glsl: mark special built-in inputs referenced by vertex stage
[mesa.git] / src / glsl / linker.cpp
1 /*
2 * Copyright © 2010 Intel Corporation
3 *
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:
10 *
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
13 * Software.
14 *
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.
22 */
23
24 /**
25 * \file linker.cpp
26 * GLSL linker implementation
27 *
28 * Given a set of shaders that are to be linked to generate a final program,
29 * there are three distinct stages.
30 *
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
33 * together.
34 *
35 * - Undefined references in each shader are resolve to definitions in
36 * another shader.
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.
41 *
42 * The result, in the terminology of the GLSL spec, is a set of shader
43 * executables for each processing unit.
44 *
45 * After the first stage is complete, a series of semantic checks are performed
46 * on each of the shader executables.
47 *
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
51 * \c gl_FragColor.
52 *
53 * In the final stage individual shader executables are linked to create a
54 * complete exectuable.
55 *
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.
63 *
64 * \author Ian Romanick <ian.d.romanick@intel.com>
65 */
66
67 #include <ctype.h>
68 #include "main/core.h"
69 #include "glsl_symbol_table.h"
70 #include "glsl_parser_extras.h"
71 #include "ir.h"
72 #include "program.h"
73 #include "program/hash_table.h"
74 #include "linker.h"
75 #include "link_varyings.h"
76 #include "ir_optimization.h"
77 #include "ir_rvalue_visitor.h"
78 #include "ir_uniform.h"
79
80 #include "main/shaderobj.h"
81 #include "main/enums.h"
82
83
84 void linker_error(gl_shader_program *, const char *, ...);
85
86 namespace {
87
88 /**
89 * Visitor that determines whether or not a variable is ever written.
90 */
91 class find_assignment_visitor : public ir_hierarchical_visitor {
92 public:
93 find_assignment_visitor(const char *name)
94 : name(name), found(false)
95 {
96 /* empty */
97 }
98
99 virtual ir_visitor_status visit_enter(ir_assignment *ir)
100 {
101 ir_variable *const var = ir->lhs->variable_referenced();
102
103 if (strcmp(name, var->name) == 0) {
104 found = true;
105 return visit_stop;
106 }
107
108 return visit_continue_with_parent;
109 }
110
111 virtual ir_visitor_status visit_enter(ir_call *ir)
112 {
113 foreach_two_lists(formal_node, &ir->callee->parameters,
114 actual_node, &ir->actual_parameters) {
115 ir_rvalue *param_rval = (ir_rvalue *) actual_node;
116 ir_variable *sig_param = (ir_variable *) formal_node;
117
118 if (sig_param->data.mode == ir_var_function_out ||
119 sig_param->data.mode == ir_var_function_inout) {
120 ir_variable *var = param_rval->variable_referenced();
121 if (var && strcmp(name, var->name) == 0) {
122 found = true;
123 return visit_stop;
124 }
125 }
126 }
127
128 if (ir->return_deref != NULL) {
129 ir_variable *const var = ir->return_deref->variable_referenced();
130
131 if (strcmp(name, var->name) == 0) {
132 found = true;
133 return visit_stop;
134 }
135 }
136
137 return visit_continue_with_parent;
138 }
139
140 bool variable_found()
141 {
142 return found;
143 }
144
145 private:
146 const char *name; /**< Find writes to a variable with this name. */
147 bool found; /**< Was a write to the variable found? */
148 };
149
150
151 /**
152 * Visitor that determines whether or not a variable is ever read.
153 */
154 class find_deref_visitor : public ir_hierarchical_visitor {
155 public:
156 find_deref_visitor(const char *name)
157 : name(name), found(false)
158 {
159 /* empty */
160 }
161
162 virtual ir_visitor_status visit(ir_dereference_variable *ir)
163 {
164 if (strcmp(this->name, ir->var->name) == 0) {
165 this->found = true;
166 return visit_stop;
167 }
168
169 return visit_continue;
170 }
171
172 bool variable_found() const
173 {
174 return this->found;
175 }
176
177 private:
178 const char *name; /**< Find writes to a variable with this name. */
179 bool found; /**< Was a write to the variable found? */
180 };
181
182
183 class geom_array_resize_visitor : public ir_hierarchical_visitor {
184 public:
185 unsigned num_vertices;
186 gl_shader_program *prog;
187
188 geom_array_resize_visitor(unsigned num_vertices, gl_shader_program *prog)
189 {
190 this->num_vertices = num_vertices;
191 this->prog = prog;
192 }
193
194 virtual ~geom_array_resize_visitor()
195 {
196 /* empty */
197 }
198
199 virtual ir_visitor_status visit(ir_variable *var)
200 {
201 if (!var->type->is_array() || var->data.mode != ir_var_shader_in)
202 return visit_continue;
203
204 unsigned size = var->type->length;
205
206 /* Generate a link error if the shader has declared this array with an
207 * incorrect size.
208 */
209 if (size && size != this->num_vertices) {
210 linker_error(this->prog, "size of array %s declared as %u, "
211 "but number of input vertices is %u\n",
212 var->name, size, this->num_vertices);
213 return visit_continue;
214 }
215
216 /* Generate a link error if the shader attempts to access an input
217 * array using an index too large for its actual size assigned at link
218 * time.
219 */
220 if (var->data.max_array_access >= this->num_vertices) {
221 linker_error(this->prog, "geometry shader accesses element %i of "
222 "%s, but only %i input vertices\n",
223 var->data.max_array_access, var->name, this->num_vertices);
224 return visit_continue;
225 }
226
227 var->type = glsl_type::get_array_instance(var->type->element_type(),
228 this->num_vertices);
229 var->data.max_array_access = this->num_vertices - 1;
230
231 return visit_continue;
232 }
233
234 /* Dereferences of input variables need to be updated so that their type
235 * matches the newly assigned type of the variable they are accessing. */
236 virtual ir_visitor_status visit(ir_dereference_variable *ir)
237 {
238 ir->type = ir->var->type;
239 return visit_continue;
240 }
241
242 /* Dereferences of 2D input arrays need to be updated so that their type
243 * matches the newly assigned type of the array they are accessing. */
244 virtual ir_visitor_status visit_leave(ir_dereference_array *ir)
245 {
246 const glsl_type *const vt = ir->array->type;
247 if (vt->is_array())
248 ir->type = vt->element_type();
249 return visit_continue;
250 }
251 };
252
253 /**
254 * Visitor that determines the highest stream id to which a (geometry) shader
255 * emits vertices. It also checks whether End{Stream}Primitive is ever called.
256 */
257 class find_emit_vertex_visitor : public ir_hierarchical_visitor {
258 public:
259 find_emit_vertex_visitor(int max_allowed)
260 : max_stream_allowed(max_allowed),
261 invalid_stream_id(0),
262 invalid_stream_id_from_emit_vertex(false),
263 end_primitive_found(false),
264 uses_non_zero_stream(false)
265 {
266 /* empty */
267 }
268
269 virtual ir_visitor_status visit_leave(ir_emit_vertex *ir)
270 {
271 int stream_id = ir->stream_id();
272
273 if (stream_id < 0) {
274 invalid_stream_id = stream_id;
275 invalid_stream_id_from_emit_vertex = true;
276 return visit_stop;
277 }
278
279 if (stream_id > max_stream_allowed) {
280 invalid_stream_id = stream_id;
281 invalid_stream_id_from_emit_vertex = true;
282 return visit_stop;
283 }
284
285 if (stream_id != 0)
286 uses_non_zero_stream = true;
287
288 return visit_continue;
289 }
290
291 virtual ir_visitor_status visit_leave(ir_end_primitive *ir)
292 {
293 end_primitive_found = true;
294
295 int stream_id = ir->stream_id();
296
297 if (stream_id < 0) {
298 invalid_stream_id = stream_id;
299 invalid_stream_id_from_emit_vertex = false;
300 return visit_stop;
301 }
302
303 if (stream_id > max_stream_allowed) {
304 invalid_stream_id = stream_id;
305 invalid_stream_id_from_emit_vertex = false;
306 return visit_stop;
307 }
308
309 if (stream_id != 0)
310 uses_non_zero_stream = true;
311
312 return visit_continue;
313 }
314
315 bool error()
316 {
317 return invalid_stream_id != 0;
318 }
319
320 const char *error_func()
321 {
322 return invalid_stream_id_from_emit_vertex ?
323 "EmitStreamVertex" : "EndStreamPrimitive";
324 }
325
326 int error_stream()
327 {
328 return invalid_stream_id;
329 }
330
331 bool uses_streams()
332 {
333 return uses_non_zero_stream;
334 }
335
336 bool uses_end_primitive()
337 {
338 return end_primitive_found;
339 }
340
341 private:
342 int max_stream_allowed;
343 int invalid_stream_id;
344 bool invalid_stream_id_from_emit_vertex;
345 bool end_primitive_found;
346 bool uses_non_zero_stream;
347 };
348
349 } /* anonymous namespace */
350
351 void
352 linker_error(gl_shader_program *prog, const char *fmt, ...)
353 {
354 va_list ap;
355
356 ralloc_strcat(&prog->InfoLog, "error: ");
357 va_start(ap, fmt);
358 ralloc_vasprintf_append(&prog->InfoLog, fmt, ap);
359 va_end(ap);
360
361 prog->LinkStatus = false;
362 }
363
364
365 void
366 linker_warning(gl_shader_program *prog, const char *fmt, ...)
367 {
368 va_list ap;
369
370 ralloc_strcat(&prog->InfoLog, "warning: ");
371 va_start(ap, fmt);
372 ralloc_vasprintf_append(&prog->InfoLog, fmt, ap);
373 va_end(ap);
374
375 }
376
377
378 /**
379 * Given a string identifying a program resource, break it into a base name
380 * and an optional array index in square brackets.
381 *
382 * If an array index is present, \c out_base_name_end is set to point to the
383 * "[" that precedes the array index, and the array index itself is returned
384 * as a long.
385 *
386 * If no array index is present (or if the array index is negative or
387 * mal-formed), \c out_base_name_end, is set to point to the null terminator
388 * at the end of the input string, and -1 is returned.
389 *
390 * Only the final array index is parsed; if the string contains other array
391 * indices (or structure field accesses), they are left in the base name.
392 *
393 * No attempt is made to check that the base name is properly formed;
394 * typically the caller will look up the base name in a hash table, so
395 * ill-formed base names simply turn into hash table lookup failures.
396 */
397 long
398 parse_program_resource_name(const GLchar *name,
399 const GLchar **out_base_name_end)
400 {
401 /* Section 7.3.1 ("Program Interfaces") of the OpenGL 4.3 spec says:
402 *
403 * "When an integer array element or block instance number is part of
404 * the name string, it will be specified in decimal form without a "+"
405 * or "-" sign or any extra leading zeroes. Additionally, the name
406 * string will not include white space anywhere in the string."
407 */
408
409 const size_t len = strlen(name);
410 *out_base_name_end = name + len;
411
412 if (len == 0 || name[len-1] != ']')
413 return -1;
414
415 /* Walk backwards over the string looking for a non-digit character. This
416 * had better be the opening bracket for an array index.
417 *
418 * Initially, i specifies the location of the ']'. Since the string may
419 * contain only the ']' charcater, walk backwards very carefully.
420 */
421 unsigned i;
422 for (i = len - 1; (i > 0) && isdigit(name[i-1]); --i)
423 /* empty */ ;
424
425 if ((i == 0) || name[i-1] != '[')
426 return -1;
427
428 long array_index = strtol(&name[i], NULL, 10);
429 if (array_index < 0)
430 return -1;
431
432 *out_base_name_end = name + (i - 1);
433 return array_index;
434 }
435
436
437 void
438 link_invalidate_variable_locations(exec_list *ir)
439 {
440 foreach_in_list(ir_instruction, node, ir) {
441 ir_variable *const var = node->as_variable();
442
443 if (var == NULL)
444 continue;
445
446 /* Only assign locations for variables that lack an explicit location.
447 * Explicit locations are set for all built-in variables, generic vertex
448 * shader inputs (via layout(location=...)), and generic fragment shader
449 * outputs (also via layout(location=...)).
450 */
451 if (!var->data.explicit_location) {
452 var->data.location = -1;
453 var->data.location_frac = 0;
454 }
455
456 /* ir_variable::is_unmatched_generic_inout is used by the linker while
457 * connecting outputs from one stage to inputs of the next stage.
