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"
69 #include "glsl_parser_extras.h"
72 #include "program/hash_table.h"
74 #include "link_varyings.h"
75 #include "ir_optimization.h"
76 #include "ir_rvalue_visitor.h"
77 #include "ir_uniform.h"
79 #include "main/shaderobj.h"
80 #include "main/enums.h"
83 void linker_error(gl_shader_program *, const char *, ...);
88 * Visitor that determines whether or not a variable is ever written.
90 class find_assignment_visitor : public ir_hierarchical_visitor {
92 find_assignment_visitor(const char *name)
93 : name(name), found(false)
98 virtual ir_visitor_status visit_enter(ir_assignment *ir)
100 ir_variable *const var = ir->lhs->variable_referenced();
102 if (strcmp(name, var->name) == 0) {
107 return visit_continue_with_parent;
110 virtual ir_visitor_status visit_enter(ir_call *ir)
112 foreach_two_lists(formal_node, &ir->callee->parameters,
113 actual_node, &ir->actual_parameters) {
114 ir_rvalue *param_rval = (ir_rvalue *) actual_node;
115 ir_variable *sig_param = (ir_variable *) formal_node;
117 if (sig_param->data.mode == ir_var_function_out ||
118 sig_param->data.mode == ir_var_function_inout) {
119 ir_variable *var = param_rval->variable_referenced();
120 if (var && strcmp(name, var->name) == 0) {
127 if (ir->return_deref != NULL) {
128 ir_variable *const var = ir->return_deref->variable_referenced();
130 if (strcmp(name, var->name) == 0) {
136 return visit_continue_with_parent;
139 bool variable_found()
145 const char *name; /**< Find writes to a variable with this name. */
146 bool found; /**< Was a write to the variable found? */
151 * Visitor that determines whether or not a variable is ever read.
153 class find_deref_visitor : public ir_hierarchical_visitor {
155 find_deref_visitor(const char *name)
156 : name(name), found(false)
161 virtual ir_visitor_status visit(ir_dereference_variable *ir)
163 if (strcmp(this->name, ir->var->name) == 0) {
168 return visit_continue;
171 bool variable_found() const
177 const char *name; /**< Find writes to a variable with this name. */
178 bool found; /**< Was a write to the variable found? */
182 class geom_array_resize_visitor : public ir_hierarchical_visitor {
184 unsigned num_vertices;
185 gl_shader_program *prog;
187 geom_array_resize_visitor(unsigned num_vertices, gl_shader_program *prog)
189 this->num_vertices = num_vertices;
193 virtual ~geom_array_resize_visitor()
198 virtual ir_visitor_status visit(ir_variable *var)
200 if (!var->type->is_array() || var->data.mode != ir_var_shader_in)
201 return visit_continue;
203 unsigned size = var->type->length;
205 /* Generate a link error if the shader has declared this array with an
208 if (size && size != this->num_vertices) {
209 linker_error(this->prog, "size of array %s declared as %u, "
210 "but number of input vertices is %u\n",
211 var->name, size, this->num_vertices);
212 return visit_continue;
215 /* Generate a link error if the shader attempts to access an input
216 * array using an index too large for its actual size assigned at link
219 if (var->data.max_array_access >= this->num_vertices) {
220 linker_error(this->prog, "geometry shader accesses element %i of "
221 "%s, but only %i input vertices\n",
222 var->data.max_array_access, var->name, this->num_vertices);
223 return visit_continue;
226 var->type = glsl_type::get_array_instance(var->type->element_type(),
228 var->data.max_array_access = this->num_vertices - 1;
230 return visit_continue;
233 /* Dereferences of input variables need to be updated so that their type
234 * matches the newly assigned type of the variable they are accessing. */
235 virtual ir_visitor_status visit(ir_dereference_variable *ir)
237 ir->type = ir->var->type;
238 return visit_continue;
241 /* Dereferences of 2D input arrays need to be updated so that their type
242 * matches the newly assigned type of the array they are accessing. */
243 virtual ir_visitor_status visit_leave(ir_dereference_array *ir)
245 const glsl_type *const vt = ir->array->type;
247 ir->type = vt->element_type();
248 return visit_continue;
253 * Visitor that determines the highest stream id to which a (geometry) shader
254 * emits vertices. It also checks whether End{Stream}Primitive is ever called.
256 class find_emit_vertex_visitor : public ir_hierarchical_visitor {
258 find_emit_vertex_visitor(int max_allowed)
259 : max_stream_allowed(max_allowed),
260 invalid_stream_id(0),
261 invalid_stream_id_from_emit_vertex(false),
262 end_primitive_found(false),
263 uses_non_zero_stream(false)
268 virtual ir_visitor_status visit_leave(ir_emit_vertex *ir)
270 int stream_id = ir->stream_id();
273 invalid_stream_id = stream_id;
274 invalid_stream_id_from_emit_vertex = true;
278 if (stream_id > max_stream_allowed) {
279 invalid_stream_id = stream_id;
280 invalid_stream_id_from_emit_vertex = true;
285 uses_non_zero_stream = true;
287 return visit_continue;
290 virtual ir_visitor_status visit_leave(ir_end_primitive *ir)
292 end_primitive_found = true;
294 int stream_id = ir->stream_id();
297 invalid_stream_id = stream_id;
298 invalid_stream_id_from_emit_vertex = false;
302 if (stream_id > max_stream_allowed) {
303 invalid_stream_id = stream_id;
304 invalid_stream_id_from_emit_vertex = false;
309 uses_non_zero_stream = true;
311 return visit_continue;
316 return invalid_stream_id != 0;
319 const char *error_func()
321 return invalid_stream_id_from_emit_vertex ?
322 "EmitStreamVertex" : "EndStreamPrimitive";
327 return invalid_stream_id;
332 return uses_non_zero_stream;
335 bool uses_end_primitive()
337 return end_primitive_found;
341 int max_stream_allowed;
342 int invalid_stream_id;
343 bool invalid_stream_id_from_emit_vertex;
344 bool end_primitive_found;
345 bool uses_non_zero_stream;
348 /* Class that finds array derefs and check if indexes are dynamic. */
349 class dynamic_sampler_array_indexing_visitor : public ir_hierarchical_visitor
352 dynamic_sampler_array_indexing_visitor() :
353 dynamic_sampler_array_indexing(false)
357 ir_visitor_status visit_enter(ir_dereference_array *ir)
359 if (!ir->variable_referenced())
360 return visit_continue;
362 if (!ir->variable_referenced()->type->contains_sampler())
363 return visit_continue;
365 if (!ir->array_index->constant_expression_value()) {
366 dynamic_sampler_array_indexing = true;
369 return visit_continue;
372 bool uses_dynamic_sampler_array_indexing()
374 return dynamic_sampler_array_indexing;
378 bool dynamic_sampler_array_indexing;
381 } /* anonymous namespace */
384 linker_error(gl_shader_program *prog, const char *fmt, ...)
388 ralloc_strcat(&prog->InfoLog, "error: ");
390 ralloc_vasprintf_append(&prog->InfoLog, fmt, ap);
393 prog->LinkStatus = false;
398 linker_warning(gl_shader_program *prog, const char *fmt, ...)
402 ralloc_strcat(&prog->InfoLog, "warning: ");
404 ralloc_vasprintf_append(&prog->InfoLog, fmt, ap);
411 * Given a string identifying a program resource, break it into a base name
412 * and an optional array index in square brackets.
414 * If an array index is present, \c out_base_name_end is set to point to the
415 * "[" that precedes the array index, and the array index itself is returned
418 * If no array index is present (or if the array index is negative or
419 * mal-formed), \c out_base_name_end, is set to point to the null terminator
420 * at the end of the input string, and -1 is returned.
422 * Only the final array index is parsed; if the string contains other array
423 * indices (or structure field accesses), they are left in the base name.
425 * No attempt is made to check that the base name is properly formed;
426 * typically the caller will look up the base name in a hash table, so
427 * ill-formed base names simply turn into hash table lookup failures.
430 parse_program_resource_name(const GLchar *name,
431 const GLchar **out_base_name_end)
433 /* Section 7.3.1 ("Program Interfaces") of the OpenGL 4.3 spec says:
435 * "When an integer array element or block instance number is part of
436 * the name string, it will be specified in decimal form without a "+"
437 * or "-" sign or any extra leading zeroes. Additionally, the name
438 * string will not include white space anywhere in the string."
441 const size_t len = strlen(name);
442 *out_base_name_end = name + len;
444 if (len == 0 || name[len-1] != ']')
447 /* Walk backwards over the string looking for a non-digit character. This
448 * had better be the opening bracket for an array index.
450 * Initially, i specifies the location of the ']'. Since the string may
451 * contain only the ']' charcater, walk backwards very carefully.
454 for (i = len - 1; (i > 0) && isdigit(name[i-1]); --i)
457 if ((i == 0) || name[i-1] != '[')
460 long array_index = strtol(&name[i], NULL, 10);
464 *out_base_name_end = name + (i - 1);
470 link_invalidate_variable_locations(exec_list *ir)
472 foreach_in_list(ir_instruction, node, ir) {
473 ir_variable *const var = node->as_variable();
478 /* Only assign locations for variables that lack an explicit location.
479 * Explicit locations are set for all built-in variables, generic vertex
480 * shader inputs (via layout(location=...)), and generic fragment shader
481 * outputs (also via layout(location=...)).
483 if (!var->data.explicit_location) {
484 var->data.location = -1;
485 var->data.location_frac = 0;
488 /* ir_variable::is_unmatched_generic_inout is used by the linker while
489 * connecting outputs from one stage to inputs of the next stage.
491 * There are two implicit assumptions here. First, we assume that any
492 * built-in variable (i.e., non-generic in or out) will have
493 * explicit_location set. Second, we assume that any generic in or out
494 * will not have explicit_location set.
496 * This second assumption will only be valid until
497 * GL_ARB_separate_shader_objects is supported. When that extension is
498 * implemented, this function will need some modifications.
500 if (!var->data.explicit_location) {
501 var->data.is_unmatched_generic_inout = 1;
503 var->data.is_unmatched_generic_inout = 0;
510 * Set UsesClipDistance and ClipDistanceArraySize based on the given shader.
512 * Also check for errors based on incorrect usage of gl_ClipVertex and
515 * Return false if an error was reported.
518 analyze_clip_usage(struct gl_shader_program *prog,
519 struct gl_shader *shader, GLboolean *UsesClipDistance,
520 GLuint *ClipDistanceArraySize)
522 *ClipDistanceArraySize = 0;
524 if (!prog->IsES && prog->Version >= 130) {
525 /* From section 7.1 (Vertex Shader Special Variables) of the
528 * "It is an error for a shader to statically write both
529 * gl_ClipVertex and gl_ClipDistance."
531 * This does not apply to GLSL ES shaders, since GLSL ES defines neither
532 * gl_ClipVertex nor gl_ClipDistance.
