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>
67 #include "main/core.h"
68 #include "glsl_symbol_table.h"
71 #include "program/hash_table.h"
73 #include "ir_optimization.h"
76 #include "main/shaderobj.h"
80 * Visitor that determines whether or not a variable is ever written.
82 class find_assignment_visitor : public ir_hierarchical_visitor {
84 find_assignment_visitor(const char *name)
85 : name(name), found(false)
90 virtual ir_visitor_status visit_enter(ir_assignment *ir)
92 ir_variable *const var = ir->lhs->variable_referenced();
94 if (strcmp(name, var->name) == 0) {
99 return visit_continue_with_parent;
102 virtual ir_visitor_status visit_enter(ir_call *ir)
104 exec_list_iterator sig_iter = ir->get_callee()->parameters.iterator();
105 foreach_iter(exec_list_iterator, iter, *ir) {
106 ir_rvalue *param_rval = (ir_rvalue *)iter.get();
107 ir_variable *sig_param = (ir_variable *)sig_iter.get();
109 if (sig_param->mode == ir_var_out ||
110 sig_param->mode == ir_var_inout) {
111 ir_variable *var = param_rval->variable_referenced();
112 if (var && strcmp(name, var->name) == 0) {
120 return visit_continue_with_parent;
123 bool variable_found()
129 const char *name; /**< Find writes to a variable with this name. */
130 bool found; /**< Was a write to the variable found? */
135 * Visitor that determines whether or not a variable is ever read.
137 class find_deref_visitor : public ir_hierarchical_visitor {
139 find_deref_visitor(const char *name)
140 : name(name), found(false)
145 virtual ir_visitor_status visit(ir_dereference_variable *ir)
147 if (strcmp(this->name, ir->var->name) == 0) {
152 return visit_continue;
155 bool variable_found() const
161 const char *name; /**< Find writes to a variable with this name. */
162 bool found; /**< Was a write to the variable found? */
167 linker_error_printf(gl_shader_program *prog, const char *fmt, ...)
171 ralloc_strcat(&prog->InfoLog, "error: ");
173 ralloc_vasprintf_append(&prog->InfoLog, fmt, ap);
179 invalidate_variable_locations(gl_shader *sh, enum ir_variable_mode mode,
182 foreach_list(node, sh->ir) {
183 ir_variable *const var = ((ir_instruction *) node)->as_variable();
185 if ((var == NULL) || (var->mode != (unsigned) mode))
188 /* Only assign locations for generic attributes / varyings / etc.
190 if ((var->location >= generic_base) && !var->explicit_location)
197 * Determine the number of attribute slots required for a particular type
199 * This code is here because it implements the language rules of a specific
200 * GLSL version. Since it's a property of the language and not a property of
201 * types in general, it doesn't really belong in glsl_type.
204 count_attribute_slots(const glsl_type *t)
206 /* From page 31 (page 37 of the PDF) of the GLSL 1.50 spec:
208 * "A scalar input counts the same amount against this limit as a vec4,
209 * so applications may want to consider packing groups of four
210 * unrelated float inputs together into a vector to better utilize the
211 * capabilities of the underlying hardware. A matrix input will use up
212 * multiple locations. The number of locations used will equal the
213 * number of columns in the matrix."
215 * The spec does not explicitly say how arrays are counted. However, it
216 * should be safe to assume the total number of slots consumed by an array
217 * is the number of entries in the array multiplied by the number of slots
218 * consumed by a single element of the array.
222 return t->array_size() * count_attribute_slots(t->element_type());
225 return t->matrix_columns;
232 * Verify that a vertex shader executable meets all semantic requirements
234 * \param shader Vertex shader executable to be verified
237 validate_vertex_shader_executable(struct gl_shader_program *prog,
238 struct gl_shader *shader)
243 find_assignment_visitor find("gl_Position");
244 find.run(shader->ir);
245 if (!find.variable_found()) {
246 linker_error_printf(prog,
247 "vertex shader does not write to `gl_Position'\n");
256 * Verify that a fragment shader executable meets all semantic requirements
258 * \param shader Fragment shader executable to be verified
261 validate_fragment_shader_executable(struct gl_shader_program *prog,
262 struct gl_shader *shader)
267 find_assignment_visitor frag_color("gl_FragColor");
268 find_assignment_visitor frag_data("gl_FragData");
270 frag_color.run(shader->ir);
271 frag_data.run(shader->ir);
273 if (frag_color.variable_found() && frag_data.variable_found()) {
274 linker_error_printf(prog, "fragment shader writes to both "
275 "`gl_FragColor' and `gl_FragData'\n");
284 * Generate a string describing the mode of a variable
287 mode_string(const ir_variable *var)
291 return (var->read_only) ? "global constant" : "global variable";
293 case ir_var_uniform: return "uniform";
294 case ir_var_in: return "shader input";
295 case ir_var_out: return "shader output";
296 case ir_var_inout: return "shader inout";
298 case ir_var_const_in:
299 case ir_var_temporary:
301 assert(!"Should not get here.");
302 return "invalid variable";
308 * Perform validation of global variables used across multiple shaders
311 cross_validate_globals(struct gl_shader_program *prog,
312 struct gl_shader **shader_list,
313 unsigned num_shaders,
316 /* Examine all of the uniforms in all of the shaders and cross validate
319 glsl_symbol_table variables;
320 for (unsigned i = 0; i < num_shaders; i++) {
321 if (shader_list[i] == NULL)
324 foreach_list(node, shader_list[i]->ir) {
325 ir_variable *const var = ((ir_instruction *) node)->as_variable();
330 if (uniforms_only && (var->mode != ir_var_uniform))
333 /* Don't cross validate temporaries that are at global scope. These
334 * will eventually get pulled into the shaders 'main'.