458 *
459 * There are two implicit assumptions here. First, we assume that any
460 * built-in variable (i.e., non-generic in or out) will have
461 * explicit_location set. Second, we assume that any generic in or out
462 * will not have explicit_location set.
463 *
464 * This second assumption will only be valid until
465 * GL_ARB_separate_shader_objects is supported. When that extension is
466 * implemented, this function will need some modifications.
467 */
468 if (!var->data.explicit_location) {
469 var->data.is_unmatched_generic_inout = 1;
470 } else {
471 var->data.is_unmatched_generic_inout = 0;
472 }
473 }
474 }
475
476
477 /**
478 * Set UsesClipDistance and ClipDistanceArraySize based on the given shader.
479 *
480 * Also check for errors based on incorrect usage of gl_ClipVertex and
481 * gl_ClipDistance.
482 *
483 * Return false if an error was reported.
484 */
485 static void
486 analyze_clip_usage(struct gl_shader_program *prog,
487 struct gl_shader *shader, GLboolean *UsesClipDistance,
488 GLuint *ClipDistanceArraySize)
489 {
490 *ClipDistanceArraySize = 0;
491
492 if (!prog->IsES && prog->Version >= 130) {
493 /* From section 7.1 (Vertex Shader Special Variables) of the
494 * GLSL 1.30 spec:
495 *
496 * "It is an error for a shader to statically write both
497 * gl_ClipVertex and gl_ClipDistance."
498 *
499 * This does not apply to GLSL ES shaders, since GLSL ES defines neither
500 * gl_ClipVertex nor gl_ClipDistance.
501 */
502 find_assignment_visitor clip_vertex("gl_ClipVertex");
503 find_assignment_visitor clip_distance("gl_ClipDistance");
504
505 clip_vertex.run(shader->ir);
506 clip_distance.run(shader->ir);
507 if (clip_vertex.variable_found() && clip_distance.variable_found()) {
508 linker_error(prog, "%s shader writes to both `gl_ClipVertex' "
509 "and `gl_ClipDistance'\n",
510 _mesa_shader_stage_to_string(shader->Stage));
511 return;
512 }
513 *UsesClipDistance = clip_distance.variable_found();
514 ir_variable *clip_distance_var =
515 shader->symbols->get_variable("gl_ClipDistance");
516 if (clip_distance_var)
517 *ClipDistanceArraySize = clip_distance_var->type->length;
518 } else {
519 *UsesClipDistance = false;
520 }
521 }
522
523
524 /**
525 * Verify that a vertex shader executable meets all semantic requirements.
526 *
527 * Also sets prog->Vert.UsesClipDistance and prog->Vert.ClipDistanceArraySize
528 * as a side effect.
529 *
530 * \param shader Vertex shader executable to be verified
531 */
532 void
533 validate_vertex_shader_executable(struct gl_shader_program *prog,
534 struct gl_shader *shader)
535 {
536 if (shader == NULL)
537 return;
538
539 /* From the GLSL 1.10 spec, page 48:
540 *
541 * "The variable gl_Position is available only in the vertex
542 * language and is intended for writing the homogeneous vertex
543 * position. All executions of a well-formed vertex shader
544 * executable must write a value into this variable. [...] The
545 * variable gl_Position is available only in the vertex
546 * language and is intended for writing the homogeneous vertex
547 * position. All executions of a well-formed vertex shader
548 * executable must write a value into this variable."
549 *
550 * while in GLSL 1.40 this text is changed to:
551 *
552 * "The variable gl_Position is available only in the vertex
553 * language and is intended for writing the homogeneous vertex
554 * position. It can be written at any time during shader
555 * execution. It may also be read back by a vertex shader
556 * after being written. This value will be used by primitive
557 * assembly, clipping, culling, and other fixed functionality
558 * operations, if present, that operate on primitives after
559 * vertex processing has occurred. Its value is undefined if
560 * the vertex shader executable does not write gl_Position."
561 *
562 * All GLSL ES Versions are similar to GLSL 1.40--failing to write to
563 * gl_Position is not an error.
564 */
565 if (prog->Version < (prog->IsES ? 300 : 140)) {
566 find_assignment_visitor find("gl_Position");
567 find.run(shader->ir);
568 if (!find.variable_found()) {
569 if (prog->IsES) {
570 linker_warning(prog,
571 "vertex shader does not write to `gl_Position'."
572 "It's value is undefined. \n");
573 } else {
574 linker_error(prog,
575 "vertex shader does not write to `gl_Position'. \n");
576 }
577 return;
578 }
579 }
580
581 analyze_clip_usage(prog, shader, &prog->Vert.UsesClipDistance,
582 &prog->Vert.ClipDistanceArraySize);
583 }
584
585
586 /**
587 * Verify that a fragment shader executable meets all semantic requirements
588 *
589 * \param shader Fragment shader executable to be verified
590 */
591 void
592 validate_fragment_shader_executable(struct gl_shader_program *prog,
593 struct gl_shader *shader)
594 {
595 if (shader == NULL)
596 return;
597
598 find_assignment_visitor frag_color("gl_FragColor");
599 find_assignment_visitor frag_data("gl_FragData");
600
601 frag_color.run(shader->ir);
602 frag_data.run(shader->ir);
603
604 if (frag_color.variable_found() && frag_data.variable_found()) {
605 linker_error(prog, "fragment shader writes to both "
606 "`gl_FragColor' and `gl_FragData'\n");
607 }
608 }
609
610 /**
611 * Verify that a geometry shader executable meets all semantic requirements
612 *
613 * Also sets prog->Geom.VerticesIn, prog->Geom.UsesClipDistance, and
614 * prog->Geom.ClipDistanceArraySize as a side effect.
615 *
616 * \param shader Geometry shader executable to be verified
617 */
618 void
619 validate_geometry_shader_executable(struct gl_shader_program *prog,
620 struct gl_shader *shader)
621 {
622 if (shader == NULL)
623 return;
624
625 unsigned num_vertices = vertices_per_prim(prog->Geom.InputType);
626 prog->Geom.VerticesIn = num_vertices;
627
628 analyze_clip_usage(prog, shader, &prog->Geom.UsesClipDistance,
629 &prog->Geom.ClipDistanceArraySize);
630 }
631
632 /**
633 * Check if geometry shaders emit to non-zero streams and do corresponding
634 * validations.
635 */
636 static void
637 validate_geometry_shader_emissions(struct gl_context *ctx,
638 struct gl_shader_program *prog)
639 {
640 if (prog->_LinkedShaders[MESA_SHADER_GEOMETRY] != NULL) {
641 find_emit_vertex_visitor emit_vertex(ctx->Const.MaxVertexStreams - 1);
642 emit_vertex.run(prog->_LinkedShaders[MESA_SHADER_GEOMETRY]->ir);
643 if (emit_vertex.error()) {
644 linker_error(prog, "Invalid call %s(%d). Accepted values for the "
645 "stream parameter are in the range [0, %d].\n",
646 emit_vertex.error_func(),
647 emit_vertex.error_stream(),
648 ctx->Const.MaxVertexStreams - 1);
649 }
650 prog->Geom.UsesStreams = emit_vertex.uses_streams();
651 prog->Geom.UsesEndPrimitive = emit_vertex.uses_end_primitive();
652
653 /* From the ARB_gpu_shader5 spec:
654 *
655 * "Multiple vertex streams are supported only if the output primitive
656 * type is declared to be "points". A program will fail to link if it
657 * contains a geometry shader calling EmitStreamVertex() or
658 * EndStreamPrimitive() if its output primitive type is not "points".
659 *
660 * However, in the same spec:
661 *
662 * "The function EmitVertex() is equivalent to calling EmitStreamVertex()
663 * with <stream> set to zero."
664 *
665 * And:
666 *
667 * "The function EndPrimitive() is equivalent to calling
668 * EndStreamPrimitive() with <stream> set to zero."
669 *
670 * Since we can call EmitVertex() and EndPrimitive() when we output
671 * primitives other than points, calling EmitStreamVertex(0) or
672 * EmitEndPrimitive(0) should not produce errors. This it also what Nvidia
673 * does. Currently we only set prog->Geom.UsesStreams to TRUE when
674 * EmitStreamVertex() or EmitEndPrimitive() are called with a non-zero
675 * stream.
676 */
677 if (prog->Geom.UsesStreams && prog->Geom.OutputType != GL_POINTS) {
678 linker_error(prog, "EmitStreamVertex(n) and EndStreamPrimitive(n) "
679 "with n>0 requires point output\n");
680 }
681 }
682 }
683
684 bool
685 validate_intrastage_arrays(struct gl_shader_program *prog,
686 ir_variable *const var,
687 ir_variable *const existing)
688 {
689 /* Consider the types to be "the same" if both types are arrays
690 * of the same type and one of the arrays is implicitly sized.
691 * In addition, set the type of the linked variable to the
692 * explicitly sized array.
693 */
694 if (var->type->is_array() && existing->type->is_array() &&
695 (var->type->fields.array == existing->type->fields.array) &&
696 ((var->type->length == 0)|| (existing->type->length == 0))) {
697 if (var->type->length != 0) {
698 if (var->type->length <= existing->data.max_array_access) {
699 linker_error(prog, "%s `%s' declared as type "
700 "`%s' but outermost dimension has an index"
701 " of `%i'\n",
702 mode_string(var),
703 var->name, var->type->name,
704 existing->data.max_array_access);
705 }
706 existing->type = var->type;
707 return true;
708 } else if (existing->type->length != 0) {
709 if(existing->type->length <= var->data.max_array_access) {
710 linker_error(prog, "%s `%s' declared as type "
711 "`%s' but outermost dimension has an index"
712 " of `%i'\n",
713 mode_string(var),
714 var->name, existing->type->name,
715 var->data.max_array_access);
716 }
717 return true;
718 }
719 }
720 return false;
721 }
722
723
724 /**
725 * Perform validation of global variables used across multiple shaders
726 */
727 void
728 cross_validate_globals(struct gl_shader_program *prog,
729 struct gl_shader **shader_list,
730 unsigned num_shaders,
731 bool uniforms_only)
732 {
733 /* Examine all of the uniforms in all of the shaders and cross validate
734 * them.
735 */
736 glsl_symbol_table variables;
737 for (unsigned i = 0; i < num_shaders; i++) {
738 if (shader_list[i] == NULL)
739 continue;
740
741 foreach_in_list(ir_instruction, node, shader_list[i]->ir) {
742 ir_variable *const var = node->as_variable();
743
744 if (var == NULL)
745 continue;
746
747 if (uniforms_only && (var->data.mode != ir_var_uniform))
748 continue;
749
750 /* Don't cross validate temporaries that are at global scope. These
751 * will eventually get pulled into the shaders 'main'.
752 */
753 if (var->data.mode == ir_var_temporary)
754 continue;
755
756 /* If a global with this name has already been seen, verify that the
757 * new instance has the same type. In addition, if the globals have
758 * initializers, the values of the initializers must be the same.
759 */
760 ir_variable *const existing = variables.get_variable(var->name);
761 if (existing != NULL) {
762 /* Check if types match. Interface blocks have some special
763 * rules so we handle those elsewhere.
764 */
765 if (var->type != existing->type &&
766 !var->is_interface_instance()) {
767 if (!validate_intrastage_arrays(prog, var, existing)) {
768 if (var->type->is_record() && existing->type->is_record()
769 && existing->type->record_compare(var->type)) {
770 existing->type = var->type;
771 } else {
772 linker_error(prog, "%s `%s' declared as type "
773 "`%s' and type `%s'\n",
774 mode_string(var),
775 var->name, var->type->name,
776 existing->type->name);
777 return;
778 }
779 }
780 }
781
782 if (var->data.explicit_location) {
783 if (existing->data.explicit_location
784 && (var->data.location != existing->data.location)) {
785 linker_error(prog, "explicit locations for %s "
786 "`%s' have differing values\n",
787 mode_string(var), var->name);
788 return;
789 }
790
791 existing->data.location = var->data.location;
792 existing->data.explicit_location = true;
793 }
794
795 /* From the GLSL 4.20 specification:
796 * "A link error will result if two compilation units in a program
797 * specify different integer-constant bindings for the same
798 * opaque-uniform name. However, it is not an error to specify a
799 * binding on some but not all declarations for the same name"
800 */
801 if (var->data.explicit_binding) {
802 if (existing->data.explicit_binding &&
803 var->data.binding != existing->data.binding) {
804 linker_error(prog, "explicit bindings for %s "
805 "`%s' have differing values\n",
806 mode_string(var), var->name);
807 return;
808 }
809
810 existing->data.binding = var->data.binding;
811 existing->data.explicit_binding = true;
812 }
813
814 if (var->type->contains_atomic() &&
815 var->data.atomic.offset != existing->data.atomic.offset) {
816 linker_error(prog, "offset specifications for %s "
817 "`%s' have differing values\n",
818 mode_string(var), var->name);
819 return;
820 }
821
822 /* Validate layout qualifiers for gl_FragDepth.
