X-Git-Url: https://git.libre-soc.org/?a=blobdiff_plain;f=src%2Fglsl%2Flinker.cpp;h=3d1b8a2f70368e07dd10399630875f5401dfe5b9;hb=42fd9c2ebb8a9eecf69d6a7a2bc0d778f49285de;hp=c5c8c9cdd6382fc2b5d1eb6a2110d5353dd5b362;hpb=bfd7c9ac228c7ed8aec04c3b3aa33f40ee00b035;p=mesa.git

diff --git a/src/glsl/linker.cpp b/src/glsl/linker.cpp
index c5c8c9cdd63..3d1b8a2f703 100644
--- a/src/glsl/linker.cpp
+++ b/src/glsl/linker.cpp
@@ -68,10 +68,6 @@
 #include <cstdarg>
 #include <climits>
 
-extern "C" {
-#include <talloc.h>
-}
-
 #include "main/core.h"
 #include "glsl_symbol_table.h"
 #include "ir.h"
@@ -80,6 +76,10 @@ extern "C" {
 #include "linker.h"
 #include "ir_optimization.h"
 
+extern "C" {
+#include "main/shaderobj.h"
+}
+
 /**
  * Visitor that determines whether or not a variable is ever written.
  */
@@ -172,9 +172,9 @@ linker_error_printf(gl_shader_program *prog, const char *fmt, ...)
 {
    va_list ap;
 
-   prog->InfoLog = talloc_strdup_append(prog->InfoLog, "error: ");
+   ralloc_strcat(&prog->InfoLog, "error: ");
    va_start(ap, fmt);
-   prog->InfoLog = talloc_vasprintf_append(prog->InfoLog, fmt, ap);
+   ralloc_vasprintf_append(&prog->InfoLog, fmt, ap);
    va_end(ap);
 }
 
@@ -191,7 +191,7 @@ invalidate_variable_locations(gl_shader *sh, enum ir_variable_mode mode,
 
       /* Only assign locations for generic attributes / varyings / etc.
        */
-      if (var->location >= generic_base)
+      if ((var->location >= generic_base) && !var->explicit_location)
 	  var->location = -1;
    }
 }
@@ -321,6 +321,9 @@ cross_validate_globals(struct gl_shader_program *prog,
     */
    glsl_symbol_table variables;
    for (unsigned i = 0; i < num_shaders; i++) {
+      if (shader_list[i] == NULL)
+	 continue;
+
       foreach_list(node, shader_list[i]->ir) {
 	 ir_variable *const var = ((ir_instruction *) node)->as_variable();
 
@@ -343,14 +346,68 @@ cross_validate_globals(struct gl_shader_program *prog,
 	 ir_variable *const existing = variables.get_variable(var->name);
 	 if (existing != NULL) {
 	    if (var->type != existing->type) {
-	       linker_error_printf(prog, "%s `%s' declared as type "
-				   "`%s' and type `%s'\n",
-				   mode_string(var),
-				   var->name, var->type->name,
-				   existing->type->name);
-	       return false;
+	       /* Consider the types to be "the same" if both types are arrays
+		* of the same type and one of the arrays is implicitly sized.
+		* In addition, set the type of the linked variable to the
+		* explicitly sized array.
+		*/
+	       if (var->type->is_array()
+		   && existing->type->is_array()
+		   && (var->type->fields.array == existing->type->fields.array)
+		   && ((var->type->length == 0)
+		       || (existing->type->length == 0))) {
+		  if (var->type->length != 0) {
+		     existing->type = var->type;
+		  }
+	       } else {
+		  linker_error_printf(prog, "%s `%s' declared as type "
+				      "`%s' and type `%s'\n",
+				      mode_string(var),
+				      var->name, var->type->name,
+				      existing->type->name);
+		  return false;
+	       }
 	    }
 
