nir: Add goto_if jump instruction
[mesa.git] / src / compiler / nir / nir_split_vars.c
index 2ff8257020310a1b132c3017b567651f73147ce3..6aee3109d4121a2b8f94455d4a513dfec97c7234 100644 (file)
 #include "nir_deref.h"
 #include "nir_vla.h"
 
+#include "util/set.h"
 #include "util/u_math.h"
 
+static struct set *
+get_complex_used_vars(nir_shader *shader, void *mem_ctx)
+{
+   struct set *complex_vars = _mesa_pointer_set_create(mem_ctx);
+
+   nir_foreach_function(function, shader) {
+      if (!function->impl)
+         continue;
+
+      nir_foreach_block(block, function->impl) {
+         nir_foreach_instr(instr, block) {
+            if (instr->type != nir_instr_type_deref)
+               continue;
+
+            nir_deref_instr *deref = nir_instr_as_deref(instr);
+
+            /* We only need to consider var derefs because
+             * nir_deref_instr_has_complex_use is recursive.
+             */
+            if (deref->deref_type == nir_deref_type_var &&
+                nir_deref_instr_has_complex_use(deref))
+               _mesa_set_add(complex_vars, deref->var);
+         }
+      }
+   }
+
+   return complex_vars;
+}
 
 struct split_var_state {
    void *mem_ctx;
@@ -127,7 +156,9 @@ static bool
 split_var_list_structs(nir_shader *shader,
                        nir_function_impl *impl,
                        struct exec_list *vars,
+                       nir_variable_mode mode,
                        struct hash_table *var_field_map,
+                       struct set **complex_vars,
                        void *mem_ctx)
 {
    struct split_var_state state = {
@@ -142,15 +173,27 @@ split_var_list_structs(nir_shader *shader,
    /* To avoid list confusion (we'll be adding things as we split variables),
     * pull all of the variables we plan to split off of the list
     */
-   nir_foreach_variable_safe(var, vars) {
+   nir_foreach_variable_in_list_safe(var, vars) {
+      if (var->data.mode != mode)
+         continue;
+
       if (!glsl_type_is_struct_or_ifc(glsl_without_array(var->type)))
          continue;
 
+      if (*complex_vars == NULL)
+         *complex_vars = get_complex_used_vars(shader, mem_ctx);
+
+      /* We can't split a variable that's referenced with deref that has any
+       * sort of complex usage.
+       */
+      if (_mesa_set_search(*complex_vars, var))
+         continue;
+
       exec_node_remove(&var->node);
       exec_list_push_tail(&split_vars, &var->node);
    }
 
-   nir_foreach_variable(var, &split_vars) {
+   nir_foreach_variable_in_list(var, &split_vars) {
       state.base_var = var;
 
       struct field *root_field = ralloc(mem_ctx, struct field);
@@ -258,14 +301,18 @@ nir_split_struct_vars(nir_shader *shader, nir_variable_mode modes)
    void *mem_ctx = ralloc_context(NULL);
    struct hash_table *var_field_map =
       _mesa_pointer_hash_table_create(mem_ctx);
+   struct set *complex_vars = NULL;
 
    assert((modes & (nir_var_shader_temp | nir_var_function_temp)) == modes);
 
    bool has_global_splits = false;
    if (modes & nir_var_shader_temp) {
       has_global_splits = split_var_list_structs(shader, NULL,
-                                                 &shader->globals,
-                                                 var_field_map, mem_ctx);
+                                                 &shader->variables,
+                                                 nir_var_shader_temp,
+                                                 var_field_map,
+                                                 &complex_vars,
+                                                 mem_ctx);
    }
 
    bool progress = false;
@@ -277,7 +324,10 @@ nir_split_struct_vars(nir_shader *shader, nir_variable_mode modes)
       if (modes & nir_var_function_temp) {
          has_local_splits = split_var_list_structs(shader, function->impl,
                                                    &function->impl->locals,
-                                                   var_field_map, mem_ctx);
+                                                   nir_var_function_temp,
+                                                   var_field_map,
+                                                   &complex_vars,
+                                                   mem_ctx);
       }
 
       if (has_global_splits || has_local_splits) {
@@ -287,6 +337,8 @@ nir_split_struct_vars(nir_shader *shader, nir_variable_mode modes)
          nir_metadata_preserve(function->impl, nir_metadata_block_index |
                                                nir_metadata_dominance);
          progress = true;
+      } else {
+         nir_metadata_preserve(function->impl, nir_metadata_all);
       }
    }
 
