[Mesa-dev] [PATCH 2/2] glsl: Only call do_common_optimization once

[Marek Olšák]

Interesting idea. Also added this:

diff --git a/src/compiler/glsl/linker.cpp b/src/compiler/glsl/linker.cpp
index cc28b26..295d10b 100644
--- a/src/compiler/glsl/linker.cpp
+++ b/src/compiler/glsl/linker.cpp
@@ -4985,10 +4985,9 @@ link_shaders(struct gl_context *ctx, struct
gl_shader_program *prog)
          lower_tess_level(prog->_LinkedShaders[i]);
       }

-      while (do_common_optimization(prog->_LinkedShaders[i]->ir, true, false,
-                                    &ctx->Const.ShaderCompilerOptions[i],
-                                    ctx->Const.NativeIntegers))
-         ;
+      do_common_optimization(prog->_LinkedShaders[i]->ir, true, false,
+                            &ctx->Const.ShaderCompilerOptions[i],
+                            ctx->Const.NativeIntegers);

       lower_const_arrays_to_uniforms(prog->_LinkedShaders[i]->ir, i);
       propagate_invariance(prog->_LinkedShaders[i]->ir);
diff --git a/src/mesa/main/ff_fragment_shader.cpp
b/src/mesa/main/ff_fragment_shader.cpp
index 9aa1dc1..3890c54 100644
--- a/src/mesa/main/ff_fragment_shader.cpp
+++ b/src/mesa/main/ff_fragment_shader.cpp
@@ -1251,9 +1251,9 @@ create_new_program(struct gl_context *ctx,
struct state_key *key)
    const struct gl_shader_compiler_options *options =
       &ctx->Const.ShaderCompilerOptions[MESA_SHADER_FRAGMENT];

-   while (do_common_optimization(p.shader->ir, false, false, options,
-                                 ctx->Const.NativeIntegers))
-      ;
+   do_common_optimization(p.shader->ir, false, false, options,
+                         ctx->Const.NativeIntegers);
+
    reparent_ir(p.shader->ir, p.shader->ir);

    p.shader->CompileStatus = true;
------------------

I've tested my private shader-db (26011 shaders, many games). I see
only 2 regressions on radeonsi, both are changes in the code size:
 WORST REGRESSIONS - Code Size
Before     After     Delta Percentage
 shaders/private/alien_isolation/788.shader_test [0]
 8784      9080       296    3.37 %
 shaders/private/unigine_tropics/264.shader_test [0]
 1008      1012         4    0.40 %

That's very good. Drivers not using LLVM might see worse regressions though.

Now, here are the compile time deltas.
- GLSL source -> TGSI: -11.5% time
- GLSL source -> GCN: -6% time

Overall, I like it.

Marek


On Sun, Nov 13, 2016 at 2:29 AM, Thomas Helland
<thomashelland90 at gmail.com> wrote:
> It turns out that only 8% of shaders benefit from more than one run
> of the do_common_optimization function, now that it has the loop.
> Doing only one run makes no difference to instruction count with NIR.
>
> 8324 shaders were run
> 8128 benefit from 1 calls to do_common_optimization
>  679 benefit from 2 calls to do_common_optimization
>   69 benefit from 3 calls to do_common_optimization
> ---
>  src/compiler/glsl/glsl_parser_extras.cpp | 5 ++---
>  1 file changed, 2 insertions(+), 3 deletions(-)
>
> diff --git a/src/compiler/glsl/glsl_parser_extras.cpp b/src/compiler/glsl/glsl_parser_extras.cpp
> index a5a0837..a1a1f90 100644
> --- a/src/compiler/glsl/glsl_parser_extras.cpp
> +++ b/src/compiler/glsl/glsl_parser_extras.cpp
> @@ -1944,9 +1944,8 @@ _mesa_glsl_compile_shader(struct gl_context *ctx, struct gl_shader *shader,
>        /* Do some optimization at compile time to reduce shader IR size
>         * and reduce later work if the same shader is linked multiple times
>         */
> -      while (do_common_optimization(shader->ir, false, false, options,
> -                                    ctx->Const.NativeIntegers))
> -         ;
> +      do_common_optimization(shader->ir, false, false, options,
> +                                    ctx->Const.NativeIntegers);
>
>        validate_ir_tree(shader->ir);
>
> --
> 2.9.3
>
> _______________________________________________
> mesa-dev mailing list
> mesa-dev at lists.freedesktop.org
> https://lists.freedesktop.org/mailman/listinfo/mesa-dev