Re: [Gomp-discuss] Open64's implementation of OpenMP

[Pop Sébastian]

On Tue, Feb 04, 2003 at 12:46:37AM +0100, Steven Bosscher wrote:
> Op ma 03-02-2003, om 23:16 schreef Diego Novillo:
> > On Mon, 03 Feb 2003, Pop Sébastian wrote:
> > 
> > > So I think that here is where we will "parse" OMP pragmas arguments...
> > > Another extension will just build an equivalent tree node as 
> > > for Open64's OPC_PRAGMA node.  
> > > 
> > > I don't think that implementing another mini-C front-end based on 
> > > treelang 
> > > will help our design.  
> > > 
> > Agreed.  Once identified, adding the OMP #pragmas to GCC should
> > be fairly straightforward.  I can do the initial implementation.
> > Unfortunately, I won't have much time to go over the OpenMP specs
> > for a while.
> 
> It's not entirely trivial to do this, you know.  GCC's implementation is
> actually very "pragma-unfriendly".  All pragmas that GCC currently can
> handle are pragmas that put no restrictions on whatever syntax follows
> the pragma itself.  There's no interaction between the #pragma handlers
> and the parser AFAICT.
> 

c-lex handles it for us:

A simple example:

----
/* #pragma weak name [= value] */ 

static void 
handle_pragma_weak (dummy) 
     cpp_reader *dummy ATTRIBUTE_UNUSED; 
{ 
  tree name, value, x, decl; 
  enum cpp_ttype t; 
 
  value = 0; 
 
  if (c_lex (&name) != CPP_NAME) 
    GCC_BAD ("malformed #pragma weak, ignored"); 
  t = c_lex (&x); 
  if (t == CPP_EQ) 
    { 
      if (c_lex (&value) != CPP_NAME) 
        GCC_BAD ("malformed #pragma weak, ignored"); 
      t = c_lex (&x); 
    } 
  if (t != CPP_EOF) 
    warning ("junk at end of #pragma weak"); 
 
  decl = identifier_global_value (name); 
  if (decl && TREE_CODE_CLASS (TREE_CODE (decl)) == 'd') 
    { 
      apply_pragma_weak (decl, value); 
      if (value) 
        assemble_alias (decl, value); 
    } 
  else 
    pending_weaks = tree_cons (name, value, pending_weaks); 
} 
----

In this you see c-lex be called everytime we need a tree node representing 
either the name, or value based on the lex analysis of the string in pragma.
Thus I see no reason why we couldn't have something like that for OMP.

Diego, could we keep patches for OMP in tree-ssa?