534 find_assignment_visitor clip_vertex("gl_ClipVertex");
535 find_assignment_visitor clip_distance("gl_ClipDistance");
537 clip_vertex.run(shader->ir);
538 clip_distance.run(shader->ir);
539 if (clip_vertex.variable_found() && clip_distance.variable_found()) {
540 linker_error(prog, "%s shader writes to both `gl_ClipVertex' "
541 "and `gl_ClipDistance'\n",
542 _mesa_shader_stage_to_string(shader->Stage));
545 *UsesClipDistance = clip_distance.variable_found();
546 ir_variable *clip_distance_var =
547 shader->symbols->get_variable("gl_ClipDistance");
548 if (clip_distance_var)
549 *ClipDistanceArraySize = clip_distance_var->type->length;
551 *UsesClipDistance = false;
557 * Verify that a vertex shader executable meets all semantic requirements.
559 * Also sets prog->Vert.UsesClipDistance and prog->Vert.ClipDistanceArraySize
562 * \param shader Vertex shader executable to be verified
565 validate_vertex_shader_executable(struct gl_shader_program *prog,
566 struct gl_shader *shader)
571 /* From the GLSL 1.10 spec, page 48:
573 * "The variable gl_Position is available only in the vertex
574 * language and is intended for writing the homogeneous vertex
575 * position. All executions of a well-formed vertex shader
576 * executable must write a value into this variable. [...] The
577 * variable gl_Position is available only in the vertex
578 * language and is intended for writing the homogeneous vertex
579 * position. All executions of a well-formed vertex shader
580 * executable must write a value into this variable."
582 * while in GLSL 1.40 this text is changed to:
584 * "The variable gl_Position is available only in the vertex
585 * language and is intended for writing the homogeneous vertex
586 * position. It can be written at any time during shader
587 * execution. It may also be read back by a vertex shader
588 * after being written. This value will be used by primitive
589 * assembly, clipping, culling, and other fixed functionality
590 * operations, if present, that operate on primitives after
591 * vertex processing has occurred. Its value is undefined if
592 * the vertex shader executable does not write gl_Position."
594 * All GLSL ES Versions are similar to GLSL 1.40--failing to write to
595 * gl_Position is not an error.
597 if (prog->Version < (prog->IsES ? 300 : 140)) {
598 find_assignment_visitor find("gl_Position");
599 find.run(shader->ir);
600 if (!find.variable_found()) {
603 "vertex shader does not write to `gl_Position'."
604 "It's value is undefined. \n");
607 "vertex shader does not write to `gl_Position'. \n");
613 analyze_clip_usage(prog, shader, &prog->Vert.UsesClipDistance,
614 &prog->Vert.ClipDistanceArraySize);
619 * Verify that a fragment shader executable meets all semantic requirements
621 * \param shader Fragment shader executable to be verified
624 validate_fragment_shader_executable(struct gl_shader_program *prog,
625 struct gl_shader *shader)
630 find_assignment_visitor frag_color("gl_FragColor");
631 find_assignment_visitor frag_data("gl_FragData");
633 frag_color.run(shader->ir);
634 frag_data.run(shader->ir);
636 if (frag_color.variable_found() && frag_data.variable_found()) {
637 linker_error(prog, "fragment shader writes to both "
638 "`gl_FragColor' and `gl_FragData'\n");
643 * Verify that a geometry shader executable meets all semantic requirements
645 * Also sets prog->Geom.VerticesIn, prog->Geom.UsesClipDistance, and
646 * prog->Geom.ClipDistanceArraySize as a side effect.
648 * \param shader Geometry shader executable to be verified
651 validate_geometry_shader_executable(struct gl_shader_program *prog,
652 struct gl_shader *shader)
657 unsigned num_vertices = vertices_per_prim(prog->Geom.InputType);
658 prog->Geom.VerticesIn = num_vertices;
660 analyze_clip_usage(prog, shader, &prog->Geom.UsesClipDistance,
661 &prog->Geom.ClipDistanceArraySize);
665 * Check if geometry shaders emit to non-zero streams and do corresponding
669 validate_geometry_shader_emissions(struct gl_context *ctx,
670 struct gl_shader_program *prog)
672 if (prog->_LinkedShaders[MESA_SHADER_GEOMETRY] != NULL) {
673 find_emit_vertex_visitor emit_vertex(ctx->Const.MaxVertexStreams - 1);
674 emit_vertex.run(prog->_LinkedShaders[MESA_SHADER_GEOMETRY]->ir);
675 if (emit_vertex.error()) {
676 linker_error(prog, "Invalid call %s(%d). Accepted values for the "
677 "stream parameter are in the range [0, %d].\n",
678 emit_vertex.error_func(),
679 emit_vertex.error_stream(),
680 ctx->Const.MaxVertexStreams - 1);
682 prog->Geom.UsesStreams = emit_vertex.uses_streams();
683 prog->Geom.UsesEndPrimitive = emit_vertex.uses_end_primitive();
685 /* From the ARB_gpu_shader5 spec:
687 * "Multiple vertex streams are supported only if the output primitive
688 * type is declared to be "points". A program will fail to link if it
689 * contains a geometry shader calling EmitStreamVertex() or
690 * EndStreamPrimitive() if its output primitive type is not "points".
692 * However, in the same spec:
694 * "The function EmitVertex() is equivalent to calling EmitStreamVertex()
695 * with <stream> set to zero."
699 * "The function EndPrimitive() is equivalent to calling
700 * EndStreamPrimitive() with <stream> set to zero."
702 * Since we can call EmitVertex() and EndPrimitive() when we output
703 * primitives other than points, calling EmitStreamVertex(0) or
704 * EmitEndPrimitive(0) should not produce errors. This it also what Nvidia
705 * does. Currently we only set prog->Geom.UsesStreams to TRUE when
706 * EmitStreamVertex() or EmitEndPrimitive() are called with a non-zero
709 if (prog->Geom.UsesStreams && prog->Geom.OutputType != GL_POINTS) {
710 linker_error(prog, "EmitStreamVertex(n) and EndStreamPrimitive(n) "
711 "with n>0 requires point output\n");
718 * Perform validation of global variables used across multiple shaders
721 cross_validate_globals(struct gl_shader_program *prog,
722 struct gl_shader **shader_list,
723 unsigned num_shaders,
726 /* Examine all of the uniforms in all of the shaders and cross validate
729 glsl_symbol_table variables;
730 for (unsigned i = 0; i < num_shaders; i++) {
731 if (shader_list[i] == NULL)
734 foreach_in_list(ir_instruction, node, shader_list[i]->ir) {
735 ir_variable *const var = node->as_variable();
740 if (uniforms_only && (var->data.mode != ir_var_uniform))
743 /* Don't cross validate temporaries that are at global scope. These
744 * will eventually get pulled into the shaders 'main'.
746 if (var->data.mode == ir_var_temporary)
749 /* If a global with this name has already been seen, verify that the
750 * new instance has the same type. In addition, if the globals have
751 * initializers, the values of the initializers must be the same.
753 ir_variable *const existing = variables.get_variable(var->name);
754 if (existing != NULL) {
755 if (var->type != existing->type) {
756 /* Consider the types to be "the same" if both types are arrays
757 * of the same type and one of the arrays is implicitly sized.
758 * In addition, set the type of the linked variable to the
759 * explicitly sized array.
761 if (var->type->is_array()
762 && existing->type->is_array()
763 && (var->type->fields.array == existing->type->fields.array)
764 && ((var->type->length == 0)
765 || (existing->type->length == 0))) {
766 if (var->type->length != 0) {
767 if (var->type->length <= existing->data.max_array_access) {
768 linker_error(prog, "%s `%s' declared as type "
769 "`%s' but outermost dimension has an index"
772 var->name, var->type->name,
773 existing->data.max_array_access);
776 existing->type = var->type;
777 } else if (existing->type->length != 0
778 && existing->type->length <=
779 var->data.max_array_access) {
780 linker_error(prog, "%s `%s' declared as type "
781 "`%s' but outermost dimension has an index"
784 var->name, existing->type->name,
785 var->data.max_array_access);
788 } else if (var->type->is_record()
789 && existing->type->is_record()
790 && existing->type->record_compare(var->type)) {
791 existing->type = var->type;
793 linker_error(prog, "%s `%s' declared as type "
794 "`%s' and type `%s'\n",
796 var->name, var->type->name,
797 existing->type->name);
802 if (var->data.explicit_location) {
803 if (existing->data.explicit_location
804 && (var->data.location != existing->data.location)) {
805 linker_error(prog, "explicit locations for %s "
806 "`%s' have differing values\n",
807 mode_string(var), var->name);
811 existing->data.location = var->data.location;
812 existing->data.explicit_location = true;
815 /* From the GLSL 4.20 specification:
816 * "A link error will result if two compilation units in a program
817 * specify different integer-constant bindings for the same
818 * opaque-uniform name. However, it is not an error to specify a
819 * binding on some but not all declarations for the same name"
821 if (var->data.explicit_binding) {
822 if (existing->data.explicit_binding &&
823 var->data.binding != existing->data.binding) {
824 linker_error(prog, "explicit bindings for %s "
825 "`%s' have differing values\n",
826 mode_string(var), var->name);
830 existing->data.binding = var->data.binding;
831 existing->data.explicit_binding = true;
834 if (var->type->contains_atomic() &&
835 var->data.atomic.offset != existing->data.atomic.offset) {
836 linker_error(prog, "offset specifications for %s "
837 "`%s' have differing values\n",
838 mode_string(var), var->name);
842 /* Validate layout qualifiers for gl_FragDepth.
844 * From the AMD/ARB_conservative_depth specs:
846 * "If gl_FragDepth is redeclared in any fragment shader in a
847 * program, it must be redeclared in all fragment shaders in
848 * that program that have static assignments to
849 * gl_FragDepth. All redeclarations of gl_FragDepth in all
850 * fragment shaders in a single program must have the same set
853 if (strcmp(var->name, "gl_FragDepth") == 0) {
854 bool layout_declared = var->data.depth_layout != ir_depth_layout_none;
855 bool layout_differs =
856 var->data.depth_layout != existing->data.depth_layout;
858 if (layout_declared && layout_differs) {
860 "All redeclarations of gl_FragDepth in all "
861 "fragment shaders in a single program must have "
862 "the same set of qualifiers.\n");
865 if (var->data.used && layout_differs) {
867 "If gl_FragDepth is redeclared with a layout "
868 "qualifier in any fragment shader, it must be "
869 "redeclared with the same layout qualifier in "
870 "all fragment shaders that have assignments to "
875 /* Page 35 (page 41 of the PDF) of the GLSL 4.20 spec says:
877 * "If a shared global has multiple initializers, the
878 * initializers must all be constant expressions, and they
879 * must all have the same value. Otherwise, a link error will
880 * result. (A shared global having only one initializer does
881 * not require that initializer to be a constant expression.)"
883 * Previous to 4.20 the GLSL spec simply said that initializers
884 * must have the same value. In this case of non-constant
885 * initializers, this was impossible to determine. As a result,
886 * no vendor actually implemented that behavior. The 4.20
887 * behavior matches the implemented behavior of at least one other
888 * vendor, so we'll implement that for all GLSL versions.
890 if (var->constant_initializer != NULL) {
891 if (existing->constant_initializer != NULL) {
892 if (!var->constant_initializer->has_value(existing->constant_initializer)) {
893 linker_error(prog, "initializers for %s "
894 "`%s' have differing values\n",
895 mode_string(var), var->name);
899 /* If the first-seen instance of a particular uniform did not
900 * have an initializer but a later instance does, copy the
901 * initializer to the version stored in the symbol table.
