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->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 if (ir->return_deref != NULL) {
121 ir_variable *const var = ir->return_deref->variable_referenced();
123 if (strcmp(name, var->name) == 0) {
129 return visit_continue_with_parent;
132 bool variable_found()
138 const char *name; /**< Find writes to a variable with this name. */
139 bool found; /**< Was a write to the variable found? */
144 * Visitor that determines whether or not a variable is ever read.
146 class find_deref_visitor : public ir_hierarchical_visitor {
148 find_deref_visitor(const char *name)
149 : name(name), found(false)
154 virtual ir_visitor_status visit(ir_dereference_variable *ir)
156 if (strcmp(this->name, ir->var->name) == 0) {
161 return visit_continue;
164 bool variable_found() const
170 const char *name; /**< Find writes to a variable with this name. */
171 bool found; /**< Was a write to the variable found? */
176 linker_error(gl_shader_program *prog, const char *fmt, ...)
180 ralloc_strcat(&prog->InfoLog, "error: ");
182 ralloc_vasprintf_append(&prog->InfoLog, fmt, ap);
185 prog->LinkStatus = false;
190 linker_warning(gl_shader_program *prog, const char *fmt, ...)
194 ralloc_strcat(&prog->InfoLog, "error: ");
196 ralloc_vasprintf_append(&prog->InfoLog, fmt, ap);
203 link_invalidate_variable_locations(gl_shader *sh, enum ir_variable_mode mode,
206 foreach_list(node, sh->ir) {
207 ir_variable *const var = ((ir_instruction *) node)->as_variable();
209 if ((var == NULL) || (var->mode != (unsigned) mode))
212 /* Only assign locations for generic attributes / varyings / etc.
214 if ((var->location >= generic_base) && !var->explicit_location)
221 * Determine the number of attribute slots required for a particular type
223 * This code is here because it implements the language rules of a specific
224 * GLSL version. Since it's a property of the language and not a property of
225 * types in general, it doesn't really belong in glsl_type.
228 count_attribute_slots(const glsl_type *t)
230 /* From page 31 (page 37 of the PDF) of the GLSL 1.50 spec:
232 * "A scalar input counts the same amount against this limit as a vec4,
233 * so applications may want to consider packing groups of four
234 * unrelated float inputs together into a vector to better utilize the
235 * capabilities of the underlying hardware. A matrix input will use up
236 * multiple locations. The number of locations used will equal the
237 * number of columns in the matrix."
239 * The spec does not explicitly say how arrays are counted. However, it
240 * should be safe to assume the total number of slots consumed by an array
241 * is the number of entries in the array multiplied by the number of slots
242 * consumed by a single element of the array.
246 return t->array_size() * count_attribute_slots(t->element_type());
249 return t->matrix_columns;
256 * Verify that a vertex shader executable meets all semantic requirements.
258 * Also sets prog->Vert.UsesClipDistance and prog->Vert.ClipDistanceArraySize
261 * \param shader Vertex shader executable to be verified
264 validate_vertex_shader_executable(struct gl_shader_program *prog,
265 struct gl_shader *shader)
270 /* From the GLSL 1.10 spec, page 48:
272 * "The variable gl_Position is available only in the vertex
273 * language and is intended for writing the homogeneous vertex
274 * position. All executions of a well-formed vertex shader
275 * executable must write a value into this variable. [...] The
276 * variable gl_Position is available only in the vertex
277 * language and is intended for writing the homogeneous vertex
278 * position. All executions of a well-formed vertex shader
279 * executable must write a value into this variable."
281 * while in GLSL 1.40 this text is changed to:
283 * "The variable gl_Position is available only in the vertex
284 * language and is intended for writing the homogeneous vertex
285 * position. It can be written at any time during shader
286 * execution. It may also be read back by a vertex shader
287 * after being written. This value will be used by primitive
288 * assembly, clipping, culling, and other fixed functionality
289 * operations, if present, that operate on primitives after
290 * vertex processing has occurred. Its value is undefined if
291 * the vertex shader executable does not write gl_Position."
293 if (prog->Version < 140) {
294 find_assignment_visitor find("gl_Position");
295 find.run(shader->ir);
296 if (!find.variable_found()) {
297 linker_error(prog, "vertex shader does not write to `gl_Position'\n");
302 prog->Vert.ClipDistanceArraySize = 0;
304 if (prog->Version >= 130) {
305 /* From section 7.1 (Vertex Shader Special Variables) of the
308 * "It is an error for a shader to statically write both
309 * gl_ClipVertex and gl_ClipDistance."
311 find_assignment_visitor clip_vertex("gl_ClipVertex");
312 find_assignment_visitor clip_distance("gl_ClipDistance");
314 clip_vertex.run(shader->ir);
315 clip_distance.run(shader->ir);
316 if (clip_vertex.variable_found() && clip_distance.variable_found()) {
317 linker_error(prog, "vertex shader writes to both `gl_ClipVertex' "
318 "and `gl_ClipDistance'\n");
321 prog->Vert.UsesClipDistance = clip_distance.variable_found();
322 ir_variable *clip_distance_var =
323 shader->symbols->get_variable("gl_ClipDistance");
324 if (clip_distance_var)
325 prog->Vert.ClipDistanceArraySize = clip_distance_var->type->length;
333 * Verify that a fragment shader executable meets all semantic requirements
335 * \param shader Fragment shader executable to be verified
338 validate_fragment_shader_executable(struct gl_shader_program *prog,
339 struct gl_shader *shader)
344 find_assignment_visitor frag_color("gl_FragColor");
345 find_assignment_visitor frag_data("gl_FragData");
347 frag_color.run(shader->ir);
348 frag_data.run(shader->ir);
350 if (frag_color.variable_found() && frag_data.variable_found()) {
351 linker_error(prog, "fragment shader writes to both "
352 "`gl_FragColor' and `gl_FragData'\n");
361 * Generate a string describing the mode of a variable
364 mode_string(const ir_variable *var)
368 return (var->read_only) ? "global constant" : "global variable";
370 case ir_var_uniform: return "uniform";
371 case ir_var_in: return "shader input";
372 case ir_var_out: return "shader output";
373 case ir_var_inout: return "shader inout";
375 case ir_var_const_in:
376 case ir_var_temporary:
378 assert(!"Should not get here.");
379 return "invalid variable";
385 * Perform validation of global variables used across multiple shaders
388 cross_validate_globals(struct gl_shader_program *prog,
389 struct gl_shader **shader_list,
390 unsigned num_shaders,
393 /* Examine all of the uniforms in all of the shaders and cross validate
396 glsl_symbol_table variables;
397 for (unsigned i = 0; i < num_shaders; i++) {
398 if (shader_list[i] == NULL)
401 foreach_list(node, shader_list[i]->ir) {
402 ir_variable *const var = ((ir_instruction *) node)->as_variable();
407 if (uniforms_only && (var->mode != ir_var_uniform))
410 /* Don't cross validate temporaries that are at global scope. These
411 * will eventually get pulled into the shaders 'main'.
413 if (var->mode == ir_var_temporary)
416 /* If a global with this name has already been seen, verify that the
417 * new instance has the same type. In addition, if the globals have
418 * initializers, the values of the initializers must be the same.
420 ir_variable *const existing = variables.get_variable(var->name);
421 if (existing != NULL) {
422 if (var->type != existing->type) {
423 /* Consider the types to be "the same" if both types are arrays
424 * of the same type and one of the arrays is implicitly sized.
425 * In addition, set the type of the linked variable to the
426 * explicitly sized array.
428 if (var->type->is_array()
429 && existing->type->is_array()
430 && (var->type->fields.array == existing->type->fields.array)
431 && ((var->type->length == 0)
432 || (existing->type->length == 0))) {
433 if (var->type->length != 0) {
434 existing->type = var->type;
437 linker_error(prog, "%s `%s' declared as type "
438 "`%s' and type `%s'\n",
440 var->name, var->type->name,
441 existing->type->name);
446 if (var->explicit_location) {
447 if (existing->explicit_location
448 && (var->location != existing->location)) {
449 linker_error(prog, "explicit locations for %s "
450 "`%s' have differing values\n",
451 mode_string(var), var->name);
455 existing->location = var->location;
456 existing->explicit_location = true;
459 /* Validate layout qualifiers for gl_FragDepth.
