ir3, freedreno, turnip: Lower io earlier

On the gallium side, set the NIR option to leave IO lowered after
linking when using GLSL. On the turnip side, move up nir_lower_io to
as early as currently possible. Further turnip passes will have to be
converted to intrinsics before we can switch to using the new linker.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/39328>

src/freedreno/ir3/ir3_compiler.c

@@ -137,6 +137,8 @@ static const nir_shader_compiler_options ir3_base_options = {
 
    .per_view_unique_driver_locations = true,
    .compact_view_index = true,
+
+   .io_options = nir_io_has_intrinsics,
 };
 
 struct ir3_compiler *

src/freedreno/ir3/ir3_nir.c

@@ -659,9 +659,36 @@ ir3_finalize_nir(struct ir3_compiler *compiler,
       mesa_logi("----------------------");
    }
 
-   if (s->info.stage == MESA_SHADER_GEOMETRY)
+   /* For vertex inputs, we expect them to all be at the top. FS inputs are also
+    * more optimal at the top.
+    */
+   if (s->info.stage == MESA_SHADER_VERTEX ||
+       s->info.stage == MESA_SHADER_FRAGMENT)
+      NIR_PASS(_, s, nir_opt_move_to_top, nir_move_to_top_input_loads);
+
+   if (s->info.stage == MESA_SHADER_GEOMETRY) {
+      /* nir_unlower_io_to_vars expects constant indirect offsets to be folded
+       * in.
+       */
+      NIR_PASS(_, s, nir_opt_constant_folding);
+      NIR_PASS(_, s, nir_opt_dce);
+
+      /* GS lowering works most easily with variables, so temporarily switch
+       * inputs/outputs to variables and then switch back after the lowering is
+       * done.
+       */
+      NIR_PASS(_, s, nir_unlower_io_to_vars, false);
+      /* nir_lower_io doesn't work with compact variables and non-constant
+       * indices, so clean up output of unlower_io_to_vars.
+       */
+      NIR_PASS(_, s, nir_opt_constant_folding);
+      NIR_PASS(_, s, nir_opt_dce);
+
       NIR_PASS(_, s, ir3_nir_lower_gs);
 
+      ir3_nir_lower_io(s);
+   }
+
    NIR_PASS(_, s, nir_lower_frexp);
    NIR_PASS(_, s, nir_lower_amul, ir3_glsl_type_size);
 

src/freedreno/vulkan/tu_shader.cc

@@ -2949,6 +2949,8 @@ tu_shader_create(struct tu_device *dev,
       NIR_PASS(_, nir, nir_lower_mem_access_bit_sizes, &options);
    }
 
+   ir3_nir_lower_io(nir);
+
    struct ir3_const_allocations const_allocs = {};
    NIR_PASS(_, nir, tu_lower_io, dev, shader, layout,
             key->read_only_input_attachments, key->dynamic_renderpass,
@@ -2961,8 +2963,6 @@ tu_shader_create(struct tu_device *dev,
 
    ir3_finalize_nir(dev->compiler, &nir_options, nir);
 
-   ir3_nir_lower_io(nir);
-
    /* This has to happen after finalizing, so that we know the final bitsize
     * after vectorizing.
     */

src/gallium/drivers/freedreno/ir3/ir3_gallium.c

@@ -461,9 +461,11 @@ ir3_screen_finalize_nir(struct pipe_screen *pscreen, struct nir_shader *nir,
 
    MESA_TRACE_FUNC();
 
-   ir3_nir_lower_io_vars_to_temporaries(nir);
+   if (!nir->info.io_lowered) {
+      ir3_nir_lower_io_vars_to_temporaries(nir);
+      ir3_nir_lower_io(nir);
+   }
    ir3_finalize_nir(screen->compiler, &options, nir);
-   ir3_nir_lower_io(nir);
 }
 
 static void

src/gallium/drivers/zink/ci/zink-tu-a750-fails.txt

@@ -168,9 +168,10 @@ spec@ext_framebuffer_multisample@interpolation 4 non-centroid-deriv-disabled,Fai
 spec@ext_framebuffer_multisample@interpolation 4 non-centroid-disabled,Fail
 spec@ext_framebuffer_multisample@sample-alpha-to-coverage 2 color,Fail
 spec@ext_framebuffer_multisample@sample-alpha-to-coverage 4 color,Fail
-spec@ext_transform_feedback@max-varyings,Fail
-spec@ext_transform_feedback@max-varyings@max-varying-arrays-of-arrays,Fail
-spec@ext_transform_feedback@max-varyings@max-varying-single-dimension-array,Fail
+
+# see https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/37602#note_3294118
+spec@ext_transform_feedback@max-varyings,Crash
+
 spec@ext_transform_feedback@tessellation quad_strip wireframe,Fail
 spec@ext_transform_feedback@tessellation quads wireframe,Fail
 spec@ext_transform_feedback@tessellation triangle_fan flat_first,Fail