2 * Copyright © 2010 Intel Corporation
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
21 * DEALINGS IN THE SOFTWARE.
26 * GLSL linker implementation
28 * Given a set of shaders that are to be linked to generate a final program,
29 * there are three distinct stages.
31 * In the first stage shaders are partitioned into groups based on the shader
32 * type. All shaders of a particular type (e.g., vertex shaders) are linked
35 * - Undefined references in each shader are resolve to definitions in
37 * - Types and qualifiers of uniforms, outputs, and global variables defined
38 * in multiple shaders with the same name are verified to be the same.
39 * - Initializers for uniforms and global variables defined
40 * in multiple shaders with the same name are verified to be the same.
42 * The result, in the terminology of the GLSL spec, is a set of shader
43 * executables for each processing unit.
45 * After the first stage is complete, a series of semantic checks are performed
46 * on each of the shader executables.
48 * - Each shader executable must define a \c main function.
49 * - Each vertex shader executable must write to \c gl_Position.
50 * - Each fragment shader executable must write to either \c gl_FragData or
53 * In the final stage individual shader executables are linked to create a
54 * complete exectuable.
56 * - Types of uniforms defined in multiple shader stages with the same name
57 * are verified to be the same.
58 * - Initializers for uniforms defined in multiple shader stages with the
59 * same name are verified to be the same.
60 * - Types and qualifiers of outputs defined in one stage are verified to
61 * be the same as the types and qualifiers of inputs defined with the same
62 * name in a later stage.
64 * \author Ian Romanick <ian.d.romanick@intel.com>
75 #include "main/core.h"
76 #include "glsl_symbol_table.h"
79 #include "program/hash_table.h"
81 #include "ir_optimization.h"
84 #include "main/shaderobj.h"
88 * Visitor that determines whether or not a variable is ever written.
90 class find_assignment_visitor : public ir_hierarchical_visitor {
92 find_assignment_visitor(const char *name)
93 : name(name), found(false)
98 virtual ir_visitor_status visit_enter(ir_assignment *ir)
100 ir_variable *const var = ir->lhs->variable_referenced();
102 if (strcmp(name, var->name) == 0) {
107 return visit_continue_with_parent;
110 virtual ir_visitor_status visit_enter(ir_call *ir)
112 exec_list_iterator sig_iter = ir->get_callee()->parameters.iterator();
113 foreach_iter(exec_list_iterator, iter, *ir) {
114 ir_rvalue *param_rval = (ir_rvalue *)iter.get();
115 ir_variable *sig_param = (ir_variable *)sig_iter.get();
117 if (sig_param->mode == ir_var_out ||
118 sig_param->mode == ir_var_inout) {
119 ir_variable *var = param_rval->variable_referenced();
120 if (var && strcmp(name, var->name) == 0) {
128 return visit_continue_with_parent;
131 bool variable_found()
137 const char *name; /**< Find writes to a variable with this name. */
138 bool found; /**< Was a write to the variable found? */
143 * Visitor that determines whether or not a variable is ever read.
145 class find_deref_visitor : public ir_hierarchical_visitor {
147 find_deref_visitor(const char *name)
148 : name(name), found(false)
153 virtual ir_visitor_status visit(ir_dereference_variable *ir)
155 if (strcmp(this->name, ir->var->name) == 0) {
160 return visit_continue;
163 bool variable_found() const
169 const char *name; /**< Find writes to a variable with this name. */
170 bool found; /**< Was a write to the variable found? */
175 linker_error_printf(gl_shader_program *prog, const char *fmt, ...)
179 prog->InfoLog = talloc_strdup_append(prog->InfoLog, "error: ");
181 prog->InfoLog = talloc_vasprintf_append(prog->InfoLog, fmt, ap);
187 invalidate_variable_locations(gl_shader *sh, enum ir_variable_mode mode,
190 foreach_list(node, sh->ir) {
191 ir_variable *const var = ((ir_instruction *) node)->as_variable();
193 if ((var == NULL) || (var->mode != (unsigned) mode))
196 /* Only assign locations for generic attributes / varyings / etc.
198 if ((var->location >= generic_base) && !var->explicit_location)
205 * Determine the number of attribute slots required for a particular type
207 * This code is here because it implements the language rules of a specific
208 * GLSL version. Since it's a property of the language and not a property of
209 * types in general, it doesn't really belong in glsl_type.
212 count_attribute_slots(const glsl_type *t)
214 /* From page 31 (page 37 of the PDF) of the GLSL 1.50 spec:
216 * "A scalar input counts the same amount against this limit as a vec4,
217 * so applications may want to consider packing groups of four
218 * unrelated float inputs together into a vector to better utilize the
219 * capabilities of the underlying hardware. A matrix input will use up
220 * multiple locations. The number of locations used will equal the
221 * number of columns in the matrix."
223 * The spec does not explicitly say how arrays are counted. However, it
224 * should be safe to assume the total number of slots consumed by an array
225 * is the number of entries in the array multiplied by the number of slots
226 * consumed by a single element of the array.
