glsl: allow linking of tessellation shaders.
[mesa.git] / src / glsl / link_varyings.cpp
index a9b15f6f7b197ce7b8cf3fb2fec7b2ae64e1edc6..9174e9c2f2da510bd5695b28175ff77bbebd597b 100644 (file)
@@ -54,10 +54,16 @@ cross_validate_types_and_qualifiers(struct gl_shader_program *prog,
    /* Check that the types match between stages.
     */
    const glsl_type *type_to_match = input->type;
-   if (consumer_stage == MESA_SHADER_GEOMETRY) {
-      assert(type_to_match->is_array()); /* Enforced by ast_to_hir */
-      type_to_match = type_to_match->element_type();
+
+   /* VS -> GS, VS -> TCS, VS -> TES, TES -> GS */
+   const bool extra_array_level = (producer_stage == MESA_SHADER_VERTEX &&
+                                   consumer_stage != MESA_SHADER_FRAGMENT) ||
+                                  consumer_stage == MESA_SHADER_GEOMETRY;
+   if (extra_array_level) {
+      assert(type_to_match->is_array());
+      type_to_match = type_to_match->fields.array;
    }
+
    if (type_to_match != output->type) {
       /* There is a bit of a special case for gl_TexCoord.  This
        * built-in is unsized by default.  Applications that variable
@@ -77,8 +83,7 @@ cross_validate_types_and_qualifiers(struct gl_shader_program *prog,
        *     correspondence between the vertex language and the
        *     fragment language."
        */
-      if (!output->type->is_array()
-          || (strncmp("gl_", output->name, 3) != 0)) {
+      if (!output->type->is_array() || !is_gl_identifier(output->name)) {
          linker_error(prog,
                       "%s shader output `%s' declared as type `%s', "
                       "but %s shader input declared as type `%s'\n",
@@ -117,7 +122,19 @@ cross_validate_types_and_qualifiers(struct gl_shader_program *prog,
       return;
    }
 
-   if (input->data.invariant != output->data.invariant) {
+   if (input->data.patch != output->data.patch) {
+      linker_error(prog,
+                   "%s shader output `%s' %s patch qualifier, "
+                   "but %s shader input %s patch qualifier\n",
+                   _mesa_shader_stage_to_string(producer_stage),
+                   output->name,
+                   (output->data.patch) ? "has" : "lacks",
+                   _mesa_shader_stage_to_string(consumer_stage),
+                   (input->data.patch) ? "has" : "lacks");
+      return;
+   }
+
+   if (!prog->IsES && input->data.invariant != output->data.invariant) {
       linker_error(prog,
                    "%s shader output `%s' %s invariant qualifier, "
                    "but %s shader input %s invariant qualifier\n",
@@ -129,7 +146,17 @@ cross_validate_types_and_qualifiers(struct gl_shader_program *prog,
       return;
    }
 
-   if (input->data.interpolation != output->data.interpolation) {
+   /* GLSL >= 4.40 removes text requiring interpolation qualifiers
+    * to match cross stage, they must only match within the same stage.
+    *
+    * From page 84 (page 90 of the PDF) of the GLSL 4.40 spec:
+    *
+    *     "It is a link-time error if, within the same stage, the interpolation
+    *     qualifiers of variables of the same name do not match.
+    *
+    */
+   if (input->data.interpolation != output->data.interpolation &&
+       prog->Version < 440) {
       linker_error(prog,
                    "%s shader output `%s' specifies %s "
                    "interpolation qualifier, "
@@ -172,16 +199,36 @@ cross_validate_outputs_to_inputs(struct gl_shader_program *prog,
                                 gl_shader *producer, gl_shader *consumer)
 {
    glsl_symbol_table parameters;
+   ir_variable *explicit_locations[MAX_VARYING] = { NULL, };
 
    /* Find all shader outputs in the "producer" stage.
     */
-   foreach_list(node, producer->ir) {
-      ir_variable *const var = ((ir_instruction *) node)->as_variable();
+   foreach_in_list(ir_instruction, node, producer->ir) {
+      ir_variable *const var = node->as_variable();
 
       if ((var == NULL) || (var->data.mode != ir_var_shader_out))
         continue;
 