336 if (var->mode == ir_var_temporary)
339 /* If a global with this name has already been seen, verify that the
340 * new instance has the same type. In addition, if the globals have
341 * initializers, the values of the initializers must be the same.
343 ir_variable *const existing = variables.get_variable(var->name);
344 if (existing != NULL) {
345 if (var->type != existing->type) {
346 /* Consider the types to be "the same" if both types are arrays
347 * of the same type and one of the arrays is implicitly sized.
348 * In addition, set the type of the linked variable to the
349 * explicitly sized array.
351 if (var->type->is_array()
352 && existing->type->is_array()
353 && (var->type->fields.array == existing->type->fields.array)
354 && ((var->type->length == 0)
355 || (existing->type->length == 0))) {
356 if (var->type->length != 0) {
357 existing->type = var->type;
360 linker_error_printf(prog, "%s `%s' declared as type "
361 "`%s' and type `%s'\n",
363 var->name, var->type->name,
364 existing->type->name);
369 if (var->explicit_location) {
370 if (existing->explicit_location
371 && (var->location != existing->location)) {
372 linker_error_printf(prog, "explicit locations for %s "
373 "`%s' have differing values\n",
374 mode_string(var), var->name);
378 existing->location = var->location;
379 existing->explicit_location = true;
382 /* Validate layout qualifiers for gl_FragDepth.
384 * From the AMD_conservative_depth spec:
385 * "If gl_FragDepth is redeclared in any fragment shader in
386 * a program, it must be redeclared in all fragment shaders in that
387 * program that have static assignments to gl_FragDepth. All
388 * redeclarations of gl_FragDepth in all fragment shaders in
389 * a single program must have the same set of qualifiers."
391 if (strcmp(var->name, "gl_FragDepth") == 0) {
392 bool layout_declared = var->depth_layout != ir_depth_layout_none;
393 bool layout_differs = var->depth_layout != existing->depth_layout;
394 if (layout_declared && layout_differs) {
395 linker_error_printf(prog,
396 "All redeclarations of gl_FragDepth in all fragment shaders "
397 "in a single program must have the same set of qualifiers.");
399 if (var->used && layout_differs) {
400 linker_error_printf(prog,
401 "If gl_FragDepth is redeclared with a layout qualifier in"
402 "any fragment shader, it must be redeclared with the same"
403 "layout qualifier in all fragment shaders that have"
404 "assignments to gl_FragDepth");
408 /* FINISHME: Handle non-constant initializers.
410 if (var->constant_value != NULL) {
411 if (existing->constant_value != NULL) {
412 if (!var->constant_value->has_value(existing->constant_value)) {
413 linker_error_printf(prog, "initializers for %s "
414 "`%s' have differing values\n",
415 mode_string(var), var->name);
419 /* If the first-seen instance of a particular uniform did not
420 * have an initializer but a later instance does, copy the
421 * initializer to the version stored in the symbol table.
423 /* FINISHME: This is wrong. The constant_value field should
424 * FINISHME: not be modified! Imagine a case where a shader
425 * FINISHME: without an initializer is linked in two different
426 * FINISHME: programs with shaders that have differing
427 * FINISHME: initializers. Linking with the first will
428 * FINISHME: modify the shader, and linking with the second
429 * FINISHME: will fail.
431 existing->constant_value =
432 var->constant_value->clone(ralloc_parent(existing), NULL);
435 if (existing->invariant != var->invariant) {
436 linker_error_printf(prog, "declarations for %s `%s' have "
437 "mismatching invariant qualifiers\n",
438 mode_string(var), var->name);
441 if (existing->centroid != var->centroid) {
442 linker_error_printf(prog, "declarations for %s `%s' have "
443 "mismatching centroid qualifiers\n",
444 mode_string(var), var->name);
448 variables.add_variable(var);
457 * Perform validation of uniforms used across multiple shader stages
460 cross_validate_uniforms(struct gl_shader_program *prog)
462 return cross_validate_globals(prog, prog->_LinkedShaders,
463 MESA_SHADER_TYPES, true);
468 * Validate that outputs from one stage match inputs of another
471 cross_validate_outputs_to_inputs(struct gl_shader_program *prog,
472 gl_shader *producer, gl_shader *consumer)
474 glsl_symbol_table parameters;
475 /* FINISHME: Figure these out dynamically. */
476 const char *const producer_stage = "vertex";
477 const char *const consumer_stage = "fragment";
479 /* Find all shader outputs in the "producer" stage.
481 foreach_list(node, producer->ir) {
482 ir_variable *const var = ((ir_instruction *) node)->as_variable();
484 /* FINISHME: For geometry shaders, this should also look for inout
485 * FINISHME: variables.
487 if ((var == NULL) || (var->mode != ir_var_out))
490 parameters.add_variable(var);
494 /* Find all shader inputs in the "consumer" stage. Any variables that have
495 * matching outputs already in the symbol table must have the same type and
498 foreach_list(node, consumer->ir) {
499 ir_variable *const input = ((ir_instruction *) node)->as_variable();
501 /* FINISHME: For geometry shaders, this should also look for inout
502 * FINISHME: variables.
504 if ((input == NULL) || (input->mode != ir_var_in))
507 ir_variable *const output = parameters.get_variable(input->name);
508 if (output != NULL) {
509 /* Check that the types match between stages.
511 if (input->type != output->type) {
512 /* There is a bit of a special case for gl_TexCoord. This
513 * built-in is unsized by default. Appliations that variable
514 * access it must redeclare it with a size. There is some
515 * language in the GLSL spec that implies the fragment shader
516 * and vertex shader do not have to agree on this size. Other
517 * driver behave this way, and one or two applications seem to
520 * Neither declaration needs to be modified here because the array
521 * sizes are fixed later when update_array_sizes is called.