230 return t->array_size() * count_attribute_slots(t->element_type());
233 return t->matrix_columns;
240 * Verify that a vertex shader executable meets all semantic requirements
242 * \param shader Vertex shader executable to be verified
245 validate_vertex_shader_executable(struct gl_shader_program *prog,
246 struct gl_shader *shader)
251 find_assignment_visitor find("gl_Position");
252 find.run(shader->ir);
253 if (!find.variable_found()) {
254 linker_error_printf(prog,
255 "vertex shader does not write to `gl_Position'\n");
264 * Verify that a fragment shader executable meets all semantic requirements
266 * \param shader Fragment shader executable to be verified
269 validate_fragment_shader_executable(struct gl_shader_program *prog,
270 struct gl_shader *shader)
275 find_assignment_visitor frag_color("gl_FragColor");
276 find_assignment_visitor frag_data("gl_FragData");
278 frag_color.run(shader->ir);
279 frag_data.run(shader->ir);
281 if (frag_color.variable_found() && frag_data.variable_found()) {
282 linker_error_printf(prog, "fragment shader writes to both "
283 "`gl_FragColor' and `gl_FragData'\n");
292 * Generate a string describing the mode of a variable
295 mode_string(const ir_variable *var)
299 return (var->read_only) ? "global constant" : "global variable";
301 case ir_var_uniform: return "uniform";
302 case ir_var_in: return "shader input";
303 case ir_var_out: return "shader output";
304 case ir_var_inout: return "shader inout";
306 case ir_var_temporary:
308 assert(!"Should not get here.");
309 return "invalid variable";
315 * Perform validation of global variables used across multiple shaders
318 cross_validate_globals(struct gl_shader_program *prog,
319 struct gl_shader **shader_list,
320 unsigned num_shaders,
323 /* Examine all of the uniforms in all of the shaders and cross validate
326 glsl_symbol_table variables;
327 for (unsigned i = 0; i < num_shaders; i++) {
328 if (shader_list[i] == NULL)
331 foreach_list(node, shader_list[i]->ir) {
332 ir_variable *const var = ((ir_instruction *) node)->as_variable();
337 if (uniforms_only && (var->mode != ir_var_uniform))
340 /* Don't cross validate temporaries that are at global scope. These
341 * will eventually get pulled into the shaders 'main'.
343 if (var->mode == ir_var_temporary)
346 /* If a global with this name has already been seen, verify that the
347 * new instance has the same type. In addition, if the globals have
348 * initializers, the values of the initializers must be the same.
350 ir_variable *const existing = variables.get_variable(var->name);
351 if (existing != NULL) {
352 if (var->type != existing->type) {
353 /* Consider the types to be "the same" if both types are arrays
354 * of the same type and one of the arrays is implicitly sized.
355 * In addition, set the type of the linked variable to the
356 * explicitly sized array.
358 if (var->type->is_array()
359 && existing->type->is_array()
360 && (var->type->fields.array == existing->type->fields.array)
361 && ((var->type->length == 0)
362 || (existing->type->length == 0))) {
363 if (existing->type->length == 0)
364 existing->type = var->type;
366 linker_error_printf(prog, "%s `%s' declared as type "
367 "`%s' and type `%s'\n",
369 var->name, var->type->name,
370 existing->type->name);
375 if (var->explicit_location) {
376 if (existing->explicit_location
377 && (var->location != existing->location)) {
378 linker_error_printf(prog, "explicit locations for %s "
379 "`%s' have differing values\n",
380 mode_string(var), var->name);
384 existing->location = var->location;
385 existing->explicit_location = true;
388 /* FINISHME: Handle non-constant initializers.
390 if (var->constant_value != NULL) {
391 if (existing->constant_value != NULL) {
392 if (!var->constant_value->has_value(existing->constant_value)) {
393 linker_error_printf(prog, "initializers for %s "
394 "`%s' have differing values\n",
395 mode_string(var), var->name);
399 /* If the first-seen instance of a particular uniform did not
400 * have an initializer but a later instance does, copy the
401 * initializer to the version stored in the symbol table.
403 /* FINISHME: This is wrong. The constant_value field should
404 * FINISHME: not be modified! Imagine a case where a shader
405 * FINISHME: without an initializer is linked in two different
406 * FINISHME: programs with shaders that have differing
407 * FINISHME: initializers. Linking with the first will
408 * FINISHME: modify the shader, and linking with the second
409 * FINISHME: will fail.
411 existing->constant_value =
412 var->constant_value->clone(talloc_parent(existing), NULL);
415 if (existing->invariant != var->invariant) {
416 linker_error_printf(prog, "declarations for %s `%s' have "
417 "mismatching invariant qualifiers\n",
418 mode_string(var), var->name);
422 variables.add_variable(var);
431 * Perform validation of uniforms used across multiple shader stages
434 cross_validate_uniforms(struct gl_shader_program *prog)
436 return cross_validate_globals(prog, prog->_LinkedShaders,
437 MESA_SHADER_TYPES, true);
442 * Validate that outputs from one stage match inputs of another
445 cross_validate_outputs_to_inputs(struct gl_shader_program *prog,
446 gl_shader *producer, gl_shader *consumer)
448 glsl_symbol_table parameters;
449 /* FINISHME: Figure these out dynamically. */
450 const char *const producer_stage = "vertex";
451 const char *const consumer_stage = "fragment";
453 /* Find all shader outputs in the "producer" stage.
455 foreach_list(node, producer->ir) {
456 ir_variable *const var = ((ir_instruction *) node)->as_variable();
458 /* FINISHME: For geometry shaders, this should also look for inout
459 * FINISHME: variables.
461 if ((var == NULL) || (var->mode != ir_var_out))
464 parameters.add_variable(var);
468 /* Find all shader inputs in the "consumer" stage. Any variables that have
469 * matching outputs already in the symbol table must have the same type and
472 foreach_list(node, consumer->ir) {
473 ir_variable *const input = ((ir_instruction *) node)->as_variable();
475 /* FINISHME: For geometry shaders, this should also look for inout
476 * FINISHME: variables.