-      parameters.add_variable(var);
+      if (!var->data.explicit_location
+          || var->data.location < VARYING_SLOT_VAR0)
+         parameters.add_variable(var);
+      else {
+         /* User-defined varyings with explicit locations are handled
+          * differently because they do not need to have matching names.
+          */
+         const unsigned idx = var->data.location - VARYING_SLOT_VAR0;
+
+         if (explicit_locations[idx] != NULL) {
+            linker_error(prog,
+                         "%s shader has multiple outputs explicitly "
+                         "assigned to location %d\n",
+                         _mesa_shader_stage_to_string(producer->Stage),
+                         idx);
+            return;
+         }
+
+         explicit_locations[idx] = var;
+      }
    }
 
 
@@ -193,8 +240,8 @@ cross_validate_outputs_to_inputs(struct gl_shader_program *prog,
     * should be arrays and the type of the array element should match the type
     * of the corresponding producer output.
     */
-   foreach_list(node, consumer->ir) {
-      ir_variable *const input = ((ir_instruction *) node)->as_variable();
+   foreach_in_list(ir_instruction, node, consumer->ir) {
+      ir_variable *const input = node->as_variable();
 
       if ((input == NULL) || (input->data.mode != ir_var_shader_in))
         continue;
@@ -220,10 +267,43 @@ cross_validate_outputs_to_inputs(struct gl_shader_program *prog,
                                              front_color, back_color,
                                              consumer->Stage, producer->Stage);
       } else {
-         ir_variable *const output = parameters.get_variable(input->name);
+         /* The rules for connecting inputs and outputs change in the presence
+          * of explicit locations.  In this case, we no longer care about the
+          * names of the variables.  Instead, we care only about the
+          * explicitly assigned location.
+          */
+         ir_variable *output = NULL;
+         if (input->data.explicit_location
+             && input->data.location >= VARYING_SLOT_VAR0) {
+            output = explicit_locations[input->data.location - VARYING_SLOT_VAR0];
+
+            if (output == NULL) {
+               linker_error(prog,
+                            "%s shader input `%s' with explicit location "
+                            "has no matching output\n",
+                            _mesa_shader_stage_to_string(consumer->Stage),
+                            input->name);
+            }
+         } else {
+            output = parameters.get_variable(input->name);
+         }
+
          if (output != NULL) {
             cross_validate_types_and_qualifiers(prog, input, output,
                                                 consumer->Stage, producer->Stage);
+         } else {
+            /* Check for input vars with unmatched output vars in prev stage
+             * taking into account that interface blocks could have a matching
+             * output but with different name, so we ignore them.
+             */
+            assert(!input->data.assigned);
+            if (input->data.used && !input->get_interface_type() &&
+                !input->data.explicit_location && !prog->SeparateShader)
+               linker_error(prog,
+                            "%s shader input `%s' "
+                            "has no matching output in the previous stage\n",
+                            _mesa_shader_stage_to_string(consumer->Stage),
+                            input->name);
          }
       }
    }
@@ -252,6 +332,7 @@ tfeedback_decl::init(struct gl_context *ctx, const void *mem_ctx,
    this->skip_components = 0;
    this->next_buffer_separator = false;
    this->matched_candidate = NULL;
+   this->stream_id = 0;
 
    if (ctx->Extensions.ARB_transform_feedback3) {
       /* Parse gl_NextBuffer. */
@@ -278,6 +359,11 @@ tfeedback_decl::init(struct gl_context *ctx, const void *mem_ctx,
    const char *base_name_end;
    long subscript = parse_program_resource_name(input, &base_name_end);
    this->var_name = ralloc_strndup(mem_ctx, input, base_name_end - input);
+   if (this->var_name == NULL) {
+      _mesa_error_no_memory(__func__);
+      return;
+   }
+
    if (subscript >= 0) {
       this->array_subscript = subscript;
       this->is_subscripted = true;
@@ -289,7 +375,7 @@ tfeedback_decl::init(struct gl_context *ctx, const void *mem_ctx,
     * class must behave specially to account for the fact that gl_ClipDistance
     * is converted from a float[8] to a vec4[2].
     */
-   if (ctx->ShaderCompilerOptions[MESA_SHADER_VERTEX].LowerClipDistance &&
+   if (ctx->Const.ShaderCompilerOptions[MESA_SHADER_VERTEX].LowerClipDistance &&
        strcmp(this->var_name, "gl_ClipDistance") == 0) {
       this->is_clip_distance_mesa = true;
    }
@@ -316,8 +402,8 @@ tfeedback_decl::is_same(const tfeedback_decl &x, const tfeedback_decl &y)
 
