mesa/st: add nir pass for lowering builtin uniforms

Signed-off-by: Rob Clark <[email protected]>
Reviewed-by: Eric Anholt <[email protected]>

2 files changed, 276 insertions, 0 deletions

diff --git a/src/mesa/state_tracker/st_nir.h b/src/mesa/state_tracker/st_nir.h
new file mode 100644
index 00000000000..1192981bda7
--- /dev/null
+++ b/src/mesa/state_tracker/st_nir.h
@@ -0,0 +1,31 @@
+/*
+ * Copyright © 2016 Red Hat
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the next
+ * paragraph) shall be included in all copies or substantial portions of the
+ * Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+#ifndef ST_NIR_H
+#define ST_NIR_H
+
+typedef struct nir_shader nir_shader;
+
+void st_nir_lower_builtin(nir_shader *shader);
+
+#endif /* ST_NIR_H */
diff --git a/src/mesa/state_tracker/st_nir_lower_builtin.c b/src/mesa/state_tracker/st_nir_lower_builtin.c
new file mode 100644
index 00000000000..b4da9017d5c
--- /dev/null
+++ b/src/mesa/state_tracker/st_nir_lower_builtin.c
@@ -0,0 +1,245 @@
+/*
+ * Copyright © 2016 Red Hat
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the next
+ * paragraph) shall be included in all copies or substantial portions of the
+ * Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+/* Lowering pass that lowers accesses to built-in uniform variables.
+ * Built-in uniforms are not necessarily packed the same way that
+ * normal uniform structs are, for example:
+ *
+ *    struct gl_FogParameters {
+ *       vec4 color;
+ *       float density;
+ *       float start;
+ *       float end;
+ *       float scale;
+ *    };
+ *
+ * is packed into vec4[2], whereas the same struct would be packed
+ * (by gallium), as vec4[5] if it where not built-in.  Because of
+ * this, we need to replace (for example) access like:
+ *
+ *    vec1 ssa_1 = intrinsic load_var () (gl_Fog.start) ()
+ *
+ * with:
+ *
+ *    vec4 ssa_2 = intrinsic load_var () (fog.params) ()
+ *    vec1 ssa_1 = ssa_2.y
+ *
+ * with appropriate substitutions in the uniform variables list:
+ *
+ *    decl_var uniform INTERP_QUALIFIER_NONE gl_FogParameters gl_Fog (0, 0)
+ *
+ * would become:
+ *
+ *    decl_var uniform INTERP_QUALIFIER_NONE vec4 state.fog.color (0, 0)
+ *    decl_var uniform INTERP_QUALIFIER_NONE vec4 state.fog.params (0, 1)
+ *
+ * See in particular 'struct gl_builtin_uniform_element'.
+ */
+
+#include "compiler/nir/nir.h"
+#include "compiler/nir/nir_builder.h"
+#include "st_nir.h"
+#include "compiler/glsl/ir.h"
+#include "uniforms.h"
+#include "program/prog_instruction.h"
+
+typedef struct {
+   nir_shader *shader;
+   nir_builder builder;
+   void *mem_ctx;
+} lower_builtin_state;
+
+static const struct gl_builtin_uniform_element *
+get_element(const struct gl_builtin_uniform_desc *desc, nir_deref_var *deref)
+{
+   nir_deref *tail = &deref->deref;
+
+   if ((desc->num_elements == 1) && (desc->elements[0].field == NULL))
+      return NULL;
+
+   /* we handle arrays in get_variable(): */
+   if (tail->child->deref_type == nir_deref_type_array)
+      tail = tail->child;
+
+   /* don't need to deal w/ non-struct or array of non-struct: */
+   if (!tail->child)
+      return NULL;
+
+   if (tail->child->deref_type != nir_deref_type_struct)
+      return NULL;
+
+   nir_deref_struct *deref_struct = nir_deref_as_struct(tail->child);
+
+   assert(deref_struct->index < desc->num_elements);
+
+   return &desc->elements[deref_struct->index];
+}
+
+static nir_variable *
+get_variable(lower_builtin_state *state, nir_deref_var *deref,
+             const struct gl_builtin_uniform_element *element)
+{
+   nir_shader *shader = state->shader;
+   int tokens[STATE_LENGTH];
+
+   memcpy(tokens, element->tokens, sizeof(tokens));
+
+   if (deref->deref.child->deref_type == nir_deref_type_array) {
+      nir_deref_array *darr = nir_deref_as_array(deref->deref.child);
+
+      assert(darr->deref_array_type == nir_deref_array_type_direct);
+
+      /* we need to fixup the array index slot: */