903 /* FINISHME: This is wrong. The constant_value field should
904 * FINISHME: not be modified! Imagine a case where a shader
905 * FINISHME: without an initializer is linked in two different
906 * FINISHME: programs with shaders that have differing
907 * FINISHME: initializers. Linking with the first will
908 * FINISHME: modify the shader, and linking with the second
909 * FINISHME: will fail.
911 existing->constant_initializer =
912 var->constant_initializer->clone(ralloc_parent(existing),
917 if (var->data.has_initializer) {
918 if (existing->data.has_initializer
919 && (var->constant_initializer == NULL
920 || existing->constant_initializer == NULL)) {
922 "shared global variable `%s' has multiple "
923 "non-constant initializers.\n",
928 /* Some instance had an initializer, so keep track of that. In
929 * this location, all sorts of initializers (constant or
930 * otherwise) will propagate the existence to the variable
931 * stored in the symbol table.
933 existing->data.has_initializer = true;
936 if (existing->data.invariant != var->data.invariant) {
937 linker_error(prog, "declarations for %s `%s' have "
938 "mismatching invariant qualifiers\n",
939 mode_string(var), var->name);
942 if (existing->data.centroid != var->data.centroid) {
943 linker_error(prog, "declarations for %s `%s' have "
944 "mismatching centroid qualifiers\n",
945 mode_string(var), var->name);
948 if (existing->data.sample != var->data.sample) {
949 linker_error(prog, "declarations for %s `%s` have "
950 "mismatching sample qualifiers\n",
951 mode_string(var), var->name);
955 variables.add_variable(var);
962 * Perform validation of uniforms used across multiple shader stages
965 cross_validate_uniforms(struct gl_shader_program *prog)
967 cross_validate_globals(prog, prog->_LinkedShaders,
968 MESA_SHADER_STAGES, true);
972 * Accumulates the array of prog->UniformBlocks and checks that all
973 * definitons of blocks agree on their contents.
976 interstage_cross_validate_uniform_blocks(struct gl_shader_program *prog)
978 unsigned max_num_uniform_blocks = 0;
979 for (unsigned i = 0; i < MESA_SHADER_STAGES; i++) {
980 if (prog->_LinkedShaders[i])
981 max_num_uniform_blocks += prog->_LinkedShaders[i]->NumUniformBlocks;
984 for (unsigned i = 0; i < MESA_SHADER_STAGES; i++) {
985 struct gl_shader *sh = prog->_LinkedShaders[i];
987 prog->UniformBlockStageIndex[i] = ralloc_array(prog, int,
988 max_num_uniform_blocks);
989 for (unsigned int j = 0; j < max_num_uniform_blocks; j++)
990 prog->UniformBlockStageIndex[i][j] = -1;
995 for (unsigned int j = 0; j < sh->NumUniformBlocks; j++) {
996 int index = link_cross_validate_uniform_block(prog,
997 &prog->UniformBlocks,
998 &prog->NumUniformBlocks,
999 &sh->UniformBlocks[j]);
1002 linker_error(prog, "uniform block `%s' has mismatching definitions\n",
1003 sh->UniformBlocks[j].Name);
1007 prog->UniformBlockStageIndex[i][index] = j;
1016 * Populates a shaders symbol table with all global declarations
1019 populate_symbol_table(gl_shader *sh)
1021 sh->symbols = new(sh) glsl_symbol_table;
1023 foreach_in_list(ir_instruction, inst, sh->ir) {
1027 if ((func = inst->as_function()) != NULL) {
1028 sh->symbols->add_function(func);
1029 } else if ((var = inst->as_variable()) != NULL) {
1030 if (var->data.mode != ir_var_temporary)
1031 sh->symbols->add_variable(var);
1038 * Remap variables referenced in an instruction tree
1040 * This is used when instruction trees are cloned from one shader and placed in
1041 * another. These trees will contain references to \c ir_variable nodes that
1042 * do not exist in the target shader. This function finds these \c ir_variable
1043 * references and replaces the references with matching variables in the target
1046 * If there is no matching variable in the target shader, a clone of the
1047 * \c ir_variable is made and added to the target shader. The new variable is
1048 * added to \b both the instruction stream and the symbol table.
1050 * \param inst IR tree that is to be processed.
1051 * \param symbols Symbol table containing global scope symbols in the
1053 * \param instructions Instruction stream where new variable declarations
1057 remap_variables(ir_instruction *inst, struct gl_shader *target,
1060 class remap_visitor : public ir_hierarchical_visitor {
1062 remap_visitor(struct gl_shader *target,
1065 this->target = target;
1066 this->symbols = target->symbols;
1067 this->instructions = target->ir;
1068 this->temps = temps;
1071 virtual ir_visitor_status visit(ir_dereference_variable *ir)
1073 if (ir->var->data.mode == ir_var_temporary) {
1074 ir_variable *var = (ir_variable *) hash_table_find(temps, ir->var);
1076 assert(var != NULL);
1078 return visit_continue;
1081 ir_variable *const existing =
1082 this->symbols->get_variable(ir->var->name);
1083 if (existing != NULL)
1086 ir_variable *copy = ir->var->clone(this->target, NULL);
1088 this->symbols->add_variable(copy);
1089 this->instructions->push_head(copy);
1093 return visit_continue;
1097 struct gl_shader *target;
1098 glsl_symbol_table *symbols;
1099 exec_list *instructions;
1103 remap_visitor v(target, temps);
1110 * Move non-declarations from one instruction stream to another
1112 * The intended usage pattern of this function is to pass the pointer to the
1113 * head sentinel of a list (i.e., a pointer to the list cast to an \c exec_node
1114 * pointer) for \c last and \c false for \c make_copies on the first
1115 * call. Successive calls pass the return value of the previous call for
1116 * \c last and \c true for \c make_copies.
1118 * \param instructions Source instruction stream
1119 * \param last Instruction after which new instructions should be
1120 * inserted in the target instruction stream
1121 * \param make_copies Flag selecting whether instructions in \c instructions
1122 * should be copied (via \c ir_instruction::clone) into the
1123 * target list or moved.
1126 * The new "last" instruction in the target instruction stream. This pointer
1127 * is suitable for use as the \c last parameter of a later call to this
1131 move_non_declarations(exec_list *instructions, exec_node *last,
1132 bool make_copies, gl_shader *target)
1134 hash_table *temps = NULL;
1137 temps = hash_table_ctor(0, hash_table_pointer_hash,
1138 hash_table_pointer_compare);
1140 foreach_in_list_safe(ir_instruction, inst, instructions) {
1141 if (inst->as_function())
1144 ir_variable *var = inst->as_variable();
1145 if ((var != NULL) && (var->data.mode != ir_var_temporary))
1148 assert(inst->as_assignment()
1150 || inst->as_if() /* for initializers with the ?: operator */
1151 || ((var != NULL) && (var->data.mode == ir_var_temporary)));
1154 inst = inst->clone(target, NULL);
1157 hash_table_insert(temps, inst, var);
1159 remap_variables(inst, target, temps);
1164 last->insert_after(inst);
1169 hash_table_dtor(temps);
1175 * Get the function signature for main from a shader
1177 ir_function_signature *
1178 link_get_main_function_signature(gl_shader *sh)
1180 ir_function *const f = sh->symbols->get_function("main");
1182 exec_list void_parameters;
1184 /* Look for the 'void main()' signature and ensure that it's defined.
1185 * This keeps the linker from accidentally pick a shader that just
1186 * contains a prototype for main.
1188 * We don't have to check for multiple definitions of main (in multiple
1189 * shaders) because that would have already been caught above.
1191 ir_function_signature *sig =
1192 f->matching_signature(NULL, &void_parameters, false);
1193 if ((sig != NULL) && sig->is_defined) {
1203 * This class is only used in link_intrastage_shaders() below but declaring
1204 * it inside that function leads to compiler warnings with some versions of
1207 class array_sizing_visitor : public ir_hierarchical_visitor {
1209 array_sizing_visitor()
1210 : mem_ctx(ralloc_context(NULL)),
1211 unnamed_interfaces(hash_table_ctor(0, hash_table_pointer_hash,
1212 hash_table_pointer_compare))
1216 ~array_sizing_visitor()
1218 hash_table_dtor(this->unnamed_interfaces);
1219 ralloc_free(this->mem_ctx);
1222 virtual ir_visitor_status visit(ir_variable *var)
1224 fixup_type(&var->type, var->data.max_array_access);
1225 if (var->type->is_interface()) {
1226 if (interface_contains_unsized_arrays(var->type)) {
1227 const glsl_type *new_type =
1228 resize_interface_members(var->type,
1229 var->get_max_ifc_array_access());
1230 var->type = new_type;
1231 var->change_interface_type(new_type);
1233 } else if (var->type->is_array() &&
1234 var->type->fields.array->is_interface()) {
1235 if (interface_contains_unsized_arrays(var->type->fields.array)) {
1236 const glsl_type *new_type =
1237 resize_interface_members(var->type->fields.array,
1238 var->get_max_ifc_array_access());
1239 var->change_interface_type(new_type);
1241 glsl_type::get_array_instance(new_type, var->type->length);
1243 } else if (const glsl_type *ifc_type = var->get_interface_type()) {
1244 /* Store a pointer to the variable in the unnamed_interfaces
1247 ir_variable **interface_vars = (ir_variable **)
1248 hash_table_find(this->unnamed_interfaces, ifc_type);
1249 if (interface_vars == NULL) {
1250 interface_vars = rzalloc_array(mem_ctx, ir_variable *,
1252 hash_table_insert(this->unnamed_interfaces, interface_vars,
1255 unsigned index = ifc_type->field_index(var->name);
1256 assert(index < ifc_type->length);
1257 assert(interface_vars[index] == NULL);
1258 interface_vars[index] = var;
1260 return visit_continue;
1264 * For each unnamed interface block that was discovered while running the
1265 * visitor, adjust the interface type to reflect the newly assigned array
1266 * sizes, and fix up the ir_variable nodes to point to the new interface
1269 void fixup_unnamed_interface_types()
1271 hash_table_call_foreach(this->unnamed_interfaces,
1272 fixup_unnamed_interface_type, NULL);
1277 * If the type pointed to by \c type represents an unsized array, replace
1278 * it with a sized array whose size is determined by max_array_access.
1280 static void fixup_type(const glsl_type **type, unsigned max_array_access)
1282 if ((*type)->is_unsized_array()) {
1283 *type = glsl_type::get_array_instance((*type)->fields.array,
1284 max_array_access + 1);
1285 assert(*type != NULL);
1290 * Determine whether the given interface type contains unsized arrays (if
1291 * it doesn't, array_sizing_visitor doesn't need to process it).
1293 static bool interface_contains_unsized_arrays(const glsl_type *type)
1295 for (unsigned i = 0; i < type->length; i++) {
1296 const glsl_type *elem_type = type->fields.structure[i].type;
1297 if (elem_type->is_unsized_array())
1304 * Create a new interface type based on the given type, with unsized arrays
1305 * replaced by sized arrays whose size is determined by
1306 * max_ifc_array_access.