 
 /**
- * Assign a location for this tfeedback_decl object based on the transform
- * feedback candidate found by find_candidate.
+ * Assign a location and stream ID for this tfeedback_decl object based on the
+ * transform feedback candidate found by find_candidate.
  *
  * If an error occurs, the error is reported through linker_error() and false
  * is returned.
@@ -398,6 +484,11 @@ tfeedback_decl::assign_location(struct gl_context *ctx,
       return false;
    }
 
+   /* Only transform feedback varyings can be assigned to non-zero streams,
+    * so assign the stream id here.
+    */
+   this->stream_id = this->matched_candidate->toplevel_var->data.stream;
+
    return true;
 }
 
@@ -456,6 +547,7 @@ tfeedback_decl::store(struct gl_context *ctx, struct gl_shader_program *prog,
       info->Outputs[info->NumOutputs].ComponentOffset = location_frac;
       info->Outputs[info->NumOutputs].OutputRegister = location;
       info->Outputs[info->NumOutputs].NumComponents = output_size;
+      info->Outputs[info->NumOutputs].StreamId = stream_id;
       info->Outputs[info->NumOutputs].OutputBuffer = buffer;
       info->Outputs[info->NumOutputs].DstOffset = info->BufferStride[buffer];
       ++info->NumOutputs;
@@ -589,10 +681,27 @@ store_tfeedback_info(struct gl_context *ctx, struct gl_shader_program *prog,
    }
    else {
       /* GL_INVERLEAVED_ATTRIBS */
+      int buffer_stream_id = -1;
       for (unsigned i = 0; i < num_tfeedback_decls; ++i) {
          if (tfeedback_decls[i].is_next_buffer_separator()) {
             num_buffers++;
+            buffer_stream_id = -1;
             continue;
+         } else if (buffer_stream_id == -1)  {
+            /* First varying writing to this buffer: remember its stream */
+            buffer_stream_id = (int) tfeedback_decls[i].get_stream_id();
+         } else if (buffer_stream_id !=
+                    (int) tfeedback_decls[i].get_stream_id()) {
+            /* Varying writes to the same buffer from a different stream */
+            linker_error(prog,
+                         "Transform feedback can't capture varyings belonging "
+                         "to different vertex streams in a single buffer. "
+                         "Varying %s writes to buffer from stream %u, other "
+                         "varyings in the same buffer write from stream %u.",
+                         tfeedback_decls[i].name(),
+                         tfeedback_decls[i].get_stream_id(),
+                         buffer_stream_id);
+            return false;
          }
 
          if (!tfeedback_decls[i].store(ctx, prog,
@@ -622,7 +731,7 @@ public:
    ~varying_matches();
    void record(ir_variable *producer_var, ir_variable *consumer_var);
    unsigned assign_locations();
-   void store_locations(unsigned producer_base, unsigned consumer_base) const;
+   void store_locations() const;
 
 private:
    /**
@@ -746,7 +855,10 @@ varying_matches::~varying_matches()
 void
 varying_matches::record(ir_variable *producer_var, ir_variable *consumer_var)
 {
-   if (!producer_var->data.is_unmatched_generic_inout) {
+   assert(producer_var != NULL || consumer_var != NULL);
+
+   if ((producer_var && !producer_var->data.is_unmatched_generic_inout)
+       || (consumer_var && !consumer_var->data.is_unmatched_generic_inout)) {
       /* Either a location already exists for this variable (since it is part
        * of fixed functionality), or it has already been recorded as part of a
        * previous match.
@@ -764,9 +876,11 @@ varying_matches::record(ir_variable *producer_var, ir_variable *consumer_var)
        * regardless of where they appear.  We can trivially satisfy that
        * requirement by changing the interpolation type to flat here.
        */
-      producer_var->data.centroid = false;
-      producer_var->data.sample = false;
-      producer_var->data.interpolation = INTERP_QUALIFIER_FLAT;
+      if (producer_var) {
+         producer_var->data.centroid = false;
+         producer_var->data.sample = false;
+         producer_var->data.interpolation = INTERP_QUALIFIER_FLAT;
+      }
 
