Subject: publishing
From: Craig Chambers (chambers@cs.washington.edu)
Date: Sat Feb 24 2001 - 17:32:11 PST
I'm publishing some improvements to Whirlwind. My main accomplishment
is getting Whirlwind to compile towers on the 1.2.1 stdlib with
optimization. Several bug fixes and some miscellaneous clean-ups were
involved.
*) I stopped the CFG construction pass breaking down InitNodes into a
VarDeclNode followed by an InternalInitAssnNode. Now InitNodes are
allowed in all slices. InitNode has been renamed VarDeclInitNode to
emphasize that it's an initializing var declaration. I removed
InternalInitAssnNode, and merged SourceInitAssnNode into InitAssnNode
(which represents a source-level ":=="-type assignment). I renamed
NonInitAssnNode as UpdateAssnNode (which represents a source-level
":="-type assignment).
*) I found and fixed a problem with the IR slice compression code.
There is a link from FormalVarDeclNodes to the scope in which they're
declared, which is a nested scope. Compression of AST nodes in the
outer scope (including the FormalVarDecls of top-level methods &
functions) keeps all the AST nodes, and tries to flush any derived
info from them that shouldn't be there, but this pointer from a
FormalVarDecl to a nested scope wasn't getting flushed. The nested
scope pointed to all VarDecls etc. in the nested scope, which have
links to all other nodes in the nested scope, so by not breaking this
link, all of the ASTs etc. of nested scopes were being retained in the
checkpoint, and during compilation. Now checkpoints stay small
(almost the same size as after doing globals), and the Whirlwind
process size doesn't grow linearly as files are compiled. As part of
this, I had to ensure that the formals of MethodCaseAddNodes (which
are in the predicate test scope) and the formals of the body of the
method (which are part of a function expression) are made distinct
when using the method sugar that leaves the function implicit.
*) To help me find the space leak, I modified the output of the
object_size_histogram operation to sort the output by size before
printing it out.
*) I discovered and fixed a problem with how representations of
DataflowMergeNode were being computed. In the process, I renamed
DataflowMergeNode as PhiNode. Also, PhiNodes no longer have a VarDecl
embedded in them, since a single PhiNode can be reused for multiple
variables.
*) I moved a bunch of pred/succ edge/node manipulation functions from
WindIRSlice to Sorin's new WindIRGraphSlice object, which only the
CFG, DFG, and CDFG (and their Numbered versions) subclass. I
introduced a corresponding WindIRGraphSliceKind object. Both of these
I parameterized by the IREdge subclass for that graph, which is fixed
for a particular subclass of WindIRGraphSlice[Kind]. This makes a lot
of these edge/node manipulation functions typecheck reasonably.
*) Also, graph printing, visiting, and verification, and dot-graph
construction are now limited to WindIRGraphSlices, allowing their
interfaces to become simpler.
*) A couple operations are defined by both the AST and the graph
slices, so I created WindIRGraphLikeSlice[Kind] objects, and moved the
operations from WindIRSlice[Kind] to there. More happiness in
typechecking.
*) I realized some things about how operands are handled during graph
printing, which allowed me to get rid of the ugly code for printing
operands of nodes with lvalue operands.
*) I now include builder-driver in Whirlwind, so that it gets
typechecked. It still can also be compiled by itself. (Maybe there
should be a builder-test file that's most of what builder-driver is
now, and a separate builder-driver file that just includes the
necessary files and then invokes some driver test function.)
*) I removed builder-ast, as builders are now restricted to work only
for WindIRGraphSlices, which ASTs are not.
*) The typechecker has some bugs w/ dynamic, which forced me to stick
in some weird and unnecessary casts in builder-dfg and
builder-driver. When the typechecker bug is fixed, we can remove
these casts.
-- Craig
_______________________________________________
Cecil mailing list
Cecil@cs.washington.edu
http://majordomo.cs.washington.edu/mailman/listinfo/cecil
This archive was generated by hypermail 2b25 : Sat Feb 24 2001 - 17:33:04 PST