glsl,gallium: add pipe_caps::glsl_bindless_handles_are_32bit

to lower bindless handles to 32 bits before nir_opt_varyings, so that
the high 32 bits of (input) loads of bindless handles are eliminated early.

Reviewed-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/41170>

src/compiler/glsl/gl_nir.h

@@ -34,12 +34,14 @@ struct nir_shader;
 struct gl_constants;
 struct gl_linked_shader;
 struct gl_shader_program;
+struct pipe_caps;
 
 bool gl_nir_lower_atomics(nir_shader *shader,
                           const struct gl_shader_program *shader_program,
                           bool use_binding_as_idx);
 
-bool gl_nir_lower_images(nir_shader *shader, bool bindless_only);
+bool gl_nir_lower_images(nir_shader *shader, const struct pipe_caps *caps,
+                         bool bindless_only);
 bool gl_nir_lower_samplers(nir_shader *shader,
                            const struct gl_shader_program *shader_program);
 bool gl_nir_lower_samplers_as_deref(nir_shader *shader,
@@ -64,8 +66,6 @@ void gl_nir_lower_packed_varyings(const struct gl_constants *consts,
                                   bool disable_varying_packing,
                                   bool disable_xfb_packing, bool xfb_enabled);
 
-void gl_nir_inline_functions(nir_shader *shader);
-
 #ifdef __cplusplus
 }
 #endif

src/compiler/glsl/gl_nir_linker.c

@@ -138,8 +138,8 @@ replace_tex_src(nir_tex_src *dst, nir_tex_src_type src_type, nir_def *src_def,
    list_addtail(&dst->src.use_link, &dst->src.ssa->uses);
 }
 
-void
-gl_nir_inline_functions(nir_shader *shader)
+static void
+gl_nir_inline_functions(const struct pipe_caps *caps, nir_shader *shader)
 {
    /* We have to lower away local constant initializers right before we
     * inline functions.  That way they get properly initialized at the top
@@ -181,6 +181,10 @@ gl_nir_inline_functions(nir_shader *shader)
                if (!nir_deref_mode_is(deref, nir_var_uniform) ||
                    nir_deref_instr_get_variable(deref)->data.bindless) {
                   nir_def *load = nir_load_deref(&b, deref);
+
+                  if (caps->glsl_bindless_handles_are_32bit)
+                     load = nir_u2u32(&b, load);
+
                   replace_tex_src(&intr->src[0], nir_tex_src_texture_handle,
                                   load, instr);
                   replace_tex_src(&intr->src[1], nir_tex_src_sampler_handle,
@@ -1371,7 +1375,7 @@ preprocess_shader(const struct pipe_screen *screen,
    NIR_PASS(_, nir, nir_opt_barrier_modes);
 
    /* before buffers and vars_to_ssa */
-   NIR_PASS(_, nir, gl_nir_lower_images, true);
+   NIR_PASS(_, nir, gl_nir_lower_images, &screen->caps, true);
 
    if (prog->nir->info.stage == MESA_SHADER_COMPUTE ||
        prog->nir->info.stage == MESA_SHADER_TASK ||
@@ -3951,7 +3955,8 @@ gl_nir_link_glsl(struct gl_context *ctx, struct gl_shader_program *prog)
       if (!prog->data->LinkStatus)
          goto done;
 
-      gl_nir_inline_functions(prog->_LinkedShaders[i]->Program->nir);
+      gl_nir_inline_functions(&ctx->screen->caps,
+                              prog->_LinkedShaders[i]->Program->nir);
    }
 
    resize_tes_inputs(consts, prog);

src/compiler/glsl/gl_nir_lower_images.c

@@ -35,6 +35,12 @@
 #include "compiler/nir/nir_deref.h"
 
 #include "compiler/glsl/gl_nir.h"
+#include "pipe/p_defines.h"
+
+typedef struct {
+   const struct pipe_caps *caps;
+   bool bindless_only;
+} lower_images_options;
 
 static void
 type_size_align_1(const struct glsl_type *type, unsigned *size, unsigned *align)
@@ -56,8 +62,7 @@ lower_instr(nir_builder *b, nir_instr *instr, void *cb_data)
    if (instr->type != nir_instr_type_intrinsic)
       return false;
 
