Subject: publishing
From: Craig Chambers (chambers@cs.washington.edu)
Date: Sun Nov 12 2000 - 11:57:57 PST
I'm publishing a bunch of useful improvements to Whirlwind.
*) I reorganized some of the graph visit code to allow multiple kinds
of visits, each of which can be run in parallel with other visit kinds
(but not with other instances of the same kind on the same IR), as
well as a general visit kind that can be run in parallel arbitrarily.
I've changed the printing visits to be a different kind than the
others, so that you can print the graph while you're in the middle of
some other visit, e.g. during an analysis.
*) I implemented dot-based graph visualization. There are three new
options, show_dot_cfgs, show_dot_dfgs, and show_dot_cdfgs, which if
turned on will ask you whether you want to display the graph. Try it
out. It has already revealed that my dataflow graph building
algorithm breaks for references to global variables.... (Perhaps just
those defined textually in a different file. I think the fix to this
and other known DFG problems is to leave in variable references as
root DFG nodes whenever there's no definition computed for the
variable. (This may still not work in the face of conditional local
assignments to a variable, but it should work a lot better.)) I
haven't tested this exhaustively, but it at least works to some
extent. Let me know if you find problems.
*) I reorganized how nested IR nodes are compiled. Now, each IR is
compiled to completion, running all the various passes, before
proceeding on to any nested IRs. This makes each pass easier to
handle, since it doesn't recursively perform that pass on all nested
IRs. It also makes the resulting graph visualizations easier to deal
with. It also made it easier to support processing the global IRs
separately from all the nested IRs.
*) IRs now explicitly store their nested IRs, and some existing code
has been rewritten to explicitly use these lists.
*) I factored out the scope-building code from the name-resolution
code.
*) I spent a lot of time rationalizing the compress & invalidate code.
*) I wrote some code to automatically indent phase time reporting
within their enclosing phases, but it's not well tested.
*) I fixed what I think is a bug with analysis of replacement graphs
that have dangling outgoing edges. The framework now sets their edge
info explicitly to the analysis's top info.
*) I explicitly insert Enter and ExitScopeNodes around the top-level
statements, so they don't need to be special-cased in the codegen
analysis. (Todd, you should copy my simplifications into your code.)
*) I did some reformatting of the WIL parser code, just so it conforms
more to "standard" (i.e. "my") Cecil style.
*) I defined standard print[_recursively] and verify[_recursively]
operations on slices, which just forward to the preexisting
slice-specific functions.
*) I added some more comments. :)
There's a known bug in this publish (and probably in previous systems
too): if you modify/touch a file and then try to recompile, the
incremental reconstruction of the global scope info crashes with
duplicate declaration errors. I have to do some work to handle
incremental rebuilds after programming changes. For now, Just Say No
(to input program changes). Or do "invalidate" before doing any
graphs/typecheck/make operation where the program has changed.
-- Craig
This archive was generated by hypermail 2b25 : Sun Nov 12 2000 - 11:58:03 PST