523 * From page 48 (page 54 of the PDF) of the GLSL 1.10 spec:
525 * "Unlike user-defined varying variables, the built-in
526 * varying variables don't have a strict one-to-one
527 * correspondence between the vertex language and the
528 * fragment language."
530 if (!output->type->is_array()
531 || (strncmp("gl_", output->name, 3) != 0)) {
532 linker_error_printf(prog,
533 "%s shader output `%s' declared as "
534 "type `%s', but %s shader input declared "
536 producer_stage, output->name,
538 consumer_stage, input->type->name);
543 /* Check that all of the qualifiers match between stages.
545 if (input->centroid != output->centroid) {
546 linker_error_printf(prog,
547 "%s shader output `%s' %s centroid qualifier, "
548 "but %s shader input %s centroid qualifier\n",
551 (output->centroid) ? "has" : "lacks",
553 (input->centroid) ? "has" : "lacks");
557 if (input->invariant != output->invariant) {
558 linker_error_printf(prog,
559 "%s shader output `%s' %s invariant qualifier, "
560 "but %s shader input %s invariant qualifier\n",
563 (output->invariant) ? "has" : "lacks",
565 (input->invariant) ? "has" : "lacks");
569 if (input->interpolation != output->interpolation) {
570 linker_error_printf(prog,
571 "%s shader output `%s' specifies %s "
572 "interpolation qualifier, "
573 "but %s shader input specifies %s "
574 "interpolation qualifier\n",
577 output->interpolation_string(),
579 input->interpolation_string());
590 * Populates a shaders symbol table with all global declarations
593 populate_symbol_table(gl_shader *sh)
595 sh->symbols = new(sh) glsl_symbol_table;
597 foreach_list(node, sh->ir) {
598 ir_instruction *const inst = (ir_instruction *) node;
602 if ((func = inst->as_function()) != NULL) {
603 sh->symbols->add_function(func);
604 } else if ((var = inst->as_variable()) != NULL) {
605 sh->symbols->add_variable(var);
612 * Remap variables referenced in an instruction tree
614 * This is used when instruction trees are cloned from one shader and placed in
615 * another. These trees will contain references to \c ir_variable nodes that
616 * do not exist in the target shader. This function finds these \c ir_variable
617 * references and replaces the references with matching variables in the target
620 * If there is no matching variable in the target shader, a clone of the
621 * \c ir_variable is made and added to the target shader. The new variable is
622 * added to \b both the instruction stream and the symbol table.
624 * \param inst IR tree that is to be processed.
625 * \param symbols Symbol table containing global scope symbols in the
627 * \param instructions Instruction stream where new variable declarations
631 remap_variables(ir_instruction *inst, struct gl_shader *target,
634 class remap_visitor : public ir_hierarchical_visitor {
636 remap_visitor(struct gl_shader *target,
639 this->target = target;
640 this->symbols = target->symbols;
641 this->instructions = target->ir;
645 virtual ir_visitor_status visit(ir_dereference_variable *ir)
647 if (ir->var->mode == ir_var_temporary) {
648 ir_variable *var = (ir_variable *) hash_table_find(temps, ir->var);
652 return visit_continue;
655 ir_variable *const existing =
656 this->symbols->get_variable(ir->var->name);
657 if (existing != NULL)
660 ir_variable *copy = ir->var->clone(this->target, NULL);
662 this->symbols->add_variable(copy);
663 this->instructions->push_head(copy);
667 return visit_continue;
671 struct gl_shader *target;
672 glsl_symbol_table *symbols;
673 exec_list *instructions;
677 remap_visitor v(target, temps);
684 * Move non-declarations from one instruction stream to another
686 * The intended usage pattern of this function is to pass the pointer to the
687 * head sentinel of a list (i.e., a pointer to the list cast to an \c exec_node
688 * pointer) for \c last and \c false for \c make_copies on the first
689 * call. Successive calls pass the return value of the previous call for
690 * \c last and \c true for \c make_copies.
692 * \param instructions Source instruction stream
693 * \param last Instruction after which new instructions should be
694 * inserted in the target instruction stream
695 * \param make_copies Flag selecting whether instructions in \c instructions
696 * should be copied (via \c ir_instruction::clone) into the
697 * target list or moved.
700 * The new "last" instruction in the target instruction stream. This pointer
701 * is suitable for use as the \c last parameter of a later call to this
705 move_non_declarations(exec_list *instructions, exec_node *last,
706 bool make_copies, gl_shader *target)
708 hash_table *temps = NULL;
711 temps = hash_table_ctor(0, hash_table_pointer_hash,
712 hash_table_pointer_compare);
714 foreach_list_safe(node, instructions) {
715 ir_instruction *inst = (ir_instruction *) node;
717 if (inst->as_function())
720 ir_variable *var = inst->as_variable();
721 if ((var != NULL) && (var->mode != ir_var_temporary))
724 assert(inst->as_assignment()
725 || ((var != NULL) && (var->mode == ir_var_temporary)));
728 inst = inst->clone(target, NULL);
731 hash_table_insert(temps, inst, var);
733 remap_variables(inst, target, temps);
738 last->insert_after(inst);
743 hash_table_dtor(temps);
749 * Get the function signature for main from a shader
751 static ir_function_signature *
752 get_main_function_signature(gl_shader *sh)
754 ir_function *const f = sh->symbols->get_function("main");
756 exec_list void_parameters;
758 /* Look for the 'void main()' signature and ensure that it's defined.
759 * This keeps the linker from accidentally pick a shader that just
760 * contains a prototype for main.
762 * We don't have to check for multiple definitions of main (in multiple
763 * shaders) because that would have already been caught above.
765 ir_function_signature *sig = f->matching_signature(&void_parameters);
766 if ((sig != NULL) && sig->is_defined) {
776 * Combine a group of shaders for a single stage to generate a linked shader
779 * If this function is supplied a single shader, it is cloned, and the new
780 * shader is returned.