@@ -321,17 +373,32 @@ struct array_var_info {
 };
 
 static bool
-init_var_list_array_infos(struct exec_list *vars,
+init_var_list_array_infos(nir_shader *shader,
+                          struct exec_list *vars,
+                          nir_variable_mode mode,
                           struct hash_table *var_info_map,
+                          struct set **complex_vars,
                           void *mem_ctx)
 {
    bool has_array = false;
 
-   nir_foreach_variable(var, vars) {
+   nir_foreach_variable_in_list(var, vars) {
+      if (var->data.mode != mode)
+         continue;
+
       int num_levels = num_array_levels_in_array_of_vector_type(var->type);
       if (num_levels <= 0)
          continue;
 
+      if (*complex_vars == NULL)
+         *complex_vars = get_complex_used_vars(shader, mem_ctx);
+
+      /* We can't split a variable that's referenced with deref that has any
+       * sort of complex usage.
+       */
+      if (_mesa_set_search(*complex_vars, var))
+         continue;
+
       struct array_var_info *info =
          rzalloc_size(mem_ctx, sizeof(*info) +
                                num_levels * sizeof(info->levels[0]));
@@ -372,8 +439,11 @@ get_array_deref_info(nir_deref_instr *deref,
    if (!(deref->mode & modes))
       return NULL;
 
-   return get_array_var_info(nir_deref_instr_get_variable(deref),
-                             var_info_map);
+   nir_variable *var = nir_deref_instr_get_variable(deref);
+   if (var == NULL)
+      return NULL;
+
+   return get_array_var_info(var, var_info_map);
 }
 
 static void
@@ -477,13 +547,17 @@ static bool
 split_var_list_arrays(nir_shader *shader,
                       nir_function_impl *impl,
                       struct exec_list *vars,
+                      nir_variable_mode mode,
                       struct hash_table *var_info_map,
                       void *mem_ctx)
 {
    struct exec_list split_vars;
    exec_list_make_empty(&split_vars);
 
-   nir_foreach_variable_safe(var, vars) {
+   nir_foreach_variable_in_list_safe(var, vars) {
+      if (var->data.mode != mode)
+         continue;
+
       struct array_var_info *info = get_array_var_info(var, var_info_map);
       if (!info)
          continue;
@@ -527,7 +601,7 @@ split_var_list_arrays(nir_shader *shader,
       }
    }
 
-   nir_foreach_variable(var, &split_vars) {
+   nir_foreach_variable_in_list(var, &split_vars) {
       struct array_var_info *info = get_array_var_info(var, var_info_map);
       create_split_array_vars(info, 0, &info->root_split, var->name,
                               shader, impl, mem_ctx);
@@ -791,13 +865,18 @@ nir_split_array_vars(nir_shader *shader, nir_variable_mode modes)
 {
    void *mem_ctx = ralloc_context(NULL);
    struct hash_table *var_info_map = _mesa_pointer_hash_table_create(mem_ctx);
+   struct set *complex_vars = NULL;
 
    assert((modes & (nir_var_shader_temp | nir_var_function_temp)) == modes);
 
    bool has_global_array = false;
    if (modes & nir_var_shader_temp) {
-      has_global_array = init_var_list_array_infos(&shader->globals,
-                                                   var_info_map, mem_ctx);
+      has_global_array = init_var_list_array_infos(shader,
+                                                   &shader->variables,
+                                                   nir_var_shader_temp,
+                                                   var_info_map,
+                                                   &complex_vars,
+                                                   mem_ctx);
    }
 
    bool has_any_array = false;
@@ -807,8 +886,12 @@ nir_split_array_vars(nir_shader *shader, nir_variable_mode modes)
 
       bool has_local_array = false;
       if (modes & nir_var_function_temp) {
-         has_local_array = init_var_list_array_infos(&function->impl->locals,
-                                                     var_info_map, mem_ctx);
+         has_local_array = init_var_list_array_infos(shader,
+                                                     &function->impl->locals,
+                                                     nir_var_function_temp,
+                                                     var_info_map,
+                                                     &complex_vars,
+                                                     mem_ctx);
       }
 
