Subject: publishing
From: Craig Chambers (chambers@cs.washington.edu)
Date: Sat Apr 27 2002 - 00:34:11 PDT
I'm publishing some more significant improvements to Whirlwind. The
hightlights are that interface GF invocations (those that can't be
compiled into regular single-inheritance-based vtbls) and exceptions
both are implemented, as well as numerous other small glitches. As a
result, Whirlwind now compiles and runs almost all of our
non-multi-threaded Java benchmarks! (There is still a bug in pizza
that I will look at, and I haven't checked anything with
optimization.)
More details:
*) In a previous publish, I got vtbl-based dispatching to work. These
are GFs that do single-dispatching over classes that have no multiple
inheritance from objects outside the domain of the GF. In Java,
interface methods can violate this constraint. I added a second
dispatching strategy that allocates index positions globally, and
used this for all single-dispatches that can't be handled with
regular vtbls. Each class then has 2 vtbl arrays, one regular, one
using global allocation (the global vtbl is referenced from position
0 of the local one). The global vtbl can waste a lot of space, but
this is limited to the number of non-regular-vtbl GFs. In the
process, I introduced a vtbl_entry abstract superclass of method_node
and a new empty_vtbl_entry, which replaced the old ugly hacks
involving not_defined testing. I expanded the gf.wil test file to
exercise this.
*) I got exception handling to work. I revised the old scheme to
represent different possible exn calling conventions, and I fully
implemented two possible exn returning conventions: one based on
PairType returns a la Vortex, and one based on a simpler global
variable strategy. By default, the global variable strategy is used,
and it's been tested most. I modified the hierarchy of IRNodes for
Raises and FunExnHandlers. I expanded the exn.wil test file to
exercise this.
*) I regularized and documented the macros available inside C/C++
primitives for doing things like return, raise, and send.
*) I fixed a nasty bug in C codegen where if the last IRNode generated
was a backward branch to some earlier label, it was omitted. So
functions that should have ended with a backward branch just fell off
the end instead. Oops....
*) I generalized some of the builder support routines, and fixed some
bugs, to make builders work with the slightly richer kinds of
replacement graphs now being built.
*) I added two kinds of helper IRNodes: NoteNodes and CommentNodes.
Notes allow diagnostic messages to be printed out when running a
compiled program, under ifdef control. Comments just insert helpful
comments into the generated .c file. I use these for sends,
exception operations, and global vtbl dispatches so far.
*) I hit a bug which made me realize that some of the unique_id
generator shared fields weren't being checkpointed, so nodes weren't
actually remaining unique. I fixed this.
*) I added a small tweak to the global name resolution code to avoid
invalidating files unnecessarily the first time they're processed.
*) I removed some out-of-date partial support for calling GFs
directly.
*) I fixed a number of bugs in the Java front-end, including (a)
ignoring GF introductions in interfaces; (b) not properly accounting
for the possibility that two independent superclasses/interfaces
might introduce the same GF, which would then need to be merged; and
(c) not properly accounting for the strange situation involving
interfaces where a value can be cast to something that's not a
descendant or ancestor, but merely a sibling. Of course, I
regenerated all the Java .wil files to reflect these changes.
*) I had to change the size of the length field of Java arrays from an
unsigned short to an unsigned int. Some benchmarks allocate arrays
bigger than 64K entries....
*) I fixed a bug in the arraycopy native method that related to the
elems field in an array being at different offsets depending on
whether the type of the elems requires different alignment
constraints.
*) I fixed a bug to skip over the array store type check if storing
null.
*) I fixed up the interfaces in the file-ops primitive functions to
properly report that they may raise exns.
*) I added a System.setOut method, not normally present in JDK1.0.2,
just for pizza's benefit.
*) I added the Java benchmarks to the CVS tree, or at least the
control files and tree structures in which the .rtl and .wil files
are generated. This makes it easy for other people to create their
own versions of the benchmarks and test them.
*) I modified the toba benchmark to take a -classpath argument. This
allows us to run it without having to remember to first change the
CLASSPATH to something ancient. I modified the configs.perl file to
reflect this change.
*) I improved pm to better handle not compiling on the machine that
would normally be the link machine.
*) I improved pm and MakefileRules to allow passing arguments to the
underlying C compiles (e.g. -D commands to enable NoteNode
diagnostics).
*) I updated compileMachineNames and linkMachineName to reflect the
new Solaris8 machines.
-- Craig
_______________________________________________
Cecil mailing list
Cecil@cs.washington.edu
http://majordomo.cs.washington.edu/mailman/listinfo/cecil
This archive was generated by hypermail 2b25 : Sat Apr 27 2002 - 00:32:10 PDT