478 if ((input == NULL) || (input->mode != ir_var_in))
481 ir_variable *const output = parameters.get_variable(input->name);
482 if (output != NULL) {
483 /* Check that the types match between stages.
485 if (input->type != output->type) {
486 linker_error_printf(prog,
487 "%s shader output `%s' declared as "
488 "type `%s', but %s shader input declared "
490 producer_stage, output->name,
492 consumer_stage, input->type->name);
496 /* Check that all of the qualifiers match between stages.
498 if (input->centroid != output->centroid) {
499 linker_error_printf(prog,
500 "%s shader output `%s' %s centroid qualifier, "
501 "but %s shader input %s centroid qualifier\n",
504 (output->centroid) ? "has" : "lacks",
506 (input->centroid) ? "has" : "lacks");
510 if (input->invariant != output->invariant) {
511 linker_error_printf(prog,
512 "%s shader output `%s' %s invariant qualifier, "
513 "but %s shader input %s invariant qualifier\n",
516 (output->invariant) ? "has" : "lacks",
518 (input->invariant) ? "has" : "lacks");
522 if (input->interpolation != output->interpolation) {
523 linker_error_printf(prog,
524 "%s shader output `%s' specifies %s "
525 "interpolation qualifier, "
526 "but %s shader input specifies %s "
527 "interpolation qualifier\n",
530 output->interpolation_string(),
532 input->interpolation_string());
543 * Populates a shaders symbol table with all global declarations
546 populate_symbol_table(gl_shader *sh)
548 sh->symbols = new(sh) glsl_symbol_table;
550 foreach_list(node, sh->ir) {
551 ir_instruction *const inst = (ir_instruction *) node;
555 if ((func = inst->as_function()) != NULL) {
556 sh->symbols->add_function(func);
557 } else if ((var = inst->as_variable()) != NULL) {
558 sh->symbols->add_variable(var);
565 * Remap variables referenced in an instruction tree
567 * This is used when instruction trees are cloned from one shader and placed in
568 * another. These trees will contain references to \c ir_variable nodes that
569 * do not exist in the target shader. This function finds these \c ir_variable
570 * references and replaces the references with matching variables in the target
573 * If there is no matching variable in the target shader, a clone of the
574 * \c ir_variable is made and added to the target shader. The new variable is
575 * added to \b both the instruction stream and the symbol table.
577 * \param inst IR tree that is to be processed.
578 * \param symbols Symbol table containing global scope symbols in the
580 * \param instructions Instruction stream where new variable declarations
584 remap_variables(ir_instruction *inst, struct gl_shader *target,
587 class remap_visitor : public ir_hierarchical_visitor {
589 remap_visitor(struct gl_shader *target,
592 this->target = target;
593 this->symbols = target->symbols;
594 this->instructions = target->ir;
598 virtual ir_visitor_status visit(ir_dereference_variable *ir)
600 if (ir->var->mode == ir_var_temporary) {
601 ir_variable *var = (ir_variable *) hash_table_find(temps, ir->var);
605 return visit_continue;
608 ir_variable *const existing =
609 this->symbols->get_variable(ir->var->name);
610 if (existing != NULL)
613 ir_variable *copy = ir->var->clone(this->target, NULL);
615 this->symbols->add_variable(copy);
616 this->instructions->push_head(copy);
620 return visit_continue;
624 struct gl_shader *target;
625 glsl_symbol_table *symbols;
626 exec_list *instructions;
630 remap_visitor v(target, temps);
637 * Move non-declarations from one instruction stream to another
639 * The intended usage pattern of this function is to pass the pointer to the
640 * head sentinel of a list (i.e., a pointer to the list cast to an \c exec_node
641 * pointer) for \c last and \c false for \c make_copies on the first
642 * call. Successive calls pass the return value of the previous call for
643 * \c last and \c true for \c make_copies.
645 * \param instructions Source instruction stream
646 * \param last Instruction after which new instructions should be
647 * inserted in the target instruction stream
648 * \param make_copies Flag selecting whether instructions in \c instructions
649 * should be copied (via \c ir_instruction::clone) into the
650 * target list or moved.
653 * The new "last" instruction in the target instruction stream. This pointer
654 * is suitable for use as the \c last parameter of a later call to this
658 move_non_declarations(exec_list *instructions, exec_node *last,
659 bool make_copies, gl_shader *target)
661 hash_table *temps = NULL;
664 temps = hash_table_ctor(0, hash_table_pointer_hash,
665 hash_table_pointer_compare);
667 foreach_list_safe(node, instructions) {
668 ir_instruction *inst = (ir_instruction *) node;
670 if (inst->as_function())
673 ir_variable *var = inst->as_variable();
674 if ((var != NULL) && (var->mode != ir_var_temporary))
677 assert(inst->as_assignment()
678 || ((var != NULL) && (var->mode == ir_var_temporary)));
681 inst = inst->clone(target, NULL);
684 hash_table_insert(temps, inst, var);
686 remap_variables(inst, target, temps);
691 last->insert_after(inst);
696 hash_table_dtor(temps);
702 * Get the function signature for main from a shader
704 static ir_function_signature *
705 get_main_function_signature(gl_shader *sh)
707 ir_function *const f = sh->symbols->get_function("main");
709 exec_list void_parameters;
711 /* Look for the 'void main()' signature and ensure that it's defined.
712 * This keeps the linker from accidentally pick a shader that just
713 * contains a prototype for main.
