[llvm-commits] PATCH: Add a few intrinsic definitions for r600

[Tom Stellard]

On Tue, Mar 20, 2012 at 12:00:07PM -0700, Bill Wendling wrote:
> On Mar 20, 2012, at 7:00 AM, Tom Stellard wrote:
> 
> > On Tue, Mar 20, 2012 at 02:27:21AM -0700, Bill Wendling wrote:
> >> Hi Tom,
> >> 
> >> Excuse my ignorance, but what is r600?
> >> 
> >> -bw
> > 
> > R600 is a target for the previous generation of AMD GPUs
> > (HD2XXX-HD6XXX).  There is an R600 LLVM backend being developed as part
> > of the Mesa project for compiling OpenCL kernels.  Latest code is here:
> > http://cgit.freedesktop.org/~tstellar/mesa/ in the clover-r600-master
> > branch.
> > 
> Thanks for the explanation. I guess my main concern is that we don't have an R600 back-end in the LLVM source tree. I'm hesitant to approve this because of that (others may disagree, though). Are there plans to add such a back-end to LLVM?
> 
> -bw
> 
> 

Hi,

I would really prefer to have the back-end live with the LLVM project,
it would make my life much easier.  However, without bug fix releases
of LLVM (e.g. 3.0.1), I don't think this is practical.

My main concern with this is that we'll want to release a stable version
of Mesa that uses the r600 back-end in LLVM 3.0 (or some other stable LLVM
version), but we'll find bugs in the 3.0 version of the back-end.
At this point we'll have two choices:

1. Fix the bugs in the current SVN development branch of LLVM and release
a stable version of Mesa that depends on an unstable version of LLVM.

2. Release a stable version of Mesa that depends on a stable version of
LLVM, but that has bugs we cannot fix.

Neither of these situations are very desirable.

The r600 back-end is very tightly integrated with the rest of the Mesa's
compute stack.  A single bug in the back-end could render the entire stack
unusable, which is why it is important to be able to release a stable
version of the back-end and a stable version of Mesa at the same time.

The bottom line is: If we make a stable release of Mesa we need to be
able to fix all bugs that turn up, and I don't think this is possible
with the r600 back-end in the LLVM source tree.

Is there any possibility that the LLVM project would start doing bug fix
releases?  Has this ever been discussed in the past?

-Tom