+	    if (var->explicit_location) {
+	       if (existing->explicit_location
+		   && (var->location != existing->location)) {
+		     linker_error_printf(prog, "explicit locations for %s "
+					 "`%s' have differing values\n",
+					 mode_string(var), var->name);
+		     return false;
+	       }
+
+	       existing->location = var->location;
+	       existing->explicit_location = true;
+	    }
+
+        /* Validate layout qualifiers for gl_FragDepth.
+         *
+         * From the AMD_conservative_depth spec:
+         *    "If gl_FragDepth is redeclared in any fragment shader in
+         *    a program, it must be redeclared in all fragment shaders in that
+         *    program that have static assignments to gl_FragDepth. All
+         *    redeclarations of gl_FragDepth in all fragment shaders in
+         *    a single program must have the same set of qualifiers."
+         */
+        if (strcmp(var->name, "gl_FragDepth") == 0) {
+           bool layout_declared = var->depth_layout != ir_depth_layout_none;
+           bool layout_differs = var->depth_layout != existing->depth_layout;
+           if (layout_declared && layout_differs) {
+              linker_error_printf(prog,
+                 "All redeclarations of gl_FragDepth in all fragment shaders "
+                 "in a single program must have the same set of qualifiers.");
+           }
+           if (var->used && layout_differs) {
+              linker_error_printf(prog,
+                    "If gl_FragDepth is redeclared with a layout qualifier in"
+                    "any fragment shader, it must be redeclared with the same"
+                    "layout qualifier in all fragment shaders that have"
+                    "assignments to gl_FragDepth");
+           }
+        }
+
 	    /* FINISHME: Handle non-constant initializers.
 	     */
 	    if (var->constant_value != NULL) {
@@ -375,10 +432,23 @@ cross_validate_globals(struct gl_shader_program *prog,
 		   * FINISHME: will fail.
 		   */
 		  existing->constant_value =
-		     var->constant_value->clone(talloc_parent(existing), NULL);
+		     var->constant_value->clone(ralloc_parent(existing), NULL);
+	    }
+
+	    if (existing->invariant != var->invariant) {
+	       linker_error_printf(prog, "declarations for %s `%s' have "
+	                           "mismatching invariant qualifiers\n",
+	                           mode_string(var), var->name);
+	       return false;
 	    }
+            if (existing->centroid != var->centroid) {
+               linker_error_printf(prog, "declarations for %s `%s' have "
+                                   "mismatching centroid qualifiers\n",
+                                   mode_string(var), var->name);
+               return false;
+            }
 	 } else
-	    variables.add_variable(var->name, var);
+	    variables.add_variable(var);
       }
    }
 
@@ -393,7 +463,7 @@ bool
 cross_validate_uniforms(struct gl_shader_program *prog)
 {
    return cross_validate_globals(prog, prog->_LinkedShaders,
-				 prog->_NumLinkedShaders, true);
+				 MESA_SHADER_TYPES, true);
 }
 
 
@@ -420,7 +490,7 @@ cross_validate_outputs_to_inputs(struct gl_shader_program *prog,
       if ((var == NULL) || (var->mode != ir_var_out))
 	 continue;
 
-      parameters.add_variable(var->name, var);
+      parameters.add_variable(var);
    }
 
 
@@ -442,14 +512,35 @@ cross_validate_outputs_to_inputs(struct gl_shader_program *prog,
 	 /* Check that the types match between stages.
 	  */
 	 if (input->type != output->type) {
-	    linker_error_printf(prog,
-				"%s shader output `%s' delcared as "
-				"type `%s', but %s shader input declared "
-				"as type `%s'\n",
-				producer_stage, output->name,
-				output->type->name,
-				consumer_stage, input->type->name);
-	    return false;
+	    /* There is a bit of a special case for gl_TexCoord.  This
+	     * built-in is unsized by default.  Appliations that variable
+	     * access it must redeclare it with a size.  There is some
+	     * language in the GLSL spec that implies the fragment shader
+	     * and vertex shader do not have to agree on this size.  Other
+	     * driver behave this way, and one or two applications seem to
+	     * rely on it.
+	     *
+	     * Neither declaration needs to be modified here because the array
+	     * sizes are fixed later when update_array_sizes is called.
+	     *
+	     * From page 48 (page 54 of the PDF) of the GLSL 1.10 spec:
+	     *
+	     *     "Unlike user-defined varying variables, the built-in
+	     *     varying variables don't have a strict one-to-one
+	     *     correspondence between the vertex language and the
+	     *     fragment language."
+	     */
+	    if (!output->type->is_array()
+		|| (strncmp("gl_", output->name, 3) != 0)) {
+	       linker_error_printf(prog,
+				   "%s shader output `%s' declared as "
+				   "type `%s', but %s shader input declared "
+				   "as type `%s'\n",
+				   producer_stage, output->name,
+				   output->type->name,
+				   consumer_stage, input->type->name);
+	       return false;
+	    }
 	 }
 
