Subject: publishing
From: Craig Chambers (chambers@cs.washington.edu)
Date: Fri Nov 23 2001 - 17:52:54 PST
I'm publishing a new Whirlwind. The result: the towers Java benchmark
runs (if compiled without optimization)! The main new facility
vtbl-based GF dispatching, which is implemented for GFs that have a
particular pattern of method and inheritance. Some GFs break the
vtbl-dispatching pattern, either by having weird dispatching
predicates (never in Java programs) or by being affected by multiple
inheritance in bad ways (happens in the towers benchmark on a dozen
GFs defined on interfaces like Dictionary and Enumeration and
Runnable); these GFs will die if ever invoked at run-time.
There are some serious problems with the new vtbl manipulation with
respect to incremental recompilation in Whirlwind. In fact, you can't
even recompile a file containing a vtbl; you have to restart the
compiler or at least invalidate its data structures in order to
recompile a file with a vtbl.
In addition, towers compiled w/ optimizations has C-compile errors.
I'll work on these problems next, but I wanted to publish all this as
a checkpoint first.
Other things:
*) I got all the passes like scope building, name resolution, and rep
computing to work on FragmentIRs, aka replacement graphs. Now (at
least in the couple of cases I've enabled) one can build replacement
graphs without any of this derived information, and without hardwired
"pseudoslices", and instead let the regular code in Whirlwind
construct all this derived information in the same way that it's
constructed for regular graphs, from CFGs. In the future, I want to
remove all the other places that hard-wire this information, and drop
all the notion of pseudoslices, to make replacement graph construction
a lot simpler.
One thing I did to make this work was breaking up the
replace_node_with_graph operation into two: a
keep_slice_after_replacement test method that decides whether to keep
a slice, and also creates the slice on the replacement graph if it
wasn't present and it's being kept, and a second
replace_node_with_graph operation that does the actual graph splicing.
Things go badly wrong if a slice is tried to be computed on the
replacement graph after its nodes have been spliced as part of some
other slice.
I also fixed some problems with the builder code and with several of
the replacement graphs.
*) I factored out all the object-related lowering into
lower-objs.cecil.
*) I split rep-check.cecil into 4 files, because it was getting very
big.
*) I added some flags to class_nodes to reflect uses of MI in
subclasses, which is queried by the vtbl layout algorithm. The class
hierarchy building code sets these flags. I also fixed a bug in the <
method for class_nodes and changed the convention for clearing marks
so that each algorithm clears its own marks before it begins, not
after it finishes (it's more robust this way).
*) I modified the interface to graph node iteration so that the border
nodes are now visited. Borders are still skipped when doing dataflow
analysis of nodes in a graph. I also changed some code to more
clearly distinguish between heads (nodes w/o predecessors) and roots
(heads plus the border nodes before incoming_edges).
*) I added in some code to verify each replacement graph, under
control of the verify_each_transformation option.
*) I fixed a problem with C code generation of pointers to arrays of
pointers, in some combination I can't remember any more.
*) I cleaned up some code by removing the if_replace_graph_action and
if_continue_action control structures.
*) I fixed some problems with printing nodes w/ DFG but not AST
operands.
*) I added a (void*) cast to the output C code for Vortex debugging
data, to fix a problem with some compilers reported by one of our
outside users.
*) The typechecker noticed that Keunwoo's code assumed that a method
value(maybe, if_none_closure) existed. It didn't, but it seemed
useful so I added it, and converted several existing occurrences of
if_some_none to value.
*) I extended the regression test to include some examples that
actually can be implemented with vtbls (all the previous tests were
too complicated to use vtbls). I also added some tests for things
that I discovered were problems when compiling the towers benchmark.
And I added a test function that contains an infinite loop, to verify
that the reverse analysis passes work correctly even for graphs with
no exits (they appear to).
*) I modified the Java front-end to generate some @= dispatch
specification in place of @<, for methods that should be generated for
each class (e.g. getClass, basic_clone). The vtbl layout algorithm
also identified a bunch of missing methods (exhaustiveness checking!)
for helper classes like the classes of class objects and array
objects. I added the methods to the native libraries. I regenerated
all the Java .wil files.
*) I fixed a nasty bug in the arraycopy native method in Java.
*) I changed Java to use 4-byte unsigned integers for array lengths,
to try to avoid a bunch of truncation operations in the generated
code. (Whirlwind didn't have to change at all!)
*) I fixed bugs in the PTR +, -, and diff operations in the WIL
runtime system.
-- Craig
_______________________________________________
Cecil mailing list
Cecil@cs.washington.edu
http://majordomo.cs.washington.edu/mailman/listinfo/cecil
This archive was generated by hypermail 2b25 : Fri Nov 23 2001 - 17:53:04 PST