782 static struct gl_shader *
783 link_intrastage_shaders(void *mem_ctx,
784 struct gl_context *ctx,
785 struct gl_shader_program *prog,
786 struct gl_shader **shader_list,
787 unsigned num_shaders)
789 /* Check that global variables defined in multiple shaders are consistent.
791 if (!cross_validate_globals(prog, shader_list, num_shaders, false))
794 /* Check that there is only a single definition of each function signature
795 * across all shaders.
797 for (unsigned i = 0; i < (num_shaders - 1); i++) {
798 foreach_list(node, shader_list[i]->ir) {
799 ir_function *const f = ((ir_instruction *) node)->as_function();
804 for (unsigned j = i + 1; j < num_shaders; j++) {
805 ir_function *const other =
806 shader_list[j]->symbols->get_function(f->name);
808 /* If the other shader has no function (and therefore no function
809 * signatures) with the same name, skip to the next shader.
814 foreach_iter (exec_list_iterator, iter, *f) {
815 ir_function_signature *sig =
816 (ir_function_signature *) iter.get();
818 if (!sig->is_defined || sig->is_builtin)
821 ir_function_signature *other_sig =
822 other->exact_matching_signature(& sig->parameters);
824 if ((other_sig != NULL) && other_sig->is_defined
825 && !other_sig->is_builtin) {
826 linker_error_printf(prog,
827 "function `%s' is multiply defined",
836 /* Find the shader that defines main, and make a clone of it.
838 * Starting with the clone, search for undefined references. If one is
839 * found, find the shader that defines it. Clone the reference and add
840 * it to the shader. Repeat until there are no undefined references or
841 * until a reference cannot be resolved.
843 gl_shader *main = NULL;
844 for (unsigned i = 0; i < num_shaders; i++) {
845 if (get_main_function_signature(shader_list[i]) != NULL) {
846 main = shader_list[i];
852 linker_error_printf(prog, "%s shader lacks `main'\n",
853 (shader_list[0]->Type == GL_VERTEX_SHADER)
854 ? "vertex" : "fragment");
858 gl_shader *linked = ctx->Driver.NewShader(NULL, 0, main->Type);
859 linked->ir = new(linked) exec_list;
860 clone_ir_list(mem_ctx, linked->ir, main->ir);
862 populate_symbol_table(linked);
864 /* The a pointer to the main function in the final linked shader (i.e., the
865 * copy of the original shader that contained the main function).
867 ir_function_signature *const main_sig = get_main_function_signature(linked);
869 /* Move any instructions other than variable declarations or function
870 * declarations into main.
872 exec_node *insertion_point =
873 move_non_declarations(linked->ir, (exec_node *) &main_sig->body, false,
876 for (unsigned i = 0; i < num_shaders; i++) {
877 if (shader_list[i] == main)
880 insertion_point = move_non_declarations(shader_list[i]->ir,
881 insertion_point, true, linked);
884 /* Resolve initializers for global variables in the linked shader.
886 unsigned num_linking_shaders = num_shaders;
887 for (unsigned i = 0; i < num_shaders; i++)
888 num_linking_shaders += shader_list[i]->num_builtins_to_link;
890 gl_shader **linking_shaders =
891 (gl_shader **) calloc(num_linking_shaders, sizeof(gl_shader *));
893 memcpy(linking_shaders, shader_list,
894 sizeof(linking_shaders[0]) * num_shaders);
896 unsigned idx = num_shaders;
897 for (unsigned i = 0; i < num_shaders; i++) {
898 memcpy(&linking_shaders[idx], shader_list[i]->builtins_to_link,
899 sizeof(linking_shaders[0]) * shader_list[i]->num_builtins_to_link);
900 idx += shader_list[i]->num_builtins_to_link;
903 assert(idx == num_linking_shaders);
905 if (!link_function_calls(prog, linked, linking_shaders,
906 num_linking_shaders)) {
907 ctx->Driver.DeleteShader(ctx, linked);
911 free(linking_shaders);
913 /* Make a pass over all variable declarations to ensure that arrays with
914 * unspecified sizes have a size specified. The size is inferred from the
915 * max_array_access field.
917 if (linked != NULL) {
918 class array_sizing_visitor : public ir_hierarchical_visitor {
920 virtual ir_visitor_status visit(ir_variable *var)
922 if (var->type->is_array() && (var->type->length == 0)) {
923 const glsl_type *type =
924 glsl_type::get_array_instance(var->type->fields.array,
925 var->max_array_access + 1);
927 assert(type != NULL);
931 return visit_continue;
942 struct uniform_node {
944 struct gl_uniform *u;
949 * Update the sizes of linked shader uniform arrays to the maximum
952 * From page 81 (page 95 of the PDF) of the OpenGL 2.1 spec:
954 * If one or more elements of an array are active,
955 * GetActiveUniform will return the name of the array in name,
956 * subject to the restrictions listed above. The type of the array
957 * is returned in type. The size parameter contains the highest
958 * array element index used, plus one. The compiler or linker
959 * determines the highest index used. There will be only one
960 * active uniform reported by the GL per uniform array.