       if (consumer_var) {
          consumer_var->data.centroid = false;
@@ -781,24 +895,28 @@ varying_matches::record(ir_variable *producer_var, ir_variable *consumer_var)
          realloc(this->matches,
                  sizeof(*this->matches) * this->matches_capacity);
    }
+
+   const ir_variable *const var = (producer_var != NULL)
+      ? producer_var : consumer_var;
+
    this->matches[this->num_matches].packing_class
-      = this->compute_packing_class(producer_var);
+      = this->compute_packing_class(var);
    this->matches[this->num_matches].packing_order
-      = this->compute_packing_order(producer_var);
+      = this->compute_packing_order(var);
    if (this->disable_varying_packing) {
-      unsigned slots = producer_var->type->is_array()
-         ? (producer_var->type->length
-            * producer_var->type->fields.array->matrix_columns)
-         : producer_var->type->matrix_columns;
+      unsigned slots = var->type->is_array()
+         ? (var->type->length * var->type->fields.array->matrix_columns)
+         : var->type->matrix_columns;
       this->matches[this->num_matches].num_components = 4 * slots;
    } else {
       this->matches[this->num_matches].num_components
-         = producer_var->type->component_slots();
+         = var->type->component_slots();
    }
    this->matches[this->num_matches].producer_var = producer_var;
    this->matches[this->num_matches].consumer_var = consumer_var;
    this->num_matches++;
-   producer_var->data.is_unmatched_generic_inout = 0;
+   if (producer_var)
+      producer_var->data.is_unmatched_generic_inout = 0;
    if (consumer_var)
       consumer_var->data.is_unmatched_generic_inout = 0;
 }
@@ -842,8 +960,7 @@ varying_matches::assign_locations()
  * assignments that were made by varying_matches::assign_locations().
  */
 void
-varying_matches::store_locations(unsigned producer_base,
-                                 unsigned consumer_base) const
+varying_matches::store_locations() const
 {
    for (unsigned i = 0; i < this->num_matches; i++) {
       ir_variable *producer_var = this->matches[i].producer_var;
@@ -852,11 +969,14 @@ varying_matches::store_locations(unsigned producer_base,
       unsigned slot = generic_location / 4;
       unsigned offset = generic_location % 4;
 
-      producer_var->data.location = producer_base + slot;
-      producer_var->data.location_frac = offset;
+      if (producer_var) {
+         producer_var->data.location = VARYING_SLOT_VAR0 + slot;
+         producer_var->data.location_frac = offset;
+      }
+
       if (consumer_var) {
          assert(consumer_var->data.location == -1);
-         consumer_var->data.location = consumer_base + slot;
+         consumer_var->data.location = VARYING_SLOT_VAR0 + slot;
          consumer_var->data.location_frac = offset;
       }
    }
@@ -887,7 +1007,8 @@ varying_matches::compute_packing_class(const ir_variable *var)
     *
     * Therefore, the packing class depends only on the interpolation type.
     */
-   unsigned packing_class = var->data.centroid | (var->data.sample << 1);
+   unsigned packing_class = var->data.centroid | (var->data.sample << 1) |
+                            (var->data.patch << 2);
    packing_class *= 4;
    packing_class += var->data.interpolation;
    return packing_class;
@@ -943,7 +1064,7 @@ varying_matches::match_comparator(const void *x_generic, const void *y_generic)
  * varyings, but excludes variables such as gl_FrontFacing and gl_FragCoord.
  */
 static bool
-is_varying_var(gl_shader_stage stage, const ir_variable *var)
+var_counts_against_varying_limit(gl_shader_stage stage, const ir_variable *var)
 {
    /* Only fragment shaders will take a varying variable as an input */
    if (stage == MESA_SHADER_FRAGMENT &&
@@ -996,10 +1117,8 @@ private:
    virtual void visit_field(const glsl_type *type, const char *name,
                             bool row_major)
    {
-      assert(!type->is_record());
-      assert(!(type->is_array() && type->fields.array->is_record()));
-      assert(!type->is_interface());
-      assert(!(type->is_array() && type->fields.array->is_interface()));
+      assert(!type->without_array()->is_record());
+      assert(!type->without_array()->is_interface());
 
       (void) row_major;
 
@@ -1037,6 +1156,157 @@ private:
 };
 