715 * We don't have to check for multiple definitions of main (in multiple
716 * shaders) because that would have already been caught above.
718 ir_function_signature *sig = f->matching_signature(&void_parameters);
719 if ((sig != NULL) && sig->is_defined) {
729 * Combine a group of shaders for a single stage to generate a linked shader
732 * If this function is supplied a single shader, it is cloned, and the new
733 * shader is returned.
735 static struct gl_shader *
736 link_intrastage_shaders(void *mem_ctx,
737 struct gl_context *ctx,
738 struct gl_shader_program *prog,
739 struct gl_shader **shader_list,
740 unsigned num_shaders)
742 /* Check that global variables defined in multiple shaders are consistent.
744 if (!cross_validate_globals(prog, shader_list, num_shaders, false))
747 /* Check that there is only a single definition of each function signature
748 * across all shaders.
750 for (unsigned i = 0; i < (num_shaders - 1); i++) {
751 foreach_list(node, shader_list[i]->ir) {
752 ir_function *const f = ((ir_instruction *) node)->as_function();
757 for (unsigned j = i + 1; j < num_shaders; j++) {
758 ir_function *const other =
759 shader_list[j]->symbols->get_function(f->name);
761 /* If the other shader has no function (and therefore no function
762 * signatures) with the same name, skip to the next shader.
767 foreach_iter (exec_list_iterator, iter, *f) {
768 ir_function_signature *sig =
769 (ir_function_signature *) iter.get();
771 if (!sig->is_defined || sig->is_builtin)
774 ir_function_signature *other_sig =
775 other->exact_matching_signature(& sig->parameters);
777 if ((other_sig != NULL) && other_sig->is_defined
778 && !other_sig->is_builtin) {
779 linker_error_printf(prog,
780 "function `%s' is multiply defined",
789 /* Find the shader that defines main, and make a clone of it.
791 * Starting with the clone, search for undefined references. If one is
792 * found, find the shader that defines it. Clone the reference and add
793 * it to the shader. Repeat until there are no undefined references or
794 * until a reference cannot be resolved.
796 gl_shader *main = NULL;
797 for (unsigned i = 0; i < num_shaders; i++) {
798 if (get_main_function_signature(shader_list[i]) != NULL) {
799 main = shader_list[i];
805 linker_error_printf(prog, "%s shader lacks `main'\n",
806 (shader_list[0]->Type == GL_VERTEX_SHADER)
807 ? "vertex" : "fragment");
811 gl_shader *linked = ctx->Driver.NewShader(NULL, 0, main->Type);
812 linked->ir = new(linked) exec_list;
813 clone_ir_list(mem_ctx, linked->ir, main->ir);
815 populate_symbol_table(linked);
817 /* The a pointer to the main function in the final linked shader (i.e., the
818 * copy of the original shader that contained the main function).
820 ir_function_signature *const main_sig = get_main_function_signature(linked);
822 /* Move any instructions other than variable declarations or function
823 * declarations into main.
825 exec_node *insertion_point =
826 move_non_declarations(linked->ir, (exec_node *) &main_sig->body, false,
829 for (unsigned i = 0; i < num_shaders; i++) {
830 if (shader_list[i] == main)
833 insertion_point = move_non_declarations(shader_list[i]->ir,
834 insertion_point, true, linked);
837 /* Resolve initializers for global variables in the linked shader.
839 unsigned num_linking_shaders = num_shaders;
840 for (unsigned i = 0; i < num_shaders; i++)
841 num_linking_shaders += shader_list[i]->num_builtins_to_link;
843 gl_shader **linking_shaders =
844 (gl_shader **) calloc(num_linking_shaders, sizeof(gl_shader *));
846 memcpy(linking_shaders, shader_list,
847 sizeof(linking_shaders[0]) * num_shaders);
849 unsigned idx = num_shaders;
850 for (unsigned i = 0; i < num_shaders; i++) {
851 memcpy(&linking_shaders[idx], shader_list[i]->builtins_to_link,
852 sizeof(linking_shaders[0]) * shader_list[i]->num_builtins_to_link);
853 idx += shader_list[i]->num_builtins_to_link;
856 assert(idx == num_linking_shaders);
858 if (!link_function_calls(prog, linked, linking_shaders,
859 num_linking_shaders)) {
860 ctx->Driver.DeleteShader(ctx, linked);
864 free(linking_shaders);
870 struct uniform_node {
872 struct gl_uniform *u;
877 * Update the sizes of linked shader uniform arrays to the maximum
880 * From page 81 (page 95 of the PDF) of the OpenGL 2.1 spec:
882 * If one or more elements of an array are active,
883 * GetActiveUniform will return the name of the array in name,
884 * subject to the restrictions listed above. The type of the array
885 * is returned in type. The size parameter contains the highest
886 * array element index used, plus one. The compiler or linker
887 * determines the highest index used. There will be only one
888 * active uniform reported by the GL per uniform array.
892 update_array_sizes(struct gl_shader_program *prog)
894 for (unsigned i = 0; i < MESA_SHADER_TYPES; i++) {
895 if (prog->_LinkedShaders[i] == NULL)
898 foreach_list(node, prog->_LinkedShaders[i]->ir) {
899 ir_variable *const var = ((ir_instruction *) node)->as_variable();
901 if ((var == NULL) || (var->mode != ir_var_uniform &&
902 var->mode != ir_var_in &&
903 var->mode != ir_var_out) ||
904 !var->type->is_array())
907 unsigned int size = var->max_array_access;
908 for (unsigned j = 0; j < MESA_SHADER_TYPES; j++) {
909 if (prog->_LinkedShaders[j] == NULL)
912 foreach_list(node2, prog->_LinkedShaders[j]->ir) {
913 ir_variable *other_var = ((ir_instruction *) node2)->as_variable();
917 if (strcmp(var->name, other_var->name) == 0 &&
918 other_var->max_array_access > size) {
919 size = other_var->max_array_access;
924 if (size + 1 != var->type->fields.array->length) {
925 var->type = glsl_type::get_array_instance(var->type->fields.array,
927 /* FINISHME: We should update the types of array
928 * dereferences of this variable now.