1308 static const glsl_type *
1309 resize_interface_members(const glsl_type *type,
1310 const unsigned *max_ifc_array_access)
1312 unsigned num_fields = type->length;
1313 glsl_struct_field *fields = new glsl_struct_field[num_fields];
1314 memcpy(fields, type->fields.structure,
1315 num_fields * sizeof(*fields));
1316 for (unsigned i = 0; i < num_fields; i++) {
1317 fixup_type(&fields[i].type, max_ifc_array_access[i]);
1319 glsl_interface_packing packing =
1320 (glsl_interface_packing) type->interface_packing;
1321 const glsl_type *new_ifc_type =
1322 glsl_type::get_interface_instance(fields, num_fields,
1323 packing, type->name);
1325 return new_ifc_type;
1328 static void fixup_unnamed_interface_type(const void *key, void *data,
1331 const glsl_type *ifc_type = (const glsl_type *) key;
1332 ir_variable **interface_vars = (ir_variable **) data;
1333 unsigned num_fields = ifc_type->length;
1334 glsl_struct_field *fields = new glsl_struct_field[num_fields];
1335 memcpy(fields, ifc_type->fields.structure,
1336 num_fields * sizeof(*fields));
1337 bool interface_type_changed = false;
1338 for (unsigned i = 0; i < num_fields; i++) {
1339 if (interface_vars[i] != NULL &&
1340 fields[i].type != interface_vars[i]->type) {
1341 fields[i].type = interface_vars[i]->type;
1342 interface_type_changed = true;
1345 if (!interface_type_changed) {
1349 glsl_interface_packing packing =
1350 (glsl_interface_packing) ifc_type->interface_packing;
1351 const glsl_type *new_ifc_type =
1352 glsl_type::get_interface_instance(fields, num_fields, packing,
1355 for (unsigned i = 0; i < num_fields; i++) {
1356 if (interface_vars[i] != NULL)
1357 interface_vars[i]->change_interface_type(new_ifc_type);
1362 * Memory context used to allocate the data in \c unnamed_interfaces.
1367 * Hash table from const glsl_type * to an array of ir_variable *'s
1368 * pointing to the ir_variables constituting each unnamed interface block.
1370 hash_table *unnamed_interfaces;
1374 * Performs the cross-validation of layout qualifiers specified in
1375 * redeclaration of gl_FragCoord for the attached fragment shaders,
1376 * and propagates them to the linked FS and linked shader program.
1379 link_fs_input_layout_qualifiers(struct gl_shader_program *prog,
1380 struct gl_shader *linked_shader,
1381 struct gl_shader **shader_list,
1382 unsigned num_shaders)
1384 linked_shader->redeclares_gl_fragcoord = false;
1385 linked_shader->uses_gl_fragcoord = false;
1386 linked_shader->origin_upper_left = false;
1387 linked_shader->pixel_center_integer = false;
1389 if (linked_shader->Stage != MESA_SHADER_FRAGMENT ||
1390 (prog->Version < 150 && !prog->ARB_fragment_coord_conventions_enable))
1393 for (unsigned i = 0; i < num_shaders; i++) {
1394 struct gl_shader *shader = shader_list[i];
1395 /* From the GLSL 1.50 spec, page 39:
1397 * "If gl_FragCoord is redeclared in any fragment shader in a program,
1398 * it must be redeclared in all the fragment shaders in that program
1399 * that have a static use gl_FragCoord."
1401 if ((linked_shader->redeclares_gl_fragcoord
1402 && !shader->redeclares_gl_fragcoord
1403 && shader->uses_gl_fragcoord)
1404 || (shader->redeclares_gl_fragcoord
1405 && !linked_shader->redeclares_gl_fragcoord
1406 && linked_shader->uses_gl_fragcoord)) {
1407 linker_error(prog, "fragment shader defined with conflicting "
1408 "layout qualifiers for gl_FragCoord\n");
1411 /* From the GLSL 1.50 spec, page 39:
1413 * "All redeclarations of gl_FragCoord in all fragment shaders in a
1414 * single program must have the same set of qualifiers."
1416 if (linked_shader->redeclares_gl_fragcoord && shader->redeclares_gl_fragcoord
1417 && (shader->origin_upper_left != linked_shader->origin_upper_left
1418 || shader->pixel_center_integer != linked_shader->pixel_center_integer)) {
1419 linker_error(prog, "fragment shader defined with conflicting "
1420 "layout qualifiers for gl_FragCoord\n");
1423 /* Update the linked shader state. Note that uses_gl_fragcoord should
1424 * accumulate the results. The other values should replace. If there
1425 * are multiple redeclarations, all the fields except uses_gl_fragcoord
1426 * are already known to be the same.
1428 if (shader->redeclares_gl_fragcoord || shader->uses_gl_fragcoord) {
1429 linked_shader->redeclares_gl_fragcoord =
1430 shader->redeclares_gl_fragcoord;
1431 linked_shader->uses_gl_fragcoord = linked_shader->uses_gl_fragcoord
1432 || shader->uses_gl_fragcoord;
1433 linked_shader->origin_upper_left = shader->origin_upper_left;
1434 linked_shader->pixel_center_integer = shader->pixel_center_integer;
1440 * Performs the cross-validation of geometry shader max_vertices and
1441 * primitive type layout qualifiers for the attached geometry shaders,
1442 * and propagates them to the linked GS and linked shader program.
1445 link_gs_inout_layout_qualifiers(struct gl_shader_program *prog,
1446 struct gl_shader *linked_shader,
1447 struct gl_shader **shader_list,
1448 unsigned num_shaders)
1450 linked_shader->Geom.VerticesOut = 0;
1451 linked_shader->Geom.Invocations = 0;
1452 linked_shader->Geom.InputType = PRIM_UNKNOWN;
1453 linked_shader->Geom.OutputType = PRIM_UNKNOWN;
1455 /* No in/out qualifiers defined for anything but GLSL 1.50+
1456 * geometry shaders so far.
1458 if (linked_shader->Stage != MESA_SHADER_GEOMETRY || prog->Version < 150)
1461 /* From the GLSL 1.50 spec, page 46:
1463 * "All geometry shader output layout declarations in a program
1464 * must declare the same layout and same value for
1465 * max_vertices. There must be at least one geometry output
1466 * layout declaration somewhere in a program, but not all
1467 * geometry shaders (compilation units) are required to
1471 for (unsigned i = 0; i < num_shaders; i++) {
1472 struct gl_shader *shader = shader_list[i];
1474 if (shader->Geom.InputType != PRIM_UNKNOWN) {
1475 if (linked_shader->Geom.InputType != PRIM_UNKNOWN &&
1476 linked_shader->Geom.InputType != shader->Geom.InputType) {
1477 linker_error(prog, "geometry shader defined with conflicting "
1481 linked_shader->Geom.InputType = shader->Geom.InputType;
1484 if (shader->Geom.OutputType != PRIM_UNKNOWN) {
1485 if (linked_shader->Geom.OutputType != PRIM_UNKNOWN &&
1486 linked_shader->Geom.OutputType != shader->Geom.OutputType) {
1487 linker_error(prog, "geometry shader defined with conflicting "
1491 linked_shader->Geom.OutputType = shader->Geom.OutputType;
1494 if (shader->Geom.VerticesOut != 0) {
1495 if (linked_shader->Geom.VerticesOut != 0 &&
1496 linked_shader->Geom.VerticesOut != shader->Geom.VerticesOut) {
1497 linker_error(prog, "geometry shader defined with conflicting "
1498 "output vertex count (%d and %d)\n",
1499 linked_shader->Geom.VerticesOut,
1500 shader->Geom.VerticesOut);
1503 linked_shader->Geom.VerticesOut = shader->Geom.VerticesOut;
1506 if (shader->Geom.Invocations != 0) {
1507 if (linked_shader->Geom.Invocations != 0 &&
1508 linked_shader->Geom.Invocations != shader->Geom.Invocations) {
1509 linker_error(prog, "geometry shader defined with conflicting "
1510 "invocation count (%d and %d)\n",
1511 linked_shader->Geom.Invocations,
1512 shader->Geom.Invocations);
1515 linked_shader->Geom.Invocations = shader->Geom.Invocations;
1519 /* Just do the intrastage -> interstage propagation right now,
1520 * since we already know we're in the right type of shader program
1523 if (linked_shader->Geom.InputType == PRIM_UNKNOWN) {
1525 "geometry shader didn't declare primitive input type\n");
1528 prog->Geom.InputType = linked_shader->Geom.InputType;
1530 if (linked_shader->Geom.OutputType == PRIM_UNKNOWN) {
1532 "geometry shader didn't declare primitive output type\n");
1535 prog->Geom.OutputType = linked_shader->Geom.OutputType;
1537 if (linked_shader->Geom.VerticesOut == 0) {
1539 "geometry shader didn't declare max_vertices\n");
1542 prog->Geom.VerticesOut = linked_shader->Geom.VerticesOut;
1544 if (linked_shader->Geom.Invocations == 0)
1545 linked_shader->Geom.Invocations = 1;
1547 prog->Geom.Invocations = linked_shader->Geom.Invocations;
1552 * Perform cross-validation of compute shader local_size_{x,y,z} layout
1553 * qualifiers for the attached compute shaders, and propagate them to the
1554 * linked CS and linked shader program.
1557 link_cs_input_layout_qualifiers(struct gl_shader_program *prog,
1558 struct gl_shader *linked_shader,
1559 struct gl_shader **shader_list,
1560 unsigned num_shaders)
1562 for (int i = 0; i < 3; i++)
1563 linked_shader->Comp.LocalSize[i] = 0;
1565 /* This function is called for all shader stages, but it only has an effect
1566 * for compute shaders.
1568 if (linked_shader->Stage != MESA_SHADER_COMPUTE)
1571 /* From the ARB_compute_shader spec, in the section describing local size
1574 * If multiple compute shaders attached to a single program object
1575 * declare local work-group size, the declarations must be identical;
1576 * otherwise a link-time error results. Furthermore, if a program
1577 * object contains any compute shaders, at least one must contain an
1578 * input layout qualifier specifying the local work sizes of the
1579 * program, or a link-time error will occur.
1581 for (unsigned sh = 0; sh < num_shaders; sh++) {
1582 struct gl_shader *shader = shader_list[sh];
1584 if (shader->Comp.LocalSize[0] != 0) {
1585 if (linked_shader->Comp.LocalSize[0] != 0) {
1586 for (int i = 0; i < 3; i++) {
1587 if (linked_shader->Comp.LocalSize[i] !=
1588 shader->Comp.LocalSize[i]) {
1589 linker_error(prog, "compute shader defined with conflicting "
1595 for (int i = 0; i < 3; i++)
1596 linked_shader->Comp.LocalSize[i] = shader->Comp.LocalSize[i];
1600 /* Just do the intrastage -> interstage propagation right now,
1601 * since we already know we're in the right type of shader program
1604 if (linked_shader->Comp.LocalSize[0] == 0) {
1605 linker_error(prog, "compute shader didn't declare local size\n");
1608 for (int i = 0; i < 3; i++)
1609 prog->Comp.LocalSize[i] = linked_shader->Comp.LocalSize[i];
1614 * Combine a group of shaders for a single stage to generate a linked shader
1617 * If this function is supplied a single shader, it is cloned, and the new
1618 * shader is returned.
1620 static struct gl_shader *
1621 link_intrastage_shaders(void *mem_ctx,
1622 struct gl_context *ctx,
1623 struct gl_shader_program *prog,
1624 struct gl_shader **shader_list,
1625 unsigned num_shaders)
1627 struct gl_uniform_block *uniform_blocks = NULL;
1629 /* Check that global variables defined in multiple shaders are consistent.