 
+namespace linker {
+
+bool
+populate_consumer_input_sets(void *mem_ctx, exec_list *ir,
+                             hash_table *consumer_inputs,
+                             hash_table *consumer_interface_inputs,
+                             ir_variable *consumer_inputs_with_locations[VARYING_SLOT_MAX])
+{
+   memset(consumer_inputs_with_locations,
+          0,
+          sizeof(consumer_inputs_with_locations[0]) * VARYING_SLOT_MAX);
+
+   foreach_in_list(ir_instruction, node, ir) {
+      ir_variable *const input_var = node->as_variable();
+
+      if ((input_var != NULL) && (input_var->data.mode == ir_var_shader_in)) {
+         if (input_var->type->is_interface())
+            return false;
+
+         if (input_var->data.explicit_location) {
+            /* assign_varying_locations only cares about finding the
+             * ir_variable at the start of a contiguous location block.
+             *
+             *     - For !producer, consumer_inputs_with_locations isn't used.
+             *
+             *     - For !consumer, consumer_inputs_with_locations is empty.
+             *
+             * For consumer && producer, if you were trying to set some
+             * ir_variable to the middle of a location block on the other side
+             * of producer/consumer, cross_validate_outputs_to_inputs() should
+             * be link-erroring due to either type mismatch or location
+             * overlaps.  If the variables do match up, then they've got a
+             * matching data.location and you only looked at
+             * consumer_inputs_with_locations[var->data.location], not any
+             * following entries for the array/structure.
+             */
+            consumer_inputs_with_locations[input_var->data.location] =
+               input_var;
+         } else if (input_var->get_interface_type() != NULL) {
+            char *const iface_field_name =
+               ralloc_asprintf(mem_ctx, "%s.%s",
+                               input_var->get_interface_type()->name,
+                               input_var->name);
+            hash_table_insert(consumer_interface_inputs, input_var,
+                              iface_field_name);
+         } else {
+            hash_table_insert(consumer_inputs, input_var,
+                              ralloc_strdup(mem_ctx, input_var->name));
+         }
+      }
+   }
+
+   return true;
+}
+
+/**
+ * Find a variable from the consumer that "matches" the specified variable
+ *
+ * This function only finds inputs with names that match.  There is no
+ * validation (here) that the types, etc. are compatible.
+ */
+ir_variable *
+get_matching_input(void *mem_ctx,
+                   const ir_variable *output_var,
+                   hash_table *consumer_inputs,
+                   hash_table *consumer_interface_inputs,
+                   ir_variable *consumer_inputs_with_locations[VARYING_SLOT_MAX])
+{
+   ir_variable *input_var;
+
+   if (output_var->data.explicit_location) {
+      input_var = consumer_inputs_with_locations[output_var->data.location];
+   } else if (output_var->get_interface_type() != NULL) {
+      char *const iface_field_name =
+         ralloc_asprintf(mem_ctx, "%s.%s",
+                         output_var->get_interface_type()->name,
+                         output_var->name);
+      input_var =
+         (ir_variable *) hash_table_find(consumer_interface_inputs,
+                                         iface_field_name);
+   } else {
+      input_var =
+         (ir_variable *) hash_table_find(consumer_inputs, output_var->name);
+   }
+
+   return (input_var == NULL || input_var->data.mode != ir_var_shader_in)
+      ? NULL : input_var;
+}
+
+}
+
+static int
+io_variable_cmp(const void *_a, const void *_b)
+{
+   const ir_variable *const a = *(const ir_variable **) _a;
+   const ir_variable *const b = *(const ir_variable **) _b;
+
+   if (a->data.explicit_location && b->data.explicit_location)
+      return b->data.location - a->data.location;
+
+   if (a->data.explicit_location && !b->data.explicit_location)
+      return 1;
+
+   if (!a->data.explicit_location && b->data.explicit_location)
+      return -1;
+
+   return -strcmp(a->name, b->name);
+}
+
+/**
+ * Sort the shader IO variables into canonical order
+ */
+static void
+canonicalize_shader_io(exec_list *ir, enum ir_variable_mode io_mode)
+{
+   ir_variable *var_table[MAX_PROGRAM_OUTPUTS * 4];
+   unsigned num_variables = 0;
+
+   foreach_in_list(ir_instruction, node, ir) {
+      ir_variable *const var = node->as_variable();
+
+      if (var == NULL || var->data.mode != io_mode)
+         continue;
+
+      /* If we have already encountered more I/O variables that could
+       * successfully link, bail.
+       */
+      if (num_variables == ARRAY_SIZE(var_table))
+         return;
+
+      var_table[num_variables++] = var;
+   }
+
+   if (num_variables == 0)
+      return;
+
+   /* Sort the list in reverse order (io_variable_cmp handles this).  Later
+    * we're going to push the variables on to the IR list as a stack, so we
+    * want the last variable (in canonical order) to be first in the list.
+    */
+   qsort(var_table, num_variables, sizeof(var_table[0]), io_variable_cmp);
+
+   /* Remove the variable from it's current location in the IR, and put it at
+    * the front.
+    */
+   for (unsigned i = 0; i < num_variables; i++) {
+      var_table[i]->remove();
+      ir->push_head(var_table[i]);
+   }
+}
+
 /**
  * Assign locations for all variables that are produced in one pipeline stage
  * (the "producer") and consumed in the next stage (the "consumer").
@@ -1069,8 +1339,6 @@ assign_varying_locations(struct gl_context *ctx,
                          tfeedback_decl *tfeedback_decls,
                          unsigned gs_input_vertices)
 {
-   const unsigned producer_base = VARYING_SLOT_VAR0;
-   const unsigned consumer_base = VARYING_SLOT_VAR0;
    varying_matches matches(ctx->Const.DisableVaryingPacking,
                            consumer && consumer->Stage == MESA_SHADER_FRAGMENT);
    hash_table *tfeedback_candidates
@@ -1079,67 +1347,96 @@ assign_varying_locations(struct gl_context *ctx,
       = hash_table_ctor(0, hash_table_string_hash, hash_table_string_compare);
    hash_table *consumer_interface_inputs
       = hash_table_ctor(0, hash_table_string_hash, hash_table_string_compare);
+   ir_variable *consumer_inputs_with_locations[VARYING_SLOT_MAX] = {
+      NULL,
+   };
 
