Subject: publishing
From: Craig Chambers (chambers@cs.washington.edu)
Date: Mon Oct 09 2000 - 15:23:30 PDT
I'm publishing some improvements to the Whirlwind sources.
*) I've gotten representation lowering and checking to work both with
and without building the DFG (the DFG is built when programs are
compiled w/ optimization). Now I consider this part of Whirlwind to
be a stable base on which Keunwoo can do lowering of operations, and
Sorin can look at alternative IRs (e.g. the control dependence graph)
and then getting the dataflow analysis & transformation engine up and
running, and Todd can look at programming interesting abstract value
analysis and transformation (an analysis subsuming constant
propagation and class analysis).
*) As part of the above, I introduced a new IR slice, CDFG, and a
topologically numbered variation, which is the union of the control
and dataflow graphs, for analyses (like lowering and rep-checking)
that want to visit all nodes in either slice, exactly once each.
*) I changed how rep slices are computed on demand. Instead of asking
each existing slice if it can compute the desired slice, and if so at
what cost, I changed the code to have each slice know how to compute
itself in terms of some simpler slice(s). The new way is simpler and
works in more situations.
*) I changed how compression and invalidation is handled, so that each
slice gets a chance, modularly, to compress out any info that it has
decorated on the IR nodes.
*) I expanded the various graph verification tests.
*) I fixed a couple of omitted functions in list_keyed_set. (FYI, I'm
now futzing with getting implementation-side typechecking in Vortex to
work, so that we all can test better our implementations.)
-- Craig
This archive was generated by hypermail 2b25 : Mon Oct 09 2000 - 15:23:37 PDT