823 *
824 * From the AMD/ARB_conservative_depth specs:
825 *
826 * "If gl_FragDepth is redeclared in any fragment shader in a
827 * program, it must be redeclared in all fragment shaders in
828 * that program that have static assignments to
829 * gl_FragDepth. All redeclarations of gl_FragDepth in all
830 * fragment shaders in a single program must have the same set
831 * of qualifiers."
832 */
833 if (strcmp(var->name, "gl_FragDepth") == 0) {
834 bool layout_declared = var->data.depth_layout != ir_depth_layout_none;
835 bool layout_differs =
836 var->data.depth_layout != existing->data.depth_layout;
837
838 if (layout_declared && layout_differs) {
839 linker_error(prog,
840 "All redeclarations of gl_FragDepth in all "
841 "fragment shaders in a single program must have "
842 "the same set of qualifiers.\n");
843 }
844
845 if (var->data.used && layout_differs) {
846 linker_error(prog,
847 "If gl_FragDepth is redeclared with a layout "
848 "qualifier in any fragment shader, it must be "
849 "redeclared with the same layout qualifier in "
850 "all fragment shaders that have assignments to "
851 "gl_FragDepth\n");
852 }
853 }
854
855 /* Page 35 (page 41 of the PDF) of the GLSL 4.20 spec says:
856 *
857 * "If a shared global has multiple initializers, the
858 * initializers must all be constant expressions, and they
859 * must all have the same value. Otherwise, a link error will
860 * result. (A shared global having only one initializer does
861 * not require that initializer to be a constant expression.)"
862 *
863 * Previous to 4.20 the GLSL spec simply said that initializers
864 * must have the same value. In this case of non-constant
865 * initializers, this was impossible to determine. As a result,
866 * no vendor actually implemented that behavior. The 4.20
867 * behavior matches the implemented behavior of at least one other
868 * vendor, so we'll implement that for all GLSL versions.
869 */
870 if (var->constant_initializer != NULL) {
871 if (existing->constant_initializer != NULL) {
872 if (!var->constant_initializer->has_value(existing->constant_initializer)) {
873 linker_error(prog, "initializers for %s "
874 "`%s' have differing values\n",
875 mode_string(var), var->name);
876 return;
877 }
878 } else {
879 /* If the first-seen instance of a particular uniform did not
880 * have an initializer but a later instance does, copy the
881 * initializer to the version stored in the symbol table.
882 */
883 /* FINISHME: This is wrong. The constant_value field should
884 * FINISHME: not be modified! Imagine a case where a shader
885 * FINISHME: without an initializer is linked in two different
886 * FINISHME: programs with shaders that have differing
887 * FINISHME: initializers. Linking with the first will
888 * FINISHME: modify the shader, and linking with the second
889 * FINISHME: will fail.
890 */
891 existing->constant_initializer =
892 var->constant_initializer->clone(ralloc_parent(existing),
893 NULL);
894 }
895 }
896
897 if (var->data.has_initializer) {
898 if (existing->data.has_initializer
899 && (var->constant_initializer == NULL
900 || existing->constant_initializer == NULL)) {
901 linker_error(prog,
902 "shared global variable `%s' has multiple "
903 "non-constant initializers.\n",
904 var->name);
905 return;
906 }
907
908 /* Some instance had an initializer, so keep track of that. In
909 * this location, all sorts of initializers (constant or
910 * otherwise) will propagate the existence to the variable
911 * stored in the symbol table.
912 */
913 existing->data.has_initializer = true;
914 }
915
916 if (existing->data.invariant != var->data.invariant) {
917 linker_error(prog, "declarations for %s `%s' have "
918 "mismatching invariant qualifiers\n",
919 mode_string(var), var->name);
920 return;
921 }
922 if (existing->data.centroid != var->data.centroid) {
923 linker_error(prog, "declarations for %s `%s' have "
924 "mismatching centroid qualifiers\n",
925 mode_string(var), var->name);
926 return;
927 }
928 if (existing->data.sample != var->data.sample) {
929 linker_error(prog, "declarations for %s `%s` have "
930 "mismatching sample qualifiers\n",
931 mode_string(var), var->name);
932 return;
933 }
934 } else
935 variables.add_variable(var);
936 }
937 }
938 }
939
940
941 /**
942 * Perform validation of uniforms used across multiple shader stages
943 */
944 void
945 cross_validate_uniforms(struct gl_shader_program *prog)
946 {
947 cross_validate_globals(prog, prog->_LinkedShaders,
948 MESA_SHADER_STAGES, true);
949 }
950
951 /**
952 * Accumulates the array of prog->UniformBlocks and checks that all
953 * definitons of blocks agree on their contents.
954 */
955 static bool
956 interstage_cross_validate_uniform_blocks(struct gl_shader_program *prog)
957 {
958 unsigned max_num_uniform_blocks = 0;
959 for (unsigned i = 0; i < MESA_SHADER_STAGES; i++) {
960 if (prog->_LinkedShaders[i])
961 max_num_uniform_blocks += prog->_LinkedShaders[i]->NumUniformBlocks;
962 }
963
964 for (unsigned i = 0; i < MESA_SHADER_STAGES; i++) {
965 struct gl_shader *sh = prog->_LinkedShaders[i];
966
967 prog->UniformBlockStageIndex[i] = ralloc_array(prog, int,
968 max_num_uniform_blocks);
969 for (unsigned int j = 0; j < max_num_uniform_blocks; j++)
970 prog->UniformBlockStageIndex[i][j] = -1;
971
972 if (sh == NULL)
973 continue;
974
975 for (unsigned int j = 0; j < sh->NumUniformBlocks; j++) {
976 int index = link_cross_validate_uniform_block(prog,
977 &prog->UniformBlocks,
978 &prog->NumUniformBlocks,
979 &sh->UniformBlocks[j]);
980
981 if (index == -1) {
982 linker_error(prog, "uniform block `%s' has mismatching definitions\n",
983 sh->UniformBlocks[j].Name);
984 return false;
985 }
986
987 prog->UniformBlockStageIndex[i][index] = j;
988 }
989 }
990
991 return true;
992 }
993
994
995 /**
996 * Populates a shaders symbol table with all global declarations
997 */
998 static void
999 populate_symbol_table(gl_shader *sh)
1000 {
1001 sh->symbols = new(sh) glsl_symbol_table;
1002
1003 foreach_in_list(ir_instruction, inst, sh->ir) {
1004 ir_variable *var;
1005 ir_function *func;
1006
1007 if ((func = inst->as_function()) != NULL) {
1008 sh->symbols->add_function(func);
1009 } else if ((var = inst->as_variable()) != NULL) {
1010 if (var->data.mode != ir_var_temporary)
1011 sh->symbols->add_variable(var);
1012 }
1013 }
1014 }
1015
1016
1017 /**
1018 * Remap variables referenced in an instruction tree
1019 *
1020 * This is used when instruction trees are cloned from one shader and placed in
1021 * another. These trees will contain references to \c ir_variable nodes that
1022 * do not exist in the target shader. This function finds these \c ir_variable
1023 * references and replaces the references with matching variables in the target
1024 * shader.
1025 *
1026 * If there is no matching variable in the target shader, a clone of the
1027 * \c ir_variable is made and added to the target shader. The new variable is
1028 * added to \b both the instruction stream and the symbol table.
1029 *
1030 * \param inst IR tree that is to be processed.
1031 * \param symbols Symbol table containing global scope symbols in the
1032 * linked shader.
1033 * \param instructions Instruction stream where new variable declarations
1034 * should be added.
1035 */
1036 void
1037 remap_variables(ir_instruction *inst, struct gl_shader *target,
1038 hash_table *temps)
1039 {
1040 class remap_visitor : public ir_hierarchical_visitor {
1041 public:
1042 remap_visitor(struct gl_shader *target,
1043 hash_table *temps)
1044 {
1045 this->target = target;
1046 this->symbols = target->symbols;
1047 this->instructions = target->ir;
1048 this->temps = temps;
1049 }
1050
1051 virtual ir_visitor_status visit(ir_dereference_variable *ir)
1052 {
1053 if (ir->var->data.mode == ir_var_temporary) {
1054 ir_variable *var = (ir_variable *) hash_table_find(temps, ir->var);
1055
1056 assert(var != NULL);
1057 ir->var = var;
1058 return visit_continue;
1059 }
1060
1061 ir_variable *const existing =
1062 this->symbols->get_variable(ir->var->name);
1063 if (existing != NULL)
1064 ir->var = existing;
1065 else {
1066 ir_variable *copy = ir->var->clone(this->target, NULL);
1067
1068 this->symbols->add_variable(copy);
1069 this->instructions->push_head(copy);
1070 ir->var = copy;
1071 }
1072
1073 return visit_continue;
1074 }
1075
1076 private:
1077 struct gl_shader *target;
1078 glsl_symbol_table *symbols;
1079 exec_list *instructions;
1080 hash_table *temps;
1081 };
1082
1083 remap_visitor v(target, temps);
1084
1085 inst->accept(&v);
1086 }
1087
1088
1089 /**
1090 * Move non-declarations from one instruction stream to another
1091 *
1092 * The intended usage pattern of this function is to pass the pointer to the
1093 * head sentinel of a list (i.e., a pointer to the list cast to an \c exec_node
1094 * pointer) for \c last and \c false for \c make_copies on the first
1095 * call. Successive calls pass the return value of the previous call for
1096 * \c last and \c true for \c make_copies.
1097 *
1098 * \param instructions Source instruction stream
1099 * \param last Instruction after which new instructions should be
1100 * inserted in the target instruction stream
1101 * \param make_copies Flag selecting whether instructions in \c instructions
1102 * should be copied (via \c ir_instruction::clone) into the
1103 * target list or moved.
1104 *
1105 * \return
1106 * The new "last" instruction in the target instruction stream. This pointer
1107 * is suitable for use as the \c last parameter of a later call to this
1108 * function.
1109 */
1110 exec_node *
1111 move_non_declarations(exec_list *instructions, exec_node *last,
1112 bool make_copies, gl_shader *target)
1113 {
1114 hash_table *temps = NULL;
1115
1116 if (make_copies)
1117 temps = hash_table_ctor(0, hash_table_pointer_hash,
1118 hash_table_pointer_compare);
1119
1120 foreach_in_list_safe(ir_instruction, inst, instructions) {
1121 if (inst->as_function())
1122 continue;
1123
1124 ir_variable *var = inst->as_variable();
1125 if ((var != NULL) && (var->data.mode != ir_var_temporary))
1126 continue;
1127
1128 assert(inst->as_assignment()
1129 || inst->as_call()
1130 || inst->as_if() /* for initializers with the ?: operator */
1131 || ((var != NULL) && (var->data.mode == ir_var_temporary)));
1132
1133 if (make_copies) {
1134 inst = inst->clone(target, NULL);
1135
1136 if (var != NULL)
1137 hash_table_insert(temps, inst, var);
1138 else
1139 remap_variables(inst, target, temps);
1140 } else {
1141 inst->remove();
1142 }
1143
1144 last->insert_after(inst);
1145 last = inst;
1146 }
1147
1148 if (make_copies)
1149 hash_table_dtor(temps);
1150
1151 return last;
1152 }
1153
1154 /**
1155 * Get the function signature for main from a shader
1156 */
1157 ir_function_signature *
1158 link_get_main_function_signature(gl_shader *sh)
1159 {
1160 ir_function *const f = sh->symbols->get_function("main");
1161 if (f != NULL) {
1162 exec_list void_parameters;
1163
1164 /* Look for the 'void main()' signature and ensure that it's defined.
