[Mesa-dev] [PATCH 00/24] RadeonSI moving away from legacy intrinsics & more

[Dave Airlie]

On 26 February 2017 at 09:57, Marek Olšák <maraeo at gmail.com> wrote:
> Hi,
>
> This series depends on the patch adding attributes at call sites:
> https://patchwork.freedesktop.org/series/20100/
>
> The only legacy intrinsics still in use are:
> - llvm.SI.load.const
> - llvm.SI.tbuffer.store (when ADD_TID = 1)
> - llvm.AMDGPU.kill
> - llvm.AMDGPU.kilp
>
> A lot of the code that builds LLVM intrinsics is moved from radeonsi
> to ac_llvm_build.c.
>
> This series also makes use of callsite attributes. It uses readnone for
> loads from read-only memory, and inaccessiblememonly for stores to
> write-only memory, taking into account the GLSL restrict modifier, and
> when read-only memory is implied by the lack of stores in a shader, and
> when write-only memory is implied by the lack of loads. It also sets
> noalias on descriptor arrays. There is not much difference in shader-db
> from these changes.
>
> Lastly, TC L2 is enabled for tessellation offchip stores. It should have
> been enabled from the beginning, but it wasn't.
>
> Please review.

For the following (mostly the ones that touch ac):
1, 2, 3, 4, 5, 6, 7, 9, 13, 14, 16, 18, 20, 22, 23, 24

Reviewed-by: Dave Airlie <airlied at redhat.com>

I can probably ack the radeonsi ones but I don't always understand all the code.

Dave.