461 * From the AMD/ARB_conservative_depth specs:
463 * "If gl_FragDepth is redeclared in any fragment shader in a
464 * program, it must be redeclared in all fragment shaders in
465 * that program that have static assignments to
466 * gl_FragDepth. All redeclarations of gl_FragDepth in all
467 * fragment shaders in a single program must have the same set
470 if (strcmp(var->name, "gl_FragDepth") == 0) {
471 bool layout_declared = var->depth_layout != ir_depth_layout_none;
472 bool layout_differs =
473 var->depth_layout != existing->depth_layout;
475 if (layout_declared && layout_differs) {
477 "All redeclarations of gl_FragDepth in all "
478 "fragment shaders in a single program must have "
479 "the same set of qualifiers.");
482 if (var->used && layout_differs) {
484 "If gl_FragDepth is redeclared with a layout "
485 "qualifier in any fragment shader, it must be "
486 "redeclared with the same layout qualifier in "
487 "all fragment shaders that have assignments to "
492 /* Page 35 (page 41 of the PDF) of the GLSL 4.20 spec says:
494 * "If a shared global has multiple initializers, the
495 * initializers must all be constant expressions, and they
496 * must all have the same value. Otherwise, a link error will
497 * result. (A shared global having only one initializer does
498 * not require that initializer to be a constant expression.)"
500 * Previous to 4.20 the GLSL spec simply said that initializers
501 * must have the same value. In this case of non-constant
502 * initializers, this was impossible to determine. As a result,
503 * no vendor actually implemented that behavior. The 4.20
504 * behavior matches the implemented behavior of at least one other
505 * vendor, so we'll implement that for all GLSL versions.
507 if (var->constant_initializer != NULL) {
508 if (existing->constant_initializer != NULL) {
509 if (!var->constant_initializer->has_value(existing->constant_initializer)) {
510 linker_error(prog, "initializers for %s "
511 "`%s' have differing values\n",
512 mode_string(var), var->name);
516 /* If the first-seen instance of a particular uniform did not
517 * have an initializer but a later instance does, copy the
518 * initializer to the version stored in the symbol table.
520 /* FINISHME: This is wrong. The constant_value field should
521 * FINISHME: not be modified! Imagine a case where a shader
522 * FINISHME: without an initializer is linked in two different
523 * FINISHME: programs with shaders that have differing
524 * FINISHME: initializers. Linking with the first will
525 * FINISHME: modify the shader, and linking with the second
526 * FINISHME: will fail.
528 existing->constant_initializer =
529 var->constant_initializer->clone(ralloc_parent(existing),
534 if (var->has_initializer) {
535 if (existing->has_initializer
536 && (var->constant_initializer == NULL
537 || existing->constant_initializer == NULL)) {
539 "shared global variable `%s' has multiple "
540 "non-constant initializers.\n",
545 /* Some instance had an initializer, so keep track of that. In
546 * this location, all sorts of initializers (constant or
547 * otherwise) will propagate the existence to the variable
548 * stored in the symbol table.
550 existing->has_initializer = true;
553 if (existing->invariant != var->invariant) {
554 linker_error(prog, "declarations for %s `%s' have "
555 "mismatching invariant qualifiers\n",
556 mode_string(var), var->name);
559 if (existing->centroid != var->centroid) {
560 linker_error(prog, "declarations for %s `%s' have "
561 "mismatching centroid qualifiers\n",
562 mode_string(var), var->name);
566 variables.add_variable(var);
575 * Perform validation of uniforms used across multiple shader stages
578 cross_validate_uniforms(struct gl_shader_program *prog)
580 return cross_validate_globals(prog, prog->_LinkedShaders,
581 MESA_SHADER_TYPES, true);
586 * Validate that outputs from one stage match inputs of another
589 cross_validate_outputs_to_inputs(struct gl_shader_program *prog,
590 gl_shader *producer, gl_shader *consumer)
592 glsl_symbol_table parameters;
593 /* FINISHME: Figure these out dynamically. */
594 const char *const producer_stage = "vertex";
595 const char *const consumer_stage = "fragment";
597 /* Find all shader outputs in the "producer" stage.
599 foreach_list(node, producer->ir) {
600 ir_variable *const var = ((ir_instruction *) node)->as_variable();
602 /* FINISHME: For geometry shaders, this should also look for inout
603 * FINISHME: variables.
605 if ((var == NULL) || (var->mode != ir_var_out))
608 parameters.add_variable(var);
612 /* Find all shader inputs in the "consumer" stage. Any variables that have
613 * matching outputs already in the symbol table must have the same type and
616 foreach_list(node, consumer->ir) {
617 ir_variable *const input = ((ir_instruction *) node)->as_variable();
619 /* FINISHME: For geometry shaders, this should also look for inout
620 * FINISHME: variables.
622 if ((input == NULL) || (input->mode != ir_var_in))
625 ir_variable *const output = parameters.get_variable(input->name);
626 if (output != NULL) {
627 /* Check that the types match between stages.
629 if (input->type != output->type) {
630 /* There is a bit of a special case for gl_TexCoord. This
631 * built-in is unsized by default. Applications that variable
632 * access it must redeclare it with a size. There is some
633 * language in the GLSL spec that implies the fragment shader
634 * and vertex shader do not have to agree on this size. Other
635 * driver behave this way, and one or two applications seem to
638 * Neither declaration needs to be modified here because the array
639 * sizes are fixed later when update_array_sizes is called.
641 * From page 48 (page 54 of the PDF) of the GLSL 1.10 spec:
643 * "Unlike user-defined varying variables, the built-in
644 * varying variables don't have a strict one-to-one
645 * correspondence between the vertex language and the
646 * fragment language."
648 if (!output->type->is_array()
649 || (strncmp("gl_", output->name, 3) != 0)) {
651 "%s shader output `%s' declared as type `%s', "
652 "but %s shader input declared as type `%s'\n",
653 producer_stage, output->name,
655 consumer_stage, input->type->name);
660 /* Check that all of the qualifiers match between stages.
662 if (input->centroid != output->centroid) {
664 "%s shader output `%s' %s centroid qualifier, "
665 "but %s shader input %s centroid qualifier\n",
668 (output->centroid) ? "has" : "lacks",
670 (input->centroid) ? "has" : "lacks");
674 if (input->invariant != output->invariant) {
676 "%s shader output `%s' %s invariant qualifier, "
677 "but %s shader input %s invariant qualifier\n",
680 (output->invariant) ? "has" : "lacks",
682 (input->invariant) ? "has" : "lacks");
686 if (input->interpolation != output->interpolation) {
688 "%s shader output `%s' specifies %s "
689 "interpolation qualifier, "
690 "but %s shader input specifies %s "
691 "interpolation qualifier\n",
694 output->interpolation_string(),
696 input->interpolation_string());
707 * Populates a shaders symbol table with all global declarations
710 populate_symbol_table(gl_shader *sh)
712 sh->symbols = new(sh) glsl_symbol_table;
714 foreach_list(node, sh->ir) {
715 ir_instruction *const inst = (ir_instruction *) node;
719 if ((func = inst->as_function()) != NULL) {
720 sh->symbols->add_function(func);
721 } else if ((var = inst->as_variable()) != NULL) {
722 sh->symbols->add_variable(var);
729 * Remap variables referenced in an instruction tree
731 * This is used when instruction trees are cloned from one shader and placed in
732 * another. These trees will contain references to \c ir_variable nodes that
733 * do not exist in the target shader. This function finds these \c ir_variable
734 * references and replaces the references with matching variables in the target
737 * If there is no matching variable in the target shader, a clone of the
738 * \c ir_variable is made and added to the target shader. The new variable is
739 * added to \b both the instruction stream and the symbol table.
741 * \param inst IR tree that is to be processed.
742 * \param symbols Symbol table containing global scope symbols in the
744 * \param instructions Instruction stream where new variable declarations
748 remap_variables(ir_instruction *inst, struct gl_shader *target,
751 class remap_visitor : public ir_hierarchical_visitor {
753 remap_visitor(struct gl_shader *target,
756 this->target = target;
757 this->symbols = target->symbols;
758 this->instructions = target->ir;
762 virtual ir_visitor_status visit(ir_dereference_variable *ir)
764 if (ir->var->mode == ir_var_temporary) {
765 ir_variable *var = (ir_variable *) hash_table_find(temps, ir->var);
769 return visit_continue;
772 ir_variable *const existing =
773 this->symbols->get_variable(ir->var->name);
774 if (existing != NULL)
777 ir_variable *copy = ir->var->clone(this->target, NULL);
779 this->symbols->add_variable(copy);
780 this->instructions->push_head(copy);
784 return visit_continue;
788 struct gl_shader *target;
789 glsl_symbol_table *symbols;
790 exec_list *instructions;
794 remap_visitor v(target, temps);
801 * Move non-declarations from one instruction stream to another
803 * The intended usage pattern of this function is to pass the pointer to the
804 * head sentinel of a list (i.e., a pointer to the list cast to an \c exec_node
805 * pointer) for \c last and \c false for \c make_copies on the first
806 * call. Successive calls pass the return value of the previous call for
807 * \c last and \c true for \c make_copies.