 	 /* Check that all of the qualifiers match between stages.
@@ -512,9 +603,9 @@ populate_symbol_table(gl_shader *sh)
       ir_function *func;
 
       if ((func = inst->as_function()) != NULL) {
-	 sh->symbols->add_function(func->name, func);
+	 sh->symbols->add_function(func);
       } else if ((var = inst->as_variable()) != NULL) {
-	 sh->symbols->add_variable(var->name, var);
+	 sh->symbols->add_variable(var);
       }
    }
 }
@@ -571,7 +662,7 @@ remap_variables(ir_instruction *inst, struct gl_shader *target,
 	 else {
 	    ir_variable *copy = ir->var->clone(this->target, NULL);
 
-	    this->symbols->add_variable(copy->name, copy);
+	    this->symbols->add_variable(copy);
 	    this->instructions->push_head(copy);
 	    ir->var = copy;
 	 }
@@ -692,7 +783,8 @@ get_main_function_signature(gl_shader *sh)
  * shader is returned.
  */
 static struct gl_shader *
-link_intrastage_shaders(GLcontext *ctx,
+link_intrastage_shaders(void *mem_ctx,
+			struct gl_context *ctx,
 			struct gl_shader_program *prog,
 			struct gl_shader **shader_list,
 			unsigned num_shaders)
@@ -726,14 +818,14 @@ link_intrastage_shaders(GLcontext *ctx,
 	       ir_function_signature *sig =
 		  (ir_function_signature *) iter.get();
 
-	       if (!sig->is_defined || sig->is_built_in)
+	       if (!sig->is_defined || sig->is_builtin)
 		  continue;
 
 	       ir_function_signature *other_sig =
 		  other->exact_matching_signature(& sig->parameters);
 
 	       if ((other_sig != NULL) && other_sig->is_defined
-		   && !other_sig->is_built_in) {
+		   && !other_sig->is_builtin) {
 		  linker_error_printf(prog,
 				      "function `%s' is multiply defined",
 				      f->name);
@@ -766,9 +858,9 @@ link_intrastage_shaders(GLcontext *ctx,
       return NULL;
    }
 
-   gl_shader *const linked = ctx->Driver.NewShader(NULL, 0, main->Type);
+   gl_shader *linked = ctx->Driver.NewShader(NULL, 0, main->Type);
    linked->ir = new(linked) exec_list;
-   clone_ir_list(linked, linked->ir, main->ir);
+   clone_ir_list(mem_ctx, linked->ir, main->ir);
 
    populate_symbol_table(linked);
 
@@ -813,10 +905,39 @@ link_intrastage_shaders(GLcontext *ctx,
 
    assert(idx == num_linking_shaders);
 
-   link_function_calls(prog, linked, linking_shaders, num_linking_shaders);
+   if (!link_function_calls(prog, linked, linking_shaders,
+			    num_linking_shaders)) {
+      ctx->Driver.DeleteShader(ctx, linked);
+      linked = NULL;
+   }
 
    free(linking_shaders);
 
+   /* Make a pass over all variable declarations to ensure that arrays with
+    * unspecified sizes have a size specified.  The size is inferred from the
+    * max_array_access field.
+    */
+   if (linked != NULL) {
+      class array_sizing_visitor : public ir_hierarchical_visitor {
+      public:
+	 virtual ir_visitor_status visit(ir_variable *var)
+	 {
+	    if (var->type->is_array() && (var->type->length == 0)) {
+	       const glsl_type *type =
+		  glsl_type::get_array_instance(var->type->fields.array,
+						var->max_array_access);
+
+	       assert(type != NULL);
+	       var->type = type;
+	    }
+
+	    return visit_continue;
+	 }
+      } v;
+
+      v.run(linked->ir);
+   }
+
    return linked;
 }
 