936 add_uniform(void *mem_ctx, exec_list *uniforms, struct hash_table *ht,
937 const char *name, const glsl_type *type, GLenum shader_type,
938 unsigned *next_shader_pos, unsigned *total_uniforms)
940 if (type->is_record()) {
941 for (unsigned int i = 0; i < type->length; i++) {
942 const glsl_type *field_type = type->fields.structure[i].type;
943 char *field_name = talloc_asprintf(mem_ctx, "%s.%s", name,
944 type->fields.structure[i].name);
946 add_uniform(mem_ctx, uniforms, ht, field_name, field_type,
947 shader_type, next_shader_pos, total_uniforms);
950 uniform_node *n = (uniform_node *) hash_table_find(ht, name);
951 unsigned int vec4_slots;
952 const glsl_type *array_elem_type = NULL;
954 if (type->is_array()) {
955 array_elem_type = type->fields.array;
956 /* Array of structures. */
957 if (array_elem_type->is_record()) {
958 for (unsigned int i = 0; i < type->length; i++) {
959 char *elem_name = talloc_asprintf(mem_ctx, "%s[%d]", name, i);
960 add_uniform(mem_ctx, uniforms, ht, elem_name, array_elem_type,
961 shader_type, next_shader_pos, total_uniforms);
967 /* Fix the storage size of samplers at 1 vec4 each. Be sure to pad out
968 * vectors to vec4 slots.
970 if (type->is_array()) {
971 if (array_elem_type->is_sampler())
972 vec4_slots = type->length;
974 vec4_slots = type->length * array_elem_type->matrix_columns;
975 } else if (type->is_sampler()) {
978 vec4_slots = type->matrix_columns;
982 n = (uniform_node *) calloc(1, sizeof(struct uniform_node));
983 n->u = (gl_uniform *) calloc(1, sizeof(struct gl_uniform));
984 n->slots = vec4_slots;
986 n->u->Name = strdup(name);
993 hash_table_insert(ht, n, name);
994 uniforms->push_tail(& n->link);
997 switch (shader_type) {
998 case GL_VERTEX_SHADER:
999 n->u->VertPos = *next_shader_pos;
1001 case GL_FRAGMENT_SHADER:
1002 n->u->FragPos = *next_shader_pos;
1004 case GL_GEOMETRY_SHADER:
1005 n->u->GeomPos = *next_shader_pos;
1009 (*next_shader_pos) += vec4_slots;
1014 assign_uniform_locations(struct gl_shader_program *prog)
1018 unsigned total_uniforms = 0;
1019 hash_table *ht = hash_table_ctor(32, hash_table_string_hash,
1020 hash_table_string_compare);
1021 void *mem_ctx = talloc_new(NULL);
1023 for (unsigned i = 0; i < MESA_SHADER_TYPES; i++) {
1024 if (prog->_LinkedShaders[i] == NULL)
1027 unsigned next_position = 0;
1029 foreach_list(node, prog->_LinkedShaders[i]->ir) {
1030 ir_variable *const var = ((ir_instruction *) node)->as_variable();
1032 if ((var == NULL) || (var->mode != ir_var_uniform))
1035 if (strncmp(var->name, "gl_", 3) == 0) {
1036 /* At the moment, we don't allocate uniform locations for
1037 * builtin uniforms. It's permitted by spec, and we'll
1038 * likely switch to doing that at some point, but not yet.
1043 var->location = next_position;
1044 add_uniform(mem_ctx, &uniforms, ht, var->name, var->type,
1045 prog->_LinkedShaders[i]->Type,
1046 &next_position, &total_uniforms);
1050 talloc_free(mem_ctx);
1052 gl_uniform_list *ul = (gl_uniform_list *)
1053 calloc(1, sizeof(gl_uniform_list));
1055 ul->Size = total_uniforms;
1056 ul->NumUniforms = total_uniforms;
1057 ul->Uniforms = (gl_uniform *) calloc(total_uniforms, sizeof(gl_uniform));
1061 for (uniform_node *node = (uniform_node *) uniforms.head
1062 ; node->link.next != NULL
1064 next = (uniform_node *) node->link.next;
1066 node->link.remove();
1067 memcpy(&ul->Uniforms[idx], node->u, sizeof(gl_uniform));
1074 hash_table_dtor(ht);
1076 prog->Uniforms = ul;
1081 * Find a contiguous set of available bits in a bitmask
1083 * \param used_mask Bits representing used (1) and unused (0) locations
1084 * \param needed_count Number of contiguous bits needed.
1087 * Base location of the available bits on success or -1 on failure.
1090 find_available_slots(unsigned used_mask, unsigned needed_count)
1092 unsigned needed_mask = (1 << needed_count) - 1;
1093 const int max_bit_to_test = (8 * sizeof(used_mask)) - needed_count;
1095 /* The comparison to 32 is redundant, but without it GCC emits "warning:
1096 * cannot optimize possibly infinite loops" for the loop below.