809 * \param instructions Source instruction stream
810 * \param last Instruction after which new instructions should be
811 * inserted in the target instruction stream
812 * \param make_copies Flag selecting whether instructions in \c instructions
813 * should be copied (via \c ir_instruction::clone) into the
814 * target list or moved.
817 * The new "last" instruction in the target instruction stream. This pointer
818 * is suitable for use as the \c last parameter of a later call to this
822 move_non_declarations(exec_list *instructions, exec_node *last,
823 bool make_copies, gl_shader *target)
825 hash_table *temps = NULL;
828 temps = hash_table_ctor(0, hash_table_pointer_hash,
829 hash_table_pointer_compare);
831 foreach_list_safe(node, instructions) {
832 ir_instruction *inst = (ir_instruction *) node;
834 if (inst->as_function())
837 ir_variable *var = inst->as_variable();
838 if ((var != NULL) && (var->mode != ir_var_temporary))
841 assert(inst->as_assignment()
843 || ((var != NULL) && (var->mode == ir_var_temporary)));
846 inst = inst->clone(target, NULL);
849 hash_table_insert(temps, inst, var);
851 remap_variables(inst, target, temps);
856 last->insert_after(inst);
861 hash_table_dtor(temps);
867 * Get the function signature for main from a shader
869 static ir_function_signature *
870 get_main_function_signature(gl_shader *sh)
872 ir_function *const f = sh->symbols->get_function("main");
874 exec_list void_parameters;
876 /* Look for the 'void main()' signature and ensure that it's defined.
877 * This keeps the linker from accidentally pick a shader that just
878 * contains a prototype for main.
880 * We don't have to check for multiple definitions of main (in multiple
881 * shaders) because that would have already been caught above.
883 ir_function_signature *sig = f->matching_signature(&void_parameters);
884 if ((sig != NULL) && sig->is_defined) {
894 * This class is only used in link_intrastage_shaders() below but declaring
895 * it inside that function leads to compiler warnings with some versions of
898 class array_sizing_visitor : public ir_hierarchical_visitor {
900 virtual ir_visitor_status visit(ir_variable *var)
902 if (var->type->is_array() && (var->type->length == 0)) {
903 const glsl_type *type =
904 glsl_type::get_array_instance(var->type->fields.array,
905 var->max_array_access + 1);
906 assert(type != NULL);
909 return visit_continue;
915 * Combine a group of shaders for a single stage to generate a linked shader
918 * If this function is supplied a single shader, it is cloned, and the new
919 * shader is returned.
921 static struct gl_shader *
922 link_intrastage_shaders(void *mem_ctx,
923 struct gl_context *ctx,
924 struct gl_shader_program *prog,
925 struct gl_shader **shader_list,
926 unsigned num_shaders)
928 /* Check that global variables defined in multiple shaders are consistent.
930 if (!cross_validate_globals(prog, shader_list, num_shaders, false))
933 /* Check that there is only a single definition of each function signature
934 * across all shaders.
936 for (unsigned i = 0; i < (num_shaders - 1); i++) {
937 foreach_list(node, shader_list[i]->ir) {
938 ir_function *const f = ((ir_instruction *) node)->as_function();
943 for (unsigned j = i + 1; j < num_shaders; j++) {
944 ir_function *const other =
945 shader_list[j]->symbols->get_function(f->name);
947 /* If the other shader has no function (and therefore no function
948 * signatures) with the same name, skip to the next shader.
953 foreach_iter (exec_list_iterator, iter, *f) {
954 ir_function_signature *sig =
955 (ir_function_signature *) iter.get();
957 if (!sig->is_defined || sig->is_builtin)
960 ir_function_signature *other_sig =
961 other->exact_matching_signature(& sig->parameters);
963 if ((other_sig != NULL) && other_sig->is_defined
964 && !other_sig->is_builtin) {
965 linker_error(prog, "function `%s' is multiply defined",
974 /* Find the shader that defines main, and make a clone of it.
976 * Starting with the clone, search for undefined references. If one is
977 * found, find the shader that defines it. Clone the reference and add
978 * it to the shader. Repeat until there are no undefined references or
979 * until a reference cannot be resolved.
981 gl_shader *main = NULL;
982 for (unsigned i = 0; i < num_shaders; i++) {
983 if (get_main_function_signature(shader_list[i]) != NULL) {
984 main = shader_list[i];
990 linker_error(prog, "%s shader lacks `main'\n",
991 (shader_list[0]->Type == GL_VERTEX_SHADER)
992 ? "vertex" : "fragment");
996 gl_shader *linked = ctx->Driver.NewShader(NULL, 0, main->Type);
997 linked->ir = new(linked) exec_list;
998 clone_ir_list(mem_ctx, linked->ir, main->ir);
1000 populate_symbol_table(linked);
1002 /* The a pointer to the main function in the final linked shader (i.e., the
1003 * copy of the original shader that contained the main function).
1005 ir_function_signature *const main_sig = get_main_function_signature(linked);
1007 /* Move any instructions other than variable declarations or function
1008 * declarations into main.
1010 exec_node *insertion_point =
1011 move_non_declarations(linked->ir, (exec_node *) &main_sig->body, false,
1014 for (unsigned i = 0; i < num_shaders; i++) {
1015 if (shader_list[i] == main)
1018 insertion_point = move_non_declarations(shader_list[i]->ir,
1019 insertion_point, true, linked);
1022 /* Resolve initializers for global variables in the linked shader.
1024 unsigned num_linking_shaders = num_shaders;
1025 for (unsigned i = 0; i < num_shaders; i++)
1026 num_linking_shaders += shader_list[i]->num_builtins_to_link;
1028 gl_shader **linking_shaders =
1029 (gl_shader **) calloc(num_linking_shaders, sizeof(gl_shader *));
1031 memcpy(linking_shaders, shader_list,
1032 sizeof(linking_shaders[0]) * num_shaders);
1034 unsigned idx = num_shaders;
1035 for (unsigned i = 0; i < num_shaders; i++) {
1036 memcpy(&linking_shaders[idx], shader_list[i]->builtins_to_link,
1037 sizeof(linking_shaders[0]) * shader_list[i]->num_builtins_to_link);
1038 idx += shader_list[i]->num_builtins_to_link;
1041 assert(idx == num_linking_shaders);
1043 if (!link_function_calls(prog, linked, linking_shaders,
1044 num_linking_shaders)) {
1045 ctx->Driver.DeleteShader(ctx, linked);
1049 free(linking_shaders);
1052 /* At this point linked should contain all of the linked IR, so
1053 * validate it to make sure nothing went wrong.
1056 validate_ir_tree(linked->ir);
1059 /* Make a pass over all variable declarations to ensure that arrays with
1060 * unspecified sizes have a size specified. The size is inferred from the
1061 * max_array_access field.
1063 if (linked != NULL) {
1064 array_sizing_visitor v;
1073 * Update the sizes of linked shader uniform arrays to the maximum
1076 * From page 81 (page 95 of the PDF) of the OpenGL 2.1 spec:
1078 * If one or more elements of an array are active,
1079 * GetActiveUniform will return the name of the array in name,
1080 * subject to the restrictions listed above. The type of the array
1081 * is returned in type. The size parameter contains the highest
1082 * array element index used, plus one. The compiler or linker
1083 * determines the highest index used. There will be only one
1084 * active uniform reported by the GL per uniform array.
1088 update_array_sizes(struct gl_shader_program *prog)
1090 for (unsigned i = 0; i < MESA_SHADER_TYPES; i++) {
1091 if (prog->_LinkedShaders[i] == NULL)
1094 foreach_list(node, prog->_LinkedShaders[i]->ir) {
1095 ir_variable *const var = ((ir_instruction *) node)->as_variable();
1097 if ((var == NULL) || (var->mode != ir_var_uniform &&
1098 var->mode != ir_var_in &&
1099 var->mode != ir_var_out) ||
1100 !var->type->is_array())
1103 unsigned int size = var->max_array_access;
1104 for (unsigned j = 0; j < MESA_SHADER_TYPES; j++) {
1105 if (prog->_LinkedShaders[j] == NULL)
1108 foreach_list(node2, prog->_LinkedShaders[j]->ir) {
1109 ir_variable *other_var = ((ir_instruction *) node2)->as_variable();
1113 if (strcmp(var->name, other_var->name) == 0 &&
1114 other_var->max_array_access > size) {
1115 size = other_var->max_array_access;
1120 if (size + 1 != var->type->fields.array->length) {
1121 /* If this is a built-in uniform (i.e., it's backed by some
1122 * fixed-function state), adjust the number of state slots to
1123 * match the new array size. The number of slots per array entry
1124 * is not known. It seems safe to assume that the total number of
1125 * slots is an integer multiple of the number of array elements.