@@ -843,18 +964,26 @@ struct uniform_node {
 
  */
 static void
-update_uniform_array_sizes(struct gl_shader_program *prog)
+update_array_sizes(struct gl_shader_program *prog)
 {
-   for (unsigned i = 0; i < prog->_NumLinkedShaders; i++) {
+   for (unsigned i = 0; i < MESA_SHADER_TYPES; i++) {
+	 if (prog->_LinkedShaders[i] == NULL)
+	    continue;
+
       foreach_list(node, prog->_LinkedShaders[i]->ir) {
 	 ir_variable *const var = ((ir_instruction *) node)->as_variable();
 
-	 if ((var == NULL) || (var->mode != ir_var_uniform) ||
+	 if ((var == NULL) || (var->mode != ir_var_uniform &&
+			       var->mode != ir_var_in &&
+			       var->mode != ir_var_out) ||
 	     !var->type->is_array())
 	    continue;
 
 	 unsigned int size = var->max_array_access;
-	 for (unsigned j = 0; j < prog->_NumLinkedShaders; j++) {
+	 for (unsigned j = 0; j < MESA_SHADER_TYPES; j++) {
+	       if (prog->_LinkedShaders[j] == NULL)
+		  continue;
+
 	    foreach_list(node2, prog->_LinkedShaders[j]->ir) {
 	       ir_variable *other_var = ((ir_instruction *) node2)->as_variable();
 	       if (!other_var)
@@ -866,6 +995,7 @@ update_uniform_array_sizes(struct gl_shader_program *prog)
 	       }
 	    }
 	 }
+
 	 if (size + 1 != var->type->fields.array->length) {
 	    var->type = glsl_type::get_array_instance(var->type->fields.array,
 						      size + 1);
@@ -877,6 +1007,84 @@ update_uniform_array_sizes(struct gl_shader_program *prog)
    }
 }
 
+static void
+add_uniform(void *mem_ctx, exec_list *uniforms, struct hash_table *ht,
+	    const char *name, const glsl_type *type, GLenum shader_type,
+	    unsigned *next_shader_pos, unsigned *total_uniforms)
+{
+   if (type->is_record()) {
+      for (unsigned int i = 0; i < type->length; i++) {
+	 const glsl_type *field_type = type->fields.structure[i].type;
+	 char *field_name = ralloc_asprintf(mem_ctx, "%s.%s", name,
+					    type->fields.structure[i].name);
+
+	 add_uniform(mem_ctx, uniforms, ht, field_name, field_type,
+		     shader_type, next_shader_pos, total_uniforms);
+      }
+   } else {
+      uniform_node *n = (uniform_node *) hash_table_find(ht, name);
+      unsigned int vec4_slots;
+      const glsl_type *array_elem_type = NULL;
+
+      if (type->is_array()) {
+	 array_elem_type = type->fields.array;
+	 /* Array of structures. */
+	 if (array_elem_type->is_record()) {
+	    for (unsigned int i = 0; i < type->length; i++) {
+	       char *elem_name = ralloc_asprintf(mem_ctx, "%s[%d]", name, i);
+	       add_uniform(mem_ctx, uniforms, ht, elem_name, array_elem_type,
+			   shader_type, next_shader_pos, total_uniforms);
+	    }
+	    return;
+	 }
+      }
+
+      /* Fix the storage size of samplers at 1 vec4 each. Be sure to pad out
+       * vectors to vec4 slots.
+       */
+      if (type->is_array()) {
+	 if (array_elem_type->is_sampler())
+	    vec4_slots = type->length;
+	 else
+	    vec4_slots = type->length * array_elem_type->matrix_columns;
+      } else if (type->is_sampler()) {
+	 vec4_slots = 1;
+      } else {
+	 vec4_slots = type->matrix_columns;
+      }
+
+      if (n == NULL) {
+	 n = (uniform_node *) calloc(1, sizeof(struct uniform_node));
+	 n->u = (gl_uniform *) calloc(1, sizeof(struct gl_uniform));
+	 n->slots = vec4_slots;
+
+	 n->u->Name = strdup(name);
+	 n->u->Type = type;
+	 n->u->VertPos = -1;
+	 n->u->FragPos = -1;
+	 n->u->GeomPos = -1;
+	 (*total_uniforms)++;
+
+	 hash_table_insert(ht, n, name);
+	 uniforms->push_tail(& n->link);
+      }
+
+      switch (shader_type) {
+      case GL_VERTEX_SHADER:
+	 n->u->VertPos = *next_shader_pos;
+	 break;
+      case GL_FRAGMENT_SHADER:
+	 n->u->FragPos = *next_shader_pos;
+	 break;
+      case GL_GEOMETRY_SHADER:
+	 n->u->GeomPos = *next_shader_pos;
+	 break;
+      }
+
+      (*next_shader_pos) += vec4_slots;
+   }
+}
+
 void
 assign_uniform_locations(struct gl_shader_program *prog)
 {
@@ -885,10 +1093,12 @@ assign_uniform_locations(struct gl_shader_program *prog)
    unsigned total_uniforms = 0;
    hash_table *ht = hash_table_ctor(32, hash_table_string_hash,
 				    hash_table_string_compare);
+   void *mem_ctx = ralloc_context(NULL);
 
