Subject: publishing
From: Craig Chambers (chambers@cs.washington.edu)
Date: Thu Mar 22 2001 - 15:53:57 PST
I'm publishing some stuff. The main accomplishment is to get all the
various Java benchmarks to be compiled successfully by Whirlwind, w/
and w/o optimization.
*) It turns out that the bug in the javac and javadoc benchmarks was
actually in the input .class files: one of the input files didn't even
pass the verifier! Whirlwind's rep checking identified the bogus
code. It was a one-line patch to the file to fix things, but if we
regenerate these .wil files again, we have to redo the patch by hand.
Or get new .class files for these programs. There's a file named MODS
in each of the benchmark directories that need some sort of
hand-modification.
*) I found and fixed a bug in the DFG construction, where global
variable references weren't being handled properly in the face of
partially available local assignments. I changed things to leave all
global variable reads (IDExprNodes) unoptimized, as their own dataflow
root. This also allowed me to skip adding in all global variable
declarations to the reaching def info propagated during sparse DFG
construction.
*) I discovered a performance bottleneck in the code to generate
typedefs for representations during C codegen, and fixed it.
Essentially, added some simple caching to a recursive function that
turned it from something massively exponential into something
approximately linear.
*) I made the representation computing & checking code verify that all
expressions have defined representations (not UndefinedRep). I had to
fix some small problems to make this pass. For example,
PredicateNodes are no longer ValueNodes (with a result
representation), but just plain IRNodes, and lowering has to visit
predicates to lower the expressions embedded in them.
*) I added some code to report phase times for the four new
optimization passes. I also factored out topological numbering into
its own phase. I also added a method phase timing, which prints out
phase times for slow-to-compile methods, as Vortex did.
*) I added the new tests directory to the standard search path for
.wil files.
*) I added some code to print out the process size at convenient
intervals.
*) I changed dot printing code to use the alternative PrintVisitSet
object, so that dot graphs can be constructed during some other
traversal of the same graph.
*) I modified the Jil Java front-end to generate a call to a
predefined load or store function to implement an array load or store,
rather than generating a bunch of code in-line. This makes .wil files
simpler, and somewhat easier/faster to compile when not applying
optimization. (Some class init functions contain long series of array
store bytecodes, to construct static arrays, and the long in-line
implementation of these was not making some parts of Whirlwind happy.)
I added arrays.wil in the Java/wil-stdlib/native directory to define
these functions, and collect together the other array-oriented
definitions.
*) To allow Jil and Javelin to coexist, I modified Jil to look for
JilFiles before Files. Now Files can be used for Javelin, and
JilFiles for Jil, in the same benchmark directory. (Javelin's Files
typically includes the whole Java stdlib files, while Jil's JilFiles
only includes the benchmark application's files.)
*) I regenerated all the Java .wil files, including the Java
benchmarks.
*) As always, I made numerous small reformattings and clean-ups.
-- Craig
_______________________________________________
Cecil mailing list
Cecil@cs.washington.edu
http://majordomo.cs.washington.edu/mailman/listinfo/cecil
This archive was generated by hypermail 2b25 : Thu Mar 22 2001 - 15:54:03 PST