1126 * Determine the number of slots per array element by dividing by
1127 * the old (total) size.
1129 if (var->num_state_slots > 0) {
1130 var->num_state_slots = (size + 1)
1131 * (var->num_state_slots / var->type->length);
1134 var->type = glsl_type::get_array_instance(var->type->fields.array,
1136 /* FINISHME: We should update the types of array
1137 * dereferences of this variable now.
1145 * Find a contiguous set of available bits in a bitmask.
1147 * \param used_mask Bits representing used (1) and unused (0) locations
1148 * \param needed_count Number of contiguous bits needed.
1151 * Base location of the available bits on success or -1 on failure.
1154 find_available_slots(unsigned used_mask, unsigned needed_count)
1156 unsigned needed_mask = (1 << needed_count) - 1;
1157 const int max_bit_to_test = (8 * sizeof(used_mask)) - needed_count;
1159 /* The comparison to 32 is redundant, but without it GCC emits "warning:
1160 * cannot optimize possibly infinite loops" for the loop below.
1162 if ((needed_count == 0) || (max_bit_to_test < 0) || (max_bit_to_test > 32))
1165 for (int i = 0; i <= max_bit_to_test; i++) {
1166 if ((needed_mask & ~used_mask) == needed_mask)
1177 * Assign locations for either VS inputs for FS outputs
1179 * \param prog Shader program whose variables need locations assigned
1180 * \param target_index Selector for the program target to receive location
1181 * assignmnets. Must be either \c MESA_SHADER_VERTEX or
1182 * \c MESA_SHADER_FRAGMENT.
1183 * \param max_index Maximum number of generic locations. This corresponds
1184 * to either the maximum number of draw buffers or the
1185 * maximum number of generic attributes.
1188 * If locations are successfully assigned, true is returned. Otherwise an
1189 * error is emitted to the shader link log and false is returned.
1192 assign_attribute_or_color_locations(gl_shader_program *prog,
1193 unsigned target_index,
1196 /* Mark invalid locations as being used.
1198 unsigned used_locations = (max_index >= 32)
1199 ? ~0 : ~((1 << max_index) - 1);
1201 assert((target_index == MESA_SHADER_VERTEX)
1202 || (target_index == MESA_SHADER_FRAGMENT));
1204 gl_shader *const sh = prog->_LinkedShaders[target_index];
1208 /* Operate in a total of four passes.
1210 * 1. Invalidate the location assignments for all vertex shader inputs.
1212 * 2. Assign locations for inputs that have user-defined (via
1213 * glBindVertexAttribLocation) locations and outputs that have
1214 * user-defined locations (via glBindFragDataLocation).
1216 * 3. Sort the attributes without assigned locations by number of slots
1217 * required in decreasing order. Fragmentation caused by attribute
1218 * locations assigned by the application may prevent large attributes
1219 * from having enough contiguous space.
1221 * 4. Assign locations to any inputs without assigned locations.
1224 const int generic_base = (target_index == MESA_SHADER_VERTEX)
1225 ? (int) VERT_ATTRIB_GENERIC0 : (int) FRAG_RESULT_DATA0;
1227 const enum ir_variable_mode direction =
1228 (target_index == MESA_SHADER_VERTEX) ? ir_var_in : ir_var_out;
1231 link_invalidate_variable_locations(sh, direction, generic_base);
1233 /* Temporary storage for the set of attributes that need locations assigned.
1239 /* Used below in the call to qsort. */
1240 static int compare(const void *a, const void *b)
1242 const temp_attr *const l = (const temp_attr *) a;
1243 const temp_attr *const r = (const temp_attr *) b;
1245 /* Reversed because we want a descending order sort below. */
1246 return r->slots - l->slots;
1250 unsigned num_attr = 0;
1252 foreach_list(node, sh->ir) {
1253 ir_variable *const var = ((ir_instruction *) node)->as_variable();
1255 if ((var == NULL) || (var->mode != (unsigned) direction))
1258 if (var->explicit_location) {
1259 if ((var->location >= (int)(max_index + generic_base))
1260 || (var->location < 0)) {
1262 "invalid explicit location %d specified for `%s'\n",
1264 ? var->location : var->location - generic_base,
1268 } else if (target_index == MESA_SHADER_VERTEX) {
1271 if (prog->AttributeBindings->get(binding, var->name)) {
1272 assert(binding >= VERT_ATTRIB_GENERIC0);
1273 var->location = binding;
1275 } else if (target_index == MESA_SHADER_FRAGMENT) {
1278 if (prog->FragDataBindings->get(binding, var->name)) {
1279 assert(binding >= FRAG_RESULT_DATA0);
1280 var->location = binding;
1284 /* If the variable is not a built-in and has a location statically
1285 * assigned in the shader (presumably via a layout qualifier), make sure
1286 * that it doesn't collide with other assigned locations. Otherwise,
1287 * add it to the list of variables that need linker-assigned locations.
1289 const unsigned slots = count_attribute_slots(var->type);
1290 if (var->location != -1) {
1291 if (var->location >= generic_base) {
1292 /* From page 61 of the OpenGL 4.0 spec:
1294 * "LinkProgram will fail if the attribute bindings assigned
1295 * by BindAttribLocation do not leave not enough space to
1296 * assign a location for an active matrix attribute or an
1297 * active attribute array, both of which require multiple
1298 * contiguous generic attributes."
1300 * Previous versions of the spec contain similar language but omit
1301 * the bit about attribute arrays.
1303 * Page 61 of the OpenGL 4.0 spec also says:
1305 * "It is possible for an application to bind more than one
1306 * attribute name to the same location. This is referred to as
1307 * aliasing. This will only work if only one of the aliased
1308 * attributes is active in the executable program, or if no
1309 * path through the shader consumes more than one attribute of
1310 * a set of attributes aliased to the same location. A link
1311 * error can occur if the linker determines that every path
1312 * through the shader consumes multiple aliased attributes,
1313 * but implementations are not required to generate an error
1316 * These two paragraphs are either somewhat contradictory, or I
1317 * don't fully understand one or both of them.
1319 /* FINISHME: The code as currently written does not support
1320 * FINISHME: attribute location aliasing (see comment above).
1322 /* Mask representing the contiguous slots that will be used by
1325 const unsigned attr = var->location - generic_base;
1326 const unsigned use_mask = (1 << slots) - 1;
1328 /* Generate a link error if the set of bits requested for this
1329 * attribute overlaps any previously allocated bits.
1331 if ((~(use_mask << attr) & used_locations) != used_locations) {
1332 const char *const string = (target_index == MESA_SHADER_VERTEX)
1333 ? "vertex shader input" : "fragment shader output";
1335 "insufficient contiguous locations "
1336 "available for %s `%s'", string,
1341 used_locations |= (use_mask << attr);
1347 to_assign[num_attr].slots = slots;
1348 to_assign[num_attr].var = var;
1352 /* If all of the attributes were assigned locations by the application (or
1353 * are built-in attributes with fixed locations), return early. This should
1354 * be the common case.
1359 qsort(to_assign, num_attr, sizeof(to_assign[0]), temp_attr::compare);
1361 if (target_index == MESA_SHADER_VERTEX) {
1362 /* VERT_ATTRIB_GENERIC0 is a pseudo-alias for VERT_ATTRIB_POS. It can
1363 * only be explicitly assigned by via glBindAttribLocation. Mark it as
1364 * reserved to prevent it from being automatically allocated below.