1165 * This keeps the linker from accidentally pick a shader that just
1166 * contains a prototype for main.
1167 *
1168 * We don't have to check for multiple definitions of main (in multiple
1169 * shaders) because that would have already been caught above.
1170 */
1171 ir_function_signature *sig =
1172 f->matching_signature(NULL, &void_parameters, false);
1173 if ((sig != NULL) && sig->is_defined) {
1174 return sig;
1175 }
1176 }
1177
1178 return NULL;
1179 }
1180
1181
1182 /**
1183 * This class is only used in link_intrastage_shaders() below but declaring
1184 * it inside that function leads to compiler warnings with some versions of
1185 * gcc.
1186 */
1187 class array_sizing_visitor : public ir_hierarchical_visitor {
1188 public:
1189 array_sizing_visitor()
1190 : mem_ctx(ralloc_context(NULL)),
1191 unnamed_interfaces(hash_table_ctor(0, hash_table_pointer_hash,
1192 hash_table_pointer_compare))
1193 {
1194 }
1195
1196 ~array_sizing_visitor()
1197 {
1198 hash_table_dtor(this->unnamed_interfaces);
1199 ralloc_free(this->mem_ctx);
1200 }
1201
1202 virtual ir_visitor_status visit(ir_variable *var)
1203 {
1204 fixup_type(&var->type, var->data.max_array_access);
1205 if (var->type->is_interface()) {
1206 if (interface_contains_unsized_arrays(var->type)) {
1207 const glsl_type *new_type =
1208 resize_interface_members(var->type,
1209 var->get_max_ifc_array_access());
1210 var->type = new_type;
1211 var->change_interface_type(new_type);
1212 }
1213 } else if (var->type->is_array() &&
1214 var->type->fields.array->is_interface()) {
1215 if (interface_contains_unsized_arrays(var->type->fields.array)) {
1216 const glsl_type *new_type =
1217 resize_interface_members(var->type->fields.array,
1218 var->get_max_ifc_array_access());
1219 var->change_interface_type(new_type);
1220 var->type =
1221 glsl_type::get_array_instance(new_type, var->type->length);
1222 }
1223 } else if (const glsl_type *ifc_type = var->get_interface_type()) {
1224 /* Store a pointer to the variable in the unnamed_interfaces
1225 * hashtable.
1226 */
1227 ir_variable **interface_vars = (ir_variable **)
1228 hash_table_find(this->unnamed_interfaces, ifc_type);
1229 if (interface_vars == NULL) {
1230 interface_vars = rzalloc_array(mem_ctx, ir_variable *,
1231 ifc_type->length);
1232 hash_table_insert(this->unnamed_interfaces, interface_vars,
1233 ifc_type);
1234 }
1235 unsigned index = ifc_type->field_index(var->name);
1236 assert(index < ifc_type->length);
1237 assert(interface_vars[index] == NULL);
1238 interface_vars[index] = var;
1239 }
1240 return visit_continue;
1241 }
1242
1243 /**
1244 * For each unnamed interface block that was discovered while running the
1245 * visitor, adjust the interface type to reflect the newly assigned array
1246 * sizes, and fix up the ir_variable nodes to point to the new interface
1247 * type.
1248 */
1249 void fixup_unnamed_interface_types()
1250 {
1251 hash_table_call_foreach(this->unnamed_interfaces,
1252 fixup_unnamed_interface_type, NULL);
1253 }
1254
1255 private:
1256 /**
1257 * If the type pointed to by \c type represents an unsized array, replace
1258 * it with a sized array whose size is determined by max_array_access.
1259 */
1260 static void fixup_type(const glsl_type **type, unsigned max_array_access)
1261 {
1262 if ((*type)->is_unsized_array()) {
1263 *type = glsl_type::get_array_instance((*type)->fields.array,
1264 max_array_access + 1);
1265 assert(*type != NULL);
1266 }
1267 }
1268
1269 /**
1270 * Determine whether the given interface type contains unsized arrays (if
1271 * it doesn't, array_sizing_visitor doesn't need to process it).
1272 */
1273 static bool interface_contains_unsized_arrays(const glsl_type *type)
1274 {
1275 for (unsigned i = 0; i < type->length; i++) {
1276 const glsl_type *elem_type = type->fields.structure[i].type;
1277 if (elem_type->is_unsized_array())
1278 return true;
1279 }
1280 return false;
1281 }
1282
1283 /**
1284 * Create a new interface type based on the given type, with unsized arrays
1285 * replaced by sized arrays whose size is determined by
1286 * max_ifc_array_access.
1287 */
1288 static const glsl_type *
1289 resize_interface_members(const glsl_type *type,
1290 const unsigned *max_ifc_array_access)
1291 {
1292 unsigned num_fields = type->length;
1293 glsl_struct_field *fields = new glsl_struct_field[num_fields];
1294 memcpy(fields, type->fields.structure,
1295 num_fields * sizeof(*fields));
1296 for (unsigned i = 0; i < num_fields; i++) {
1297 fixup_type(&fields[i].type, max_ifc_array_access[i]);
1298 }
1299 glsl_interface_packing packing =
1300 (glsl_interface_packing) type->interface_packing;
1301 const glsl_type *new_ifc_type =
1302 glsl_type::get_interface_instance(fields, num_fields,
1303 packing, type->name);
1304 delete [] fields;
1305 return new_ifc_type;
1306 }
1307
1308 static void fixup_unnamed_interface_type(const void *key, void *data,
1309 void *)
1310 {
1311 const glsl_type *ifc_type = (const glsl_type *) key;
1312 ir_variable **interface_vars = (ir_variable **) data;
1313 unsigned num_fields = ifc_type->length;
1314 glsl_struct_field *fields = new glsl_struct_field[num_fields];
1315 memcpy(fields, ifc_type->fields.structure,
1316 num_fields * sizeof(*fields));
1317 bool interface_type_changed = false;
1318 for (unsigned i = 0; i < num_fields; i++) {
1319 if (interface_vars[i] != NULL &&
1320 fields[i].type != interface_vars[i]->type) {
1321 fields[i].type = interface_vars[i]->type;
1322 interface_type_changed = true;
1323 }
1324 }
1325 if (!interface_type_changed) {
1326 delete [] fields;
1327 return;
1328 }
1329 glsl_interface_packing packing =
1330 (glsl_interface_packing) ifc_type->interface_packing;
1331 const glsl_type *new_ifc_type =
1332 glsl_type::get_interface_instance(fields, num_fields, packing,
1333 ifc_type->name);
1334 delete [] fields;
1335 for (unsigned i = 0; i < num_fields; i++) {
1336 if (interface_vars[i] != NULL)
1337 interface_vars[i]->change_interface_type(new_ifc_type);
1338 }
1339 }
1340
1341 /**
1342 * Memory context used to allocate the data in \c unnamed_interfaces.
1343 */
1344 void *mem_ctx;
1345
1346 /**
1347 * Hash table from const glsl_type * to an array of ir_variable *'s
1348 * pointing to the ir_variables constituting each unnamed interface block.
1349 */
1350 hash_table *unnamed_interfaces;
1351 };
1352
1353 /**
1354 * Performs the cross-validation of layout qualifiers specified in
1355 * redeclaration of gl_FragCoord for the attached fragment shaders,
1356 * and propagates them to the linked FS and linked shader program.
1357 */
1358 static void
1359 link_fs_input_layout_qualifiers(struct gl_shader_program *prog,
1360 struct gl_shader *linked_shader,
1361 struct gl_shader **shader_list,
1362 unsigned num_shaders)
1363 {
1364 linked_shader->redeclares_gl_fragcoord = false;
1365 linked_shader->uses_gl_fragcoord = false;
1366 linked_shader->origin_upper_left = false;
1367 linked_shader->pixel_center_integer = false;
1368
1369 if (linked_shader->Stage != MESA_SHADER_FRAGMENT ||
1370 (prog->Version < 150 && !prog->ARB_fragment_coord_conventions_enable))
1371 return;
1372
1373 for (unsigned i = 0; i < num_shaders; i++) {
1374 struct gl_shader *shader = shader_list[i];
1375 /* From the GLSL 1.50 spec, page 39:
1376 *
1377 * "If gl_FragCoord is redeclared in any fragment shader in a program,
1378 * it must be redeclared in all the fragment shaders in that program
1379 * that have a static use gl_FragCoord."
1380 */
1381 if ((linked_shader->redeclares_gl_fragcoord
1382 && !shader->redeclares_gl_fragcoord
1383 && shader->uses_gl_fragcoord)
1384 || (shader->redeclares_gl_fragcoord
1385 && !linked_shader->redeclares_gl_fragcoord
1386 && linked_shader->uses_gl_fragcoord)) {
1387 linker_error(prog, "fragment shader defined with conflicting "
1388 "layout qualifiers for gl_FragCoord\n");
1389 }
1390
1391 /* From the GLSL 1.50 spec, page 39:
1392 *
1393 * "All redeclarations of gl_FragCoord in all fragment shaders in a
1394 * single program must have the same set of qualifiers."
1395 */
1396 if (linked_shader->redeclares_gl_fragcoord && shader->redeclares_gl_fragcoord
1397 && (shader->origin_upper_left != linked_shader->origin_upper_left
1398 || shader->pixel_center_integer != linked_shader->pixel_center_integer)) {
1399 linker_error(prog, "fragment shader defined with conflicting "
1400 "layout qualifiers for gl_FragCoord\n");
1401 }
1402
1403 /* Update the linked shader state.  Note that uses_gl_fragcoord should
1404 * accumulate the results.  The other values should replace.  If there
1405 * are multiple redeclarations, all the fields except uses_gl_fragcoord
1406 * are already known to be the same.
1407 */
1408 if (shader->redeclares_gl_fragcoord || shader->uses_gl_fragcoord) {
1409 linked_shader->redeclares_gl_fragcoord =
1410 shader->redeclares_gl_fragcoord;
1411 linked_shader->uses_gl_fragcoord = linked_shader->uses_gl_fragcoord
1412 || shader->uses_gl_fragcoord;
1413 linked_shader->origin_upper_left = shader->origin_upper_left;
1414 linked_shader->pixel_center_integer = shader->pixel_center_integer;
1415 }
1416
1417 linked_shader->EarlyFragmentTests |= shader->EarlyFragmentTests;
1418 }
1419 }
1420
1421 /**
1422 * Performs the cross-validation of geometry shader max_vertices and
1423 * primitive type layout qualifiers for the attached geometry shaders,
1424 * and propagates them to the linked GS and linked shader program.
1425 */
1426 static void
1427 link_gs_inout_layout_qualifiers(struct gl_shader_program *prog,
1428 struct gl_shader *linked_shader,
1429 struct gl_shader **shader_list,
1430 unsigned num_shaders)
1431 {
1432 linked_shader->Geom.VerticesOut = 0;
1433 linked_shader->Geom.Invocations = 0;
1434 linked_shader->Geom.InputType = PRIM_UNKNOWN;
1435 linked_shader->Geom.OutputType = PRIM_UNKNOWN;
1436
1437 /* No in/out qualifiers defined for anything but GLSL 1.50+
1438 * geometry shaders so far.
1439 */
1440 if (linked_shader->Stage != MESA_SHADER_GEOMETRY || prog->Version < 150)
1441 return;
1442
1443 /* From the GLSL 1.50 spec, page 46:
1444 *
1445 * "All geometry shader output layout declarations in a program
1446 * must declare the same layout and same value for
1447 * max_vertices. There must be at least one geometry output
1448 * layout declaration somewhere in a program, but not all
1449 * geometry shaders (compilation units) are required to
1450 * declare it."
