This is the mail archive of the mailing list for the GCC project.

Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: [gomp] parallel binding reorg

On September 27, 2005 21:43, Richard Henderson wrote:

> It'll probably fall flat on its face when it gets to the omp_parallel
> node, which has sub-nodes.  Our CFG doesn't handle that anymore,
> remember?
I missed one 'C'.  I meant Concurrent CFG.  Yes, it would need to handle 
the *_BODY nodes.  Which is OK, because those give you natural regions.

> But if you munge that somehow, the actual gimplification should go
> ok. We'll have not yet set decl_value_expr on the inner shared decls,
> so they'll stay intact during gimplification.
That should be good enough.

> Your analysis will have to be able to handle a many-to-one mapping of
> decls to objects.  But that should be trivial in comparison to
> everything else, really.
Only if it's going across parallel sections, though.  But, yes, in 
principle it doesn't sound too bad.

> > Hmm?  What is missed by walk_stmts?
> The fact that it relies on gimple, and this block is generic?
It uses walk_tree at the bottom.  It really doesn't miss anything.  
Otherwise we would've been missing replacements in the previous 
incarnation of omp lowering.  But it was mostly curiosity, I certainly 
don't mind.

Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]