964 update_array_sizes(struct gl_shader_program *prog)
966 for (unsigned i = 0; i < MESA_SHADER_TYPES; i++) {
967 if (prog->_LinkedShaders[i] == NULL)
970 foreach_list(node, prog->_LinkedShaders[i]->ir) {
971 ir_variable *const var = ((ir_instruction *) node)->as_variable();
973 if ((var == NULL) || (var->mode != ir_var_uniform &&
974 var->mode != ir_var_in &&
975 var->mode != ir_var_out) ||
976 !var->type->is_array())
979 unsigned int size = var->max_array_access;
980 for (unsigned j = 0; j < MESA_SHADER_TYPES; j++) {
981 if (prog->_LinkedShaders[j] == NULL)
984 foreach_list(node2, prog->_LinkedShaders[j]->ir) {
985 ir_variable *other_var = ((ir_instruction *) node2)->as_variable();
989 if (strcmp(var->name, other_var->name) == 0 &&
990 other_var->max_array_access > size) {
991 size = other_var->max_array_access;
996 if (size + 1 != var->type->fields.array->length) {
997 var->type = glsl_type::get_array_instance(var->type->fields.array,
999 /* FINISHME: We should update the types of array
1000 * dereferences of this variable now.
1008 add_uniform(void *mem_ctx, exec_list *uniforms, struct hash_table *ht,
1009 const char *name, const glsl_type *type, GLenum shader_type,
1010 unsigned *next_shader_pos, unsigned *total_uniforms)
1012 if (type->is_record()) {
1013 for (unsigned int i = 0; i < type->length; i++) {
1014 const glsl_type *field_type = type->fields.structure[i].type;
1015 char *field_name = ralloc_asprintf(mem_ctx, "%s.%s", name,
1016 type->fields.structure[i].name);
1018 add_uniform(mem_ctx, uniforms, ht, field_name, field_type,
1019 shader_type, next_shader_pos, total_uniforms);
1022 uniform_node *n = (uniform_node *) hash_table_find(ht, name);
1023 unsigned int vec4_slots;
1024 const glsl_type *array_elem_type = NULL;
1026 if (type->is_array()) {
1027 array_elem_type = type->fields.array;
1028 /* Array of structures. */
1029 if (array_elem_type->is_record()) {
1030 for (unsigned int i = 0; i < type->length; i++) {
1031 char *elem_name = ralloc_asprintf(mem_ctx, "%s[%d]", name, i);
1032 add_uniform(mem_ctx, uniforms, ht, elem_name, array_elem_type,
1033 shader_type, next_shader_pos, total_uniforms);
1039 /* Fix the storage size of samplers at 1 vec4 each. Be sure to pad out
1040 * vectors to vec4 slots.
1042 if (type->is_array()) {
1043 if (array_elem_type->is_sampler())
1044 vec4_slots = type->length;
1046 vec4_slots = type->length * array_elem_type->matrix_columns;
1047 } else if (type->is_sampler()) {
1050 vec4_slots = type->matrix_columns;
1054 n = (uniform_node *) calloc(1, sizeof(struct uniform_node));
1055 n->u = (gl_uniform *) calloc(1, sizeof(struct gl_uniform));
1056 n->slots = vec4_slots;
1058 n->u->Name = strdup(name);
1063 (*total_uniforms)++;
1065 hash_table_insert(ht, n, name);
1066 uniforms->push_tail(& n->link);
1069 switch (shader_type) {
1070 case GL_VERTEX_SHADER:
1071 n->u->VertPos = *next_shader_pos;
1073 case GL_FRAGMENT_SHADER:
1074 n->u->FragPos = *next_shader_pos;
1076 case GL_GEOMETRY_SHADER:
1077 n->u->GeomPos = *next_shader_pos;
1081 (*next_shader_pos) += vec4_slots;
1086 assign_uniform_locations(struct gl_shader_program *prog)
1090 unsigned total_uniforms = 0;
1091 hash_table *ht = hash_table_ctor(32, hash_table_string_hash,
1092 hash_table_string_compare);
1093 void *mem_ctx = ralloc_context(NULL);
1095 for (unsigned i = 0; i < MESA_SHADER_TYPES; i++) {
1096 if (prog->_LinkedShaders[i] == NULL)
1099 unsigned next_position = 0;
1101 foreach_list(node, prog->_LinkedShaders[i]->ir) {
1102 ir_variable *const var = ((ir_instruction *) node)->as_variable();
1104 if ((var == NULL) || (var->mode != ir_var_uniform))
1107 if (strncmp(var->name, "gl_", 3) == 0) {
1108 /* At the moment, we don't allocate uniform locations for
1109 * builtin uniforms. It's permitted by spec, and we'll
1110 * likely switch to doing that at some point, but not yet.
1115 var->location = next_position;
1116 add_uniform(mem_ctx, &uniforms, ht, var->name, var->type,
1117 prog->_LinkedShaders[i]->Type,
1118 &next_position, &total_uniforms);
1122 ralloc_free(mem_ctx);
1124 gl_uniform_list *ul = (gl_uniform_list *)
1125 calloc(1, sizeof(gl_uniform_list));
1127 ul->Size = total_uniforms;
1128 ul->NumUniforms = total_uniforms;
1129 ul->Uniforms = (gl_uniform *) calloc(total_uniforms, sizeof(gl_uniform));
1133 for (uniform_node *node = (uniform_node *) uniforms.head
1134 ; node->link.next != NULL
1136 next = (uniform_node *) node->link.next;
1138 node->link.remove();
1139 memcpy(&ul->Uniforms[idx], node->u, sizeof(gl_uniform));
1146 hash_table_dtor(ht);
1148 prog->Uniforms = ul;
1153 * Find a contiguous set of available bits in a bitmask
1155 * \param used_mask Bits representing used (1) and unused (0) locations
1156 * \param needed_count Number of contiguous bits needed.
1159 * Base location of the available bits on success or -1 on failure.
1162 find_available_slots(unsigned used_mask, unsigned needed_count)
1164 unsigned needed_mask = (1 << needed_count) - 1;
1165 const int max_bit_to_test = (8 * sizeof(used_mask)) - needed_count;
1167 /* The comparison to 32 is redundant, but without it GCC emits "warning:
1168 * cannot optimize possibly infinite loops" for the loop below.