1631 cross_validate_globals(prog, shader_list, num_shaders, false);
1632 if (!prog->LinkStatus)
1635 /* Check that interface blocks defined in multiple shaders are consistent.
1637 validate_intrastage_interface_blocks(prog, (const gl_shader **)shader_list,
1639 if (!prog->LinkStatus)
1642 /* Link up uniform blocks defined within this stage. */
1643 const unsigned num_uniform_blocks =
1644 link_uniform_blocks(mem_ctx, prog, shader_list, num_shaders,
1646 if (!prog->LinkStatus)
1649 /* Check that there is only a single definition of each function signature
1650 * across all shaders.
1652 for (unsigned i = 0; i < (num_shaders - 1); i++) {
1653 foreach_in_list(ir_instruction, node, shader_list[i]->ir) {
1654 ir_function *const f = node->as_function();
1659 for (unsigned j = i + 1; j < num_shaders; j++) {
1660 ir_function *const other =
1661 shader_list[j]->symbols->get_function(f->name);
1663 /* If the other shader has no function (and therefore no function
1664 * signatures) with the same name, skip to the next shader.
1669 foreach_in_list(ir_function_signature, sig, &f->signatures) {
1670 if (!sig->is_defined || sig->is_builtin())
1673 ir_function_signature *other_sig =
1674 other->exact_matching_signature(NULL, &sig->parameters);
1676 if ((other_sig != NULL) && other_sig->is_defined
1677 && !other_sig->is_builtin()) {
1678 linker_error(prog, "function `%s' is multiply defined\n",
1687 /* Find the shader that defines main, and make a clone of it.
1689 * Starting with the clone, search for undefined references. If one is
1690 * found, find the shader that defines it. Clone the reference and add
1691 * it to the shader. Repeat until there are no undefined references or
1692 * until a reference cannot be resolved.
1694 gl_shader *main = NULL;
1695 for (unsigned i = 0; i < num_shaders; i++) {
1696 if (link_get_main_function_signature(shader_list[i]) != NULL) {
1697 main = shader_list[i];
1703 linker_error(prog, "%s shader lacks `main'\n",
1704 _mesa_shader_stage_to_string(shader_list[0]->Stage));
1708 gl_shader *linked = ctx->Driver.NewShader(NULL, 0, main->Type);
1709 linked->ir = new(linked) exec_list;
1710 clone_ir_list(mem_ctx, linked->ir, main->ir);
1712 linked->UniformBlocks = uniform_blocks;
1713 linked->NumUniformBlocks = num_uniform_blocks;
1714 ralloc_steal(linked, linked->UniformBlocks);
1716 link_fs_input_layout_qualifiers(prog, linked, shader_list, num_shaders);
1717 link_gs_inout_layout_qualifiers(prog, linked, shader_list, num_shaders);
1718 link_cs_input_layout_qualifiers(prog, linked, shader_list, num_shaders);
1720 populate_symbol_table(linked);
1722 /* The pointer to the main function in the final linked shader (i.e., the
1723 * copy of the original shader that contained the main function).
1725 ir_function_signature *const main_sig =
1726 link_get_main_function_signature(linked);
1728 /* Move any instructions other than variable declarations or function
1729 * declarations into main.
1731 exec_node *insertion_point =
1732 move_non_declarations(linked->ir, (exec_node *) &main_sig->body, false,
1735 for (unsigned i = 0; i < num_shaders; i++) {
1736 if (shader_list[i] == main)
1739 insertion_point = move_non_declarations(shader_list[i]->ir,
1740 insertion_point, true, linked);
1743 /* Check if any shader needs built-in functions. */
1744 bool need_builtins = false;
1745 for (unsigned i = 0; i < num_shaders; i++) {
1746 if (shader_list[i]->uses_builtin_functions) {
1747 need_builtins = true;
1753 if (need_builtins) {
1754 /* Make a temporary array one larger than shader_list, which will hold
1755 * the built-in function shader as well.
1757 gl_shader **linking_shaders = (gl_shader **)
1758 calloc(num_shaders + 1, sizeof(gl_shader *));
1760 ok = linking_shaders != NULL;
1763 memcpy(linking_shaders, shader_list, num_shaders * sizeof(gl_shader *));
1764 linking_shaders[num_shaders] = _mesa_glsl_get_builtin_function_shader();
1766 ok = link_function_calls(prog, linked, linking_shaders, num_shaders + 1);
1768 free(linking_shaders);
1770 _mesa_error_no_memory(__func__);
1773 ok = link_function_calls(prog, linked, shader_list, num_shaders);
1778 ctx->Driver.DeleteShader(ctx, linked);
1782 /* At this point linked should contain all of the linked IR, so
1783 * validate it to make sure nothing went wrong.
1785 validate_ir_tree(linked->ir);
1787 /* Set the size of geometry shader input arrays */
1788 if (linked->Stage == MESA_SHADER_GEOMETRY) {
1789 unsigned num_vertices = vertices_per_prim(prog->Geom.InputType);
1790 geom_array_resize_visitor input_resize_visitor(num_vertices, prog);
1791 foreach_in_list(ir_instruction, ir, linked->ir) {
1792 ir->accept(&input_resize_visitor);
1796 if (ctx->Const.VertexID_is_zero_based)
1797 lower_vertex_id(linked);
1799 /* Make a pass over all variable declarations to ensure that arrays with
1800 * unspecified sizes have a size specified. The size is inferred from the
1801 * max_array_access field.
1803 array_sizing_visitor v;
1805 v.fixup_unnamed_interface_types();
1811 * Update the sizes of linked shader uniform arrays to the maximum
1814 * From page 81 (page 95 of the PDF) of the OpenGL 2.1 spec:
1816 * If one or more elements of an array are active,
1817 * GetActiveUniform will return the name of the array in name,
1818 * subject to the restrictions listed above. The type of the array
1819 * is returned in type. The size parameter contains the highest
1820 * array element index used, plus one. The compiler or linker
1821 * determines the highest index used. There will be only one
1822 * active uniform reported by the GL per uniform array.
1826 update_array_sizes(struct gl_shader_program *prog)
1828 for (unsigned i = 0; i < MESA_SHADER_STAGES; i++) {
1829 if (prog->_LinkedShaders[i] == NULL)
1832 foreach_in_list(ir_instruction, node, prog->_LinkedShaders[i]->ir) {
1833 ir_variable *const var = node->as_variable();
1835 if ((var == NULL) || (var->data.mode != ir_var_uniform) ||
1836 !var->type->is_array())
1839 /* GL_ARB_uniform_buffer_object says that std140 uniforms
1840 * will not be eliminated. Since we always do std140, just
1841 * don't resize arrays in UBOs.
1843 * Atomic counters are supposed to get deterministic
1844 * locations assigned based on the declaration ordering and
1845 * sizes, array compaction would mess that up.
1847 if (var->is_in_uniform_block() || var->type->contains_atomic())
1850 unsigned int size = var->data.max_array_access;
1851 for (unsigned j = 0; j < MESA_SHADER_STAGES; j++) {
1852 if (prog->_LinkedShaders[j] == NULL)
1855 foreach_in_list(ir_instruction, node2, prog->_LinkedShaders[j]->ir) {
1856 ir_variable *other_var = node2->as_variable();
1860 if (strcmp(var->name, other_var->name) == 0 &&
1861 other_var->data.max_array_access > size) {
1862 size = other_var->data.max_array_access;
1867 if (size + 1 != var->type->length) {
1868 /* If this is a built-in uniform (i.e., it's backed by some
1869 * fixed-function state), adjust the number of state slots to
1870 * match the new array size. The number of slots per array entry
1871 * is not known. It seems safe to assume that the total number of
1872 * slots is an integer multiple of the number of array elements.
1873 * Determine the number of slots per array element by dividing by
1874 * the old (total) size.
1876 const unsigned num_slots = var->get_num_state_slots();
1877 if (num_slots > 0) {
1878 var->set_num_state_slots((size + 1)
1879 * (num_slots / var->type->length));
1882 var->type = glsl_type::get_array_instance(var->type->fields.array,
1884 /* FINISHME: We should update the types of array
1885 * dereferences of this variable now.
1893 * Find a contiguous set of available bits in a bitmask.
1895 * \param used_mask Bits representing used (1) and unused (0) locations
1896 * \param needed_count Number of contiguous bits needed.
1899 * Base location of the available bits on success or -1 on failure.
1902 find_available_slots(unsigned used_mask, unsigned needed_count)
1904 unsigned needed_mask = (1 << needed_count) - 1;
1905 const int max_bit_to_test = (8 * sizeof(used_mask)) - needed_count;
1907 /* The comparison to 32 is redundant, but without it GCC emits "warning:
1908 * cannot optimize possibly infinite loops" for the loop below.
1910 if ((needed_count == 0) || (max_bit_to_test < 0) || (max_bit_to_test > 32))
1913 for (int i = 0; i <= max_bit_to_test; i++) {
1914 if ((needed_mask & ~used_mask) == needed_mask)
1925 * Assign locations for either VS inputs or FS outputs
1927 * \param prog Shader program whose variables need locations assigned
1928 * \param target_index Selector for the program target to receive location
1929 * assignmnets. Must be either \c MESA_SHADER_VERTEX or
1930 * \c MESA_SHADER_FRAGMENT.
1931 * \param max_index Maximum number of generic locations. This corresponds
1932 * to either the maximum number of draw buffers or the
1933 * maximum number of generic attributes.
1936 * If locations are successfully assigned, true is returned. Otherwise an
1937 * error is emitted to the shader link log and false is returned.
1940 assign_attribute_or_color_locations(gl_shader_program *prog,
1941 unsigned target_index,
1944 /* Mark invalid locations as being used.
1946 unsigned used_locations = (max_index >= 32)
1947 ? ~0 : ~((1 << max_index) - 1);
1949 assert((target_index == MESA_SHADER_VERTEX)
1950 || (target_index == MESA_SHADER_FRAGMENT));
1952 gl_shader *const sh = prog->_LinkedShaders[target_index];
1956 /* Operate in a total of four passes.
1958 * 1. Invalidate the location assignments for all vertex shader inputs.
1960 * 2. Assign locations for inputs that have user-defined (via
1961 * glBindVertexAttribLocation) locations and outputs that have
1962 * user-defined locations (via glBindFragDataLocation).
1964 * 3. Sort the attributes without assigned locations by number of slots
1965 * required in decreasing order. Fragmentation caused by attribute
1966 * locations assigned by the application may prevent large attributes
1967 * from having enough contiguous space.
1969 * 4. Assign locations to any inputs without assigned locations.
1972 const int generic_base = (target_index == MESA_SHADER_VERTEX)
1973 ? (int) VERT_ATTRIB_GENERIC0 : (int) FRAG_RESULT_DATA0;
1975 const enum ir_variable_mode direction =
1976 (target_index == MESA_SHADER_VERTEX)
1977 ? ir_var_shader_in : ir_var_shader_out;
1980 /* Temporary storage for the set of attributes that need locations assigned.