1451 */
1452
1453 for (unsigned i = 0; i < num_shaders; i++) {
1454 struct gl_shader *shader = shader_list[i];
1455
1456 if (shader->Geom.InputType != PRIM_UNKNOWN) {
1457 if (linked_shader->Geom.InputType != PRIM_UNKNOWN &&
1458 linked_shader->Geom.InputType != shader->Geom.InputType) {
1459 linker_error(prog, "geometry shader defined with conflicting "
1460 "input types\n");
1461 return;
1462 }
1463 linked_shader->Geom.InputType = shader->Geom.InputType;
1464 }
1465
1466 if (shader->Geom.OutputType != PRIM_UNKNOWN) {
1467 if (linked_shader->Geom.OutputType != PRIM_UNKNOWN &&
1468 linked_shader->Geom.OutputType != shader->Geom.OutputType) {
1469 linker_error(prog, "geometry shader defined with conflicting "
1470 "output types\n");
1471 return;
1472 }
1473 linked_shader->Geom.OutputType = shader->Geom.OutputType;
1474 }
1475
1476 if (shader->Geom.VerticesOut != 0) {
1477 if (linked_shader->Geom.VerticesOut != 0 &&
1478 linked_shader->Geom.VerticesOut != shader->Geom.VerticesOut) {
1479 linker_error(prog, "geometry shader defined with conflicting "
1480 "output vertex count (%d and %d)\n",
1481 linked_shader->Geom.VerticesOut,
1482 shader->Geom.VerticesOut);
1483 return;
1484 }
1485 linked_shader->Geom.VerticesOut = shader->Geom.VerticesOut;
1486 }
1487
1488 if (shader->Geom.Invocations != 0) {
1489 if (linked_shader->Geom.Invocations != 0 &&
1490 linked_shader->Geom.Invocations != shader->Geom.Invocations) {
1491 linker_error(prog, "geometry shader defined with conflicting "
1492 "invocation count (%d and %d)\n",
1493 linked_shader->Geom.Invocations,
1494 shader->Geom.Invocations);
1495 return;
1496 }
1497 linked_shader->Geom.Invocations = shader->Geom.Invocations;
1498 }
1499 }
1500
1501 /* Just do the intrastage -> interstage propagation right now,
1502 * since we already know we're in the right type of shader program
1503 * for doing it.
1504 */
1505 if (linked_shader->Geom.InputType == PRIM_UNKNOWN) {
1506 linker_error(prog,
1507 "geometry shader didn't declare primitive input type\n");
1508 return;
1509 }
1510 prog->Geom.InputType = linked_shader->Geom.InputType;
1511
1512 if (linked_shader->Geom.OutputType == PRIM_UNKNOWN) {
1513 linker_error(prog,
1514 "geometry shader didn't declare primitive output type\n");
1515 return;
1516 }
1517 prog->Geom.OutputType = linked_shader->Geom.OutputType;
1518
1519 if (linked_shader->Geom.VerticesOut == 0) {
1520 linker_error(prog,
1521 "geometry shader didn't declare max_vertices\n");
1522 return;
1523 }
1524 prog->Geom.VerticesOut = linked_shader->Geom.VerticesOut;
1525
1526 if (linked_shader->Geom.Invocations == 0)
1527 linked_shader->Geom.Invocations = 1;
1528
1529 prog->Geom.Invocations = linked_shader->Geom.Invocations;
1530 }
1531
1532
1533 /**
1534 * Perform cross-validation of compute shader local_size_{x,y,z} layout
1535 * qualifiers for the attached compute shaders, and propagate them to the
1536 * linked CS and linked shader program.
1537 */
1538 static void
1539 link_cs_input_layout_qualifiers(struct gl_shader_program *prog,
1540 struct gl_shader *linked_shader,
1541 struct gl_shader **shader_list,
1542 unsigned num_shaders)
1543 {
1544 for (int i = 0; i < 3; i++)
1545 linked_shader->Comp.LocalSize[i] = 0;
1546
1547 /* This function is called for all shader stages, but it only has an effect
1548 * for compute shaders.
1549 */
1550 if (linked_shader->Stage != MESA_SHADER_COMPUTE)
1551 return;
1552
1553 /* From the ARB_compute_shader spec, in the section describing local size
1554 * declarations:
1555 *
1556 * If multiple compute shaders attached to a single program object
1557 * declare local work-group size, the declarations must be identical;
1558 * otherwise a link-time error results. Furthermore, if a program
1559 * object contains any compute shaders, at least one must contain an
1560 * input layout qualifier specifying the local work sizes of the
1561 * program, or a link-time error will occur.
1562 */
1563 for (unsigned sh = 0; sh < num_shaders; sh++) {
1564 struct gl_shader *shader = shader_list[sh];
1565
1566 if (shader->Comp.LocalSize[0] != 0) {
1567 if (linked_shader->Comp.LocalSize[0] != 0) {
1568 for (int i = 0; i < 3; i++) {
1569 if (linked_shader->Comp.LocalSize[i] !=
1570 shader->Comp.LocalSize[i]) {
1571 linker_error(prog, "compute shader defined with conflicting "
1572 "local sizes\n");
1573 return;
1574 }
1575 }
1576 }
1577 for (int i = 0; i < 3; i++)
1578 linked_shader->Comp.LocalSize[i] = shader->Comp.LocalSize[i];
1579 }
1580 }
1581
1582 /* Just do the intrastage -> interstage propagation right now,
1583 * since we already know we're in the right type of shader program
1584 * for doing it.
1585 */
1586 if (linked_shader->Comp.LocalSize[0] == 0) {
1587 linker_error(prog, "compute shader didn't declare local size\n");
1588 return;
1589 }
1590 for (int i = 0; i < 3; i++)
1591 prog->Comp.LocalSize[i] = linked_shader->Comp.LocalSize[i];
1592 }
1593
1594
1595 /**
1596 * Combine a group of shaders for a single stage to generate a linked shader
1597 *
1598 * \note
1599 * If this function is supplied a single shader, it is cloned, and the new
1600 * shader is returned.
1601 */
1602 static struct gl_shader *
1603 link_intrastage_shaders(void *mem_ctx,
1604 struct gl_context *ctx,
1605 struct gl_shader_program *prog,
1606 struct gl_shader **shader_list,
1607 unsigned num_shaders)
1608 {
1609 struct gl_uniform_block *uniform_blocks = NULL;
1610
1611 /* Check that global variables defined in multiple shaders are consistent.
1612 */
1613 cross_validate_globals(prog, shader_list, num_shaders, false);
1614 if (!prog->LinkStatus)
1615 return NULL;
1616
1617 /* Check that interface blocks defined in multiple shaders are consistent.
1618 */
1619 validate_intrastage_interface_blocks(prog, (const gl_shader **)shader_list,
1620 num_shaders);
1621 if (!prog->LinkStatus)
1622 return NULL;
1623
1624 /* Link up uniform blocks defined within this stage. */
1625 const unsigned num_uniform_blocks =
1626 link_uniform_blocks(mem_ctx, prog, shader_list, num_shaders,
1627 &uniform_blocks);
1628 if (!prog->LinkStatus)
1629 return NULL;
1630
1631 /* Check that there is only a single definition of each function signature
1632 * across all shaders.
1633 */
1634 for (unsigned i = 0; i < (num_shaders - 1); i++) {
1635 foreach_in_list(ir_instruction, node, shader_list[i]->ir) {
1636 ir_function *const f = node->as_function();
1637
1638 if (f == NULL)
1639 continue;
1640
1641 for (unsigned j = i + 1; j < num_shaders; j++) {
1642 ir_function *const other =
1643 shader_list[j]->symbols->get_function(f->name);
1644
1645 /* If the other shader has no function (and therefore no function
1646 * signatures) with the same name, skip to the next shader.
1647 */
1648 if (other == NULL)
1649 continue;
1650
1651 foreach_in_list(ir_function_signature, sig, &f->signatures) {
1652 if (!sig->is_defined || sig->is_builtin())
1653 continue;
1654
1655 ir_function_signature *other_sig =
1656 other->exact_matching_signature(NULL, &sig->parameters);
1657
1658 if ((other_sig != NULL) && other_sig->is_defined
1659 && !other_sig->is_builtin()) {
1660 linker_error(prog, "function `%s' is multiply defined\n",
1661 f->name);
1662 return NULL;
1663 }
1664 }
1665 }
1666 }
1667 }
1668
1669 /* Find the shader that defines main, and make a clone of it.
1670 *
1671 * Starting with the clone, search for undefined references. If one is
1672 * found, find the shader that defines it. Clone the reference and add
1673 * it to the shader. Repeat until there are no undefined references or
1674 * until a reference cannot be resolved.
1675 */
1676 gl_shader *main = NULL;
1677 for (unsigned i = 0; i < num_shaders; i++) {
1678 if (link_get_main_function_signature(shader_list[i]) != NULL) {
1679 main = shader_list[i];
1680 break;
1681 }
1682 }
1683
1684 if (main == NULL) {
1685 linker_error(prog, "%s shader lacks `main'\n",
1686 _mesa_shader_stage_to_string(shader_list[0]->Stage));
1687 return NULL;
1688 }
1689
1690 gl_shader *linked = ctx->Driver.NewShader(NULL, 0, main->Type);
1691 linked->ir = new(linked) exec_list;
1692 clone_ir_list(mem_ctx, linked->ir, main->ir);
1693
1694 linked->UniformBlocks = uniform_blocks;
1695 linked->NumUniformBlocks = num_uniform_blocks;
1696 ralloc_steal(linked, linked->UniformBlocks);
1697
1698 link_fs_input_layout_qualifiers(prog, linked, shader_list, num_shaders);
1699 link_gs_inout_layout_qualifiers(prog, linked, shader_list, num_shaders);
1700 link_cs_input_layout_qualifiers(prog, linked, shader_list, num_shaders);
1701
1702 populate_symbol_table(linked);
1703
1704 /* The pointer to the main function in the final linked shader (i.e., the
1705 * copy of the original shader that contained the main function).
1706 */
1707 ir_function_signature *const main_sig =
1708 link_get_main_function_signature(linked);
1709
1710 /* Move any instructions other than variable declarations or function
1711 * declarations into main.
1712 */
1713 exec_node *insertion_point =
1714 move_non_declarations(linked->ir, (exec_node *) &main_sig->body, false,
1715 linked);
1716
1717 for (unsigned i = 0; i < num_shaders; i++) {
1718 if (shader_list[i] == main)
1719 continue;
1720
1721 insertion_point = move_non_declarations(shader_list[i]->ir,
1722 insertion_point, true, linked);
1723 }
1724
1725 /* Check if any shader needs built-in functions. */
1726 bool need_builtins = false;
1727 for (unsigned i = 0; i < num_shaders; i++) {
1728 if (shader_list[i]->uses_builtin_functions) {
1729 need_builtins = true;
1730 break;
1731 }
1732 }
1733
1734 bool ok;
1735 if (need_builtins) {
1736 /* Make a temporary array one larger than shader_list, which will hold
1737 * the built-in function shader as well.
1738 */
1739 gl_shader **linking_shaders = (gl_shader **)
1740 calloc(num_shaders + 1, sizeof(gl_shader *));
1741
1742 ok = linking_shaders != NULL;
1743
1744 if (ok) {
1745 memcpy(linking_shaders, shader_list, num_shaders * sizeof(gl_shader *));
1746 linking_shaders[num_shaders] = _mesa_glsl_get_builtin_function_shader();
1747
1748 ok = link_function_calls(prog, linked, linking_shaders, num_shaders + 1);
1749
1750 free(linking_shaders);
1751 } else {
1752 _mesa_error_no_memory(__func__);
1753 }
1754 } else {
1755 ok = link_function_calls(prog, linked, shader_list, num_shaders);
1756 }
1757
1758
1759 if (!ok) {
1760 ctx->Driver.DeleteShader(ctx, linked);
1761 return NULL;
1762 }
1763
1764 /* At this point linked should contain all of the linked IR, so
1765 * validate it to make sure nothing went wrong.
1766 */
1767 validate_ir_tree(linked->ir);
1768
1769 /* Set the size of geometry shader input arrays */
1770 if (linked->Stage == MESA_SHADER_GEOMETRY) {
1771 unsigned num_vertices = vertices_per_prim(prog->Geom.InputType);
1772 geom_array_resize_visitor input_resize_visitor(num_vertices, prog);
1773 foreach_in_list(ir_instruction, ir, linked->ir) {
1774 ir->accept(&input_resize_visitor);
1775 }
1776 }
1777
1778 if (ctx->Const.VertexID_is_zero_based)
1779 lower_vertex_id(linked);
1780
1781 /* Make a pass over all variable declarations to ensure that arrays with
1782 * unspecified sizes have a size specified. The size is inferred from the
1783 * max_array_access field.
1784 */
1785 array_sizing_visitor v;
1786 v.run(linked->ir);
1787 v.fixup_unnamed_interface_types();
1788
1789 return linked;
1790 }
1791
1792 /**
1793 * Update the sizes of linked shader uniform arrays to the maximum
1794 * array index used.