+      switch (tokens[0]) {
+      case STATE_MODELVIEW_MATRIX:
+      case STATE_PROJECTION_MATRIX:
+      case STATE_MVP_MATRIX:
+      case STATE_TEXTURE_MATRIX:
+      case STATE_PROGRAM_MATRIX:
+      case STATE_LIGHT:
+      case STATE_LIGHTPROD:
+      case STATE_TEXGEN:
+      case STATE_TEXENV_COLOR:
+      case STATE_CLIPPLANE:
+         tokens[1] = darr->base_offset;
+         break;
+      }
+   }
+
+   char *name = _mesa_program_state_string((gl_state_index *)tokens);
+
+   nir_foreach_variable(var, &shader->uniforms)
+      if (strcmp(var->name, name) == 0)
+         return var;
+
+   /* variable doesn't exist yet, so create it: */
+   nir_variable *var =
+      nir_variable_create(shader, nir_var_uniform, glsl_vec4_type(), name);
+
+   var->num_state_slots = 1;
+   var->state_slots = ralloc_array(var, nir_state_slot, 1);
+   memcpy(var->state_slots[0].tokens, tokens,
+          sizeof(var->state_slots[0].tokens));
+
+   free(name);
+
+   return var;
+}
+
+static bool
+lower_builtin_block(lower_builtin_state *state, nir_block *block)
+{
+   nir_builder *b = &state->builder;
+
+   nir_foreach_instr_safe(instr, block) {
+      if (instr->type != nir_instr_type_intrinsic)
+         continue;
+
+      nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr);
+
+      if (intrin->intrinsic != nir_intrinsic_load_var)
+         continue;
+
+      nir_variable *var = intrin->variables[0]->var;
+      if (var->data.mode != nir_var_uniform)
+         continue;
+
+      /* built-in's will always start with "gl_" */
+      if (strncmp(var->name, "gl_", 3) != 0)
+         continue;
+
+      const struct gl_builtin_uniform_desc *desc =
+         _mesa_glsl_get_builtin_uniform_desc(var->name);
+
+      /* if no descriptor, it isn't something we need to handle specially: */
+      if (!desc)
+         continue;
+
+      const struct gl_builtin_uniform_element *element =
+         get_element(desc, intrin->variables[0]);
+
+      /* matrix elements (array_deref) do not need special handling: */
+      if (!element)
+         continue;
+
+      /* remove existing var from uniform list: */
+      exec_node_remove(&var->node);
+      /* the _self_link() ensures we can remove multiple times, rather than
+       * trying to keep track of what we have already removed:
+       */
+      exec_node_self_link(&var->node);
+
+      nir_variable *new_var =
+         get_variable(state, intrin->variables[0], element);
+
+      b->cursor = nir_before_instr(instr);
+
+      nir_ssa_def *def = nir_load_var(b, new_var);
+
+      /* swizzle the result: */
+      unsigned swiz[4];
+      for (unsigned i = 0; i < 4; i++) {
+         swiz[i] = GET_SWZ(element->swizzle, i);
+         assert(swiz[i] <= SWIZZLE_W);
+      }
+      def = nir_swizzle(b, def, swiz, intrin->num_components, true);
+
+      /* and rewrite uses of original instruction: */
+      assert(intrin->dest.is_ssa);
+      nir_ssa_def_rewrite_uses(&intrin->dest.ssa, nir_src_for_ssa(def));
+
+      /* at this point intrin should be unused.  We need to remove it
+       * (rather than waiting for DCE pass) to avoid dangling reference
+       * to remove'd var.  And we have to remove the original uniform
+       * var since we don't want it to get uniform space allocated.
+       */
+      exec_node_remove(&intrin->instr.node);
+   }
+
+   return true;
+}
+
+static void
+lower_builtin_impl(lower_builtin_state *state, nir_function_impl *impl)
+{
+   nir_builder_init(&state->builder, impl);
+   state->mem_ctx = ralloc_parent(impl);
+
+   nir_foreach_block(block, impl) {
+      lower_builtin_block(state, block);
+   }
+
+   nir_metadata_preserve(impl, nir_metadata_block_index |
+                               nir_metadata_dominance);
+}
+
+void
+st_nir_lower_builtin(nir_shader *shader)
+{
+   lower_builtin_state state;
+   state.shader = shader;
+   nir_foreach_function(function, shader) {
+      if (function->impl)
+         lower_builtin_impl(&state, function->impl);
+   }
+}