1986 /* Used below in the call to qsort. */
1987 static int compare(const void *a, const void *b)
1989 const temp_attr *const l = (const temp_attr *) a;
1990 const temp_attr *const r = (const temp_attr *) b;
1992 /* Reversed because we want a descending order sort below. */
1993 return r->slots - l->slots;
1997 unsigned num_attr = 0;
1999 foreach_in_list(ir_instruction, node, sh->ir) {
2000 ir_variable *const var = node->as_variable();
2002 if ((var == NULL) || (var->data.mode != (unsigned) direction))
2005 if (var->data.explicit_location) {
2006 if ((var->data.location >= (int)(max_index + generic_base))
2007 || (var->data.location < 0)) {
2009 "invalid explicit location %d specified for `%s'\n",
2010 (var->data.location < 0)
2011 ? var->data.location
2012 : var->data.location - generic_base,
2016 } else if (target_index == MESA_SHADER_VERTEX) {
2019 if (prog->AttributeBindings->get(binding, var->name)) {
2020 assert(binding >= VERT_ATTRIB_GENERIC0);
2021 var->data.location = binding;
2022 var->data.is_unmatched_generic_inout = 0;
2024 } else if (target_index == MESA_SHADER_FRAGMENT) {
2028 if (prog->FragDataBindings->get(binding, var->name)) {
2029 assert(binding >= FRAG_RESULT_DATA0);
2030 var->data.location = binding;
2031 var->data.is_unmatched_generic_inout = 0;
2033 if (prog->FragDataIndexBindings->get(index, var->name)) {
2034 var->data.index = index;
2039 /* If the variable is not a built-in and has a location statically
2040 * assigned in the shader (presumably via a layout qualifier), make sure
2041 * that it doesn't collide with other assigned locations. Otherwise,
2042 * add it to the list of variables that need linker-assigned locations.
2044 const unsigned slots = var->type->count_attribute_slots();
2045 if (var->data.location != -1) {
2046 if (var->data.location >= generic_base && var->data.index < 1) {
2047 /* From page 61 of the OpenGL 4.0 spec:
2049 * "LinkProgram will fail if the attribute bindings assigned
2050 * by BindAttribLocation do not leave not enough space to
2051 * assign a location for an active matrix attribute or an
2052 * active attribute array, both of which require multiple
2053 * contiguous generic attributes."
2055 * I think above text prohibits the aliasing of explicit and
2056 * automatic assignments. But, aliasing is allowed in manual
2057 * assignments of attribute locations. See below comments for
2060 * From OpenGL 4.0 spec, page 61:
2062 * "It is possible for an application to bind more than one
2063 * attribute name to the same location. This is referred to as
2064 * aliasing. This will only work if only one of the aliased
2065 * attributes is active in the executable program, or if no
2066 * path through the shader consumes more than one attribute of
2067 * a set of attributes aliased to the same location. A link
2068 * error can occur if the linker determines that every path
2069 * through the shader consumes multiple aliased attributes,
2070 * but implementations are not required to generate an error
2073 * From GLSL 4.30 spec, page 54:
2075 * "A program will fail to link if any two non-vertex shader
2076 * input variables are assigned to the same location. For
2077 * vertex shaders, multiple input variables may be assigned
2078 * to the same location using either layout qualifiers or via
2079 * the OpenGL API. However, such aliasing is intended only to
2080 * support vertex shaders where each execution path accesses
2081 * at most one input per each location. Implementations are
2082 * permitted, but not required, to generate link-time errors
2083 * if they detect that every path through the vertex shader
2084 * executable accesses multiple inputs assigned to any single
2085 * location. For all shader types, a program will fail to link
2086 * if explicit location assignments leave the linker unable
2087 * to find space for other variables without explicit
2090 * From OpenGL ES 3.0 spec, page 56:
2092 * "Binding more than one attribute name to the same location
2093 * is referred to as aliasing, and is not permitted in OpenGL
2094 * ES Shading Language 3.00 vertex shaders. LinkProgram will
2095 * fail when this condition exists. However, aliasing is
2096 * possible in OpenGL ES Shading Language 1.00 vertex shaders.
2097 * This will only work if only one of the aliased attributes
2098 * is active in the executable program, or if no path through
2099 * the shader consumes more than one attribute of a set of
2100 * attributes aliased to the same location. A link error can
2101 * occur if the linker determines that every path through the
2102 * shader consumes multiple aliased attributes, but implemen-
2103 * tations are not required to generate an error in this case."
2105 * After looking at above references from OpenGL, OpenGL ES and
2106 * GLSL specifications, we allow aliasing of vertex input variables
2107 * in: OpenGL 2.0 (and above) and OpenGL ES 2.0.
2109 * NOTE: This is not required by the spec but its worth mentioning
2110 * here that we're not doing anything to make sure that no path
2111 * through the vertex shader executable accesses multiple inputs
2112 * assigned to any single location.
2115 /* Mask representing the contiguous slots that will be used by
2118 const unsigned attr = var->data.location - generic_base;
2119 const unsigned use_mask = (1 << slots) - 1;
2120 const char *const string = (target_index == MESA_SHADER_VERTEX)
2121 ? "vertex shader input" : "fragment shader output";
2123 /* Generate a link error if the requested locations for this
2124 * attribute exceed the maximum allowed attribute location.
2126 if (attr + slots > max_index) {
2128 "insufficient contiguous locations "
2129 "available for %s `%s' %d %d %d\n", string,
2130 var->name, used_locations, use_mask, attr);
2134 /* Generate a link error if the set of bits requested for this
2135 * attribute overlaps any previously allocated bits.
2137 if ((~(use_mask << attr) & used_locations) != used_locations) {
2138 if (target_index == MESA_SHADER_FRAGMENT ||
2139 (prog->IsES && prog->Version >= 300)) {
2141 "overlapping location is assigned "
2142 "to %s `%s' %d %d %d\n", string,
2143 var->name, used_locations, use_mask, attr);
2146 linker_warning(prog,
2147 "overlapping location is assigned "
2148 "to %s `%s' %d %d %d\n", string,
2149 var->name, used_locations, use_mask, attr);
2153 used_locations |= (use_mask << attr);
2159 to_assign[num_attr].slots = slots;
2160 to_assign[num_attr].var = var;
2164 /* If all of the attributes were assigned locations by the application (or
2165 * are built-in attributes with fixed locations), return early. This should
2166 * be the common case.
2171 qsort(to_assign, num_attr, sizeof(to_assign[0]), temp_attr::compare);
2173 if (target_index == MESA_SHADER_VERTEX) {
2174 /* VERT_ATTRIB_GENERIC0 is a pseudo-alias for VERT_ATTRIB_POS. It can
2175 * only be explicitly assigned by via glBindAttribLocation. Mark it as
2176 * reserved to prevent it from being automatically allocated below.
2178 find_deref_visitor find("gl_Vertex");
2180 if (find.variable_found())
2181 used_locations |= (1 << 0);
2184 for (unsigned i = 0; i < num_attr; i++) {
2185 /* Mask representing the contiguous slots that will be used by this
2188 const unsigned use_mask = (1 << to_assign[i].slots) - 1;
2190 int location = find_available_slots(used_locations, to_assign[i].slots);
2193 const char *const string = (target_index == MESA_SHADER_VERTEX)
2194 ? "vertex shader input" : "fragment shader output";
2197 "insufficient contiguous locations "
2198 "available for %s `%s'\n",
2199 string, to_assign[i].var->name);
2203 to_assign[i].var->data.location = generic_base + location;
2204 to_assign[i].var->data.is_unmatched_generic_inout = 0;
2205 used_locations |= (use_mask << location);
2213 * Demote shader inputs and outputs that are not used in other stages
2216 demote_shader_inputs_and_outputs(gl_shader *sh, enum ir_variable_mode mode)
2218 foreach_in_list(ir_instruction, node, sh->ir) {
2219 ir_variable *const var = node->as_variable();
2221 if ((var == NULL) || (var->data.mode != int(mode)))
2224 /* A shader 'in' or 'out' variable is only really an input or output if
2225 * its value is used by other shader stages. This will cause the variable
2226 * to have a location assigned.
2228 if (var->data.is_unmatched_generic_inout) {
2229 assert(var->data.mode != ir_var_temporary);
2230 var->data.mode = ir_var_auto;
2237 * Store the gl_FragDepth layout in the gl_shader_program struct.
2240 store_fragdepth_layout(struct gl_shader_program *prog)
2242 if (prog->_LinkedShaders[MESA_SHADER_FRAGMENT] == NULL) {
2246 struct exec_list *ir = prog->_LinkedShaders[MESA_SHADER_FRAGMENT]->ir;
2248 /* We don't look up the gl_FragDepth symbol directly because if
2249 * gl_FragDepth is not used in the shader, it's removed from the IR.
2250 * However, the symbol won't be removed from the symbol table.
2252 * We're only interested in the cases where the variable is NOT removed
2255 foreach_in_list(ir_instruction, node, ir) {
2256 ir_variable *const var = node->as_variable();
2258 if (var == NULL || var->data.mode != ir_var_shader_out) {
2262 if (strcmp(var->name, "gl_FragDepth") == 0) {
2263 switch (var->data.depth_layout) {
2264 case ir_depth_layout_none:
2265 prog->FragDepthLayout = FRAG_DEPTH_LAYOUT_NONE;
2267 case ir_depth_layout_any:
2268 prog->FragDepthLayout = FRAG_DEPTH_LAYOUT_ANY;
2270 case ir_depth_layout_greater:
2271 prog->FragDepthLayout = FRAG_DEPTH_LAYOUT_GREATER;
2273 case ir_depth_layout_less:
2274 prog->FragDepthLayout = FRAG_DEPTH_LAYOUT_LESS;
2276 case ir_depth_layout_unchanged:
2277 prog->FragDepthLayout = FRAG_DEPTH_LAYOUT_UNCHANGED;
2288 * Validate the resources used by a program versus the implementation limits
2291 check_resources(struct gl_context *ctx, struct gl_shader_program *prog)
2293 for (unsigned i = 0; i < MESA_SHADER_STAGES; i++) {
2294 struct gl_shader *sh = prog->_LinkedShaders[i];
2299 if (sh->num_samplers > ctx->Const.Program[i].MaxTextureImageUnits) {
2300 linker_error(prog, "Too many %s shader texture samplers\n",
2301 _mesa_shader_stage_to_string(i));
2304 if (sh->num_uniform_components >
2305 ctx->Const.Program[i].MaxUniformComponents) {
2306 if (ctx->Const.GLSLSkipStrictMaxUniformLimitCheck) {
2307 linker_warning(prog, "Too many %s shader default uniform block "
2308 "components, but the driver will try to optimize "
2309 "them out; this is non-portable out-of-spec "
2311 _mesa_shader_stage_to_string(i));
2313 linker_error(prog, "Too many %s shader default uniform block "
2315 _mesa_shader_stage_to_string(i));
2319 if (sh->num_combined_uniform_components >
2320 ctx->Const.Program[i].MaxCombinedUniformComponents) {
2321 if (ctx->Const.GLSLSkipStrictMaxUniformLimitCheck) {
2322 linker_warning(prog, "Too many %s shader uniform components, "
2323 "but the driver will try to optimize them out; "
2324 "this is non-portable out-of-spec behavior\n",
2325 _mesa_shader_stage_to_string(i));
2327 linker_error(prog, "Too many %s shader uniform components\n",
2328 _mesa_shader_stage_to_string(i));
2333 unsigned blocks[MESA_SHADER_STAGES] = {0};
2334 unsigned total_uniform_blocks = 0;
2336 for (unsigned i = 0; i < prog->NumUniformBlocks; i++) {
2337 for (unsigned j = 0; j < MESA_SHADER_STAGES; j++) {
2338 if (prog->UniformBlockStageIndex[j][i] != -1) {
2340 total_uniform_blocks++;
2344 if (total_uniform_blocks > ctx->Const.MaxCombinedUniformBlocks) {
2345 linker_error(prog, "Too many combined uniform blocks (%d/%d)\n",
2346 prog->NumUniformBlocks,
2347 ctx->Const.MaxCombinedUniformBlocks);
2349 for (unsigned i = 0; i < MESA_SHADER_STAGES; i++) {
2350 const unsigned max_uniform_blocks =
2351 ctx->Const.Program[i].MaxUniformBlocks;
2352 if (blocks[i] > max_uniform_blocks) {
2353 linker_error(prog, "Too many %s uniform blocks (%d/%d)\n",
2354 _mesa_shader_stage_to_string(i),
2356 max_uniform_blocks);
2365 * Validate shader image resources.