1795 *
1796 * From page 81 (page 95 of the PDF) of the OpenGL 2.1 spec:
1797 *
1798 * If one or more elements of an array are active,
1799 * GetActiveUniform will return the name of the array in name,
1800 * subject to the restrictions listed above. The type of the array
1801 * is returned in type. The size parameter contains the highest
1802 * array element index used, plus one. The compiler or linker
1803 * determines the highest index used. There will be only one
1804 * active uniform reported by the GL per uniform array.
1805
1806 */
1807 static void
1808 update_array_sizes(struct gl_shader_program *prog)
1809 {
1810 for (unsigned i = 0; i < MESA_SHADER_STAGES; i++) {
1811 if (prog->_LinkedShaders[i] == NULL)
1812 continue;
1813
1814 foreach_in_list(ir_instruction, node, prog->_LinkedShaders[i]->ir) {
1815 ir_variable *const var = node->as_variable();
1816
1817 if ((var == NULL) || (var->data.mode != ir_var_uniform) ||
1818 !var->type->is_array())
1819 continue;
1820
1821 /* GL_ARB_uniform_buffer_object says that std140 uniforms
1822 * will not be eliminated. Since we always do std140, just
1823 * don't resize arrays in UBOs.
1824 *
1825 * Atomic counters are supposed to get deterministic
1826 * locations assigned based on the declaration ordering and
1827 * sizes, array compaction would mess that up.
1828 */
1829 if (var->is_in_uniform_block() || var->type->contains_atomic())
1830 continue;
1831
1832 unsigned int size = var->data.max_array_access;
1833 for (unsigned j = 0; j < MESA_SHADER_STAGES; j++) {
1834 if (prog->_LinkedShaders[j] == NULL)
1835 continue;
1836
1837 foreach_in_list(ir_instruction, node2, prog->_LinkedShaders[j]->ir) {
1838 ir_variable *other_var = node2->as_variable();
1839 if (!other_var)
1840 continue;
1841
1842 if (strcmp(var->name, other_var->name) == 0 &&
1843 other_var->data.max_array_access > size) {
1844 size = other_var->data.max_array_access;
1845 }
1846 }
1847 }
1848
1849 if (size + 1 != var->type->length) {
1850 /* If this is a built-in uniform (i.e., it's backed by some
1851 * fixed-function state), adjust the number of state slots to
1852 * match the new array size. The number of slots per array entry
1853 * is not known. It seems safe to assume that the total number of
1854 * slots is an integer multiple of the number of array elements.
1855 * Determine the number of slots per array element by dividing by
1856 * the old (total) size.
1857 */
1858 const unsigned num_slots = var->get_num_state_slots();
1859 if (num_slots > 0) {
1860 var->set_num_state_slots((size + 1)
1861 * (num_slots / var->type->length));
1862 }
1863
1864 var->type = glsl_type::get_array_instance(var->type->fields.array,
1865 size + 1);
1866 /* FINISHME: We should update the types of array
1867 * dereferences of this variable now.
1868 */
1869 }
1870 }
1871 }
1872 }
1873
1874 /**
1875 * Find a contiguous set of available bits in a bitmask.
1876 *
1877 * \param used_mask Bits representing used (1) and unused (0) locations
1878 * \param needed_count Number of contiguous bits needed.
1879 *
1880 * \return
1881 * Base location of the available bits on success or -1 on failure.
1882 */
1883 int
1884 find_available_slots(unsigned used_mask, unsigned needed_count)
1885 {
1886 unsigned needed_mask = (1 << needed_count) - 1;
1887 const int max_bit_to_test = (8 * sizeof(used_mask)) - needed_count;
1888
1889 /* The comparison to 32 is redundant, but without it GCC emits "warning:
1890 * cannot optimize possibly infinite loops" for the loop below.
1891 */
1892 if ((needed_count == 0) || (max_bit_to_test < 0) || (max_bit_to_test > 32))
1893 return -1;
1894
1895 for (int i = 0; i <= max_bit_to_test; i++) {
1896 if ((needed_mask & ~used_mask) == needed_mask)
1897 return i;
1898
1899 needed_mask <<= 1;
1900 }
1901
1902 return -1;
1903 }
1904
1905
1906 /**
1907 * Assign locations for either VS inputs or FS outputs
1908 *
1909 * \param prog Shader program whose variables need locations assigned
1910 * \param target_index Selector for the program target to receive location
1911 * assignmnets. Must be either \c MESA_SHADER_VERTEX or
1912 * \c MESA_SHADER_FRAGMENT.
1913 * \param max_index Maximum number of generic locations. This corresponds
1914 * to either the maximum number of draw buffers or the
1915 * maximum number of generic attributes.
1916 *
1917 * \return
1918 * If locations are successfully assigned, true is returned. Otherwise an
1919 * error is emitted to the shader link log and false is returned.
1920 */
1921 bool
1922 assign_attribute_or_color_locations(gl_shader_program *prog,
1923 unsigned target_index,
1924 unsigned max_index)
1925 {
1926 /* Mark invalid locations as being used.
1927 */
1928 unsigned used_locations = (max_index >= 32)
1929 ? ~0 : ~((1 << max_index) - 1);
1930
1931 assert((target_index == MESA_SHADER_VERTEX)
1932 || (target_index == MESA_SHADER_FRAGMENT));
1933
1934 gl_shader *const sh = prog->_LinkedShaders[target_index];
1935 if (sh == NULL)
1936 return true;
1937
1938 /* Operate in a total of four passes.
1939 *
1940 * 1. Invalidate the location assignments for all vertex shader inputs.
1941 *
1942 * 2. Assign locations for inputs that have user-defined (via
1943 * glBindVertexAttribLocation) locations and outputs that have
1944 * user-defined locations (via glBindFragDataLocation).
1945 *
1946 * 3. Sort the attributes without assigned locations by number of slots
1947 * required in decreasing order. Fragmentation caused by attribute
1948 * locations assigned by the application may prevent large attributes
1949 * from having enough contiguous space.
1950 *
1951 * 4. Assign locations to any inputs without assigned locations.
1952 */
1953
1954 const int generic_base = (target_index == MESA_SHADER_VERTEX)
1955 ? (int) VERT_ATTRIB_GENERIC0 : (int) FRAG_RESULT_DATA0;
1956
1957 const enum ir_variable_mode direction =
1958 (target_index == MESA_SHADER_VERTEX)
1959 ? ir_var_shader_in : ir_var_shader_out;
1960
1961
1962 /* Temporary storage for the set of attributes that need locations assigned.
1963 */
1964 struct temp_attr {
1965 unsigned slots;
1966 ir_variable *var;
1967
1968 /* Used below in the call to qsort. */
1969 static int compare(const void *a, const void *b)
1970 {
1971 const temp_attr *const l = (const temp_attr *) a;
1972 const temp_attr *const r = (const temp_attr *) b;
1973
1974 /* Reversed because we want a descending order sort below. */
1975 return r->slots - l->slots;
1976 }
1977 } to_assign[16];
1978
1979 unsigned num_attr = 0;
1980
1981 foreach_in_list(ir_instruction, node, sh->ir) {
1982 ir_variable *const var = node->as_variable();
1983
1984 if ((var == NULL) || (var->data.mode != (unsigned) direction))
1985 continue;
1986
1987 if (var->data.explicit_location) {
1988 if ((var->data.location >= (int)(max_index + generic_base))
1989 || (var->data.location < 0)) {
1990 linker_error(prog,
1991 "invalid explicit location %d specified for `%s'\n",
1992 (var->data.location < 0)
1993 ? var->data.location
1994 : var->data.location - generic_base,
1995 var->name);
1996 return false;
1997 }
1998 } else if (target_index == MESA_SHADER_VERTEX) {
1999 unsigned binding;
2000
2001 if (prog->AttributeBindings->get(binding, var->name)) {
2002 assert(binding >= VERT_ATTRIB_GENERIC0);
2003 var->data.location = binding;
2004 var->data.is_unmatched_generic_inout = 0;
2005 }
2006 } else if (target_index == MESA_SHADER_FRAGMENT) {
2007 unsigned binding;
2008 unsigned index;
2009
2010 if (prog->FragDataBindings->get(binding, var->name)) {
2011 assert(binding >= FRAG_RESULT_DATA0);
2012 var->data.location = binding;
2013 var->data.is_unmatched_generic_inout = 0;
2014
2015 if (prog->FragDataIndexBindings->get(index, var->name)) {
2016 var->data.index = index;
2017 }
2018 }
2019 }
2020
2021 /* If the variable is not a built-in and has a location statically
2022 * assigned in the shader (presumably via a layout qualifier), make sure
2023 * that it doesn't collide with other assigned locations. Otherwise,
2024 * add it to the list of variables that need linker-assigned locations.
2025 */
2026 const unsigned slots = var->type->count_attribute_slots();
2027 if (var->data.location != -1) {
2028 if (var->data.location >= generic_base && var->data.index < 1) {
2029 /* From page 61 of the OpenGL 4.0 spec:
2030 *
2031 * "LinkProgram will fail if the attribute bindings assigned
2032 * by BindAttribLocation do not leave not enough space to
2033 * assign a location for an active matrix attribute or an
2034 * active attribute array, both of which require multiple
2035 * contiguous generic attributes."
2036 *
2037 * I think above text prohibits the aliasing of explicit and
2038 * automatic assignments. But, aliasing is allowed in manual
2039 * assignments of attribute locations. See below comments for
2040 * the details.
2041 *
2042 * From OpenGL 4.0 spec, page 61:
2043 *
2044 * "It is possible for an application to bind more than one
2045 * attribute name to the same location. This is referred to as
2046 * aliasing. This will only work if only one of the aliased
2047 * attributes is active in the executable program, or if no
2048 * path through the shader consumes more than one attribute of
2049 * a set of attributes aliased to the same location. A link
2050 * error can occur if the linker determines that every path
2051 * through the shader consumes multiple aliased attributes,
2052 * but implementations are not required to generate an error
2053 * in this case."
2054 *
2055 * From GLSL 4.30 spec, page 54:
2056 *
2057 * "A program will fail to link if any two non-vertex shader
2058 * input variables are assigned to the same location. For
2059 * vertex shaders, multiple input variables may be assigned
2060 * to the same location using either layout qualifiers or via
2061 * the OpenGL API. However, such aliasing is intended only to
2062 * support vertex shaders where each execution path accesses
2063 * at most one input per each location. Implementations are
2064 * permitted, but not required, to generate link-time errors
2065 * if they detect that every path through the vertex shader
2066 * executable accesses multiple inputs assigned to any single
2067 * location. For all shader types, a program will fail to link
2068 * if explicit location assignments leave the linker unable
2069 * to find space for other variables without explicit
2070 * assignments."
2071 *
2072 * From OpenGL ES 3.0 spec, page 56:
2073 *
2074 * "Binding more than one attribute name to the same location
2075 * is referred to as aliasing, and is not permitted in OpenGL
2076 * ES Shading Language 3.00 vertex shaders. LinkProgram will
2077 * fail when this condition exists. However, aliasing is
2078 * possible in OpenGL ES Shading Language 1.00 vertex shaders.
2079 * This will only work if only one of the aliased attributes
2080 * is active in the executable program, or if no path through
2081 * the shader consumes more than one attribute of a set of
2082 * attributes aliased to the same location. A link error can
2083 * occur if the linker determines that every path through the
2084 * shader consumes multiple aliased attributes, but implemen-
2085 * tations are not required to generate an error in this case."
2086 *
2087 * After looking at above references from OpenGL, OpenGL ES and
2088 * GLSL specifications, we allow aliasing of vertex input variables
2089 * in: OpenGL 2.0 (and above) and OpenGL ES 2.0.
2090 *
2091 * NOTE: This is not required by the spec but its worth mentioning
2092 * here that we're not doing anything to make sure that no path
2093 * through the vertex shader executable accesses multiple inputs
2094 * assigned to any single location.
2095 */
2096
2097 /* Mask representing the contiguous slots that will be used by
2098 * this attribute.