-   update_uniform_array_sizes(prog);
+   for (unsigned i = 0; i < MESA_SHADER_TYPES; i++) {
+      if (prog->_LinkedShaders[i] == NULL)
+	 continue;
 
-   for (unsigned i = 0; i < prog->_NumLinkedShaders; i++) {
       unsigned next_position = 0;
 
       foreach_list(node, prog->_LinkedShaders[i]->ir) {
@@ -897,54 +1107,23 @@ assign_uniform_locations(struct gl_shader_program *prog)
 	 if ((var == NULL) || (var->mode != ir_var_uniform))
 	    continue;
 
-	 const unsigned vec4_slots = (var->component_slots() + 3) / 4;
-	 if (vec4_slots == 0) {
-	    /* If we've got a sampler or an aggregate of them, the size can
-	     * end up zero.  Don't allocate any space.
+	 if (strncmp(var->name, "gl_", 3) == 0) {
+	    /* At the moment, we don't allocate uniform locations for
+	     * builtin uniforms.  It's permitted by spec, and we'll
+	     * likely switch to doing that at some point, but not yet.
 	     */
 	    continue;
 	 }
 
-	 uniform_node *n = (uniform_node *) hash_table_find(ht, var->name);
-	 if (n == NULL) {
-	    n = (uniform_node *) calloc(1, sizeof(struct uniform_node));
-	    n->u = (gl_uniform *) calloc(vec4_slots, sizeof(struct gl_uniform));
-	    n->slots = vec4_slots;
-
-	    n->u[0].Name = strdup(var->name);
-	    for (unsigned j = 1; j < vec4_slots; j++)
-	       n->u[j].Name = strdup(var->name);
-
-	    hash_table_insert(ht, n, n->u[0].Name);
-	    uniforms.push_tail(& n->link);
-	    total_uniforms += vec4_slots;
-	 }
-
-	 if (var->constant_value != NULL)
-	    for (unsigned j = 0; j < vec4_slots; j++)
-	       n->u[j].Initialized = true;
-
 	 var->location = next_position;
-
-	 for (unsigned j = 0; j < vec4_slots; j++) {
-	    switch (prog->_LinkedShaders[i]->Type) {
-	    case GL_VERTEX_SHADER:
-	       n->u[j].VertPos = next_position;
-	       break;
-	    case GL_FRAGMENT_SHADER:
-	       n->u[j].FragPos = next_position;
-	       break;
-	    case GL_GEOMETRY_SHADER:
-	       /* FINISHME: Support geometry shaders. */
-	       assert(prog->_LinkedShaders[i]->Type != GL_GEOMETRY_SHADER);
-	       break;
-	    }
-
-	    next_position++;
-	 }
+	 add_uniform(mem_ctx, &uniforms, ht, var->name, var->type,
+		     prog->_LinkedShaders[i]->Type,
+		     &next_position, &total_uniforms);
       }
    }
 