2368 check_image_resources(struct gl_context *ctx, struct gl_shader_program *prog)
2370 unsigned total_image_units = 0;
2371 unsigned fragment_outputs = 0;
2373 if (!ctx->Extensions.ARB_shader_image_load_store)
2376 for (unsigned i = 0; i < MESA_SHADER_STAGES; i++) {
2377 struct gl_shader *sh = prog->_LinkedShaders[i];
2380 if (sh->NumImages > ctx->Const.Program[i].MaxImageUniforms)
2381 linker_error(prog, "Too many %s shader image uniforms\n",
2382 _mesa_shader_stage_to_string(i));
2384 total_image_units += sh->NumImages;
2386 if (i == MESA_SHADER_FRAGMENT) {
2387 foreach_in_list(ir_instruction, node, sh->ir) {
2388 ir_variable *var = node->as_variable();
2389 if (var && var->data.mode == ir_var_shader_out)
2390 fragment_outputs += var->type->count_attribute_slots();
2396 if (total_image_units > ctx->Const.MaxCombinedImageUniforms)
2397 linker_error(prog, "Too many combined image uniforms\n");
2399 if (total_image_units + fragment_outputs >
2400 ctx->Const.MaxCombinedImageUnitsAndFragmentOutputs)
2401 linker_error(prog, "Too many combined image uniforms and fragment outputs\n");
2406 * Initializes explicit location slots to INACTIVE_UNIFORM_EXPLICIT_LOCATION
2407 * for a variable, checks for overlaps between other uniforms using explicit
2411 reserve_explicit_locations(struct gl_shader_program *prog,
2412 string_to_uint_map *map, ir_variable *var)
2414 unsigned slots = var->type->uniform_locations();
2415 unsigned max_loc = var->data.location + slots - 1;
2417 /* Resize remap table if locations do not fit in the current one. */
2418 if (max_loc + 1 > prog->NumUniformRemapTable) {
2419 prog->UniformRemapTable =
2420 reralloc(prog, prog->UniformRemapTable,
2421 gl_uniform_storage *,
2424 if (!prog->UniformRemapTable) {
2425 linker_error(prog, "Out of memory during linking.\n");
2429 /* Initialize allocated space. */
2430 for (unsigned i = prog->NumUniformRemapTable; i < max_loc + 1; i++)
2431 prog->UniformRemapTable[i] = NULL;
2433 prog->NumUniformRemapTable = max_loc + 1;
2436 for (unsigned i = 0; i < slots; i++) {
2437 unsigned loc = var->data.location + i;
2439 /* Check if location is already used. */
2440 if (prog->UniformRemapTable[loc] == INACTIVE_UNIFORM_EXPLICIT_LOCATION) {
2442 /* Possibly same uniform from a different stage, this is ok. */
2444 if (map->get(hash_loc, var->name) && hash_loc == loc - i)
2447 /* ARB_explicit_uniform_location specification states:
2449 * "No two default-block uniform variables in the program can have
2450 * the same location, even if they are unused, otherwise a compiler
2451 * or linker error will be generated."
2454 "location qualifier for uniform %s overlaps "
2455 "previously used location\n",
2460 /* Initialize location as inactive before optimization
2461 * rounds and location assignment.
2463 prog->UniformRemapTable[loc] = INACTIVE_UNIFORM_EXPLICIT_LOCATION;
2466 /* Note, base location used for arrays. */
2467 map->put(var->data.location, var->name);
2473 * Check and reserve all explicit uniform locations, called before
2474 * any optimizations happen to handle also inactive uniforms and
2475 * inactive array elements that may get trimmed away.
2478 check_explicit_uniform_locations(struct gl_context *ctx,
2479 struct gl_shader_program *prog)
2481 if (!ctx->Extensions.ARB_explicit_uniform_location)
2484 /* This map is used to detect if overlapping explicit locations
2485 * occur with the same uniform (from different stage) or a different one.
2487 string_to_uint_map *uniform_map = new string_to_uint_map;
2490 linker_error(prog, "Out of memory during linking.\n");
2494 for (unsigned i = 0; i < MESA_SHADER_STAGES; i++) {
2495 struct gl_shader *sh = prog->_LinkedShaders[i];
2500 foreach_in_list(ir_instruction, node, sh->ir) {
2501 ir_variable *var = node->as_variable();
2502 if ((var && var->data.mode == ir_var_uniform) &&
2503 var->data.explicit_location) {
2504 if (!reserve_explicit_locations(prog, uniform_map, var)) {
2516 * This check is done to make sure we allow only constant expression
2517 * indexing and "constant-index-expression" (indexing with an expression
2518 * that includes loop induction variable).
2521 validate_sampler_array_indexing(struct gl_context *ctx,
2522 struct gl_shader_program *prog)
2524 dynamic_sampler_array_indexing_visitor v;
2525 for (unsigned i = 0; i < MESA_SHADER_STAGES; i++) {
2526 if (prog->_LinkedShaders[i] == NULL)
2529 bool no_dynamic_indexing =
2530 ctx->Const.ShaderCompilerOptions[i].EmitNoIndirectSampler;
2532 /* Search for array derefs in shader. */
2533 v.run(prog->_LinkedShaders[i]->ir);
2534 if (v.uses_dynamic_sampler_array_indexing()) {
2535 const char *msg = "sampler arrays indexed with non-constant "
2536 "expressions is forbidden in GLSL %s %u";
2537 /* Backend has indicated that it has no dynamic indexing support. */
2538 if (no_dynamic_indexing) {
2539 linker_error(prog, msg, prog->IsES ? "ES" : "", prog->Version);
2542 linker_warning(prog, msg, prog->IsES ? "ES" : "", prog->Version);
2551 link_shaders(struct gl_context *ctx, struct gl_shader_program *prog)
2553 tfeedback_decl *tfeedback_decls = NULL;
2554 unsigned num_tfeedback_decls = prog->TransformFeedback.NumVarying;
2556 void *mem_ctx = ralloc_context(NULL); // temporary linker context
2558 prog->LinkStatus = true; /* All error paths will set this to false */
2559 prog->Validated = false;
2560 prog->_Used = false;
2562 prog->ARB_fragment_coord_conventions_enable = false;
2564 /* Separate the shaders into groups based on their type.
2566 struct gl_shader **shader_list[MESA_SHADER_STAGES];
2567 unsigned num_shaders[MESA_SHADER_STAGES];
2569 for (int i = 0; i < MESA_SHADER_STAGES; i++) {
2570 shader_list[i] = (struct gl_shader **)
2571 calloc(prog->NumShaders, sizeof(struct gl_shader *));
2575 unsigned min_version = UINT_MAX;
2576 unsigned max_version = 0;
2577 const bool is_es_prog =
2578 (prog->NumShaders > 0 && prog->Shaders[0]->IsES) ? true : false;
2579 for (unsigned i = 0; i < prog->NumShaders; i++) {
2580 min_version = MIN2(min_version, prog->Shaders[i]->Version);
2581 max_version = MAX2(max_version, prog->Shaders[i]->Version);
2583 if (prog->Shaders[i]->IsES != is_es_prog) {
2584 linker_error(prog, "all shaders must use same shading "
2585 "language version\n");
2589 prog->ARB_fragment_coord_conventions_enable |=
2590 prog->Shaders[i]->ARB_fragment_coord_conventions_enable;
2592 gl_shader_stage shader_type = prog->Shaders[i]->Stage;
2593 shader_list[shader_type][num_shaders[shader_type]] = prog->Shaders[i];
2594 num_shaders[shader_type]++;
2597 /* In desktop GLSL, different shader versions may be linked together. In
2598 * GLSL ES, all shader versions must be the same.
2600 if (is_es_prog && min_version != max_version) {
2601 linker_error(prog, "all shaders must use same shading "
2602 "language version\n");
2606 prog->Version = max_version;
2607 prog->IsES = is_es_prog;
2609 /* Geometry shaders have to be linked with vertex shaders.
2611 if (num_shaders[MESA_SHADER_GEOMETRY] > 0 &&
2612 num_shaders[MESA_SHADER_VERTEX] == 0 &&
2613 !prog->SeparateShader) {
2614 linker_error(prog, "Geometry shader must be linked with "
2619 /* Compute shaders have additional restrictions. */
2620 if (num_shaders[MESA_SHADER_COMPUTE] > 0 &&
2621 num_shaders[MESA_SHADER_COMPUTE] != prog->NumShaders) {
2622 linker_error(prog, "Compute shaders may not be linked with any other "
2623 "type of shader\n");
2626 for (unsigned int i = 0; i < MESA_SHADER_STAGES; i++) {
2627 if (prog->_LinkedShaders[i] != NULL)
2628 ctx->Driver.DeleteShader(ctx, prog->_LinkedShaders[i]);
2630 prog->_LinkedShaders[i] = NULL;
2633 /* Link all shaders for a particular stage and validate the result.