       if (has_global_array || has_local_array) {
@@ -820,13 +903,15 @@ nir_split_array_vars(nir_shader *shader, nir_variable_mode modes)
    /* If we failed to find any arrays of arrays, bail early. */
    if (!has_any_array) {
       ralloc_free(mem_ctx);
+      nir_shader_preserve_all_metadata(shader);
       return false;
    }
 
    bool has_global_splits = false;
    if (modes & nir_var_shader_temp) {
       has_global_splits = split_var_list_arrays(shader, NULL,
-                                                &shader->globals,
+                                                &shader->variables,
+                                                nir_var_shader_temp,
                                                 var_info_map, mem_ctx);
    }
 
@@ -839,6 +924,7 @@ nir_split_array_vars(nir_shader *shader, nir_variable_mode modes)
       if (modes & nir_var_function_temp) {
          has_local_splits = split_var_list_arrays(shader, function->impl,
                                                   &function->impl->locals,
+                                                  nir_var_function_temp,
                                                   var_info_map, mem_ctx);
       }
 
@@ -849,6 +935,8 @@ nir_split_array_vars(nir_shader *shader, nir_variable_mode modes)
          nir_metadata_preserve(function->impl, nir_metadata_block_index |
                                                nir_metadata_dominance);
          progress = true;
+      } else {
+         nir_metadata_preserve(function->impl, nir_metadata_all);
       }
    }
 
@@ -880,6 +968,7 @@ struct vec_var_usage {
 
    /* True if there is a copy that isn't to/from a shrinkable vector */
    bool has_external_copy;
+   bool has_complex_use;
    struct set *vars_copied;
 
    unsigned num_levels;
@@ -939,6 +1028,32 @@ get_vec_deref_usage(nir_deref_instr *deref,
                             var_usage_map, add_usage_entry, mem_ctx);
 }
 
+static void
+mark_deref_if_complex(nir_deref_instr *deref,
+                      struct hash_table *var_usage_map,
+                      nir_variable_mode modes,
+                      void *mem_ctx)
+{
+   if (!(deref->mode & modes))
+      return;
+
+   /* Only bother with var derefs because nir_deref_instr_has_complex_use is
+    * recursive.
+    */
+   if (deref->deref_type != nir_deref_type_var)
+      return;
+
+   if (!nir_deref_instr_has_complex_use(deref))
+      return;
+
+   struct vec_var_usage *usage =
+      get_vec_var_usage(deref->var, var_usage_map, true, mem_ctx);
+   if (!usage)
+      return;
+
+   usage->has_complex_use = true;
+}
+
 static void
 mark_deref_used(nir_deref_instr *deref,
                 nir_component_mask_t comps_read,
@@ -952,6 +1067,8 @@ mark_deref_used(nir_deref_instr *deref,
       return;
 
    nir_variable *var = nir_deref_instr_get_variable(deref);
+   if (var == NULL)
+      return;
 
    struct vec_var_usage *usage =
       get_vec_var_usage(var, var_usage_map, true, mem_ctx);
@@ -1072,9 +1189,7 @@ get_non_self_referential_store_comps(nir_intrinsic_instr *store)
                comps &= ~(1u << i);
          }
       }
-   } else if (src_alu->op == nir_op_vec2 ||
-              src_alu->op == nir_op_vec3 ||
-              src_alu->op == nir_op_vec4) {
+   } else if (nir_op_is_vec(src_alu->op)) {
       /* If it's a vec, discount any channels that are just loads from the
        * same deref put in the same spot.
        */
@@ -1096,6 +1211,11 @@ find_used_components_impl(nir_function_impl *impl,
 {
    nir_foreach_block(block, impl) {
       nir_foreach_instr(instr, block) {
+         if (instr->type == nir_instr_type_deref) {
+            mark_deref_if_complex(nir_instr_as_deref(instr),
+                                  var_usage_map, modes, mem_ctx);
+         }
+
          if (instr->type != nir_instr_type_intrinsic)
             continue;
 