1098 if ((needed_count == 0) || (max_bit_to_test < 0) || (max_bit_to_test > 32))
1101 for (int i = 0; i <= max_bit_to_test; i++) {
1102 if ((needed_mask & ~used_mask) == needed_mask)
1113 assign_attribute_locations(gl_shader_program *prog, unsigned max_attribute_index)
1115 /* Mark invalid attribute locations as being used.
1117 unsigned used_locations = (max_attribute_index >= 32)
1118 ? ~0 : ~((1 << max_attribute_index) - 1);
1120 gl_shader *const sh = prog->_LinkedShaders[0];
1121 assert(sh->Type == GL_VERTEX_SHADER);
1123 /* Operate in a total of four passes.
1125 * 1. Invalidate the location assignments for all vertex shader inputs.
1127 * 2. Assign locations for inputs that have user-defined (via
1128 * glBindVertexAttribLocation) locatoins.
1130 * 3. Sort the attributes without assigned locations by number of slots
1131 * required in decreasing order. Fragmentation caused by attribute
1132 * locations assigned by the application may prevent large attributes
1133 * from having enough contiguous space.
1135 * 4. Assign locations to any inputs without assigned locations.
1138 invalidate_variable_locations(sh, ir_var_in, VERT_ATTRIB_GENERIC0);
1140 if (prog->Attributes != NULL) {
1141 for (unsigned i = 0; i < prog->Attributes->NumParameters; i++) {
1142 ir_variable *const var =
1143 sh->symbols->get_variable(prog->Attributes->Parameters[i].Name);
1145 /* Note: attributes that occupy multiple slots, such as arrays or
1146 * matrices, may appear in the attrib array multiple times.
1148 if ((var == NULL) || (var->location != -1))
1151 /* From page 61 of the OpenGL 4.0 spec:
1153 * "LinkProgram will fail if the attribute bindings assigned by
1154 * BindAttribLocation do not leave not enough space to assign a
1155 * location for an active matrix attribute or an active attribute
1156 * array, both of which require multiple contiguous generic
1159 * Previous versions of the spec contain similar language but omit the
1160 * bit about attribute arrays.
1162 * Page 61 of the OpenGL 4.0 spec also says:
1164 * "It is possible for an application to bind more than one
1165 * attribute name to the same location. This is referred to as
1166 * aliasing. This will only work if only one of the aliased
1167 * attributes is active in the executable program, or if no path
1168 * through the shader consumes more than one attribute of a set
1169 * of attributes aliased to the same location. A link error can
1170 * occur if the linker determines that every path through the
1171 * shader consumes multiple aliased attributes, but
1172 * implementations are not required to generate an error in this
1175 * These two paragraphs are either somewhat contradictory, or I don't
1176 * fully understand one or both of them.
1178 /* FINISHME: The code as currently written does not support attribute
1179 * FINISHME: location aliasing (see comment above).
1181 const int attr = prog->Attributes->Parameters[i].StateIndexes[0];
1182 const unsigned slots = count_attribute_slots(var->type);
1184 /* Mask representing the contiguous slots that will be used by this
1187 const unsigned use_mask = (1 << slots) - 1;
1189 /* Generate a link error if the set of bits requested for this
1190 * attribute overlaps any previously allocated bits.
1192 if ((~(use_mask << attr) & used_locations) != used_locations) {
1193 linker_error_printf(prog,
1194 "insufficient contiguous attribute locations "
1195 "available for vertex shader input `%s'",
1200 var->location = VERT_ATTRIB_GENERIC0 + attr;
1201 used_locations |= (use_mask << attr);
1205 /* Temporary storage for the set of attributes that need locations assigned.
1211 /* Used below in the call to qsort. */
1212 static int compare(const void *a, const void *b)
1214 const temp_attr *const l = (const temp_attr *) a;
1215 const temp_attr *const r = (const temp_attr *) b;
1217 /* Reversed because we want a descending order sort below. */
1218 return r->slots - l->slots;
1222 unsigned num_attr = 0;
1224 foreach_list(node, sh->ir) {
1225 ir_variable *const var = ((ir_instruction *) node)->as_variable();
1227 if ((var == NULL) || (var->mode != ir_var_in))
1230 if (var->explicit_location) {
1231 const unsigned slots = count_attribute_slots(var->type);
1232 const unsigned use_mask = (1 << slots) - 1;
1233 const int attr = var->location - VERT_ATTRIB_GENERIC0;
1235 if ((var->location >= (int)(max_attribute_index + VERT_ATTRIB_GENERIC0))
1236 || (var->location < 0)) {
1237 linker_error_printf(prog,
1238 "invalid explicit location %d specified for "
1240 (var->location < 0) ? var->location : attr,
1243 } else if (var->location >= VERT_ATTRIB_GENERIC0) {
1244 used_locations |= (use_mask << attr);
1248 /* The location was explicitly assigned, nothing to do here.
1250 if (var->location != -1)
1253 to_assign[num_attr].slots = count_attribute_slots(var->type);
1254 to_assign[num_attr].var = var;
1258 /* If all of the attributes were assigned locations by the application (or
1259 * are built-in attributes with fixed locations), return early. This should
1260 * be the common case.
1265 qsort(to_assign, num_attr, sizeof(to_assign[0]), temp_attr::compare);
1267 /* VERT_ATTRIB_GENERIC0 is a psdueo-alias for VERT_ATTRIB_POS. It can only
1268 * be explicitly assigned by via glBindAttribLocation. Mark it as reserved
1269 * to prevent it from being automatically allocated below.