2635 for (int stage = 0; stage < MESA_SHADER_STAGES; stage++) {
2636 if (num_shaders[stage] > 0) {
2637 gl_shader *const sh =
2638 link_intrastage_shaders(mem_ctx, ctx, prog, shader_list[stage],
2639 num_shaders[stage]);
2641 if (!prog->LinkStatus) {
2643 ctx->Driver.DeleteShader(ctx, sh);
2648 case MESA_SHADER_VERTEX:
2649 validate_vertex_shader_executable(prog, sh);
2651 case MESA_SHADER_GEOMETRY:
2652 validate_geometry_shader_executable(prog, sh);
2654 case MESA_SHADER_FRAGMENT:
2655 validate_fragment_shader_executable(prog, sh);
2658 if (!prog->LinkStatus) {
2660 ctx->Driver.DeleteShader(ctx, sh);
2664 _mesa_reference_shader(ctx, &prog->_LinkedShaders[stage], sh);
2668 if (num_shaders[MESA_SHADER_GEOMETRY] > 0)
2669 prog->LastClipDistanceArraySize = prog->Geom.ClipDistanceArraySize;
2670 else if (num_shaders[MESA_SHADER_VERTEX] > 0)
2671 prog->LastClipDistanceArraySize = prog->Vert.ClipDistanceArraySize;
2673 prog->LastClipDistanceArraySize = 0; /* Not used */
2675 /* Here begins the inter-stage linking phase. Some initial validation is
2676 * performed, then locations are assigned for uniforms, attributes, and
2679 cross_validate_uniforms(prog);
2680 if (!prog->LinkStatus)
2685 for (prev = 0; prev <= MESA_SHADER_FRAGMENT; prev++) {
2686 if (prog->_LinkedShaders[prev] != NULL)
2690 check_explicit_uniform_locations(ctx, prog);
2691 if (!prog->LinkStatus)
2694 /* Validate the inputs of each stage with the output of the preceding
2697 for (unsigned i = prev + 1; i <= MESA_SHADER_FRAGMENT; i++) {
2698 if (prog->_LinkedShaders[i] == NULL)
2701 validate_interstage_inout_blocks(prog, prog->_LinkedShaders[prev],
2702 prog->_LinkedShaders[i]);
2703 if (!prog->LinkStatus)
2706 cross_validate_outputs_to_inputs(prog,
2707 prog->_LinkedShaders[prev],
2708 prog->_LinkedShaders[i]);
2709 if (!prog->LinkStatus)
2715 /* Cross-validate uniform blocks between shader stages */
2716 validate_interstage_uniform_blocks(prog, prog->_LinkedShaders,
2717 MESA_SHADER_STAGES);
2718 if (!prog->LinkStatus)
2721 for (unsigned int i = 0; i < MESA_SHADER_STAGES; i++) {
2722 if (prog->_LinkedShaders[i] != NULL)
2723 lower_named_interface_blocks(mem_ctx, prog->_LinkedShaders[i]);
2726 /* Implement the GLSL 1.30+ rule for discard vs infinite loops Do
2727 * it before optimization because we want most of the checks to get
2728 * dropped thanks to constant propagation.
2730 * This rule also applies to GLSL ES 3.00.
2732 if (max_version >= (is_es_prog ? 300 : 130)) {
2733 struct gl_shader *sh = prog->_LinkedShaders[MESA_SHADER_FRAGMENT];
2735 lower_discard_flow(sh->ir);
2739 if (!interstage_cross_validate_uniform_blocks(prog))
2742 /* Do common optimization before assigning storage for attributes,
2743 * uniforms, and varyings. Later optimization could possibly make
2744 * some of that unused.
2746 for (unsigned i = 0; i < MESA_SHADER_STAGES; i++) {
2747 if (prog->_LinkedShaders[i] == NULL)
2750 detect_recursion_linked(prog, prog->_LinkedShaders[i]->ir);
2751 if (!prog->LinkStatus)
2754 if (ctx->Const.ShaderCompilerOptions[i].LowerClipDistance) {
2755 lower_clip_distance(prog->_LinkedShaders[i]);
2758 while (do_common_optimization(prog->_LinkedShaders[i]->ir, true, false,
2759 &ctx->Const.ShaderCompilerOptions[i],
2760 ctx->Const.NativeIntegers))
2763 lower_const_arrays_to_uniforms(prog->_LinkedShaders[i]->ir);
2766 /* Validation for special cases where we allow sampler array indexing
2767 * with loop induction variable. This check emits a warning or error
2768 * depending if backend can handle dynamic indexing.
2770 if ((!prog->IsES && prog->Version < 130) ||
2771 (prog->IsES && prog->Version < 300)) {
2772 if (!validate_sampler_array_indexing(ctx, prog))
2776 /* Check and validate stream emissions in geometry shaders */
2777 validate_geometry_shader_emissions(ctx, prog);
2779 /* Mark all generic shader inputs and outputs as unpaired. */
2780 for (unsigned i = MESA_SHADER_VERTEX; i <= MESA_SHADER_FRAGMENT; i++) {
2781 if (prog->_LinkedShaders[i] != NULL) {
2782 link_invalidate_variable_locations(prog->_LinkedShaders[i]->ir);
2786 /* FINISHME: The value of the max_attribute_index parameter is
2787 * FINISHME: implementation dependent based on the value of
2788 * FINISHME: GL_MAX_VERTEX_ATTRIBS. GL_MAX_VERTEX_ATTRIBS must be
2789 * FINISHME: at least 16, so hardcode 16 for now.
2791 if (!assign_attribute_or_color_locations(prog, MESA_SHADER_VERTEX, 16)) {
2795 if (!assign_attribute_or_color_locations(prog, MESA_SHADER_FRAGMENT, MAX2(ctx->Const.MaxDrawBuffers, ctx->Const.MaxDualSourceDrawBuffers))) {
2800 for (first = 0; first <= MESA_SHADER_FRAGMENT; first++) {
2801 if (prog->_LinkedShaders[first] != NULL)
2805 if (num_tfeedback_decls != 0) {
2806 /* From GL_EXT_transform_feedback:
2807 * A program will fail to link if:
2809 * * the <count> specified by TransformFeedbackVaryingsEXT is
2810 * non-zero, but the program object has no vertex or geometry
2813 if (first == MESA_SHADER_FRAGMENT) {
2814 linker_error(prog, "Transform feedback varyings specified, but "
2815 "no vertex or geometry shader is present.\n");
2819 tfeedback_decls = ralloc_array(mem_ctx, tfeedback_decl,
2820 prog->TransformFeedback.NumVarying);
2821 if (!parse_tfeedback_decls(ctx, prog, mem_ctx, num_tfeedback_decls,
2822 prog->TransformFeedback.VaryingNames,
2827 /* Linking the stages in the opposite order (from fragment to vertex)
2828 * ensures that inter-shader outputs written to in an earlier stage are
2829 * eliminated if they are (transitively) not used in a later stage.
2832 for (last = MESA_SHADER_FRAGMENT; last >= 0; last--) {
2833 if (prog->_LinkedShaders[last] != NULL)
2837 if (last >= 0 && last < MESA_SHADER_FRAGMENT) {
2838 gl_shader *const sh = prog->_LinkedShaders[last];
2840 if (first == MESA_SHADER_GEOMETRY) {
2841 /* There was no vertex shader, but we still have to assign varying
2842 * locations for use by geometry shader inputs in SSO.
2844 * If the shader is not separable (i.e., prog->SeparateShader is
2845 * false), linking will have already failed when first is
2846 * MESA_SHADER_GEOMETRY.
2848 if (!assign_varying_locations(ctx, mem_ctx, prog,
2850 num_tfeedback_decls, tfeedback_decls,
2851 prog->Geom.VerticesIn))
2855 if (num_tfeedback_decls != 0 || prog->SeparateShader) {
2856 /* There was no fragment shader, but we still have to assign varying
2857 * locations for use by transform feedback.
2859 if (!assign_varying_locations(ctx, mem_ctx, prog,
2861 num_tfeedback_decls, tfeedback_decls,
2866 do_dead_builtin_varyings(ctx, sh, NULL,
2867 num_tfeedback_decls, tfeedback_decls);
2869 if (!prog->SeparateShader)
2870 demote_shader_inputs_and_outputs(sh, ir_var_shader_out);
2872 /* Eliminate code that is now dead due to unused outputs being demoted.
2874 while (do_dead_code(sh->ir, false))
2877 else if (first == MESA_SHADER_FRAGMENT) {
2878 /* If the program only contains a fragment shader...
2880 gl_shader *const sh = prog->_LinkedShaders[first];
2882 do_dead_builtin_varyings(ctx, NULL, sh,
2883 num_tfeedback_decls, tfeedback_decls);
2885 if (prog->SeparateShader) {
2886 if (!assign_varying_locations(ctx, mem_ctx, prog,
2887 NULL /* producer */,
2889 0 /* num_tfeedback_decls */,
2890 NULL /* tfeedback_decls */,
2891 0 /* gs_input_vertices */))
2894 demote_shader_inputs_and_outputs(sh, ir_var_shader_in);
2896 while (do_dead_code(sh->ir, false))
2901 for (int i = next - 1; i >= 0; i--) {
2902 if (prog->_LinkedShaders[i] == NULL)
2905 gl_shader *const sh_i = prog->_LinkedShaders[i];
2906 gl_shader *const sh_next = prog->_LinkedShaders[next];
2907 unsigned gs_input_vertices =
2908 next == MESA_SHADER_GEOMETRY ? prog->Geom.VerticesIn : 0;
2910 if (!assign_varying_locations(ctx, mem_ctx, prog, sh_i, sh_next,
2911 next == MESA_SHADER_FRAGMENT ? num_tfeedback_decls : 0,
2912 tfeedback_decls, gs_input_vertices))
2915 do_dead_builtin_varyings(ctx, sh_i, sh_next,
2916 next == MESA_SHADER_FRAGMENT ? num_tfeedback_decls : 0,
2919 demote_shader_inputs_and_outputs(sh_i, ir_var_shader_out);
2920 demote_shader_inputs_and_outputs(sh_next, ir_var_shader_in);
2922 /* Eliminate code that is now dead due to unused outputs being demoted.
2924 while (do_dead_code(sh_i->ir, false))
2926 while (do_dead_code(sh_next->ir, false))
2929 /* This must be done after all dead varyings are eliminated. */
2930 if (!check_against_output_limit(ctx, prog, sh_i))
2932 if (!check_against_input_limit(ctx, prog, sh_next))
2938 if (!store_tfeedback_info(ctx, prog, num_tfeedback_decls, tfeedback_decls))
2941 update_array_sizes(prog);
2942 link_assign_uniform_locations(prog, ctx->Const.UniformBooleanTrue);
2943 link_assign_atomic_counter_resources(ctx, prog);
2944 store_fragdepth_layout(prog);
2946 check_resources(ctx, prog);
2947 check_image_resources(ctx, prog);
2948 link_check_atomic_counter_resources(ctx, prog);
2950 if (!prog->LinkStatus)
2953 /* OpenGL ES requires that a vertex shader and a fragment shader both be
2954 * present in a linked program. GL_ARB_ES2_compatibility doesn't say
2955 * anything about shader linking when one of the shaders (vertex or
2956 * fragment shader) is absent. So, the extension shouldn't change the
2957 * behavior specified in GLSL specification.
2959 if (!prog->SeparateShader && ctx->API == API_OPENGLES2) {
2960 if (prog->_LinkedShaders[MESA_SHADER_VERTEX] == NULL) {
2961 linker_error(prog, "program lacks a vertex shader\n");
2962 } else if (prog->_LinkedShaders[MESA_SHADER_FRAGMENT] == NULL) {
2963 linker_error(prog, "program lacks a fragment shader\n");
2967 /* FINISHME: Assign fragment shader output locations. */
2970 for (unsigned i = 0; i < MESA_SHADER_STAGES; i++) {
2971 free(shader_list[i]);
2972 if (prog->_LinkedShaders[i] == NULL)
2975 /* Do a final validation step to make sure that the IR wasn't
2976 * invalidated by any modifications performed after intrastage linking.
2978 validate_ir_tree(prog->_LinkedShaders[i]->ir);
2980 /* Retain any live IR, but trash the rest. */
2981 reparent_ir(prog->_LinkedShaders[i]->ir, prog->_LinkedShaders[i]->ir);
2983 /* The symbol table in the linked shaders may contain references to
2984 * variables that were removed (e.g., unused uniforms). Since it may
2985 * contain junk, there is no possible valid use. Delete it and set the
2988 delete prog->_LinkedShaders[i]->symbols;
2989 prog->_LinkedShaders[i]->symbols = NULL;
2992 ralloc_free(mem_ctx);