[Mesa-dev] [PATCH 10/14] glsl_to_tgsi: remove subroutine support

[Marek Olšák]

On Mon, Oct 17, 2016 at 8:21 PM, Brian Paul <brianp at vmware.com> wrote:
> The series looks OK to me.
>
> Acked-by: Brian Paul <brianp at vmware.com>
>
> I guess I have a general question though.  Over the years of debugging GL
> apps I've seen a few that generate pretty huge shaders (in terms of TGSI
> instructions).  I recall one (can't remember the app) in which a moderate
> size shader compiled into several 10s of thousands of instruction.  The
> cause was all the function call inlining.
>
> I'm wondering if there are cases where not inlining everything and using
> actual subroutines would make sense.  Has anyone ever taken a look into
> that?  Do any modern devices have a limit on program instructions?

No, there is no limit on program instructions since DX10 I think.
Shaders are just memory allocations that end with an "end"
instruction.

I guess compilation takes longer with longer shaders, but function
inlining usually decreases register usage, which increases GPU
occupancy and improves latency hiding, which means better performance.

Marek