Subject: publishing
From: Craig Chambers (chambers@cs.washington.edu)
Date: Fri Jun 15 2001 - 15:53:02 PDT
I'm publishing some Whirlwind source-code changes, before embarking on
a massive reorg of Wil representations.
*) I got my first Wil test program to compile and run. This program
includes a factorial function, so procedure declarations, calls, and
returns and conditional branches now work, and it includes primitive
code statements (embedded C++ code).
*) I did some reorganization and renaming of the concrete return and
raise nodes, and their C code generation.
*) A lot of my effort went to ensuring that primitive code blocks can
reference Wil variables using unmangled C names. This required me to
change the syntax of primitive statements to document explicitly which
Wil variables of enclosing scopes are (possibly) read, and which are
(possibly) written. This tells the C code generator to introduce
assignments from the mangled names to the unmangled names on entry,
and vice versa on exit. It also has the pleasant side-effect of
enabling the dataflow graph to be built properly in the face of
primitive blocks embedded in the middle of regular Wil code.
*) Primitive blocks are the first kind of Wil IR node that can produce
multiple results. Others may follow. To account for this notion in
the dataflow graph, I added a PrimStmtResultSelectNode to the DFG
representation, which allows a particular result to be picked out of
the many produced by a PrimStmtNode. This is a little ugly; I've
wanted to have multiple result "ports" on IR nodes analogous to IR
nodes having multiple argument "ports" instead; but this solution
causes minimal "damage" to the current scheme where all nodes have
either 0 or 1 (possibly implicitly fanned-out) result. (Probably,
output ports should become more first-class. They should replace
edges in the dataflow analysis engine worklist, to avoid replicating
AnalysisInfos along all edges fanning out from some result port.)
*) To make gotos work, I separated variable declarations in the
generated C code from their initialization. Otherwise, g++ complains
about gotos across variable initializations.
*) I fixed a bug with global variables being declared as local
variables inappropriately.
*) I changed how .h files are used in the generated C code. The rule
is that a .h file is *only* for the corresponding .c file to use, and
is where all extern declarations for anything from other files and
other declarations needed in the global scope at the start of the .c
file go. So each .c/.h file should be self-contained in terms of its
declarations of what it assumes about the outside world (aside from
runtime system declarations in WIL-defs.h, of course).
*) I added a prims.wil test program.
*) The make command now invokes the C->exec compile automatically.
Use make1 to only do the Whirlwind->C compile.
*) I changed the make target filename to omit the ".wil" suffix. I
also got rid of some of the old Vortex-related stuff for ".prims"
filename parts.
*) I fixed some dependency problems, after the file_info reorg, where
building the global info caused parsing to become out of date, and
also it never detected that the generated code was present.
*) I removed "var" annotations from field declarations in
ir-node.cecil where possible. I also removed all references to
ast_data_sources from that file.
*) I removed some mistaken comments about how compile_file_internal
works; it doesn't compile included files.
*) I futzed with some specialize pragmas for add_code, triggered by a
bug in Vortex that broke when one of the specialized-on objects is
removed from the system. I didn't fix the bug, since it's obscure and
would have required actual work on my part :), but the source code
changes I made probably ensure that other people who update won't hit
the same bug.
-- Craig
_______________________________________________
Cecil mailing list
Cecil@cs.washington.edu
http://majordomo.cs.washington.edu/mailman/listinfo/cecil
This archive was generated by hypermail 2b25 : Fri Jun 15 2001 - 15:54:07 PDT