+   ralloc_free(mem_ctx);
+
    gl_uniform_list *ul = (gl_uniform_list *)
       calloc(1, sizeof(gl_uniform_list));
 
@@ -960,8 +1139,8 @@ assign_uniform_locations(struct gl_shader_program *prog)
       next = (uniform_node *) node->link.next;
 
       node->link.remove();
-      memcpy(&ul->Uniforms[idx], node->u, sizeof(gl_uniform) * node->slots);
-      idx += node->slots;
+      memcpy(&ul->Uniforms[idx], node->u, sizeof(gl_uniform));
+      idx++;
 
       free(node->u);
       free(node);
@@ -1123,6 +1302,24 @@ assign_attribute_locations(gl_shader_program *prog, unsigned max_attribute_index
       if ((var == NULL) || (var->mode != ir_var_in))
 	 continue;
 
+      if (var->explicit_location) {
+	 const unsigned slots = count_attribute_slots(var->type);
+	 const unsigned use_mask = (1 << slots) - 1;
+	 const int attr = var->location - VERT_ATTRIB_GENERIC0;
+
+	 if ((var->location >= (int)(max_attribute_index + VERT_ATTRIB_GENERIC0))
+	     || (var->location < 0)) {
+	    linker_error_printf(prog,
+				"invalid explicit location %d specified for "
+				"`%s'\n",
+				(var->location < 0) ? var->location : attr,
+				var->name);
+	    return false;
+	 } else if (var->location >= VERT_ATTRIB_GENERIC0) {
+	    used_locations |= (use_mask << attr);
+	 }
+      }
+
       /* The location was explicitly assigned, nothing to do here.
        */
       if (var->location != -1)
@@ -1176,19 +1373,20 @@ assign_attribute_locations(gl_shader_program *prog, unsigned max_attribute_index
 
 
 /**
- * Demote shader outputs that are not read to being just plain global variables
+ * Demote shader inputs and outputs that are not used in other stages
  */
 void
-demote_unread_shader_outputs(gl_shader *sh)
+demote_shader_inputs_and_outputs(gl_shader *sh, enum ir_variable_mode mode)
 {
    foreach_list(node, sh->ir) {
       ir_variable *const var = ((ir_instruction *) node)->as_variable();
 
-      if ((var == NULL) || (var->mode != ir_var_out))
+      if ((var == NULL) || (var->mode != int(mode)))
 	 continue;
 
-      /* An 'out' variable is only really a shader output if its value is read
-       * by the following stage.
+      /* A shader 'in' or 'out' variable is only really an input or output if
+       * its value is used by other shader stages.  This will cause the variable
+       * to have a location assigned.
        */
       if (var->location == -1) {
 	 var->mode = ir_var_auto;
@@ -1234,18 +1432,25 @@ assign_varying_locations(struct gl_shader_program *prog,
 
       assert(input_var->location == -1);
 
-      /* FINISHME: Location assignment will need some changes when arrays,
-       * FINISHME: matrices, and structures are allowed as shader inputs /
-       * FINISHME: outputs.
-       */
       output_var->location = output_index;
       input_var->location = input_index;
 
-      output_index++;
-      input_index++;
-   }
+      /* FINISHME: Support for "varying" records in GLSL 1.50. */
+      assert(!output_var->type->is_record());
+
+      if (output_var->type->is_array()) {
+	 const unsigned slots = output_var->type->length
+	    * output_var->type->fields.array->matrix_columns;
 
-   demote_unread_shader_outputs(producer);
+	 output_index += slots;
+	 input_index += slots;
+      } else {
+	 const unsigned slots = output_var->type->matrix_columns;
+
+	 output_index += slots;
+	 input_index += slots;
+      }
+   }
 
    foreach_list(node, consumer->ir) {
       ir_variable *const var = ((ir_instruction *) node)->as_variable();
@@ -1283,16 +1488,18 @@ assign_varying_locations(struct gl_shader_program *prog,
 
 
 void
-link_shaders(GLcontext *ctx, struct gl_shader_program *prog)
+link_shaders(struct gl_context *ctx, struct gl_shader_program *prog)
 {
+   void *mem_ctx = ralloc_context(NULL); // temporary linker context
+
    prog->LinkStatus = false;
    prog->Validated = false;
    prog->_Used = false;
 
    if (prog->InfoLog != NULL)
-      talloc_free(prog->InfoLog);
+      ralloc_free(prog->InfoLog);
 
