Subject: publishing
From: Keunwoo Lee (klee@cs.washington.edu)
Date: Fri Jun 08 2001 - 20:46:38 PDT
Summary: Records partly work, but much is still broken. C/C++ back
end is part of the problem; I encountered, once again, a lot of
ugliness in the interface between WIL semantics and C semantics.
I am only going to describe what currently works/doesn't work in this
changelog, and possible immediate fixes. I think that I could make
some things work through valiant hackery, but it's probably a waste of
time. Thoughts on how to fix what's more deeply wrong will be in a
separate e-mail to cecil@cs.
BTW towers no longer gets through make1 without dying, because the
size methods are not implemented for certain reps. I think if you
define the ClassRep and a couple of others to be something sensible,
it would be okay.
----------------------------------------------------------------------
LOWERING
+ I believe that I have size computation working for both records and
fixed arrays, and arbitrary nesting thereof, assuming things are
well-formed---i.e., arrays contain elements of statically computable,
uniform size. It might make sense to allow elements of
non-statically-computable uniform size, but that currently breaks
loudly. So allocation should actually work for everything besides
dynamic arrays nested in other objects. But initialization is the
hard part, so...
+ "new" records of scalars can be constructed and initialized, but
they do not pass gmake for type reasons. C records use name equality:
therefore, no anonymous structure types are equivalent, nor are
pointers to such types. Therefore, we cannot even cast pointers to
anonymous structure types. Furthermore, according to g++, ANSI
disallows initialization from void *, so we're hosed either way. This
is but one example of how C's poor excuse for type checking makes it
less than ideal as an intermediate language.
+ The following initializations do NOT work:
+ Directly allocated (non-"new") records
+ Any nested composites: {records,arrays} inside {records,arrays}
These will get through make1 without crashing Whirlwind, but of
course they do not pass gmake.
+ Arrays that contain fixed arrays of varying sizes
+ Unions do not work, and now break loudly instead of silently. We're
getting to the point where we should start making unimplemented
lowerings screech in complaint. Fortunately, towers and the Java
standard library use no unions, so it won't affect us right away.
+ Added some convenience constructors, scale factor computation, etc.
for rep sizes---see reps-helpers.cecil
----------------------------------------------------------------------
INFRASTRUCTURE
+ Added a way of globally tracking nodes by ID. Useful for debugging
the cause of some random codegen'd expression. When you start up
Whirlwind, set the option "track_node_unique_ids true". Then, after
running a make pass, you can enter the debugger and ask the global
node ID table for the value of a node.
So, if you have some inscrutable codegen'd expression like
struct {R_int4 a; R_int4 b; } *V14 =
MALLOC_byte(struct {R_int4 a; R_int4 b; } *,8);
you can just type the following:
debug> ir_node_unique_id_table!14
fixed_malloc(<anon/BytesRepSize/: 0xfb0821>)
See ir-node.cecil for details (look for unique_id). Note that you
have to turn on tracking before any nodes whose IDs you want to
track are created. I didn't want to override the unique_id field,
only its initializer.
+ Added a CompositeRep to be the superclass of array and record reps
(in reps.cecil). I don't currently use this class, but I was about to
use it and I figured there was no harm in leaving it in. Notes:
+ At some point in the near future, we should actually factor the
is_scalar() method (added in one of my recent commits, defined on
a bunch of different DefinedReps) out to a ScalarRep abstract
class. Anything that didn't inherit from ScalarRep or
CompositeRep would then be implicitly a rep of unknown
"scalarity".
+ I wasn't sure whether a union was a composite. Unions have a
weird status because they can be alloc'd and initialized like a
scalar, if they're small enough. Right now UnionRep does not
inherit from CompositeRep. Perhaps it is properly of unknown
scalarity.
----------------------------------------------------------------------
TRIVIA
+ Changed & commented the size(@:PtrRep) method in reps.cecil.
+ Some more IRNode constructors, in ir-node.cecil, that do not
initialize the ast_data_sources.
+ Changed the is_fully_lowered of DerefUnaryOp and its child,
FakeDerefUnaryOp.
_______________________________________________
Cecil mailing list
Cecil@cs.washington.edu
http://majordomo.cs.washington.edu/mailman/listinfo/cecil
This archive was generated by hypermail 2b25 : Fri Jun 08 2001 - 20:47:05 PDT