1271 find_deref_visitor find("gl_Vertex");
1273 if (find.variable_found())
1274 used_locations |= (1 << 0);
1276 for (unsigned i = 0; i < num_attr; i++) {
1277 /* Mask representing the contiguous slots that will be used by this
1280 const unsigned use_mask = (1 << to_assign[i].slots) - 1;
1282 int location = find_available_slots(used_locations, to_assign[i].slots);
1285 linker_error_printf(prog,
1286 "insufficient contiguous attribute locations "
1287 "available for vertex shader input `%s'",
1288 to_assign[i].var->name);
1292 to_assign[i].var->location = VERT_ATTRIB_GENERIC0 + location;
1293 used_locations |= (use_mask << location);
1301 * Demote shader inputs and outputs that are not used in other stages
1304 demote_shader_inputs_and_outputs(gl_shader *sh, enum ir_variable_mode mode)
1306 foreach_list(node, sh->ir) {
1307 ir_variable *const var = ((ir_instruction *) node)->as_variable();
1309 if ((var == NULL) || (var->mode != int(mode)))
1312 /* A shader 'in' or 'out' variable is only really an input or output if
1313 * its value is used by other shader stages. This will cause the variable
1314 * to have a location assigned.
1316 if (var->location == -1) {
1317 var->mode = ir_var_auto;
1324 assign_varying_locations(struct gl_shader_program *prog,
1325 gl_shader *producer, gl_shader *consumer)
1327 /* FINISHME: Set dynamically when geometry shader support is added. */
1328 unsigned output_index = VERT_RESULT_VAR0;
1329 unsigned input_index = FRAG_ATTRIB_VAR0;
1331 /* Operate in a total of three passes.
1333 * 1. Assign locations for any matching inputs and outputs.
1335 * 2. Mark output variables in the producer that do not have locations as
1336 * not being outputs. This lets the optimizer eliminate them.
1338 * 3. Mark input variables in the consumer that do not have locations as
1339 * not being inputs. This lets the optimizer eliminate them.
1342 invalidate_variable_locations(producer, ir_var_out, VERT_RESULT_VAR0);
1343 invalidate_variable_locations(consumer, ir_var_in, FRAG_ATTRIB_VAR0);
1345 foreach_list(node, producer->ir) {
1346 ir_variable *const output_var = ((ir_instruction *) node)->as_variable();
1348 if ((output_var == NULL) || (output_var->mode != ir_var_out)
1349 || (output_var->location != -1))
1352 ir_variable *const input_var =
1353 consumer->symbols->get_variable(output_var->name);
1355 if ((input_var == NULL) || (input_var->mode != ir_var_in))
1358 assert(input_var->location == -1);
1360 output_var->location = output_index;
1361 input_var->location = input_index;
1363 /* FINISHME: Support for "varying" records in GLSL 1.50. */
1364 assert(!output_var->type->is_record());
1366 if (output_var->type->is_array()) {
1367 const unsigned slots = output_var->type->length
1368 * output_var->type->fields.array->matrix_columns;
1370 output_index += slots;
1371 input_index += slots;
1373 const unsigned slots = output_var->type->matrix_columns;
1375 output_index += slots;
1376 input_index += slots;
1380 foreach_list(node, consumer->ir) {
1381 ir_variable *const var = ((ir_instruction *) node)->as_variable();
1383 if ((var == NULL) || (var->mode != ir_var_in))
1386 if (var->location == -1) {
1387 if (prog->Version <= 120) {
1388 /* On page 25 (page 31 of the PDF) of the GLSL 1.20 spec:
1390 * Only those varying variables used (i.e. read) in
1391 * the fragment shader executable must be written to
1392 * by the vertex shader executable; declaring
1393 * superfluous varying variables in a vertex shader is
1396 * We interpret this text as meaning that the VS must
1397 * write the variable for the FS to read it. See
1398 * "glsl1-varying read but not written" in piglit.
1401 linker_error_printf(prog, "fragment shader varying %s not written "
1402 "by vertex shader\n.", var->name);
1403 prog->LinkStatus = false;
1406 /* An 'in' variable is only really a shader input if its
1407 * value is written by the previous stage.
1409 var->mode = ir_var_auto;
1416 link_shaders(struct gl_context *ctx, struct gl_shader_program *prog)
1418 void *mem_ctx = talloc_init("temporary linker context");
1420 prog->LinkStatus = false;
1421 prog->Validated = false;
1422 prog->_Used = false;
1424 if (prog->InfoLog != NULL)
1425 talloc_free(prog->InfoLog);
1427 prog->InfoLog = talloc_strdup(NULL, "");
1429 /* Separate the shaders into groups based on their type.
1431 struct gl_shader **vert_shader_list;
1432 unsigned num_vert_shaders = 0;
1433 struct gl_shader **frag_shader_list;
1434 unsigned num_frag_shaders = 0;
1436 vert_shader_list = (struct gl_shader **)
1437 calloc(2 * prog->NumShaders, sizeof(struct gl_shader *));
1438 frag_shader_list = &vert_shader_list[prog->NumShaders];
1440 unsigned min_version = UINT_MAX;
1441 unsigned max_version = 0;
1442 for (unsigned i = 0; i < prog->NumShaders; i++) {
1443 min_version = MIN2(min_version, prog->Shaders[i]->Version);
1444 max_version = MAX2(max_version, prog->Shaders[i]->Version);
1446 switch (prog->Shaders[i]->Type) {
1447 case GL_VERTEX_SHADER:
1448 vert_shader_list[num_vert_shaders] = prog->Shaders[i];
1451 case GL_FRAGMENT_SHADER:
1452 frag_shader_list[num_frag_shaders] = prog->Shaders[i];
1455 case GL_GEOMETRY_SHADER:
1456 /* FINISHME: Support geometry shaders. */
1457 assert(prog->Shaders[i]->Type != GL_GEOMETRY_SHADER);
1462 /* Previous to GLSL version 1.30, different compilation units could mix and
1463 * match shading language versions. With GLSL 1.30 and later, the versions
1464 * of all shaders must match.