1170 if ((needed_count == 0) || (max_bit_to_test < 0) || (max_bit_to_test > 32))
1173 for (int i = 0; i <= max_bit_to_test; i++) {
1174 if ((needed_mask & ~used_mask) == needed_mask)
1185 assign_attribute_locations(gl_shader_program *prog, unsigned max_attribute_index)
1187 /* Mark invalid attribute locations as being used.
1189 unsigned used_locations = (max_attribute_index >= 32)
1190 ? ~0 : ~((1 << max_attribute_index) - 1);
1192 gl_shader *const sh = prog->_LinkedShaders[0];
1193 assert(sh->Type == GL_VERTEX_SHADER);
1195 /* Operate in a total of four passes.
1197 * 1. Invalidate the location assignments for all vertex shader inputs.
1199 * 2. Assign locations for inputs that have user-defined (via
1200 * glBindVertexAttribLocation) locatoins.
1202 * 3. Sort the attributes without assigned locations by number of slots
1203 * required in decreasing order. Fragmentation caused by attribute
1204 * locations assigned by the application may prevent large attributes
1205 * from having enough contiguous space.
1207 * 4. Assign locations to any inputs without assigned locations.
1210 invalidate_variable_locations(sh, ir_var_in, VERT_ATTRIB_GENERIC0);
1212 if (prog->Attributes != NULL) {
1213 for (unsigned i = 0; i < prog->Attributes->NumParameters; i++) {
1214 ir_variable *const var =
1215 sh->symbols->get_variable(prog->Attributes->Parameters[i].Name);
1217 /* Note: attributes that occupy multiple slots, such as arrays or
1218 * matrices, may appear in the attrib array multiple times.
1220 if ((var == NULL) || (var->location != -1))
1223 /* From page 61 of the OpenGL 4.0 spec:
1225 * "LinkProgram will fail if the attribute bindings assigned by
1226 * BindAttribLocation do not leave not enough space to assign a
1227 * location for an active matrix attribute or an active attribute
1228 * array, both of which require multiple contiguous generic
1231 * Previous versions of the spec contain similar language but omit the
1232 * bit about attribute arrays.
1234 * Page 61 of the OpenGL 4.0 spec also says:
1236 * "It is possible for an application to bind more than one
1237 * attribute name to the same location. This is referred to as
1238 * aliasing. This will only work if only one of the aliased
1239 * attributes is active in the executable program, or if no path
1240 * through the shader consumes more than one attribute of a set
1241 * of attributes aliased to the same location. A link error can
1242 * occur if the linker determines that every path through the
1243 * shader consumes multiple aliased attributes, but
1244 * implementations are not required to generate an error in this
1247 * These two paragraphs are either somewhat contradictory, or I don't
1248 * fully understand one or both of them.
1250 /* FINISHME: The code as currently written does not support attribute
1251 * FINISHME: location aliasing (see comment above).
1253 const int attr = prog->Attributes->Parameters[i].StateIndexes[0];
1254 const unsigned slots = count_attribute_slots(var->type);
1256 /* Mask representing the contiguous slots that will be used by this
1259 const unsigned use_mask = (1 << slots) - 1;
1261 /* Generate a link error if the set of bits requested for this
1262 * attribute overlaps any previously allocated bits.
1264 if ((~(use_mask << attr) & used_locations) != used_locations) {
1265 linker_error_printf(prog,
1266 "insufficient contiguous attribute locations "
1267 "available for vertex shader input `%s'",
1272 var->location = VERT_ATTRIB_GENERIC0 + attr;
1273 used_locations |= (use_mask << attr);
1277 /* Temporary storage for the set of attributes that need locations assigned.
1283 /* Used below in the call to qsort. */
1284 static int compare(const void *a, const void *b)
1286 const temp_attr *const l = (const temp_attr *) a;
1287 const temp_attr *const r = (const temp_attr *) b;
1289 /* Reversed because we want a descending order sort below. */
1290 return r->slots - l->slots;
1294 unsigned num_attr = 0;
1296 foreach_list(node, sh->ir) {
1297 ir_variable *const var = ((ir_instruction *) node)->as_variable();
1299 if ((var == NULL) || (var->mode != ir_var_in))
1302 if (var->explicit_location) {
1303 const unsigned slots = count_attribute_slots(var->type);
1304 const unsigned use_mask = (1 << slots) - 1;
1305 const int attr = var->location - VERT_ATTRIB_GENERIC0;
1307 if ((var->location >= (int)(max_attribute_index + VERT_ATTRIB_GENERIC0))
1308 || (var->location < 0)) {
1309 linker_error_printf(prog,
1310 "invalid explicit location %d specified for "
1312 (var->location < 0) ? var->location : attr,
1315 } else if (var->location >= VERT_ATTRIB_GENERIC0) {
1316 used_locations |= (use_mask << attr);
1320 /* The location was explicitly assigned, nothing to do here.
1322 if (var->location != -1)
1325 to_assign[num_attr].slots = count_attribute_slots(var->type);
1326 to_assign[num_attr].var = var;
1330 /* If all of the attributes were assigned locations by the application (or
1331 * are built-in attributes with fixed locations), return early. This should
1332 * be the common case.
1337 qsort(to_assign, num_attr, sizeof(to_assign[0]), temp_attr::compare);
1339 /* VERT_ATTRIB_GENERIC0 is a psdueo-alias for VERT_ATTRIB_POS. It can only
1340 * be explicitly assigned by via glBindAttribLocation. Mark it as reserved
1341 * to prevent it from being automatically allocated below.