-   /* Operate in a total of three passes.
+   /* Operate in a total of four passes.
+    *
+    * 1. Sort inputs / outputs into a canonical order.  This is necessary so
+    *    that inputs / outputs of separable shaders will be assigned
+    *    predictable locations regardless of the order in which declarations
+    *    appeared in the shader source.
     *
-    * 1. Assign locations for any matching inputs and outputs.
+    * 2. Assign locations for any matching inputs and outputs.
     *
-    * 2. Mark output variables in the producer that do not have locations as
+    * 3. Mark output variables in the producer that do not have locations as
     *    not being outputs.  This lets the optimizer eliminate them.
     *
-    * 3. Mark input variables in the consumer that do not have locations as
+    * 4. Mark input variables in the consumer that do not have locations as
     *    not being inputs.  This lets the optimizer eliminate them.
     */
-
-   if (consumer) {
-      foreach_list(node, consumer->ir) {
-         ir_variable *const input_var =
-            ((ir_instruction *) node)->as_variable();
-
-         if ((input_var != NULL) && (input_var->data.mode == ir_var_shader_in)) {
-            if (input_var->get_interface_type() != NULL) {
-               char *const iface_field_name =
-                  ralloc_asprintf(mem_ctx, "%s.%s",
-                                  input_var->get_interface_type()->name,
-                                  input_var->name);
-               hash_table_insert(consumer_interface_inputs, input_var,
-                                 iface_field_name);
-            } else {
-               hash_table_insert(consumer_inputs, input_var,
-                                 ralloc_strdup(mem_ctx, input_var->name));
-            }
-         }
-      }
+   if (consumer)
+      canonicalize_shader_io(consumer->ir, ir_var_shader_in);
+
+   if (producer)
+      canonicalize_shader_io(producer->ir, ir_var_shader_out);
+
+   if (consumer
+       && !linker::populate_consumer_input_sets(mem_ctx,
+                                                consumer->ir,
+                                                consumer_inputs,
+                                                consumer_interface_inputs,
+                                                consumer_inputs_with_locations)) {
+      assert(!"populate_consumer_input_sets failed");
+      hash_table_dtor(tfeedback_candidates);
+      hash_table_dtor(consumer_inputs);
+      hash_table_dtor(consumer_interface_inputs);
+      return false;
    }
 
