[Mesa-dev] [PATCH 6/7] nir: fix missing increments of num_inputs/num_outputs

[Jason Ekstrand]

On Fri, Nov 6, 2015 at 5:10 PM, Rob Clark <robdclark at gmail.com> wrote:
> On Fri, Nov 6, 2015 at 6:39 PM, Rob Clark <robdclark at gmail.com> wrote:
>> On Fri, Nov 6, 2015 at 6:23 PM, Jason Ekstrand <jason at jlekstrand.net> wrote:
>>> On Fri, Nov 6, 2015 at 8:35 AM, Rob Clark <robdclark at gmail.com> wrote:
>>>> From: Rob Clark <robclark at freedesktop.org>
>>>>
>>>> Signed-off-by: Rob Clark <robclark at freedesktop.org>
>>>> ---
>>>>  src/glsl/nir/nir_lower_clip.c            | 2 ++
>>>>  src/glsl/nir/nir_lower_two_sided_color.c | 2 ++
>>>>  2 files changed, 4 insertions(+)
>>>>
>>>> diff --git a/src/glsl/nir/nir_lower_clip.c b/src/glsl/nir/nir_lower_clip.c
>>>> index 31ccfb2..4a91527 100644
>>>> --- a/src/glsl/nir/nir_lower_clip.c
>>>> +++ b/src/glsl/nir/nir_lower_clip.c
>>>> @@ -55,9 +55,11 @@ create_clipdist_var(nir_shader *shader, unsigned drvloc,
>>>>
>>>>     if (output) {
>>>>        exec_list_push_tail(&shader->outputs, &var->node);
>>>> +      shader->num_outputs++;
>>>>     }
>>>>     else {
>>>>        exec_list_push_tail(&shader->inputs, &var->node);
>>>> +      shader->num_inputs++;
>>>
>>> I'm not sure what I think about this.  Usually, num_inputs/outputs is
>>> set by nir_lower_io or similar.  They don't have to be vec4's.  In the
>>> i965 driver, FS inputs are in terms of floats.  Maybe tgsi_to_nir
>>> provides you those guarantees but we haven't for i965.
>>
>> hmm, what do you recommend then?  There isn't really any
>> straightforward way to run this *prior* to lower_io... tgsi->nir gives
>> me something that is already i/o lowered, and what I'm doing so far w/
>> gallium support for glsl->nir is doing lower_io (and a few other
>> steps) in gallium so that the result the driver gets is equivalent to
>> tgsi->nir..

Hrm... That does make things a bit sticky.  If you'd like, you can go
ahead and push it with

Akced-by: Jason Ekstrand <jason.ekstrand at intel.com>

> btw Jason, how do you feel about stuffing the type_size fxn ptr in
> nir_shader_compiler_options?  It shouldn't ever change over the
> lifetime of the shader, we could drop it as arg to
> nir_assign_var_locations() (err, well, replace w/ nir_shader ptr), and
> we could use it to dtrt here..

That won't work.  The type_size function we use depends on shader
stage, gen, and type of thing we're lowering.  Stuffing it into
nir_shader_compiler_options won't work.