1343 find_deref_visitor find("gl_Vertex");
1345 if (find.variable_found())
1346 used_locations |= (1 << 0);
1348 for (unsigned i = 0; i < num_attr; i++) {
1349 /* Mask representing the contiguous slots that will be used by this
1352 const unsigned use_mask = (1 << to_assign[i].slots) - 1;
1354 int location = find_available_slots(used_locations, to_assign[i].slots);
1357 linker_error_printf(prog,
1358 "insufficient contiguous attribute locations "
1359 "available for vertex shader input `%s'",
1360 to_assign[i].var->name);
1364 to_assign[i].var->location = VERT_ATTRIB_GENERIC0 + location;
1365 used_locations |= (use_mask << location);
1373 * Demote shader inputs and outputs that are not used in other stages
1376 demote_shader_inputs_and_outputs(gl_shader *sh, enum ir_variable_mode mode)
1378 foreach_list(node, sh->ir) {
1379 ir_variable *const var = ((ir_instruction *) node)->as_variable();
1381 if ((var == NULL) || (var->mode != int(mode)))
1384 /* A shader 'in' or 'out' variable is only really an input or output if
1385 * its value is used by other shader stages. This will cause the variable
1386 * to have a location assigned.
1388 if (var->location == -1) {
1389 var->mode = ir_var_auto;
1396 assign_varying_locations(struct gl_shader_program *prog,
1397 gl_shader *producer, gl_shader *consumer)
1399 /* FINISHME: Set dynamically when geometry shader support is added. */
1400 unsigned output_index = VERT_RESULT_VAR0;
1401 unsigned input_index = FRAG_ATTRIB_VAR0;
1403 /* Operate in a total of three passes.
1405 * 1. Assign locations for any matching inputs and outputs.
1407 * 2. Mark output variables in the producer that do not have locations as
1408 * not being outputs. This lets the optimizer eliminate them.
1410 * 3. Mark input variables in the consumer that do not have locations as
1411 * not being inputs. This lets the optimizer eliminate them.
1414 invalidate_variable_locations(producer, ir_var_out, VERT_RESULT_VAR0);
1415 invalidate_variable_locations(consumer, ir_var_in, FRAG_ATTRIB_VAR0);
1417 foreach_list(node, producer->ir) {
1418 ir_variable *const output_var = ((ir_instruction *) node)->as_variable();
1420 if ((output_var == NULL) || (output_var->mode != ir_var_out)
1421 || (output_var->location != -1))
1424 ir_variable *const input_var =
1425 consumer->symbols->get_variable(output_var->name);
1427 if ((input_var == NULL) || (input_var->mode != ir_var_in))
1430 assert(input_var->location == -1);
1432 output_var->location = output_index;
1433 input_var->location = input_index;
1435 /* FINISHME: Support for "varying" records in GLSL 1.50. */
1436 assert(!output_var->type->is_record());
1438 if (output_var->type->is_array()) {
1439 const unsigned slots = output_var->type->length
1440 * output_var->type->fields.array->matrix_columns;
1442 output_index += slots;
1443 input_index += slots;
1445 const unsigned slots = output_var->type->matrix_columns;
1447 output_index += slots;
1448 input_index += slots;
1452 foreach_list(node, consumer->ir) {
1453 ir_variable *const var = ((ir_instruction *) node)->as_variable();
1455 if ((var == NULL) || (var->mode != ir_var_in))
1458 if (var->location == -1) {
1459 if (prog->Version <= 120) {
1460 /* On page 25 (page 31 of the PDF) of the GLSL 1.20 spec:
1462 * Only those varying variables used (i.e. read) in
1463 * the fragment shader executable must be written to
1464 * by the vertex shader executable; declaring
1465 * superfluous varying variables in a vertex shader is
1468 * We interpret this text as meaning that the VS must
1469 * write the variable for the FS to read it. See
1470 * "glsl1-varying read but not written" in piglit.
1473 linker_error_printf(prog, "fragment shader varying %s not written "
1474 "by vertex shader\n.", var->name);
1475 prog->LinkStatus = false;
1478 /* An 'in' variable is only really a shader input if its
1479 * value is written by the previous stage.
1481 var->mode = ir_var_auto;
1488 link_shaders(struct gl_context *ctx, struct gl_shader_program *prog)
1490 void *mem_ctx = ralloc_context(NULL); // temporary linker context
1492 prog->LinkStatus = false;
1493 prog->Validated = false;
1494 prog->_Used = false;
1496 if (prog->InfoLog != NULL)
1497 ralloc_free(prog->InfoLog);
1499 prog->InfoLog = ralloc_strdup(NULL, "");
1501 /* Separate the shaders into groups based on their type.
1503 struct gl_shader **vert_shader_list;
1504 unsigned num_vert_shaders = 0;
1505 struct gl_shader **frag_shader_list;
1506 unsigned num_frag_shaders = 0;
1508 vert_shader_list = (struct gl_shader **)
1509 calloc(2 * prog->NumShaders, sizeof(struct gl_shader *));
1510 frag_shader_list = &vert_shader_list[prog->NumShaders];
1512 unsigned min_version = UINT_MAX;
1513 unsigned max_version = 0;
1514 for (unsigned i = 0; i < prog->NumShaders; i++) {
1515 min_version = MIN2(min_version, prog->Shaders[i]->Version);
1516 max_version = MAX2(max_version, prog->Shaders[i]->Version);
1518 switch (prog->Shaders[i]->Type) {
1519 case GL_VERTEX_SHADER:
1520 vert_shader_list[num_vert_shaders] = prog->Shaders[i];
1523 case GL_FRAGMENT_SHADER:
1524 frag_shader_list[num_frag_shaders] = prog->Shaders[i];
1527 case GL_GEOMETRY_SHADER:
1528 /* FINISHME: Support geometry shaders. */
1529 assert(prog->Shaders[i]->Type != GL_GEOMETRY_SHADER);
1534 /* Previous to GLSL version 1.30, different compilation units could mix and
1535 * match shading language versions. With GLSL 1.30 and later, the versions
1536 * of all shaders must match.