2099 */
2100 const unsigned attr = var->data.location - generic_base;
2101 const unsigned use_mask = (1 << slots) - 1;
2102 const char *const string = (target_index == MESA_SHADER_VERTEX)
2103 ? "vertex shader input" : "fragment shader output";
2104
2105 /* Generate a link error if the requested locations for this
2106 * attribute exceed the maximum allowed attribute location.
2107 */
2108 if (attr + slots > max_index) {
2109 linker_error(prog,
2110 "insufficient contiguous locations "
2111 "available for %s `%s' %d %d %d\n", string,
2112 var->name, used_locations, use_mask, attr);
2113 return false;
2114 }
2115
2116 /* Generate a link error if the set of bits requested for this
2117 * attribute overlaps any previously allocated bits.
2118 */
2119 if ((~(use_mask << attr) & used_locations) != used_locations) {
2120 if (target_index == MESA_SHADER_FRAGMENT ||
2121 (prog->IsES && prog->Version >= 300)) {
2122 linker_error(prog,
2123 "overlapping location is assigned "
2124 "to %s `%s' %d %d %d\n", string,
2125 var->name, used_locations, use_mask, attr);
2126 return false;
2127 } else {
2128 linker_warning(prog,
2129 "overlapping location is assigned "
2130 "to %s `%s' %d %d %d\n", string,
2131 var->name, used_locations, use_mask, attr);
2132 }
2133 }
2134
2135 used_locations |= (use_mask << attr);
2136 }
2137
2138 continue;
2139 }
2140
2141 to_assign[num_attr].slots = slots;
2142 to_assign[num_attr].var = var;
2143 num_attr++;
2144 }
2145
2146 /* If all of the attributes were assigned locations by the application (or
2147 * are built-in attributes with fixed locations), return early. This should
2148 * be the common case.
2149 */
2150 if (num_attr == 0)
2151 return true;
2152
2153 qsort(to_assign, num_attr, sizeof(to_assign[0]), temp_attr::compare);
2154
2155 if (target_index == MESA_SHADER_VERTEX) {
2156 /* VERT_ATTRIB_GENERIC0 is a pseudo-alias for VERT_ATTRIB_POS. It can
2157 * only be explicitly assigned by via glBindAttribLocation. Mark it as
2158 * reserved to prevent it from being automatically allocated below.
2159 */
2160 find_deref_visitor find("gl_Vertex");
2161 find.run(sh->ir);
2162 if (find.variable_found())
2163 used_locations |= (1 << 0);
2164 }
2165
2166 for (unsigned i = 0; i < num_attr; i++) {
2167 /* Mask representing the contiguous slots that will be used by this
2168 * attribute.
2169 */
2170 const unsigned use_mask = (1 << to_assign[i].slots) - 1;
2171
2172 int location = find_available_slots(used_locations, to_assign[i].slots);
2173
2174 if (location < 0) {
2175 const char *const string = (target_index == MESA_SHADER_VERTEX)
2176 ? "vertex shader input" : "fragment shader output";
2177
2178 linker_error(prog,
2179 "insufficient contiguous locations "
2180 "available for %s `%s'\n",
2181 string, to_assign[i].var->name);
2182 return false;
2183 }
2184
2185 to_assign[i].var->data.location = generic_base + location;
2186 to_assign[i].var->data.is_unmatched_generic_inout = 0;
2187 used_locations |= (use_mask << location);
2188 }
2189
2190 return true;
2191 }
2192
2193
2194 /**
2195 * Demote shader inputs and outputs that are not used in other stages
2196 */
2197 void
2198 demote_shader_inputs_and_outputs(gl_shader *sh, enum ir_variable_mode mode)
2199 {
2200 foreach_in_list(ir_instruction, node, sh->ir) {
2201 ir_variable *const var = node->as_variable();
2202
2203 if ((var == NULL) || (var->data.mode != int(mode)))
2204 continue;
2205
2206 /* A shader 'in' or 'out' variable is only really an input or output if
2207 * its value is used by other shader stages. This will cause the variable
2208 * to have a location assigned.
2209 */
2210 if (var->data.is_unmatched_generic_inout) {
2211 assert(var->data.mode != ir_var_temporary);
2212 var->data.mode = ir_var_auto;
2213 }
2214 }
2215 }
2216
2217
2218 /**
2219 * Store the gl_FragDepth layout in the gl_shader_program struct.
2220 */
2221 static void
2222 store_fragdepth_layout(struct gl_shader_program *prog)
2223 {
2224 if (prog->_LinkedShaders[MESA_SHADER_FRAGMENT] == NULL) {
2225 return;
2226 }
2227
2228 struct exec_list *ir = prog->_LinkedShaders[MESA_SHADER_FRAGMENT]->ir;
2229
2230 /* We don't look up the gl_FragDepth symbol directly because if
2231 * gl_FragDepth is not used in the shader, it's removed from the IR.
2232 * However, the symbol won't be removed from the symbol table.
2233 *
2234 * We're only interested in the cases where the variable is NOT removed
2235 * from the IR.
2236 */
2237 foreach_in_list(ir_instruction, node, ir) {
2238 ir_variable *const var = node->as_variable();
2239
2240 if (var == NULL || var->data.mode != ir_var_shader_out) {
2241 continue;
2242 }
2243
2244 if (strcmp(var->name, "gl_FragDepth") == 0) {
2245 switch (var->data.depth_layout) {
2246 case ir_depth_layout_none:
2247 prog->FragDepthLayout = FRAG_DEPTH_LAYOUT_NONE;
2248 return;
2249 case ir_depth_layout_any:
2250 prog->FragDepthLayout = FRAG_DEPTH_LAYOUT_ANY;
2251 return;
2252 case ir_depth_layout_greater:
2253 prog->FragDepthLayout = FRAG_DEPTH_LAYOUT_GREATER;
2254 return;
2255 case ir_depth_layout_less:
2256 prog->FragDepthLayout = FRAG_DEPTH_LAYOUT_LESS;
2257 return;
2258 case ir_depth_layout_unchanged:
2259 prog->FragDepthLayout = FRAG_DEPTH_LAYOUT_UNCHANGED;
2260 return;
2261 default:
2262 assert(0);
2263 return;
2264 }
2265 }
2266 }
2267 }
2268
2269 /**
2270 * Validate the resources used by a program versus the implementation limits
2271 */
2272 static void
2273 check_resources(struct gl_context *ctx, struct gl_shader_program *prog)
2274 {
2275 for (unsigned i = 0; i < MESA_SHADER_STAGES; i++) {
2276 struct gl_shader *sh = prog->_LinkedShaders[i];
2277
2278 if (sh == NULL)
2279 continue;
2280
2281 if (sh->num_samplers > ctx->Const.Program[i].MaxTextureImageUnits) {
2282 linker_error(prog, "Too many %s shader texture samplers\n",
2283 _mesa_shader_stage_to_string(i));
2284 }
2285
2286 if (sh->num_uniform_components >
2287 ctx->Const.Program[i].MaxUniformComponents) {
2288 if (ctx->Const.GLSLSkipStrictMaxUniformLimitCheck) {
2289 linker_warning(prog, "Too many %s shader default uniform block "
2290 "components, but the driver will try to optimize "
2291 "them out; this is non-portable out-of-spec "
2292 "behavior\n",
2293 _mesa_shader_stage_to_string(i));
2294 } else {
2295 linker_error(prog, "Too many %s shader default uniform block "
2296 "components\n",
2297 _mesa_shader_stage_to_string(i));
2298 }
2299 }
2300
2301 if (sh->num_combined_uniform_components >
2302 ctx->Const.Program[i].MaxCombinedUniformComponents) {
2303 if (ctx->Const.GLSLSkipStrictMaxUniformLimitCheck) {
2304 linker_warning(prog, "Too many %s shader uniform components, "
2305 "but the driver will try to optimize them out; "
2306 "this is non-portable out-of-spec behavior\n",
2307 _mesa_shader_stage_to_string(i));
2308 } else {
2309 linker_error(prog, "Too many %s shader uniform components\n",
2310 _mesa_shader_stage_to_string(i));
2311 }
2312 }
2313 }
2314
2315 unsigned blocks[MESA_SHADER_STAGES] = {0};
2316 unsigned total_uniform_blocks = 0;
2317
2318 for (unsigned i = 0; i < prog->NumUniformBlocks; i++) {
2319 for (unsigned j = 0; j < MESA_SHADER_STAGES; j++) {
2320 if (prog->UniformBlockStageIndex[j][i] != -1) {
2321 blocks[j]++;
2322 total_uniform_blocks++;
2323 }
2324 }
2325
2326 if (total_uniform_blocks > ctx->Const.MaxCombinedUniformBlocks) {
2327 linker_error(prog, "Too many combined uniform blocks (%d/%d)\n",
2328 prog->NumUniformBlocks,
2329 ctx->Const.MaxCombinedUniformBlocks);
2330 } else {
2331 for (unsigned i = 0; i < MESA_SHADER_STAGES; i++) {
2332 const unsigned max_uniform_blocks =
2333 ctx->Const.Program[i].MaxUniformBlocks;
2334 if (blocks[i] > max_uniform_blocks) {
2335 linker_error(prog, "Too many %s uniform blocks (%d/%d)\n",
2336 _mesa_shader_stage_to_string(i),
2337 blocks[i],
2338 max_uniform_blocks);
2339 break;
2340 }
2341 }
2342 }
2343 }
2344 }
2345
2346 /**
2347 * Validate shader image resources.
2348 */
2349 static void
2350 check_image_resources(struct gl_context *ctx, struct gl_shader_program *prog)
2351 {
2352 unsigned total_image_units = 0;
2353 unsigned fragment_outputs = 0;
2354
2355 if (!ctx->Extensions.ARB_shader_image_load_store)
2356 return;
2357
2358 for (unsigned i = 0; i < MESA_SHADER_STAGES; i++) {
2359 struct gl_shader *sh = prog->_LinkedShaders[i];
2360
2361 if (sh) {
2362 if (sh->NumImages > ctx->Const.Program[i].MaxImageUniforms)
2363 linker_error(prog, "Too many %s shader image uniforms\n",
2364 _mesa_shader_stage_to_string(i));
2365
2366 total_image_units += sh->NumImages;
2367
2368 if (i == MESA_SHADER_FRAGMENT) {
2369 foreach_in_list(ir_instruction, node, sh->ir) {
2370 ir_variable *var = node->as_variable();
2371 if (var && var->data.mode == ir_var_shader_out)
2372 fragment_outputs += var->type->count_attribute_slots();
2373 }
2374 }
2375 }
2376 }
2377
2378 if (total_image_units > ctx->Const.MaxCombinedImageUniforms)
2379 linker_error(prog, "Too many combined image uniforms\n");
2380
2381 if (total_image_units + fragment_outputs >
2382 ctx->Const.MaxCombinedImageUnitsAndFragmentOutputs)
2383 linker_error(prog, "Too many combined image uniforms and fragment outputs\n");
2384 }
2385
2386
2387 /**
2388 * Initializes explicit location slots to INACTIVE_UNIFORM_EXPLICIT_LOCATION
2389 * for a variable, checks for overlaps between other uniforms using explicit
2390 * locations.
2391 */
2392 static bool
2393 reserve_explicit_locations(struct gl_shader_program *prog,
2394 string_to_uint_map *map, ir_variable *var)
2395 {
2396 unsigned slots = var->type->uniform_locations();
2397 unsigned max_loc = var->data.location + slots - 1;
2398
2399 /* Resize remap table if locations do not fit in the current one. */
2400 if (max_loc + 1 > prog->NumUniformRemapTable) {
2401 prog->UniformRemapTable =
2402 reralloc(prog, prog->UniformRemapTable,
2403 gl_uniform_storage *,
2404 max_loc + 1);
2405
2406 if (!prog->UniformRemapTable) {
2407 linker_error(prog, "Out of memory during linking.\n");
2408 return false;
2409 }
2410
2411 /* Initialize allocated space. */
2412 for (unsigned i = prog->NumUniformRemapTable; i < max_loc + 1; i++)
2413 prog->UniformRemapTable[i] = NULL;
2414
2415 prog->NumUniformRemapTable = max_loc + 1;
2416 }
2417
2418 for (unsigned i = 0; i < slots; i++) {
2419 unsigned loc = var->data.location + i;
2420
2421 /* Check if location is already used. */
2422 if (prog->UniformRemapTable[loc] == INACTIVE_UNIFORM_EXPLICIT_LOCATION) {
2423
2424 /* Possibly same uniform from a different stage, this is ok. */
2425 unsigned hash_loc;
2426 if (map->get(hash_loc, var->name) && hash_loc == loc - i)
2427 continue;
2428
2429 /* ARB_explicit_uniform_location specification states:
2430 *
2431 * "No two default-block uniform variables in the program can have
2432 * the same location, even if they are unused, otherwise a compiler
2433 * or linker error will be generated."
