Subject: publishing
From: Craig Chambers (chambers@cs.washington.edu)
Date: Tue Oct 03 2000 - 12:39:16 PDT
I'm publishing my system, including a new vortex executable.
--- WHIRLWIND CHANGES ---
*) Most recently, I worked long and hard to get the Whirlwind compiler
moderately up and running. In particular, it now successfully
translates Java WIL code into CFGs, annotates it with representation
info, lowers any abstract representations into fully concrete ones,
checks that representations are internally consistent (rep-checking,
analogously to type-checking), and produces C target code. The C code
doesn't C-compile correctly, because there is a bunch of stuff that's
missing from Whirlwind and from the Whirlwind Java run-time system,
but we're making progress towards that. Also, I know it doesn't
correctly pass all these phases when optimization is turned on, which
currently means that def/use chains (the dataflow graph) are
constructed and used for lowering, rep-checking, and codegen; I'll be
debugging this shortly.
*) I got Whirlwind to print out a proper help message.
*) I did some indenting and other reformatting of Whirlwind source
files, to make them more readable and to bring them closer into
"standard" Cecil style (i.e., my style).
*) I made some clean-ups of the name-resolve task worklist structure.
*) I made a few misc cleanups to whirlwind's IR and naming scheme.
*) I brought the whirlwind/notes/reps.txt file more up to date
w.r.t. the changes to ir.txt.
*) I made various changes to whirlwind files, and a few vortex files,
to cut down on the number of spurious vortex files that are included
in whirlwind. Now whirlwind is a slim 207 files big! Of course, it
doesn't actually compile anything yet....
--- VORTEX CHANGES ---
*) I discovered what the problem was I discovered last spring where
some string constants were not linked right: calvin is running a more
recent version of g++ than our other suns. When I dropped calvin from
the compileMachineNames list and recompiled all my .o files, the goofy
string literal problem went away!
*) I fixed a bug where the no_optimize pragma was being silently
turned off in Cecil methods containing paren scopes.
*) I added an optimization to suppress generating all the list of
fields for maps that don't correspond to concrete objects.
*) I fixed a bug where if a field initializer expression in an object
constructor expression or declaration (e.g. the many field initializer
expressions in the various *_lang concrete objects in lang.cecil) led
to a fatal error (e.g. referencing an undefined variable), then any
later code after the initializer expression was not properly linked
into the CFG constructed, causing reverse dataflow optimizations
(e.g. dead assignment elimination) to break.
*) I also moved the get_uninit_literal functions out of a file of the
std prelude and into a file in the compiler, so that compilations of
non-compiler programs don't generate accessing-undefined-variable warnings.
*) I renamed print_address to address_print_string.
*) I added an optimization to the CSE pass to seed range analysis
information by the representation of the variable and/or the operator
computing some value. E.g. things of rep int1 (a 1-byte signed
integer) are known to be >= -128 and <= 127. This enables a few more
foldings of comparisons, which also leads to a warning message from
the C compiler being avoided. (There was a comparison we weren't
folding, which the C compiler knew to always be true (or false) based
on the C datatypes involved.)
*) I changed "int" to "wordInt" in all RTL primitives.
*) I implemented new_[i|m]_vector_init_from on arbitrary collections.
They used to only be implemented on ordered_collections.
*) I made some small tweaks to the run-time debug interface, so that
its help message only refers to commands that are possible given the
current state of the debugging information.
*) I revised the LICENSE file to be in English, not lawyerese. For
future releases, though, I'd like to rethink our licensing policies.
-- Craig
This archive was generated by hypermail 2b25 : Tue Oct 03 2000 - 15:21:52 PDT