Path: utzoo!utgpu!watmath!uunet!lll-winken!vette!brooks
From: brooks@vette.llnl.gov (Eugene Brooks)
Newsgroups: gnu.gcc.bug
Subject: Re: Future of gcc in terms of threads and shared libraries
Message-ID: <17259@lll-winken.LLNL.GOV>
Date: 24 Jan 89 17:33:11 GMT
References: <HAM.89Jan23194625@polya.Stanford.EDU>
Sender: usenet@lll-winken.LLNL.GOV
Reply-To: brooks@maddog.llnl.gov.UUCP (Eugene Brooks)
Distribution: gnu
Organization: Lawrence Livermore National Laboratory
Lines: 35

In article <HAM.89Jan23194625@polya.Stanford.EDU> ham@polya.Stanford.EDU (Peter R. Ham) writes:
>
>	At the Distributed Systems Group, we are working to extend 
>gcc to accept a new keyword "shared" to distinguish global data
>that is shared among light weight process in a task rather than private
>to a single thread.  I hear that Sequent has a similar feature in one
>of their proprietary compilers.
>
>	I would like to solicit some comments and ideas about what the
>right way to add this to gcc is.  Right now, I have added the new keyword
>and taught gcc how to generate different types of global reference for
>the memory architecture we are using here.  Now, I'm modifying gas to 
>reflect a new a.out format for this new feature.  Next, I will be tackling the
>linker.  I would also like my changes to be of sufficient quality to migrate
>their way back into the released gcc proper.
>
The keywords "shared" and "private" are implemented at storage class
modifiers of sorts in Sequent's PCC based compiler.  They may be applied
to extern or static declarations.  In addition to the keywords you can
apply a flag '-Y' to the compiler, similarly to the -fshared-data flag
for GCC now, which defaults data to shared status.  This is the reason
for the existence of the keyword "private."

With the keywords as implemented by Sequent, you can NOT for instance declare
a "pointer to shared data" which would be useful for an enhanced version of lint.
The notion of type modfiers instead of storage class modifiers would also be
useful on machines which might have seperate address spaces for shared and
private data, much like some machines have seperate address spaces for text
and data.  Such checking of pointers is very useful to anyone who has extensive
experience with parallel programming.  Defining the keywords "shared" and
"private" to be type modifiers used much like "const" and "volitile" would be
very useful in this regard.

							Eugene Brooks
							brooks@maddog.llnl.gov