1466 assert(min_version >= 100);
1467 assert(max_version <= 130);
1468 if ((max_version >= 130 || min_version == 100)
1469 && min_version != max_version) {
1470 linker_error_printf(prog, "all shaders must use same shading "
1471 "language version\n");
1475 prog->Version = max_version;
1477 for (unsigned int i = 0; i < MESA_SHADER_TYPES; i++) {
1478 if (prog->_LinkedShaders[i] != NULL)
1479 ctx->Driver.DeleteShader(ctx, prog->_LinkedShaders[i]);
1481 prog->_LinkedShaders[i] = NULL;
1484 /* Link all shaders for a particular stage and validate the result.
1486 if (num_vert_shaders > 0) {
1487 gl_shader *const sh =
1488 link_intrastage_shaders(mem_ctx, ctx, prog, vert_shader_list,
1494 if (!validate_vertex_shader_executable(prog, sh))
1497 _mesa_reference_shader(ctx, &prog->_LinkedShaders[MESA_SHADER_VERTEX],
1501 if (num_frag_shaders > 0) {
1502 gl_shader *const sh =
1503 link_intrastage_shaders(mem_ctx, ctx, prog, frag_shader_list,
1509 if (!validate_fragment_shader_executable(prog, sh))
1512 _mesa_reference_shader(ctx, &prog->_LinkedShaders[MESA_SHADER_FRAGMENT],
1516 /* Here begins the inter-stage linking phase. Some initial validation is
1517 * performed, then locations are assigned for uniforms, attributes, and
1520 if (cross_validate_uniforms(prog)) {
1523 for (prev = 0; prev < MESA_SHADER_TYPES; prev++) {
1524 if (prog->_LinkedShaders[prev] != NULL)
1528 /* Validate the inputs of each stage with the output of the preceeding
1531 for (unsigned i = prev + 1; i < MESA_SHADER_TYPES; i++) {
1532 if (prog->_LinkedShaders[i] == NULL)
1535 if (!cross_validate_outputs_to_inputs(prog,
1536 prog->_LinkedShaders[prev],
1537 prog->_LinkedShaders[i]))
1543 prog->LinkStatus = true;
1546 /* Do common optimization before assigning storage for attributes,
1547 * uniforms, and varyings. Later optimization could possibly make
1548 * some of that unused.
1550 for (unsigned i = 0; i < MESA_SHADER_TYPES; i++) {
1551 if (prog->_LinkedShaders[i] == NULL)
1554 while (do_common_optimization(prog->_LinkedShaders[i]->ir, true, 32))
1558 update_array_sizes(prog);
1560 assign_uniform_locations(prog);
1562 if (prog->_LinkedShaders[MESA_SHADER_VERTEX] != NULL) {
1563 /* FINISHME: The value of the max_attribute_index parameter is
1564 * FINISHME: implementation dependent based on the value of
1565 * FINISHME: GL_MAX_VERTEX_ATTRIBS. GL_MAX_VERTEX_ATTRIBS must be
1566 * FINISHME: at least 16, so hardcode 16 for now.
1568 if (!assign_attribute_locations(prog, 16)) {
1569 prog->LinkStatus = false;
1575 for (prev = 0; prev < MESA_SHADER_TYPES; prev++) {
1576 if (prog->_LinkedShaders[prev] != NULL)
1580 for (unsigned i = prev + 1; i < MESA_SHADER_TYPES; i++) {
1581 if (prog->_LinkedShaders[i] == NULL)
1584 assign_varying_locations(prog,
1585 prog->_LinkedShaders[prev],
1586 prog->_LinkedShaders[i]);
1590 if (prog->_LinkedShaders[MESA_SHADER_VERTEX] != NULL) {
1591 demote_shader_inputs_and_outputs(prog->_LinkedShaders[MESA_SHADER_VERTEX],
1595 if (prog->_LinkedShaders[MESA_SHADER_GEOMETRY] != NULL) {
1596 gl_shader *const sh = prog->_LinkedShaders[MESA_SHADER_GEOMETRY];
1598 demote_shader_inputs_and_outputs(sh, ir_var_in);
1599 demote_shader_inputs_and_outputs(sh, ir_var_inout);
1600 demote_shader_inputs_and_outputs(sh, ir_var_out);
1603 if (prog->_LinkedShaders[MESA_SHADER_FRAGMENT] != NULL) {
1604 gl_shader *const sh = prog->_LinkedShaders[MESA_SHADER_FRAGMENT];
1606 demote_shader_inputs_and_outputs(sh, ir_var_in);
1609 /* FINISHME: Assign fragment shader output locations. */
1612 free(vert_shader_list);
1614 for (unsigned i = 0; i < MESA_SHADER_TYPES; i++) {
1615 if (prog->_LinkedShaders[i] == NULL)
1618 /* Retain any live IR, but trash the rest. */
1619 reparent_ir(prog->_LinkedShaders[i]->ir, prog->_LinkedShaders[i]->ir);
1622 talloc_free(mem_ctx);