[cfe-dev] Adding a new type qualifier

[Chris Lattner]

On May 10, 2010, at 9:30 PM, kalyan ponnala wrote:

> Thanks for the reply john. What I want is something like the _global qualifier of opencl. The hardware architecture that I am trying to support with clang has this global address space. I am not sure about how to add an address space qualifier and parse and generate IR all the way. 
> I looked into the embedded C "address space" type qualifier, Can an expert who is familiar with openCL, tell me if this could be changed to act like the _global type qualifier. If not, I can use some suggestions too.
> 
> The following link was posted in march about the address space qualifiers, I would like to know if there was any progress on this topic. 
> http://lists.cs.uiuc.edu/pipermail/cfe-dev/2010-March/008252.html 

The opencl implementations that I'm familiar with handle this by making _global be a #define for __attribute__((address_space...

-Chris



> 
> Thanks a lot.
> 
> On Mon, May 10, 2010 at 4:28 PM, John McCall <rjmccall at apple.com> wrote:
> On May 10, 2010, at 11:50 AM, kalyan ponnala wrote:
> > I am trying to add a new qualifier to the existing clang architecture. So far I was looking into replacing one of the existing qualifiers "volatile or restrict" from the "fast qualifiers". Now I think its a bad idea to do that. Can someone tell me
> >
> > >> whats the best way to add a new type quailfier?
> 
> The first question you should ask yourself is whether this is really a qualifier.  Does it need to apply to arbitrary types?  Is it more appropriately modeled as an attribute on specific type declarations, or possibly an independent type kind?  Does it behave like a normal C qualifier, e.g. it is lost during r-value conversion?
> 
> Assuming that's true, read on.
> 
> > >> Why are there only 3 fast qualifiers added to the qualtype?
> 
> The fast qualifiers are mangled into the low bits of the type pointer;  that's why they're fast, because they don't need to be stored separately.  Other qualifiers are stored in a separately-allocated ExtQuals structure.
> 
> If we allocate Types with 8-byte alignment, we have three bits available.  Currently only const and restrict are fast, and the third bit is used to indicate that the type has extended qualifiers.
> 
> > >> How can I increase the size of the fast qualifiers, I mean how to add new qualifier to qualtype without actually disturbing the existing fast quals : const, volatile, restrict.
> 
> 
> I would strongly suggest that you avoid changing the fast-qualifiers representation.  Just add your new qualifier to the Qualifiers structure and make sure the Profile method includes it, and then you'll just need to worry about the N places in the code which assume they know about every qualifier (I consider such code to be buggy, but surprisingly that judgement doesn't by itself fix any bugs :) ).
> 
> John.
> 
> 
> 
> -- 
> Kalyan Ponnala
> phone: 8163772059
> _______________________________________________
> cfe-dev mailing list
> cfe-dev at cs.uiuc.edu
> http://lists.cs.uiuc.edu/mailman/listinfo/cfe-dev

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/cfe-dev/attachments/20100510/e5e1edb8/attachment.html>