1366 find_deref_visitor find("gl_Vertex");
1368 if (find.variable_found())
1369 used_locations |= (1 << 0);
1372 for (unsigned i = 0; i < num_attr; i++) {
1373 /* Mask representing the contiguous slots that will be used by this
1376 const unsigned use_mask = (1 << to_assign[i].slots) - 1;
1378 int location = find_available_slots(used_locations, to_assign[i].slots);
1381 const char *const string = (target_index == MESA_SHADER_VERTEX)
1382 ? "vertex shader input" : "fragment shader output";
1385 "insufficient contiguous locations "
1386 "available for %s `%s'",
1387 string, to_assign[i].var->name);
1391 to_assign[i].var->location = generic_base + location;
1392 used_locations |= (use_mask << location);
1400 * Demote shader inputs and outputs that are not used in other stages
1403 demote_shader_inputs_and_outputs(gl_shader *sh, enum ir_variable_mode mode)
1405 foreach_list(node, sh->ir) {
1406 ir_variable *const var = ((ir_instruction *) node)->as_variable();
1408 if ((var == NULL) || (var->mode != int(mode)))
1411 /* A shader 'in' or 'out' variable is only really an input or output if
1412 * its value is used by other shader stages. This will cause the variable
1413 * to have a location assigned.
1415 if (var->location == -1) {
1416 var->mode = ir_var_auto;
1423 * Data structure tracking information about a transform feedback declaration
1426 class tfeedback_decl
1429 bool init(struct gl_context *ctx, struct gl_shader_program *prog,
1430 const void *mem_ctx, const char *input);
1431 static bool is_same(const tfeedback_decl &x, const tfeedback_decl &y);
1432 bool assign_location(struct gl_context *ctx, struct gl_shader_program *prog,
1433 ir_variable *output_var);
1434 bool accumulate_num_outputs(struct gl_shader_program *prog, unsigned *count);
1435 bool store(struct gl_context *ctx, struct gl_shader_program *prog,
1436 struct gl_transform_feedback_info *info, unsigned buffer,
1437 unsigned varying, const unsigned max_outputs) const;
1441 * True if assign_location() has been called for this object.
1443 bool is_assigned() const
1445 return this->location != -1;
1449 * Determine whether this object refers to the variable var.
1451 bool matches_var(ir_variable *var) const
1453 if (this->is_clip_distance_mesa)
1454 return strcmp(var->name, "gl_ClipDistanceMESA") == 0;
1456 return strcmp(var->name, this->var_name) == 0;
1460 * The total number of varying components taken up by this variable. Only
1461 * valid if is_assigned() is true.
1463 unsigned num_components() const
1465 if (this->is_clip_distance_mesa)
1468 return this->vector_elements * this->matrix_columns * this->size;
1473 * The name that was supplied to glTransformFeedbackVaryings. Used for
1474 * error reporting and glGetTransformFeedbackVarying().
1476 const char *orig_name;
1479 * The name of the variable, parsed from orig_name.
1481 const char *var_name;
1484 * True if the declaration in orig_name represents an array.
1486 bool is_subscripted;
1489 * If is_subscripted is true, the subscript that was specified in orig_name.
1491 unsigned array_subscript;
1494 * True if the variable is gl_ClipDistance and the driver lowers
1495 * gl_ClipDistance to gl_ClipDistanceMESA.
1497 bool is_clip_distance_mesa;
1500 * The vertex shader output location that the linker assigned for this
1501 * variable. -1 if a location hasn't been assigned yet.
1506 * If location != -1, the number of vector elements in this variable, or 1
1507 * if this variable is a scalar.
1509 unsigned vector_elements;
1512 * If location != -1, the number of matrix columns in this variable, or 1
1513 * if this variable is not a matrix.
1515 unsigned matrix_columns;
1517 /** Type of the varying returned by glGetTransformFeedbackVarying() */
1521 * If location != -1, the size that should be returned by
1522 * glGetTransformFeedbackVarying().
1529 * Initialize this object based on a string that was passed to
1530 * glTransformFeedbackVaryings. If there is a parse error, the error is
1531 * reported using linker_error(), and false is returned.
1534 tfeedback_decl::init(struct gl_context *ctx, struct gl_shader_program *prog,
1535 const void *mem_ctx, const char *input)
1537 /* We don't have to be pedantic about what is a valid GLSL variable name,
1538 * because any variable with an invalid name can't exist in the IR anyway.
1541 this->location = -1;
1542 this->orig_name = input;
1543 this->is_clip_distance_mesa = false;
1545 const char *bracket = strrchr(input, '[');
1548 this->var_name = ralloc_strndup(mem_ctx, input, bracket - input);
1549 if (sscanf(bracket, "[%u]", &this->array_subscript) != 1) {
1550 linker_error(prog, "Cannot parse transform feedback varying %s", input);
1553 this->is_subscripted = true;
1555 this->var_name = ralloc_strdup(mem_ctx, input);
1556 this->is_subscripted = false;
1559 /* For drivers that lower gl_ClipDistance to gl_ClipDistanceMESA, this
1560 * class must behave specially to account for the fact that gl_ClipDistance
1561 * is converted from a float[8] to a vec4[2].
1563 if (ctx->ShaderCompilerOptions[MESA_SHADER_VERTEX].LowerClipDistance &&
1564 strcmp(this->var_name, "gl_ClipDistance") == 0) {
1565 this->is_clip_distance_mesa = true;
1573 * Determine whether two tfeedback_decl objects refer to the same variable and
1574 * array index (if applicable).
1577 tfeedback_decl::is_same(const tfeedback_decl &x, const tfeedback_decl &y)
1579 if (strcmp(x.var_name, y.var_name) != 0)
1581 if (x.is_subscripted != y.is_subscripted)
1583 if (x.is_subscripted && x.array_subscript != y.array_subscript)
1590 * Assign a location for this tfeedback_decl object based on the location
1591 * assignment in output_var.
1593 * If an error occurs, the error is reported through linker_error() and false
1597 tfeedback_decl::assign_location(struct gl_context *ctx,
1598 struct gl_shader_program *prog,
1599 ir_variable *output_var)
1601 if (output_var->type->is_array()) {
1602 /* Array variable */
1603 const unsigned matrix_cols =
1604 output_var->type->fields.array->matrix_columns;
1605 unsigned actual_array_size = this->is_clip_distance_mesa ?
1606 prog->Vert.ClipDistanceArraySize : output_var->type->array_size();
1608 if (this->is_subscripted) {
1609 /* Check array bounds. */
1610 if (this->array_subscript >= actual_array_size) {
1611 linker_error(prog, "Transform feedback varying %s has index "
1612 "%i, but the array size is %u.",
1613 this->orig_name, this->array_subscript,
1617 if (this->is_clip_distance_mesa) {
1619 output_var->location + this->array_subscript / 4;
1622 output_var->location + this->array_subscript * matrix_cols;
1626 this->location = output_var->location;
1627 this->size = actual_array_size;
1629 this->vector_elements = output_var->type->fields.array->vector_elements;
1630 this->matrix_columns = matrix_cols;
1631 if (this->is_clip_distance_mesa)
1632 this->type = GL_FLOAT;
1634 this->type = output_var->type->fields.array->gl_type;
1636 /* Regular variable (scalar, vector, or matrix) */
1637 if (this->is_subscripted) {
1638 linker_error(prog, "Transform feedback varying %s requested, "
1639 "but %s is not an array.",
1640 this->orig_name, this->var_name);
1643 this->location = output_var->location;
1645 this->vector_elements = output_var->type->vector_elements;
1646 this->matrix_columns = output_var->type->matrix_columns;
1647 this->type = output_var->type->gl_type;
1650 /* From GL_EXT_transform_feedback:
1651 * A program will fail to link if:
1653 * * the total number of components to capture in any varying
1654 * variable in <varyings> is greater than the constant
1655 * MAX_TRANSFORM_FEEDBACK_SEPARATE_COMPONENTS_EXT and the
1656 * buffer mode is SEPARATE_ATTRIBS_EXT;
1658 if (prog->TransformFeedback.BufferMode == GL_SEPARATE_ATTRIBS &&
1659 this->num_components() >
1660 ctx->Const.MaxTransformFeedbackSeparateComponents) {
1661 linker_error(prog, "Transform feedback varying %s exceeds "
1662 "MAX_TRANSFORM_FEEDBACK_SEPARATE_COMPONENTS.",
1672 tfeedback_decl::accumulate_num_outputs(struct gl_shader_program *prog,
1675 if (!this->is_assigned()) {
1676 /* From GL_EXT_transform_feedback:
1677 * A program will fail to link if:
1679 * * any variable name specified in the <varyings> array is not
1680 * declared as an output in the geometry shader (if present) or
1681 * the vertex shader (if no geometry shader is present);
1683 linker_error(prog, "Transform feedback varying %s undeclared.",
1688 unsigned translated_size = this->size;
1689 if (this->is_clip_distance_mesa)
1690 translated_size = (translated_size + 3) / 4;
1692 *count += translated_size * this->matrix_columns;
1699 * Update gl_transform_feedback_info to reflect this tfeedback_decl.
