[llvm-commits] RFC: Add R600/SI backend as an experimental target

[Justin Holewinski]

The back-end builds fine for me!

I haven't done extensive testing, but it looks like there is some issue
with the way registers are defined.  In text output, I am getting
placeholders instead of actual numbers:

jholewinski at rapture [R600]$ llc -march=r600 -mcpu=redwood <
llvm.AMDGPU.pow.ll
.text
; BB#0:
LOG_IEEE T#Index#.#Chan, T#Index#.#Chan (Pred_sel_off)
MUL NON-IEEE T#Index#.#Chan, T#Index#.#Chan, T#Index#.#Chan
EXP_IEEE T#Index#.#Chan, T#Index#.#Chan (Pred_sel_off)
RETURN


On Fri, Sep 7, 2012 at 4:35 PM, Tom Stellard <tom at stellard.net> wrote:

> Hi,
>
> Attached is a patch with an updated version of the R600/SI backend.  The
> same code can also be found in the r600-review-v9 branch here:
> http://cgit.freedesktop.org/~tstellar/llvm/
>
> Changes since last week's version include:
>  +  Lots of fixes and new features for Southern Islands GPUs.
>  +  The *RegisterInfo.td files have been updated to use newer features
>     of tablegen and they no longer need to be generated with perl
>     scripts. This change reduces the size of the patch by around ~100kb.
>
> I'm still hoping to get this accepted as an experimental backend, and I
> look forward to your comments.
>
> Thanks,
> Tom Stellard
>
>
> On Fri, Aug 31, 2012 at 03:18:17PM -0400, Tom Stellard wrote:
> > Hi,
> >
> > I've attached a patch with an updated version of the R600/SI backend.
> > This time I've squashed all the changes into a single patch and used a
> > format that is compatible with svn for easier testing/reviewing.
> >
> > Since the last version, I've fixed some build errors caused by recent
> > tablegen changes, and I've also gone through and fixed all the
> > errors caused by setting InstrInfo.guessInstructionProperties = 0.
> >
> > If you prefer git, this code can also be found in the r600-review-v8
> > branch here: http://cgit.freedesktop.org/~tstellar/llvm/
> >
> > I'm still interested in getting this code into the main tree as an
> > experimental target, so I'd appreciate any feedback on whether or not
> > this backend is ready to be accepted as an experimental target.
> >
> > Thanks,
> > Tom Stellard
> >
> > On Mon, Aug 27, 2012 at 02:46:06PM +0000, Tom Stellard wrote:
> > >
> > > Hi,
> > >
> > > Here is an updated version of the R600/SI backend that I would like to
> > > be considered for inclusion in the LLVM main tree as an experimental[1]
> > > target.
> > >
> > > Changes since the last version include:
> > >   + Added AsmPrinter
> > >   + Converted Tests to use AsmPrinter
> > >   + Replaced CodeEmitters with MCCodeEmitters
> > >   + Removed all uses of MachineOperands flags
> > >   + Enabled the If Conversion pass on R600
> > >   + Various clean ups and bug fixes.
> > >
> > > For more information about the backend, here are some frequently asked
> > > questions:
> > >
> > > ++ What is the R600/SI backend?
> > >
> > >   The R600/SI backend is a code generator for AMD GPUs and supports
> > >   the HD2XXX-HD7XXX GPU models.  It is currently being used by the
> > >   Open Source graphics drivers for AMD GPUs in the Mesa[2] project as
> > >   a compiler backend for graphics and compute shaders.
> > >
> > > ++ How will the R600/SI backend benefit LLVM?
> > >
> > >   The R600/SI backend is a native code generator for GPUs.  There are
> > >   very few completely Open Source compiler backends for GPUs, so
> > >   adding this backend to LLVM will give the project a leg up over other
> > >   compiler projects and help attract the attention of researchers and
> > >   new developers.
> > >
> > > ++ Why do the R600/SI developers want this code in the main LLVM tree?
> > >
> > >   The main reason for this is to reduce the maintenance burden on
> > >   developers so they have more time to spend improving the backend and
> > >   core LLVM code.  Supporting multiple versions of LLVM in the Mesa
> > >   tree will require a lot of ugly ifdefs, and it will be very difficult
> > >   to take advantage of new features in Tablegen.  Also, changes to the
> > >   backend will require testing against at least two version of LLVM.
> > >   Moving the backend code into the main LLVM tree will allow the
> > >   backend to evolve with the rest of the project and remove these
> > >   extra burdens from developers.
> > >
> > > ++ What is the planned development model for the R600/SI backend?
> > >
> > >   If accepted into the main tree, all development on the R600/SI will
> be based
> > >   on the public LLVM tree.  The backend has been Open Source from the
> beginning
> > >   and there are no private internal trees.  All code developed for
> this backend
> > >   is pushed to the public tree as soon as it is ready.
> > >
> > > Looking forward to you comments.
> > >
> > > Thanks,
> > > Tom Stellard
> > >
> > > [1] http://lists.cs.uiuc.edu/pipermail/llvmdev/2012-July/051929.html
> > > [2] http://www.mesa3d.org
> > > _______________________________________________
> > > llvm-commits mailing list
> > > llvm-commits at cs.uiuc.edu
> > > http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits
>
>
> _______________________________________________
> llvm-commits mailing list
> llvm-commits at cs.uiuc.edu
> http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits
>
>


-- 

Thanks,

Justin Holewinski
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20120909/102b02e5/attachment.html>