-   prog->InfoLog = talloc_strdup(NULL, "");
+   prog->InfoLog = ralloc_strdup(NULL, "");
 
    /* Separate the shaders into groups based on their type.
     */
@@ -1331,9 +1538,10 @@ link_shaders(GLcontext *ctx, struct gl_shader_program *prog)
     * match shading language versions.  With GLSL 1.30 and later, the versions
     * of all shaders must match.
     */
-   assert(min_version >= 110);
+   assert(min_version >= 100);
    assert(max_version <= 130);
-   if ((max_version >= 130) && (min_version != max_version)) {
+   if ((max_version >= 130 || min_version == 100)
+       && min_version != max_version) {
       linker_error_printf(prog, "all shaders must use same shading "
 			  "language version\n");
       goto done;
@@ -1341,39 +1549,43 @@ link_shaders(GLcontext *ctx, struct gl_shader_program *prog)
 
    prog->Version = max_version;
 
-   for (unsigned int i = 0; i < prog->_NumLinkedShaders; i++) {
-      ctx->Driver.DeleteShader(ctx, prog->_LinkedShaders[i]);
+   for (unsigned int i = 0; i < MESA_SHADER_TYPES; i++) {
+      if (prog->_LinkedShaders[i] != NULL)
+	 ctx->Driver.DeleteShader(ctx, prog->_LinkedShaders[i]);
+
+      prog->_LinkedShaders[i] = NULL;
    }
 
    /* Link all shaders for a particular stage and validate the result.
     */
-   prog->_NumLinkedShaders = 0;
    if (num_vert_shaders > 0) {
       gl_shader *const sh =
-	 link_intrastage_shaders(ctx, prog, vert_shader_list, num_vert_shaders);
+	 link_intrastage_shaders(mem_ctx, ctx, prog, vert_shader_list,
+				 num_vert_shaders);
 
       if (sh == NULL)
 	 goto done;
 
       if (!validate_vertex_shader_executable(prog, sh))
-	  goto done;
+	 goto done;
 
-      prog->_LinkedShaders[prog->_NumLinkedShaders] = sh;
-      prog->_NumLinkedShaders++;
+      _mesa_reference_shader(ctx, &prog->_LinkedShaders[MESA_SHADER_VERTEX],
+			     sh);
    }
 
    if (num_frag_shaders > 0) {
       gl_shader *const sh =
-	 link_intrastage_shaders(ctx, prog, frag_shader_list, num_frag_shaders);
+	 link_intrastage_shaders(mem_ctx, ctx, prog, frag_shader_list,
+				 num_frag_shaders);
 
       if (sh == NULL)
 	 goto done;
 
       if (!validate_fragment_shader_executable(prog, sh))
-	  goto done;
+	 goto done;
 
-      prog->_LinkedShaders[prog->_NumLinkedShaders] = sh;
-      prog->_NumLinkedShaders++;
+      _mesa_reference_shader(ctx, &prog->_LinkedShaders[MESA_SHADER_FRAGMENT],
+			     sh);
    }
 
    /* Here begins the inter-stage linking phase.  Some initial validation is
@@ -1381,14 +1593,26 @@ link_shaders(GLcontext *ctx, struct gl_shader_program *prog)
     * varyings.
     */
    if (cross_validate_uniforms(prog)) {
+      unsigned prev;
+
+      for (prev = 0; prev < MESA_SHADER_TYPES; prev++) {
+	 if (prog->_LinkedShaders[prev] != NULL)
+	    break;
+      }
+
       /* Validate the inputs of each stage with the output of the preceeding
        * stage.
        */
-      for (unsigned i = 1; i < prog->_NumLinkedShaders; i++) {
+      for (unsigned i = prev + 1; i < MESA_SHADER_TYPES; i++) {
+	 if (prog->_LinkedShaders[i] == NULL)
+	    continue;
+
 	 if (!cross_validate_outputs_to_inputs(prog,
-					       prog->_LinkedShaders[i - 1],
+					       prog->_LinkedShaders[prev],
 					       prog->_LinkedShaders[i]))
 	    goto done;
+
+	 prev = i;
       }
 