1701 * If an error occurs, the error is reported through linker_error() and false
1705 tfeedback_decl::store(struct gl_context *ctx, struct gl_shader_program *prog,
1706 struct gl_transform_feedback_info *info,
1708 unsigned varying, const unsigned max_outputs) const
1710 /* From GL_EXT_transform_feedback:
1711 * A program will fail to link if:
1713 * * the total number of components to capture is greater than
1714 * the constant MAX_TRANSFORM_FEEDBACK_INTERLEAVED_COMPONENTS_EXT
1715 * and the buffer mode is INTERLEAVED_ATTRIBS_EXT.
1717 if (prog->TransformFeedback.BufferMode == GL_INTERLEAVED_ATTRIBS &&
1718 info->BufferStride[buffer] + this->num_components() >
1719 ctx->Const.MaxTransformFeedbackInterleavedComponents) {
1720 linker_error(prog, "The MAX_TRANSFORM_FEEDBACK_INTERLEAVED_COMPONENTS "
1721 "limit has been exceeded.");
1725 unsigned translated_size = this->size;
1726 if (this->is_clip_distance_mesa)
1727 translated_size = (translated_size + 3) / 4;
1728 unsigned components_so_far = 0;
1729 for (unsigned index = 0; index < translated_size; ++index) {
1730 for (unsigned v = 0; v < this->matrix_columns; ++v) {
1731 unsigned num_components = this->vector_elements;
1732 assert(info->NumOutputs < max_outputs);
1733 info->Outputs[info->NumOutputs].ComponentOffset = 0;
1734 if (this->is_clip_distance_mesa) {
1735 if (this->is_subscripted) {
1737 info->Outputs[info->NumOutputs].ComponentOffset =
1738 this->array_subscript % 4;
1740 num_components = MIN2(4, this->size - components_so_far);
1743 info->Outputs[info->NumOutputs].OutputRegister =
1744 this->location + v + index * this->matrix_columns;
1745 info->Outputs[info->NumOutputs].NumComponents = num_components;
1746 info->Outputs[info->NumOutputs].OutputBuffer = buffer;
1747 info->Outputs[info->NumOutputs].DstOffset = info->BufferStride[buffer];
1749 info->BufferStride[buffer] += num_components;
1750 components_so_far += num_components;
1753 assert(components_so_far == this->num_components());
1755 info->Varyings[varying].Name = ralloc_strdup(prog, this->orig_name);
1756 info->Varyings[varying].Type = this->type;
1757 info->Varyings[varying].Size = this->size;
1765 * Parse all the transform feedback declarations that were passed to
1766 * glTransformFeedbackVaryings() and store them in tfeedback_decl objects.
1768 * If an error occurs, the error is reported through linker_error() and false
1772 parse_tfeedback_decls(struct gl_context *ctx, struct gl_shader_program *prog,
1773 const void *mem_ctx, unsigned num_names,
1774 char **varying_names, tfeedback_decl *decls)
1776 for (unsigned i = 0; i < num_names; ++i) {
1777 if (!decls[i].init(ctx, prog, mem_ctx, varying_names[i]))
1779 /* From GL_EXT_transform_feedback:
1780 * A program will fail to link if:
1782 * * any two entries in the <varyings> array specify the same varying
1785 * We interpret this to mean "any two entries in the <varyings> array
1786 * specify the same varying variable and array index", since transform
1787 * feedback of arrays would be useless otherwise.
1789 for (unsigned j = 0; j < i; ++j) {
1790 if (tfeedback_decl::is_same(decls[i], decls[j])) {
1791 linker_error(prog, "Transform feedback varying %s specified "
1792 "more than once.", varying_names[i]);
1802 * Assign a location for a variable that is produced in one pipeline stage
1803 * (the "producer") and consumed in the next stage (the "consumer").
1805 * \param input_var is the input variable declaration in the consumer.
1807 * \param output_var is the output variable declaration in the producer.
1809 * \param input_index is the counter that keeps track of assigned input
1810 * locations in the consumer.
1812 * \param output_index is the counter that keeps track of assigned output
1813 * locations in the producer.
1815 * It is permissible for \c input_var to be NULL (this happens if a variable
1816 * is output by the producer and consumed by transform feedback, but not
1817 * consumed by the consumer).
1819 * If the variable has already been assigned a location, this function has no
1823 assign_varying_location(ir_variable *input_var, ir_variable *output_var,
1824 unsigned *input_index, unsigned *output_index)
1826 if (output_var->location != -1) {
1827 /* Location already assigned. */
1832 assert(input_var->location == -1);
1833 input_var->location = *input_index;
1836 output_var->location = *output_index;
1838 /* FINISHME: Support for "varying" records in GLSL 1.50. */
1839 assert(!output_var->type->is_record());
1841 if (output_var->type->is_array()) {
1842 const unsigned slots = output_var->type->length
1843 * output_var->type->fields.array->matrix_columns;
1845 *output_index += slots;
1846 *input_index += slots;
1848 const unsigned slots = output_var->type->matrix_columns;
1850 *output_index += slots;
1851 *input_index += slots;
1857 * Assign locations for all variables that are produced in one pipeline stage
1858 * (the "producer") and consumed in the next stage (the "consumer").
1860 * Variables produced by the producer may also be consumed by transform
1863 * \param num_tfeedback_decls is the number of declarations indicating
1864 * variables that may be consumed by transform feedback.
1866 * \param tfeedback_decls is a pointer to an array of tfeedback_decl objects
1867 * representing the result of parsing the strings passed to
1868 * glTransformFeedbackVaryings(). assign_location() will be called for
1869 * each of these objects that matches one of the outputs of the
1872 * When num_tfeedback_decls is nonzero, it is permissible for the consumer to
1873 * be NULL. In this case, varying locations are assigned solely based on the
1874 * requirements of transform feedback.
1877 assign_varying_locations(struct gl_context *ctx,
1878 struct gl_shader_program *prog,
1879 gl_shader *producer, gl_shader *consumer,
1880 unsigned num_tfeedback_decls,
1881 tfeedback_decl *tfeedback_decls)
1883 /* FINISHME: Set dynamically when geometry shader support is added. */
1884 unsigned output_index = VERT_RESULT_VAR0;
1885 unsigned input_index = FRAG_ATTRIB_VAR0;
1887 /* Operate in a total of three passes.
1889 * 1. Assign locations for any matching inputs and outputs.
1891 * 2. Mark output variables in the producer that do not have locations as
1892 * not being outputs. This lets the optimizer eliminate them.
1894 * 3. Mark input variables in the consumer that do not have locations as
1895 * not being inputs. This lets the optimizer eliminate them.
1898 link_invalidate_variable_locations(producer, ir_var_out, VERT_RESULT_VAR0);
1900 link_invalidate_variable_locations(consumer, ir_var_in, FRAG_ATTRIB_VAR0);
1902 foreach_list(node, producer->ir) {
1903 ir_variable *const output_var = ((ir_instruction *) node)->as_variable();
1905 if ((output_var == NULL) || (output_var->mode != ir_var_out))
1908 ir_variable *input_var =
1909 consumer ? consumer->symbols->get_variable(output_var->name) : NULL;
1911 if (input_var && input_var->mode != ir_var_in)
1915 assign_varying_location(input_var, output_var, &input_index,
1919 for (unsigned i = 0; i < num_tfeedback_decls; ++i) {
1920 if (!tfeedback_decls[i].is_assigned() &&
1921 tfeedback_decls[i].matches_var(output_var)) {
1922 if (output_var->location == -1) {
1923 assign_varying_location(input_var, output_var, &input_index,
1926 if (!tfeedback_decls[i].assign_location(ctx, prog, output_var))
1932 unsigned varying_vectors = 0;
1935 foreach_list(node, consumer->ir) {
1936 ir_variable *const var = ((ir_instruction *) node)->as_variable();
1938 if ((var == NULL) || (var->mode != ir_var_in))
1941 if (var->location == -1) {
1942 if (prog->Version <= 120) {
1943 /* On page 25 (page 31 of the PDF) of the GLSL 1.20 spec:
1945 * Only those varying variables used (i.e. read) in
1946 * the fragment shader executable must be written to
1947 * by the vertex shader executable; declaring
1948 * superfluous varying variables in a vertex shader is
1951 * We interpret this text as meaning that the VS must
1952 * write the variable for the FS to read it. See
1953 * "glsl1-varying read but not written" in piglit.