-   foreach_list(node, producer->ir) {
-      ir_variable *const output_var = ((ir_instruction *) node)->as_variable();
+   if (producer) {
+      foreach_in_list(ir_instruction, node, producer->ir) {
+         ir_variable *const output_var = node->as_variable();
 
-      if ((output_var == NULL) || (output_var->data.mode != ir_var_shader_out))
-        continue;
+         if ((output_var == NULL) ||
+             (output_var->data.mode != ir_var_shader_out))
+            continue;
 
-      tfeedback_candidate_generator g(mem_ctx, tfeedback_candidates);
-      g.process(output_var);
-
-      ir_variable *input_var;
-      if (output_var->get_interface_type() != NULL) {
-         char *const iface_field_name =
-            ralloc_asprintf(mem_ctx, "%s.%s",
-                            output_var->get_interface_type()->name,
-                            output_var->name);
-         input_var =
-            (ir_variable *) hash_table_find(consumer_interface_inputs,
-                                            iface_field_name);
-      } else {
-         input_var =
-            (ir_variable *) hash_table_find(consumer_inputs, output_var->name);
+         /* Only geometry shaders can use non-zero streams */
+         assert(output_var->data.stream == 0 ||
+                (output_var->data.stream < MAX_VERTEX_STREAMS &&
+                 producer->Stage == MESA_SHADER_GEOMETRY));
+
+         tfeedback_candidate_generator g(mem_ctx, tfeedback_candidates);
+         g.process(output_var);
+
+         ir_variable *const input_var =
+            linker::get_matching_input(mem_ctx, output_var, consumer_inputs,
+                                       consumer_interface_inputs,
+                                       consumer_inputs_with_locations);
+
+         /* If a matching input variable was found, add this ouptut (and the
+          * input) to the set.  If this is a separable program and there is no
+          * consumer stage, add the output.
+          */
+         if (input_var || (prog->SeparateShader && consumer == NULL)) {
+            matches.record(output_var, input_var);
+         }
+
+         /* Only stream 0 outputs can be consumed in the next stage */
+         if (input_var && output_var->data.stream != 0) {
+            linker_error(prog, "output %s is assigned to stream=%d but "
+                         "is linked to an input, which requires stream=0",
+                         output_var->name, output_var->data.stream);
+            return false;
+         }
       }
+   } else {
+      /* If there's no producer stage, then this must be a separable program.
+       * For example, we may have a program that has just a fragment shader.
+       * Later this program will be used with some arbitrary vertex (or
+       * geometry) shader program.  This means that locations must be assigned
+       * for all the inputs.
+       */
+      foreach_in_list(ir_instruction, node, consumer->ir) {
+         ir_variable *const input_var = node->as_variable();
 
-      if (input_var && input_var->data.mode != ir_var_shader_in)
-         input_var = NULL;
+         if ((input_var == NULL) ||
+             (input_var->data.mode != ir_var_shader_in))
+            continue;
 
-      if (input_var) {
-         matches.record(output_var, input_var);
+         matches.record(NULL, input_var);
       }
    }
 
@@ -1162,7 +1459,7 @@ assign_varying_locations(struct gl_context *ctx,
    }
 
    const unsigned slots_used = matches.assign_locations();
-   matches.store_locations(producer_base, consumer_base);
+   matches.store_locations();
 
    for (unsigned i = 0; i < num_tfeedback_decls; ++i) {
       if (!tfeedback_decls[i].is_varying())
@@ -1187,21 +1484,38 @@ assign_varying_locations(struct gl_context *ctx,
        */
       assert(!ctx->Extensions.EXT_transform_feedback);
    } else {
-      lower_packed_varyings(mem_ctx, producer_base, slots_used,
-                            ir_var_shader_out, 0, producer);
+      if (producer) {
+         lower_packed_varyings(mem_ctx, slots_used, ir_var_shader_out,
+                               0, producer);
+      }
       if (consumer) {
-         lower_packed_varyings(mem_ctx, consumer_base, slots_used,
-                               ir_var_shader_in, gs_input_vertices, consumer);
+         lower_packed_varyings(mem_ctx, slots_used, ir_var_shader_in,
+                               gs_input_vertices, consumer);
       }
    }
 