@@ -1131,6 +1251,7 @@ find_used_components_impl(nir_function_impl *impl,
 
 static bool
 shrink_vec_var_list(struct exec_list *vars,
+                    nir_variable_mode mode,
                     struct hash_table *var_usage_map)
 {
    /* Initialize the components kept field of each variable.  This is the
@@ -1149,14 +1270,17 @@ shrink_vec_var_list(struct exec_list *vars,
     * Also, if we have a copy that to/from something we can't shrink, we need
     * to leave components and array_len of any wildcards alone.
     */
-   nir_foreach_variable(var, vars) {
+   nir_foreach_variable_in_list(var, vars) {
+      if (var->data.mode != mode)
+         continue;
+
       struct vec_var_usage *usage =
          get_vec_var_usage(var, var_usage_map, false, NULL);
       if (!usage)
          continue;
 
       assert(usage->comps_kept == 0);
-      if (usage->has_external_copy)
+      if (usage->has_external_copy || usage->has_complex_use)
          usage->comps_kept = usage->all_comps;
       else
          usage->comps_kept = usage->comps_read & usage->comps_written;
@@ -1165,7 +1289,8 @@ shrink_vec_var_list(struct exec_list *vars,
          struct array_level_usage *level = &usage->levels[i];
          assert(level->array_len > 0);
 
-         if (level->max_written == UINT_MAX || level->has_external_copy)
+         if (level->max_written == UINT_MAX || level->has_external_copy ||
+             usage->has_complex_use)
             continue; /* Can't shrink */
 
          unsigned max_used = MIN2(level->max_read, level->max_written);
@@ -1181,7 +1306,10 @@ shrink_vec_var_list(struct exec_list *vars,
    bool fp_progress;
    do {
       fp_progress = false;
-      nir_foreach_variable(var, vars) {
+      nir_foreach_variable_in_list(var, vars) {
+         if (var->data.mode != mode)
+            continue;
+
          struct vec_var_usage *var_usage =
             get_vec_var_usage(var, var_usage_map, false, NULL);
          if (!var_usage || !var_usage->vars_copied)
@@ -1218,7 +1346,10 @@ shrink_vec_var_list(struct exec_list *vars,
    } while (fp_progress);
 
    bool vars_shrunk = false;
-   nir_foreach_variable_safe(var, vars) {
+   nir_foreach_variable_in_list_safe(var, vars) {
+      if (var->data.mode != mode)
+         continue;
+
       struct vec_var_usage *usage =
          get_vec_var_usage(var, var_usage_map, false, NULL);
       if (!usage)
@@ -1487,8 +1618,11 @@ function_impl_has_vars_with_modes(nir_function_impl *impl,
 {
    nir_shader *shader = impl->function->shader;
 
-   if ((modes & nir_var_shader_temp) && !exec_list_is_empty(&shader->globals))
-      return true;
+   if (modes & ~nir_var_function_temp) {
+      nir_foreach_variable_with_modes(var, shader,
+                                      modes & ~nir_var_function_temp)
+         return true;
+   }
 
    if ((modes & nir_var_function_temp) && !exec_list_is_empty(&impl->locals))
       return true;
@@ -1532,12 +1666,16 @@ nir_shrink_vec_array_vars(nir_shader *shader, nir_variable_mode modes)
    }
    if (!has_vars_to_shrink) {
       ralloc_free(mem_ctx);
+      nir_shader_preserve_all_metadata(shader);
       return false;
    }
 
    bool globals_shrunk = false;
-   if (modes & nir_var_shader_temp)
-      globals_shrunk = shrink_vec_var_list(&shader->globals, var_usage_map);
+   if (modes & nir_var_shader_temp) {
+      globals_shrunk = shrink_vec_var_list(&shader->variables,
+                                           nir_var_shader_temp,
+                                           var_usage_map);
+   }
 
    bool progress = false;
    nir_foreach_function(function, shader) {
@@ -1547,6 +1685,7 @@ nir_shrink_vec_array_vars(nir_shader *shader, nir_variable_mode modes)
       bool locals_shrunk = false;
       if (modes & nir_var_function_temp) {
          locals_shrunk = shrink_vec_var_list(&function->impl->locals,
+                                             nir_var_function_temp,
                                              var_usage_map);
       }
 
@@ -1556,6 +1695,8 @@ nir_shrink_vec_array_vars(nir_shader *shader, nir_variable_mode modes)
          nir_metadata_preserve(function->impl, nir_metadata_block_index |
                                                nir_metadata_dominance);
          progress = true;
+      } else {
+         nir_metadata_preserve(function->impl, nir_metadata_all);
       }
    }