1956 linker_error(prog, "fragment shader varying %s not written "
1957 "by vertex shader\n.", var->name);
1960 /* An 'in' variable is only really a shader input if its
1961 * value is written by the previous stage.
1963 var->mode = ir_var_auto;
1965 /* The packing rules are used for vertex shader inputs are also
1966 * used for fragment shader inputs.
1968 varying_vectors += count_attribute_slots(var->type);
1973 if (ctx->API == API_OPENGLES2 || prog->Version == 100) {
1974 if (varying_vectors > ctx->Const.MaxVarying) {
1975 if (ctx->Const.GLSLSkipStrictMaxVaryingLimitCheck) {
1976 linker_warning(prog, "shader uses too many varying vectors "
1977 "(%u > %u), but the driver will try to optimize "
1978 "them out; this is non-portable out-of-spec "
1980 varying_vectors, ctx->Const.MaxVarying);
1982 linker_error(prog, "shader uses too many varying vectors "
1984 varying_vectors, ctx->Const.MaxVarying);
1989 const unsigned float_components = varying_vectors * 4;
1990 if (float_components > ctx->Const.MaxVarying * 4) {
1991 if (ctx->Const.GLSLSkipStrictMaxVaryingLimitCheck) {
1992 linker_warning(prog, "shader uses too many varying components "
1993 "(%u > %u), but the driver will try to optimize "
1994 "them out; this is non-portable out-of-spec "
1996 float_components, ctx->Const.MaxVarying * 4);
1998 linker_error(prog, "shader uses too many varying components "
2000 float_components, ctx->Const.MaxVarying * 4);
2011 * Store transform feedback location assignments into
2012 * prog->LinkedTransformFeedback based on the data stored in tfeedback_decls.
2014 * If an error occurs, the error is reported through linker_error() and false
2018 store_tfeedback_info(struct gl_context *ctx, struct gl_shader_program *prog,
2019 unsigned num_tfeedback_decls,
2020 tfeedback_decl *tfeedback_decls)
2022 bool separate_attribs_mode =
2023 prog->TransformFeedback.BufferMode == GL_SEPARATE_ATTRIBS;
2025 ralloc_free(prog->LinkedTransformFeedback.Varyings);
2026 ralloc_free(prog->LinkedTransformFeedback.Outputs);
2028 memset(&prog->LinkedTransformFeedback, 0,
2029 sizeof(prog->LinkedTransformFeedback));
2031 prog->LinkedTransformFeedback.NumBuffers =
2032 separate_attribs_mode ? num_tfeedback_decls : 1;
2034 prog->LinkedTransformFeedback.Varyings =
2036 struct gl_transform_feedback_varying_info,
2037 num_tfeedback_decls);
2039 unsigned num_outputs = 0;
2040 for (unsigned i = 0; i < num_tfeedback_decls; ++i)
2041 if (!tfeedback_decls[i].accumulate_num_outputs(prog, &num_outputs))
2044 prog->LinkedTransformFeedback.Outputs =
2046 struct gl_transform_feedback_output,
2049 for (unsigned i = 0; i < num_tfeedback_decls; ++i) {
2050 unsigned buffer = separate_attribs_mode ? i : 0;
2051 if (!tfeedback_decls[i].store(ctx, prog, &prog->LinkedTransformFeedback,
2052 buffer, i, num_outputs))
2055 assert(prog->LinkedTransformFeedback.NumOutputs == num_outputs);
2061 * Store the gl_FragDepth layout in the gl_shader_program struct.
2064 store_fragdepth_layout(struct gl_shader_program *prog)
2066 if (prog->_LinkedShaders[MESA_SHADER_FRAGMENT] == NULL) {
2070 struct exec_list *ir = prog->_LinkedShaders[MESA_SHADER_FRAGMENT]->ir;
2072 /* We don't look up the gl_FragDepth symbol directly because if
2073 * gl_FragDepth is not used in the shader, it's removed from the IR.
2074 * However, the symbol won't be removed from the symbol table.
2076 * We're only interested in the cases where the variable is NOT removed
2079 foreach_list(node, ir) {
2080 ir_variable *const var = ((ir_instruction *) node)->as_variable();
2082 if (var == NULL || var->mode != ir_var_out) {
2086 if (strcmp(var->name, "gl_FragDepth") == 0) {
2087 switch (var->depth_layout) {
2088 case ir_depth_layout_none:
2089 prog->FragDepthLayout = FRAG_DEPTH_LAYOUT_NONE;
2091 case ir_depth_layout_any:
2092 prog->FragDepthLayout = FRAG_DEPTH_LAYOUT_ANY;
2094 case ir_depth_layout_greater:
2095 prog->FragDepthLayout = FRAG_DEPTH_LAYOUT_GREATER;
2097 case ir_depth_layout_less:
2098 prog->FragDepthLayout = FRAG_DEPTH_LAYOUT_LESS;
2100 case ir_depth_layout_unchanged:
2101 prog->FragDepthLayout = FRAG_DEPTH_LAYOUT_UNCHANGED;
2112 * Validate the resources used by a program versus the implementation limits
2115 check_resources(struct gl_context *ctx, struct gl_shader_program *prog)
2117 static const char *const shader_names[MESA_SHADER_TYPES] = {
2118 "vertex", "fragment", "geometry"
2121 const unsigned max_samplers[MESA_SHADER_TYPES] = {
2122 ctx->Const.MaxVertexTextureImageUnits,
2123 ctx->Const.MaxTextureImageUnits,
2124 ctx->Const.MaxGeometryTextureImageUnits
2127 const unsigned max_uniform_components[MESA_SHADER_TYPES] = {
2128 ctx->Const.VertexProgram.MaxUniformComponents,
2129 ctx->Const.FragmentProgram.MaxUniformComponents,
2130 0 /* FINISHME: Geometry shaders. */
2133 for (unsigned i = 0; i < MESA_SHADER_TYPES; i++) {
2134 struct gl_shader *sh = prog->_LinkedShaders[i];
2139 if (sh->num_samplers > max_samplers[i]) {
2140 linker_error(prog, "Too many %s shader texture samplers",
2144 if (sh->num_uniform_components > max_uniform_components[i]) {
2145 if (ctx->Const.GLSLSkipStrictMaxUniformLimitCheck) {
2146 linker_warning(prog, "Too many %s shader uniform components, "
2147 "but the driver will try to optimize them out; "
2148 "this is non-portable out-of-spec behavior\n",
2151 linker_error(prog, "Too many %s shader uniform components",
2157 return prog->LinkStatus;
2161 link_shaders(struct gl_context *ctx, struct gl_shader_program *prog)
2163 tfeedback_decl *tfeedback_decls = NULL;
2164 unsigned num_tfeedback_decls = prog->TransformFeedback.NumVarying;
2166 void *mem_ctx = ralloc_context(NULL); // temporary linker context
2168 prog->LinkStatus = false;
2169 prog->Validated = false;
2170 prog->_Used = false;
2172 if (prog->InfoLog != NULL)
2173 ralloc_free(prog->InfoLog);
2175 prog->InfoLog = ralloc_strdup(NULL, "");
2177 /* Separate the shaders into groups based on their type.
2179 struct gl_shader **vert_shader_list;
2180 unsigned num_vert_shaders = 0;
2181 struct gl_shader **frag_shader_list;
2182 unsigned num_frag_shaders = 0;
2184 vert_shader_list = (struct gl_shader **)
2185 calloc(2 * prog->NumShaders, sizeof(struct gl_shader *));
2186 frag_shader_list = &vert_shader_list[prog->NumShaders];
2188 unsigned min_version = UINT_MAX;
2189 unsigned max_version = 0;
2190 for (unsigned i = 0; i < prog->NumShaders; i++) {
2191 min_version = MIN2(min_version, prog->Shaders[i]->Version);
2192 max_version = MAX2(max_version, prog->Shaders[i]->Version);
2194 switch (prog->Shaders[i]->Type) {
2195 case GL_VERTEX_SHADER:
2196 vert_shader_list[num_vert_shaders] = prog->Shaders[i];
2199 case GL_FRAGMENT_SHADER:
2200 frag_shader_list[num_frag_shaders] = prog->Shaders[i];
2203 case GL_GEOMETRY_SHADER:
2204 /* FINISHME: Support geometry shaders. */
2205 assert(prog->Shaders[i]->Type != GL_GEOMETRY_SHADER);
2210 /* Previous to GLSL version 1.30, different compilation units could mix and
2211 * match shading language versions. With GLSL 1.30 and later, the versions
2212 * of all shaders must match.