-   if (consumer) {
-      foreach_list(node, consumer->ir) {
-         ir_variable *const var = ((ir_instruction *) node)->as_variable();
+   if (consumer && producer) {
+      foreach_in_list(ir_instruction, node, consumer->ir) {
+         ir_variable *const var = node->as_variable();
 
          if (var && var->data.mode == ir_var_shader_in &&
              var->data.is_unmatched_generic_inout) {
-            if (prog->Version <= 120) {
+            if (prog->IsES) {
+               /*
+                * On Page 91 (Page 97 of the PDF) of the GLSL ES 1.0 spec:
+                *
+                *     If the vertex shader declares but doesn't write to a
+                *     varying and the fragment shader declares and reads it,
+                *     is this an error?
+                *
+                *     RESOLUTION: No.
+                */
+               linker_warning(prog, "%s shader varying %s not written "
+                              "by %s shader\n.",
+                              _mesa_shader_stage_to_string(consumer->Stage),
+                              var->name,
+                              _mesa_shader_stage_to_string(producer->Stage));
+            } else if (prog->Version <= 120) {
                /* On page 25 (page 31 of the PDF) of the GLSL 1.20 spec:
                 *
                 *     Only those varying variables used (i.e. read) in
@@ -1214,7 +1528,6 @@ assign_varying_locations(struct gl_context *ctx,
                 * write the variable for the FS to read it.  See
                 * "glsl1-varying read but not written" in piglit.
                 */
-
                linker_error(prog, "%s shader varying %s not written "
                             "by %s shader\n.",
                             _mesa_shader_stage_to_string(consumer->Stage),
@@ -1240,11 +1553,11 @@ check_against_output_limit(struct gl_context *ctx,
 {
    unsigned output_vectors = 0;
 
-   foreach_list(node, producer->ir) {
-      ir_variable *const var = ((ir_instruction *) node)->as_variable();
+   foreach_in_list(ir_instruction, node, producer->ir) {
+      ir_variable *const var = node->as_variable();
 
       if (var && var->data.mode == ir_var_shader_out &&
-          is_varying_var(producer->Stage, var)) {
+          var_counts_against_varying_limit(producer->Stage, var)) {
          output_vectors += var->type->count_attribute_slots();
       }
    }
@@ -1256,13 +1569,15 @@ check_against_output_limit(struct gl_context *ctx,
    const unsigned output_components = output_vectors * 4;
    if (output_components > max_output_components) {
       if (ctx->API == API_OPENGLES2 || prog->IsES)
-         linker_error(prog, "shader uses too many output vectors "
+         linker_error(prog, "%s shader uses too many output vectors "
                       "(%u > %u)\n",
+                      _mesa_shader_stage_to_string(producer->Stage),
                       output_vectors,
                       max_output_components / 4);
       else
-         linker_error(prog, "shader uses too many output components "
+         linker_error(prog, "%s shader uses too many output components "
                       "(%u > %u)\n",
+                      _mesa_shader_stage_to_string(producer->Stage),
                       output_components,
                       max_output_components);
 
@@ -1279,11 +1594,11 @@ check_against_input_limit(struct gl_context *ctx,
 {
    unsigned input_vectors = 0;
 
-   foreach_list(node, consumer->ir) {
-      ir_variable *const var = ((ir_instruction *) node)->as_variable();
+   foreach_in_list(ir_instruction, node, consumer->ir) {
+      ir_variable *const var = node->as_variable();
 
       if (var && var->data.mode == ir_var_shader_in &&
-          is_varying_var(consumer->Stage, var)) {
+          var_counts_against_varying_limit(consumer->Stage, var)) {
          input_vectors += var->type->count_attribute_slots();
       }
    }
@@ -1295,13 +1610,15 @@ check_against_input_limit(struct gl_context *ctx,
    const unsigned input_components = input_vectors * 4;
    if (input_components > max_input_components) {
       if (ctx->API == API_OPENGLES2 || prog->IsES)
-         linker_error(prog, "shader uses too many input vectors "
+         linker_error(prog, "%s shader uses too many input vectors "
                       "(%u > %u)\n",
+                      _mesa_shader_stage_to_string(consumer->Stage),
                       input_vectors,
                       max_input_components / 4);
       else
-         linker_error(prog, "shader uses too many input components "
+         linker_error(prog, "%s shader uses too many input components "
                       "(%u > %u)\n",
+                      _mesa_shader_stage_to_string(consumer->Stage),
                       input_components,
                       max_input_components);