1538 assert(min_version >= 100);
1539 assert(max_version <= 130);
1540 if ((max_version >= 130 || min_version == 100)
1541 && min_version != max_version) {
1542 linker_error_printf(prog, "all shaders must use same shading "
1543 "language version\n");
1547 prog->Version = max_version;
1549 for (unsigned int i = 0; i < MESA_SHADER_TYPES; i++) {
1550 if (prog->_LinkedShaders[i] != NULL)
1551 ctx->Driver.DeleteShader(ctx, prog->_LinkedShaders[i]);
1553 prog->_LinkedShaders[i] = NULL;
1556 /* Link all shaders for a particular stage and validate the result.
1558 if (num_vert_shaders > 0) {
1559 gl_shader *const sh =
1560 link_intrastage_shaders(mem_ctx, ctx, prog, vert_shader_list,
1566 if (!validate_vertex_shader_executable(prog, sh))
1569 _mesa_reference_shader(ctx, &prog->_LinkedShaders[MESA_SHADER_VERTEX],
1573 if (num_frag_shaders > 0) {
1574 gl_shader *const sh =
1575 link_intrastage_shaders(mem_ctx, ctx, prog, frag_shader_list,
1581 if (!validate_fragment_shader_executable(prog, sh))
1584 _mesa_reference_shader(ctx, &prog->_LinkedShaders[MESA_SHADER_FRAGMENT],
1588 /* Here begins the inter-stage linking phase. Some initial validation is
1589 * performed, then locations are assigned for uniforms, attributes, and
1592 if (cross_validate_uniforms(prog)) {
1595 for (prev = 0; prev < MESA_SHADER_TYPES; prev++) {
1596 if (prog->_LinkedShaders[prev] != NULL)
1600 /* Validate the inputs of each stage with the output of the preceeding
1603 for (unsigned i = prev + 1; i < MESA_SHADER_TYPES; i++) {
1604 if (prog->_LinkedShaders[i] == NULL)
1607 if (!cross_validate_outputs_to_inputs(prog,
1608 prog->_LinkedShaders[prev],
1609 prog->_LinkedShaders[i]))
1615 prog->LinkStatus = true;
1618 /* Do common optimization before assigning storage for attributes,
1619 * uniforms, and varyings. Later optimization could possibly make
1620 * some of that unused.
1622 for (unsigned i = 0; i < MESA_SHADER_TYPES; i++) {
1623 if (prog->_LinkedShaders[i] == NULL)
1626 while (do_common_optimization(prog->_LinkedShaders[i]->ir, true, 32))
1630 update_array_sizes(prog);
1632 assign_uniform_locations(prog);
1634 if (prog->_LinkedShaders[MESA_SHADER_VERTEX] != NULL) {
1635 /* FINISHME: The value of the max_attribute_index parameter is
1636 * FINISHME: implementation dependent based on the value of
1637 * FINISHME: GL_MAX_VERTEX_ATTRIBS. GL_MAX_VERTEX_ATTRIBS must be
1638 * FINISHME: at least 16, so hardcode 16 for now.
1640 if (!assign_attribute_locations(prog, 16)) {
1641 prog->LinkStatus = false;
1647 for (prev = 0; prev < MESA_SHADER_TYPES; prev++) {
1648 if (prog->_LinkedShaders[prev] != NULL)
1652 for (unsigned i = prev + 1; i < MESA_SHADER_TYPES; i++) {
1653 if (prog->_LinkedShaders[i] == NULL)
1656 assign_varying_locations(prog,
1657 prog->_LinkedShaders[prev],
1658 prog->_LinkedShaders[i]);
1662 if (prog->_LinkedShaders[MESA_SHADER_VERTEX] != NULL) {
1663 demote_shader_inputs_and_outputs(prog->_LinkedShaders[MESA_SHADER_VERTEX],
1667 if (prog->_LinkedShaders[MESA_SHADER_GEOMETRY] != NULL) {
1668 gl_shader *const sh = prog->_LinkedShaders[MESA_SHADER_GEOMETRY];
1670 demote_shader_inputs_and_outputs(sh, ir_var_in);
1671 demote_shader_inputs_and_outputs(sh, ir_var_inout);
1672 demote_shader_inputs_and_outputs(sh, ir_var_out);
1675 if (prog->_LinkedShaders[MESA_SHADER_FRAGMENT] != NULL) {
1676 gl_shader *const sh = prog->_LinkedShaders[MESA_SHADER_FRAGMENT];
1678 demote_shader_inputs_and_outputs(sh, ir_var_in);
1681 /* OpenGL ES requires that a vertex shader and a fragment shader both be
1682 * present in a linked program. By checking for use of shading language
1683 * version 1.00, we also catch the GL_ARB_ES2_compatibility case.
1685 if (ctx->API == API_OPENGLES2 || prog->Version == 100) {
1686 if (prog->_LinkedShaders[MESA_SHADER_VERTEX] == NULL) {
1687 linker_error_printf(prog, "program lacks a vertex shader\n");
1688 prog->LinkStatus = false;
1689 } else if (prog->_LinkedShaders[MESA_SHADER_FRAGMENT] == NULL) {
1690 linker_error_printf(prog, "program lacks a fragment shader\n");
1691 prog->LinkStatus = false;
1695 /* FINISHME: Assign fragment shader output locations. */
1698 free(vert_shader_list);
1700 for (unsigned i = 0; i < MESA_SHADER_TYPES; i++) {
1701 if (prog->_LinkedShaders[i] == NULL)
1704 /* Retain any live IR, but trash the rest. */
1705 reparent_ir(prog->_LinkedShaders[i]->ir, prog->_LinkedShaders[i]->ir);
1708 ralloc_free(mem_ctx);