2214 assert(min_version >= 100);
2215 assert(max_version <= 140);
2216 if ((max_version >= 130 || min_version == 100)
2217 && min_version != max_version) {
2218 linker_error(prog, "all shaders must use same shading "
2219 "language version\n");
2223 prog->Version = max_version;
2225 for (unsigned int i = 0; i < MESA_SHADER_TYPES; i++) {
2226 if (prog->_LinkedShaders[i] != NULL)
2227 ctx->Driver.DeleteShader(ctx, prog->_LinkedShaders[i]);
2229 prog->_LinkedShaders[i] = NULL;
2232 /* Link all shaders for a particular stage and validate the result.
2234 if (num_vert_shaders > 0) {
2235 gl_shader *const sh =
2236 link_intrastage_shaders(mem_ctx, ctx, prog, vert_shader_list,
2242 if (!validate_vertex_shader_executable(prog, sh))
2245 _mesa_reference_shader(ctx, &prog->_LinkedShaders[MESA_SHADER_VERTEX],
2249 if (num_frag_shaders > 0) {
2250 gl_shader *const sh =
2251 link_intrastage_shaders(mem_ctx, ctx, prog, frag_shader_list,
2257 if (!validate_fragment_shader_executable(prog, sh))
2260 _mesa_reference_shader(ctx, &prog->_LinkedShaders[MESA_SHADER_FRAGMENT],
2264 /* Here begins the inter-stage linking phase. Some initial validation is
2265 * performed, then locations are assigned for uniforms, attributes, and
2268 if (cross_validate_uniforms(prog)) {
2271 for (prev = 0; prev < MESA_SHADER_TYPES; prev++) {
2272 if (prog->_LinkedShaders[prev] != NULL)
2276 /* Validate the inputs of each stage with the output of the preceding
2279 for (unsigned i = prev + 1; i < MESA_SHADER_TYPES; i++) {
2280 if (prog->_LinkedShaders[i] == NULL)
2283 if (!cross_validate_outputs_to_inputs(prog,
2284 prog->_LinkedShaders[prev],
2285 prog->_LinkedShaders[i]))
2291 prog->LinkStatus = true;
2294 /* Do common optimization before assigning storage for attributes,
2295 * uniforms, and varyings. Later optimization could possibly make
2296 * some of that unused.
2298 for (unsigned i = 0; i < MESA_SHADER_TYPES; i++) {
2299 if (prog->_LinkedShaders[i] == NULL)
2302 detect_recursion_linked(prog, prog->_LinkedShaders[i]->ir);
2303 if (!prog->LinkStatus)
2306 if (ctx->ShaderCompilerOptions[i].LowerClipDistance)
2307 lower_clip_distance(prog->_LinkedShaders[i]->ir);
2309 unsigned max_unroll = ctx->ShaderCompilerOptions[i].MaxUnrollIterations;
2311 while (do_common_optimization(prog->_LinkedShaders[i]->ir, true, false, max_unroll))
2315 /* FINISHME: The value of the max_attribute_index parameter is
2316 * FINISHME: implementation dependent based on the value of
2317 * FINISHME: GL_MAX_VERTEX_ATTRIBS. GL_MAX_VERTEX_ATTRIBS must be
2318 * FINISHME: at least 16, so hardcode 16 for now.
2320 if (!assign_attribute_or_color_locations(prog, MESA_SHADER_VERTEX, 16)) {
2324 if (!assign_attribute_or_color_locations(prog, MESA_SHADER_FRAGMENT, ctx->Const.MaxDrawBuffers)) {
2329 for (prev = 0; prev < MESA_SHADER_TYPES; prev++) {
2330 if (prog->_LinkedShaders[prev] != NULL)
2334 if (num_tfeedback_decls != 0) {
2335 /* From GL_EXT_transform_feedback:
2336 * A program will fail to link if:
2338 * * the <count> specified by TransformFeedbackVaryingsEXT is
2339 * non-zero, but the program object has no vertex or geometry
2342 if (prev >= MESA_SHADER_FRAGMENT) {
2343 linker_error(prog, "Transform feedback varyings specified, but "
2344 "no vertex or geometry shader is present.");
2348 tfeedback_decls = ralloc_array(mem_ctx, tfeedback_decl,
2349 prog->TransformFeedback.NumVarying);
2350 if (!parse_tfeedback_decls(ctx, prog, mem_ctx, num_tfeedback_decls,
2351 prog->TransformFeedback.VaryingNames,
2356 for (unsigned i = prev + 1; i < MESA_SHADER_TYPES; i++) {
2357 if (prog->_LinkedShaders[i] == NULL)
2360 if (!assign_varying_locations(
2361 ctx, prog, prog->_LinkedShaders[prev], prog->_LinkedShaders[i],
2362 i == MESA_SHADER_FRAGMENT ? num_tfeedback_decls : 0,
2369 if (prev != MESA_SHADER_FRAGMENT && num_tfeedback_decls != 0) {
2370 /* There was no fragment shader, but we still have to assign varying
2371 * locations for use by transform feedback.
2373 if (!assign_varying_locations(
2374 ctx, prog, prog->_LinkedShaders[prev], NULL, num_tfeedback_decls,
2379 if (!store_tfeedback_info(ctx, prog, num_tfeedback_decls, tfeedback_decls))
2382 if (prog->_LinkedShaders[MESA_SHADER_VERTEX] != NULL) {
2383 demote_shader_inputs_and_outputs(prog->_LinkedShaders[MESA_SHADER_VERTEX],
2386 /* Eliminate code that is now dead due to unused vertex outputs being
2389 while (do_dead_code(prog->_LinkedShaders[MESA_SHADER_VERTEX]->ir, false))
2393 if (prog->_LinkedShaders[MESA_SHADER_GEOMETRY] != NULL) {
2394 gl_shader *const sh = prog->_LinkedShaders[MESA_SHADER_GEOMETRY];
2396 demote_shader_inputs_and_outputs(sh, ir_var_in);
2397 demote_shader_inputs_and_outputs(sh, ir_var_inout);
2398 demote_shader_inputs_and_outputs(sh, ir_var_out);
2400 /* Eliminate code that is now dead due to unused geometry outputs being
2403 while (do_dead_code(prog->_LinkedShaders[MESA_SHADER_GEOMETRY]->ir, false))
2407 if (prog->_LinkedShaders[MESA_SHADER_FRAGMENT] != NULL) {
2408 gl_shader *const sh = prog->_LinkedShaders[MESA_SHADER_FRAGMENT];
2410 demote_shader_inputs_and_outputs(sh, ir_var_in);
2412 /* Eliminate code that is now dead due to unused fragment inputs being
2413 * demoted. This shouldn't actually do anything other than remove
2414 * declarations of the (now unused) global variables.
2416 while (do_dead_code(prog->_LinkedShaders[MESA_SHADER_FRAGMENT]->ir, false))
2420 update_array_sizes(prog);
2421 link_assign_uniform_locations(prog);
2422 store_fragdepth_layout(prog);
2424 if (!check_resources(ctx, prog))
2427 /* OpenGL ES requires that a vertex shader and a fragment shader both be
2428 * present in a linked program. By checking for use of shading language
2429 * version 1.00, we also catch the GL_ARB_ES2_compatibility case.
2431 if (!prog->InternalSeparateShader &&
2432 (ctx->API == API_OPENGLES2 || prog->Version == 100)) {
2433 if (prog->_LinkedShaders[MESA_SHADER_VERTEX] == NULL) {
2434 linker_error(prog, "program lacks a vertex shader\n");
2435 } else if (prog->_LinkedShaders[MESA_SHADER_FRAGMENT] == NULL) {
2436 linker_error(prog, "program lacks a fragment shader\n");
2440 /* FINISHME: Assign fragment shader output locations. */
2443 free(vert_shader_list);
2445 for (unsigned i = 0; i < MESA_SHADER_TYPES; i++) {
2446 if (prog->_LinkedShaders[i] == NULL)
2449 /* Retain any live IR, but trash the rest. */
2450 reparent_ir(prog->_LinkedShaders[i]->ir, prog->_LinkedShaders[i]->ir);
2452 /* The symbol table in the linked shaders may contain references to
2453 * variables that were removed (e.g., unused uniforms). Since it may
2454 * contain junk, there is no possible valid use. Delete it and set the
2457 delete prog->_LinkedShaders[i]->symbols;
2458 prog->_LinkedShaders[i]->symbols = NULL;
2461 ralloc_free(mem_ctx);