
be improved, but they're relatively minor.
I tested this by compiling test (the small regression test suite) and
towers (the first Java benchmark), and doing lots of recompilations,
interruptions, etc.
Whirlwind changes:
*) I fixed a problem with C code gen where a local variable could be
referenced before it was defined. Now all C local variable
declarations are emitted before any other statements are
generated.
*) C code generation of TopDefNodes (the dataflow graph's
representation of an undefined value) now emits a cast to the assumed
type, to make the C compiler happier.
*) I removed the recompute_globals flag (which used to indicate
whether the global data structures needed to be recomputed), since
the slices on the ProgramWindIR object now do a better job.
*) I made recovery after interrupting slice generation more drastic:
now all information about the file is dropped. The old way was
problematic, since slice-specific information could have been
partially computed, but the slice not yet registered as present.
*) I fixed a problem with infinite recursion on invalidation of files.
*) I made the compression and invalidation functions of the
HierarchyComputed, DispatchersComputed, and DispatchersGenerated
slices just drop the slice, but not drop all the computed tables.
This allows the next pass of computing to consult the table from the
previous pass, and then trigger invalidation of compiled files only
for changes.
*) I introduced an internal declaration mode, for things like vtbls,
which are handled differently than user declarations when it comes to
checking for duplicates.
*) I introduced a gf_table object which holds information about the
gfs in the system. This gives a clean place to put methods that
manage incremental updates to gf->method information, analogously to
the class_hierarchy object.
*) I added some code to test structural equality of predicate
expressions and some regular expressions, to enable testing whether a
new method declaration has the same predicate as an old method
declaration.
*) Lots of new code to manage class hierarchies, gfs, methods, etc.
Other changes:
*) I changed some of the Cecil benchmark stdlib source code to reflect
changes we've made to the regular stdlib to support compilation on
multiple platforms.
*) I made VORTEX_SOLARIS5 be defined when compiling on the old Solaris
boxes, to make it easier for ifdefs to test for this version of
Solaris specifically. I changed some ifdefs to use it.
*) Some improvements in the ssh references in the support scripts,
e.g. pm.
*) I added some g++ compiler flags to our Makefiles so that it won't
emit warning messages and won't fail on some non-strict-ANSI usages
that we have, e.g. implicit casting from void* to a real pointer
type.
*) I added lines to the makefiles to remove the target file if it's
being recomputed. This way, if there's an error in compiling a file,
the previous version of the file won't be silently used by mistake.
-- Craig
_______________________________________________
Cecil mailing list
Cecil@cs.washington.edu
http://majordomo.cs.washington.edu/mailman/listinfo/cecil
This archive was generated by hypermail 2b25 : Sun Apr 14 2002 - 18:23:07 PDT