-   bool bindless_only = *(bool *)cb_data;
-
+   const lower_images_options *options = (const lower_images_options *)cb_data;
    nir_intrinsic_instr *intrinsic = nir_instr_as_intrinsic(instr);
 
    nir_deref_instr *deref;
@@ -81,7 +86,7 @@ lower_instr(nir_builder *b, nir_instr *instr, void *cb_data)
    }
 
    bool bindless = var->data.mode != nir_var_image || var->data.bindless;
-   if (bindless_only && !bindless)
+   if (options->bindless_only && !bindless)
       return false;
 
    b->cursor = nir_before_instr(instr);
@@ -98,6 +103,10 @@ lower_instr(nir_builder *b, nir_instr *instr, void *cb_data)
                          nir_build_deref_offset(b, deref, type_size_align_1),
                          var->data.driver_location);
    }
+
+   if (bindless && options->caps->glsl_bindless_handles_are_32bit)
+      src = nir_u2u32(b, src);
+
    nir_rewrite_image_intrinsic(intrinsic, src,
       bindless ? nir_image_intrinsic_type_bindless : nir_image_intrinsic_type_default);
    if (!bindless)
@@ -107,9 +116,15 @@ lower_instr(nir_builder *b, nir_instr *instr, void *cb_data)
 }
 
 bool
-gl_nir_lower_images(nir_shader *shader, bool bindless_only)
+gl_nir_lower_images(nir_shader *shader, const struct pipe_caps *caps,
+                    bool bindless_only)
 {
+   lower_images_options options = {
+      .caps = caps,
+      .bindless_only = bindless_only,
+   };
+
    return nir_shader_instructions_pass(shader, lower_instr,
                                        nir_metadata_control_flow,
-                                       &bindless_only);
+                                       &options);
 }

src/gallium/drivers/svga/svga_pipe_cs.c

@@ -41,7 +41,7 @@ svga_create_compute_state(struct pipe_context *pipe,
 
    assert(templ->ir_type == PIPE_SHADER_IR_NIR);
    /* nir_to_tgsi requires lowered images */
-   NIR_PASS(_, nir, gl_nir_lower_images, false);
+   NIR_PASS(_, nir, gl_nir_lower_images, NULL, false);
 
    cs->base.tokens = nir_to_tgsi((void *)nir, pipe->screen);
 

src/gallium/drivers/svga/svga_shader.c

@@ -932,7 +932,7 @@ svga_create_shader(struct pipe_context *pipe,
          .keep_double_immediates = true,
       };
       /* nir_to_tgsi requires lowered images */
-      NIR_PASS(_, nir, gl_nir_lower_images, false);
+      NIR_PASS(_, nir, gl_nir_lower_images, NULL, false);
       shader->tokens = nir_to_tgsi_options(nir, pipe->screen, &ntt_options);
    } else {
       shader->tokens = pipe_shader_state_to_tgsi_tokens(pipe->screen, templ);

src/gallium/include/pipe/p_defines.h

@@ -1062,6 +1062,7 @@ struct pipe_caps {
    bool representative_fragment_test;
    bool prefer_persp;
    bool blit_3d;
+   bool glsl_bindless_handles_are_32bit;
 
    int accelerated;
    int min_texel_offset;

src/mesa/state_tracker/st_glsl_to_nir.cpp

@@ -755,7 +755,7 @@ st_finalize_nir(struct st_context *st, struct gl_program *prog,
 
    st_nir_lower_samplers(screen, nir, shader_program, prog);
    if (!is_draw_shader && !screen->caps.nir_images_as_deref)
-      NIR_PASS(_, nir, gl_nir_lower_images, false);
+      NIR_PASS(_, nir, gl_nir_lower_images, NULL, false);
 }
 
 /**

src/mesa/state_tracker/st_nir_builtins.c

@@ -58,7 +58,7 @@ st_nir_finish_builtin_nir(struct st_context *st, nir_shader *nir)
    st_nir_lower_samplers(screen, nir, NULL, NULL);
    st_nir_lower_uniforms(st, nir);
    if (!screen->caps.nir_images_as_deref)
-      NIR_PASS(_, nir, gl_nir_lower_images, false);
+      NIR_PASS(_, nir, gl_nir_lower_images, NULL, false);
 
    assert(nir->info.stage == MESA_SHADER_COMPUTE || nir->info.io_lowered);
 

src/mesa/state_tracker/st_program.c

@@ -899,7 +899,7 @@ st_create_common_variant(struct st_context *st,
    }
 
    if (key->is_draw_shader) {
-      NIR_PASS(_, state.ir.nir, gl_nir_lower_images, false);
+      NIR_PASS(_, state.ir.nir, gl_nir_lower_images, NULL, false);
       v->base.driver_shader = draw_create_vertex_shader(st->draw, &state);
    }
    else