Subject: publishing
From: Craig Chambers (chambers@cs.washington.edu)
Date: Fri Sep 13 2002 - 10:52:12 PDT
I'm publishing a collection of miscellaneous improvements to
Whirlwind. The regression test suite passes, on several machines, w/
and w/o optimization, and the Java benchmarks all compile & run w/
optimization, at least on Solaris5.8 machines.
*) I started work on getting class-analysis-based static method lookup
implemented, which will sometimes allow static binding (aka
devirtualization) of calls of generic functions with more than one
method in them. (Right now, Whirlwind does static binding of gfs
with a single method in them, since no analysis is required in that
case.) As part of this effort:
*) I extended the class hierarchy computation with code to compute
the set of concrete descendant classes for each class.
*) I extended class analysis to use these sets to get more precise
information about any value which is known to be an instance of a
class or one of its subclasses ("cone types" in Vortex).
*) I reorganized the class analysis class-test folding code.
*) I extended constant propagation to track floating point constants,
and I added code operations to fold floating point operations. This
included adding support to standard library code to test whether a
floating point number is a NaN (since you can't test using regular
equality), and changing how the Infinity and NaN literal IR nodes
work. I fixed some problems with the representation checking code in
the face of Infinities and NaNs. I extended the nums.wil test file
with tests for these floating point operators. In the process, I
discovered (and fixed) a long-standing bug in the Cecil std library
code for round_towards_zero on floats!
*) I refined the treatment of conversion operators. I made some name
changes, including in Wil syntax. I cleaned up the treatment of char
conversions. I clarified the role of the generic convert_int
operator. There are no more "Yell at Keunwoo" error messages in
lower-ops.cecil. I augmented the unary-conversions test to actually
print out some results. I modified the Java front-end to produce the
right kind of conversion operators, and regenerated all the Java .wil
code. (I also modified the Java front-end to replace "if not(x op
y)" with "if x negated-op y", doing a little arithmetic
simplification in the front-end.)
*) I added a test file exercising nested function declarations. For
now, Whirlwind supports nested function declarations as long as the
only free variables referenced in the nested function are globals. I
fixed a bug in DFG building for references to functions (or any
non-variable declaration) as first-class values.
*) I changed the C codegen code to name functions directly, rather
than storing the C functions into a global variable, and referencing
that variable when the function is referenced. This should
streamline the code sequence generated by the C compiler for calling
functions.
*) I cleaned up the special treatment of casting nodes (which have no
dynamic effect and hence can sometimes be ignored), mainly in C
codegen.
*) I fixed a bug in C codegen for the most negative integer value in
each size.
*) I fixed a bug in rep-checking of return statements. (It's too easy
right now to accidentally use the AST operand instead of the DFG
operand.)
*) I added a bunch of support to the Wil regression test suite to
actively check for expected behavior and print a clear error message
(and a summary message) if not. I also made the output format more
consistent across tests. I used to have to manually eyeball the
output to see whether it looked right, and a few errors (e.g. the
most negative integer problem, and the floating point round_to_zero
problem) had previously escaped my notice.
*) I modified the Wil language specification document (ir.txt) to
reflect the conversion operation updates and to document some
additional macros available to prim statement block writers.
*) I updated the compile- and linkMachineNames.
-- Craig
_______________________________________________
Cecil mailing list
Cecil@cs.washington.edu
http://majordomo.cs.washington.edu/mailman/listinfo/cecil
This archive was generated by hypermail 2b25 : Fri Sep 13 2002 - 10:47:47 PDT