On Tue 11-03-2003 10:42 Lars Segerlund took OT to the next level:
In the discussion on the gcc mailing list a while ago someone
suggested that the tree structures should be restructured in a way as to
prevent duplication of information when lowering, ( thus lowering would
mainly consist of adding nodes to the tree and perhaps some infor
mation), and this would give a representation which could keep the
higher level constructs all the way to the RTL ( since the RTL could be
attached to the tree ).
[ OT: Frankly this is one of the biggest problems in GCC: It
does not have an IR that can be "lowered" in the traditional
meaning of the word. The only "real" lowering we do is
GENERIC->GIMPLE, everything else is a complete rewrite. When
GCC lowers from trees to RTL, it basically duplicates the
whole tree in a completely new representation (adding cruft
and magic all over in the process). ]
[ OT^2: Diego, do you plan to lower GIMPLE even further?
The closer you'd bring it to RTL, the easier (hence
faster, less bloated) the tree->RTL expanders could be... ]
OK, now lets try to get back ON topic... What does this discussion mean
for our project?
Seb. said that CIL is concurrency-aware. Maybe somebody can explain a
bit how it represents concurrency, e.g.: Which constructs does it have
to support concurrency, and how does it handle data in those constructs
(i.e. what is thread private, what is a semaphore, etc.). Is it enough
to support everything we need for OpenMP? And, is it reasonably small?
If so, we could model our GENERIC concurrency-awareness tree
representation to what CIL does. We could then define a superset of
these trees that can represent OpenMP pragmas in the parser, and lower
the superset to the stuff we'd have in GENERIC in the pass than
translates the C/C++ parse tree to GENERIC.
Greetz
Steevn
_______________________________________________
Gomp-discuss mailing list
address@hidden
http://mail.nongnu.org/mailman/listinfo/gomp-discuss