Subject: publishing
From: Craig Chambers (chambers@cs.washington.edu)
Date: Wed Nov 29 2000 - 12:28:30 PST
I'm publishing my system, including a new *vortex* executable. Note:
my new vortex executable won't be able to reuse old checkpoints, so if
you either adopt my new vortex executable, or build your own, then
you'll have to recompile programs (e.g. whirlwind) from scratch.
*) I changed file_infos to be indexed by the full file path name, not
just the basename (w/o the leading directory). This should fix the
long-standing bug we've had where if you move a file from one place to
another, or create a file that should shadow an existing one, Vortex
(or Whirlwind) won't correctly update things. Now it should. A lot
of little file-name-related changes throughout the system.
*) I created an initial make-whirlwind-profile script, which can be
used to build a whirlwind.nCCP profile to use in future compiles of
whirlwind with optimization. Performance of Whirlwind might
double.... make-whirlwind-profile is also a whirlwind test script, to
ensure that whirlwind still works after program changes.
*) I've generated a new vortex.nCCP and initial whirlwind.nCCP
profiles, and installed them in the standard place:
/projects/cecil/cecil/bin/pic-files/{vortex,whirlwind}.nCCP. You can
load mine into your Vortex compiler incarnation compiling Whirlwind by
doing "load_profile
/projects/cecil/cecil/bin/pic-files/whirlwind.nCCP". The profile info
will persist in your whirlwind.db checkpoints, and remains valid even
as you change the whirlwind source code from what it was when I made
the profile.
By the way, if you ever build your own copy of the vortex executable
from scratch, then you should do
Vortex> load_profile /projects/cecil/cecil/bin/pic-files/vortex.nCCP
before compiling. This drives profile-guided class prediction, which
gives about a 2x speed-up to the resulting compiler. It's really
noticable.
*) I fixed a bug Todd encountered where the CFG-building code only
demanded that the AST be built, not also the ScopesBuilt,
NamesResolved, and RepsComputed slices.
*) I added some code to vortex-commit to ensure that checked-in code
is world-readable, to support CVSWeb.
*) I removed hobbes from the compileMachineNames list, since it has a
different gcc version and I think I was having problems with string
literals caused by mixing gcc versions in my compiled code. I also
removed rhodes temporarily, since it's on loan to support and not on
the network. (We should try to get it back soon. It was only going
to be borrowed for a month, before the fall qtr.)
*) I added an inline pragma to a core array iterator method, since it
was one of the most frequently executed methods that wasn't inlined,
and it invokes an argument closure (which often benefits a lot from
being inlined). We'll see if this is a good idea.
I discovered some bugs in vortex while doing this, which I've just
worked around for the moment (preferring to invest effort in
Whirlwind). One is that 1CFA trips an assertion during make-profile;
I commented it out. Another is that the big function in
libraries-rel.cecil compiles forever, when part of the whirlwind
program when using whirlwind.nCCP data; I turned off optimization for
that function. There's also the known bug in splitting for Java
programs compiled w/ library optimization that Jonathan reported, and
that's still not fixed.
-- Craig
This archive was generated by hypermail 2b25 : Wed Nov 29 2000 - 12:28:40 PST