       prog->LinkStatus = true;
@@ -1398,33 +1622,77 @@ link_shaders(GLcontext *ctx, struct gl_shader_program *prog)
     * uniforms, and varyings.  Later optimization could possibly make
     * some of that unused.
     */
-   for (unsigned i = 0; i < prog->_NumLinkedShaders; i++) {
-      while (do_common_optimization(prog->_LinkedShaders[i]->ir, true))
+   for (unsigned i = 0; i < MESA_SHADER_TYPES; i++) {
+      if (prog->_LinkedShaders[i] == NULL)
+	 continue;
+
+      while (do_common_optimization(prog->_LinkedShaders[i]->ir, true, 32))
 	 ;
    }
 
+   update_array_sizes(prog);
+
    assign_uniform_locations(prog);
 
-   if (prog->_LinkedShaders[0]->Type == GL_VERTEX_SHADER) {
+   if (prog->_LinkedShaders[MESA_SHADER_VERTEX] != NULL) {
       /* FINISHME: The value of the max_attribute_index parameter is
        * FINISHME: implementation dependent based on the value of
        * FINISHME: GL_MAX_VERTEX_ATTRIBS.  GL_MAX_VERTEX_ATTRIBS must be
        * FINISHME: at least 16, so hardcode 16 for now.
        */
-      if (!assign_attribute_locations(prog, 16))
+      if (!assign_attribute_locations(prog, 16)) {
+	 prog->LinkStatus = false;
 	 goto done;
+      }
+   }
 
-      if (prog->_NumLinkedShaders == 1)
-	 demote_unread_shader_outputs(prog->_LinkedShaders[0]);
+   unsigned prev;
+   for (prev = 0; prev < MESA_SHADER_TYPES; prev++) {
+      if (prog->_LinkedShaders[prev] != NULL)
+	 break;
    }
 
-   for (unsigned i = 1; i < prog->_NumLinkedShaders; i++)
+   for (unsigned i = prev + 1; i < MESA_SHADER_TYPES; i++) {
+      if (prog->_LinkedShaders[i] == NULL)
+	 continue;
+
       assign_varying_locations(prog,
-			       prog->_LinkedShaders[i - 1],
+			       prog->_LinkedShaders[prev],
 			       prog->_LinkedShaders[i]);
+      prev = i;
+   }
+
+   if (prog->_LinkedShaders[MESA_SHADER_VERTEX] != NULL) {
+      demote_shader_inputs_and_outputs(prog->_LinkedShaders[MESA_SHADER_VERTEX],
+				       ir_var_out);
+   }
+
+   if (prog->_LinkedShaders[MESA_SHADER_GEOMETRY] != NULL) {
+      gl_shader *const sh = prog->_LinkedShaders[MESA_SHADER_GEOMETRY];
+
+      demote_shader_inputs_and_outputs(sh, ir_var_in);
+      demote_shader_inputs_and_outputs(sh, ir_var_inout);
+      demote_shader_inputs_and_outputs(sh, ir_var_out);
+   }
+
+   if (prog->_LinkedShaders[MESA_SHADER_FRAGMENT] != NULL) {
+      gl_shader *const sh = prog->_LinkedShaders[MESA_SHADER_FRAGMENT];
+
+      demote_shader_inputs_and_outputs(sh, ir_var_in);
+   }
 
    /* FINISHME: Assign fragment shader output locations. */
 
 done:
    free(vert_shader_list);
+
+   for (unsigned i = 0; i < MESA_SHADER_TYPES; i++) {
+      if (prog->_LinkedShaders[i] == NULL)
+	 continue;
+
+      /* Retain any live IR, but trash the rest. */
+      reparent_ir(prog->_LinkedShaders[i]->ir, prog->_LinkedShaders[i]->ir);
+   }
+
+   ralloc_free(mem_ctx);
 }