2434 */
2435 linker_error(prog,
2436 "location qualifier for uniform %s overlaps "
2437 "previously used location\n",
2438 var->name);
2439 return false;
2440 }
2441
2442 /* Initialize location as inactive before optimization
2443 * rounds and location assignment.
2444 */
2445 prog->UniformRemapTable[loc] = INACTIVE_UNIFORM_EXPLICIT_LOCATION;
2446 }
2447
2448 /* Note, base location used for arrays. */
2449 map->put(var->data.location, var->name);
2450
2451 return true;
2452 }
2453
2454 /**
2455 * Check and reserve all explicit uniform locations, called before
2456 * any optimizations happen to handle also inactive uniforms and
2457 * inactive array elements that may get trimmed away.
2458 */
2459 static void
2460 check_explicit_uniform_locations(struct gl_context *ctx,
2461 struct gl_shader_program *prog)
2462 {
2463 if (!ctx->Extensions.ARB_explicit_uniform_location)
2464 return;
2465
2466 /* This map is used to detect if overlapping explicit locations
2467 * occur with the same uniform (from different stage) or a different one.
2468 */
2469 string_to_uint_map *uniform_map = new string_to_uint_map;
2470
2471 if (!uniform_map) {
2472 linker_error(prog, "Out of memory during linking.\n");
2473 return;
2474 }
2475
2476 for (unsigned i = 0; i < MESA_SHADER_STAGES; i++) {
2477 struct gl_shader *sh = prog->_LinkedShaders[i];
2478
2479 if (!sh)
2480 continue;
2481
2482 foreach_in_list(ir_instruction, node, sh->ir) {
2483 ir_variable *var = node->as_variable();
2484 if ((var && var->data.mode == ir_var_uniform) &&
2485 var->data.explicit_location) {
2486 if (!reserve_explicit_locations(prog, uniform_map, var)) {
2487 delete uniform_map;
2488 return;
2489 }
2490 }
2491 }
2492 }
2493
2494 delete uniform_map;
2495 }
2496
2497 static bool
2498 add_program_resource(struct gl_shader_program *prog, GLenum type,
2499 const void *data, uint8_t stages)
2500 {
2501 assert(data);
2502
2503 /* If resource already exists, do not add it again. */
2504 for (unsigned i = 0; i < prog->NumProgramResourceList; i++)
2505 if (prog->ProgramResourceList[i].Data == data)
2506 return true;
2507
2508 prog->ProgramResourceList =
2509 reralloc(prog,
2510 prog->ProgramResourceList,
2511 gl_program_resource,
2512 prog->NumProgramResourceList + 1);
2513
2514 if (!prog->ProgramResourceList) {
2515 linker_error(prog, "Out of memory during linking.\n");
2516 return false;
2517 }
2518
2519 struct gl_program_resource *res =
2520 &prog->ProgramResourceList[prog->NumProgramResourceList];
2521
2522 res->Type = type;
2523 res->Data = data;
2524 res->StageReferences = stages;
2525
2526 prog->NumProgramResourceList++;
2527
2528 return true;
2529 }
2530
2531 /**
2532 * Function builds a stage reference bitmask from variable name.
2533 */
2534 static uint8_t
2535 build_stageref(struct gl_shader_program *shProg, const char *name)
2536 {
2537 uint8_t stages = 0;
2538
2539 /* Note, that we assume MAX 8 stages, if there will be more stages, type
2540 * used for reference mask in gl_program_resource will need to be changed.
2541 */
2542 assert(MESA_SHADER_STAGES < 8);
2543
2544 for (unsigned i = 0; i < MESA_SHADER_STAGES; i++) {
2545 struct gl_shader *sh = shProg->_LinkedShaders[i];
2546 if (!sh)
2547 continue;
2548 ir_variable *var = sh->symbols->get_variable(name);
2549 if (var)
2550 stages |= (1 << i);
2551 }
2552 return stages;
2553 }
2554
2555 static bool
2556 add_interface_variables(struct gl_shader_program *shProg,
2557 struct gl_shader *sh, GLenum programInterface)
2558 {
2559 foreach_in_list(ir_instruction, node, sh->ir) {
2560 ir_variable *var = node->as_variable();
2561 uint8_t mask = 0;
2562
2563 if (!var)
2564 continue;
2565
2566 switch (var->data.mode) {
2567 /* From GL 4.3 core spec, section 11.1.1 (Vertex Attributes):
2568 * "For GetActiveAttrib, all active vertex shader input variables
2569 * are enumerated, including the special built-in inputs gl_VertexID
2570 * and gl_InstanceID."
2571 */
2572 case ir_var_system_value:
2573 if (var->data.location != SYSTEM_VALUE_VERTEX_ID &&
2574 var->data.location != SYSTEM_VALUE_VERTEX_ID_ZERO_BASE &&
2575 var->data.location != SYSTEM_VALUE_INSTANCE_ID)
2576 continue;
2577 /* Mark special built-in inputs referenced by the vertex stage so
2578 * that they are considered active by the shader queries.
2579 */
2580 mask = (1 << (MESA_SHADER_VERTEX));
2581 /* FALLTHROUGH */
2582 case ir_var_shader_in:
2583 if (programInterface != GL_PROGRAM_INPUT)
2584 continue;
2585 break;
2586 case ir_var_shader_out:
2587 if (programInterface != GL_PROGRAM_OUTPUT)
2588 continue;
2589 break;
2590 default:
2591 continue;
2592 };
2593
2594 if (!add_program_resource(shProg, programInterface, var,
2595 build_stageref(shProg, var->name) | mask))
2596 return false;
2597 }
2598 return true;
2599 }
2600
2601 /**
2602 * Builds up a list of program resources that point to existing
2603 * resource data.
2604 */
2605 static void
2606 build_program_resource_list(struct gl_context *ctx,
2607 struct gl_shader_program *shProg)
2608 {
2609 /* Rebuild resource list. */
2610 if (shProg->ProgramResourceList) {
2611 ralloc_free(shProg->ProgramResourceList);
2612 shProg->ProgramResourceList = NULL;
2613 shProg->NumProgramResourceList = 0;
2614 }
2615
2616 int input_stage = MESA_SHADER_STAGES, output_stage = 0;
2617
2618 /* Determine first input and final output stage. These are used to
2619 * detect which variables should be enumerated in the resource list
2620 * for GL_PROGRAM_INPUT and GL_PROGRAM_OUTPUT.
2621 */
2622 for (unsigned i = 0; i < MESA_SHADER_STAGES; i++) {
2623 if (!shProg->_LinkedShaders[i])
2624 continue;
2625 if (input_stage == MESA_SHADER_STAGES)
2626 input_stage = i;
2627 output_stage = i;
2628 }
2629
2630 /* Empty shader, no resources. */
2631 if (input_stage == MESA_SHADER_STAGES && output_stage == 0)
2632 return;
2633
2634 /* Add inputs and outputs to the resource list. */
2635 if (!add_interface_variables(shProg, shProg->_LinkedShaders[input_stage],
2636 GL_PROGRAM_INPUT))
2637 return;
2638
2639 if (!add_interface_variables(shProg, shProg->_LinkedShaders[output_stage],
2640 GL_PROGRAM_OUTPUT))
2641 return;
2642
2643 /* Add transform feedback varyings. */
2644 if (shProg->LinkedTransformFeedback.NumVarying > 0) {
2645 for (int i = 0; i < shProg->LinkedTransformFeedback.NumVarying; i++) {
2646 uint8_t stageref =
2647 build_stageref(shProg,
2648 shProg->LinkedTransformFeedback.Varyings[i].Name);
2649 if (!add_program_resource(shProg, GL_TRANSFORM_FEEDBACK_VARYING,
2650 &shProg->LinkedTransformFeedback.Varyings[i],
2651 stageref))
2652 return;
2653 }
2654 }
2655
2656 /* Add uniforms from uniform storage. */
2657 for (unsigned i = 0; i < shProg->NumUserUniformStorage; i++) {
2658 /* Do not add uniforms internally used by Mesa. */
2659 if (shProg->UniformStorage[i].hidden)
2660 continue;
2661
2662 uint8_t stageref =
2663 build_stageref(shProg, shProg->UniformStorage[i].name);
2664 if (!add_program_resource(shProg, GL_UNIFORM,
2665 &shProg->UniformStorage[i], stageref))
2666 return;
2667 }
2668
2669 /* Add program uniform blocks. */
2670 for (unsigned i = 0; i < shProg->NumUniformBlocks; i++) {
2671 if (!add_program_resource(shProg, GL_UNIFORM_BLOCK,
2672 &shProg->UniformBlocks[i], 0))
2673 return;
2674 }
2675
2676 /* Add atomic counter buffers. */
2677 for (unsigned i = 0; i < shProg->NumAtomicBuffers; i++) {
2678 if (!add_program_resource(shProg, GL_ATOMIC_COUNTER_BUFFER,
2679 &shProg->AtomicBuffers[i], 0))
2680 return;
2681 }
2682
2683 /* TODO - following extensions will require more resource types:
2684 *
2685 * GL_ARB_shader_storage_buffer_object
2686 * GL_ARB_shader_subroutine
2687 */
2688 }
2689
2690
2691 void
2692 link_shaders(struct gl_context *ctx, struct gl_shader_program *prog)
2693 {
2694 tfeedback_decl *tfeedback_decls = NULL;
2695 unsigned num_tfeedback_decls = prog->TransformFeedback.NumVarying;
2696
2697 void *mem_ctx = ralloc_context(NULL); // temporary linker context
2698
2699 prog->LinkStatus = true; /* All error paths will set this to false */
2700 prog->Validated = false;
2701 prog->_Used = false;
2702
2703 prog->ARB_fragment_coord_conventions_enable = false;
2704
2705 /* Separate the shaders into groups based on their type.
2706 */
2707 struct gl_shader **shader_list[MESA_SHADER_STAGES];
2708 unsigned num_shaders[MESA_SHADER_STAGES];
2709
2710 for (int i = 0; i < MESA_SHADER_STAGES; i++) {
2711 shader_list[i] = (struct gl_shader **)
2712 calloc(prog->NumShaders, sizeof(struct gl_shader *));
2713 num_shaders[i] = 0;
2714 }
2715
2716 unsigned min_version = UINT_MAX;
2717 unsigned max_version = 0;
2718 const bool is_es_prog =
2719 (prog->NumShaders > 0 && prog->Shaders[0]->IsES) ? true : false;
2720 for (unsigned i = 0; i < prog->NumShaders; i++) {
2721 min_version = MIN2(min_version, prog->Shaders[i]->Version);
2722 max_version = MAX2(max_version, prog->Shaders[i]->Version);
2723
2724 if (prog->Shaders[i]->IsES != is_es_prog) {
2725 linker_error(prog, "all shaders must use same shading "
2726 "language version\n");
2727 goto done;
2728 }
2729
2730 if (prog->Shaders[i]->ARB_fragment_coord_conventions_enable) {
2731 prog->ARB_fragment_coord_conventions_enable = true;
2732 }
2733
2734 gl_shader_stage shader_type = prog->Shaders[i]->Stage;
2735 shader_list[shader_type][num_shaders[shader_type]] = prog->Shaders[i];
2736 num_shaders[shader_type]++;
2737 }
2738
2739 /* In desktop GLSL, different shader versions may be linked together. In
2740 * GLSL ES, all shader versions must be the same.
2741 */
2742 if (is_es_prog && min_version != max_version) {
2743 linker_error(prog, "all shaders must use same shading "
2744 "language version\n");
2745 goto done;
2746 }
2747
2748 prog->Version = max_version;
2749 prog->IsES = is_es_prog;
2750
2751 /* Geometry shaders have to be linked with vertex shaders.
2752 */
2753 if (num_shaders[MESA_SHADER_GEOMETRY] > 0 &&
2754 num_shaders[MESA_SHADER_VERTEX] == 0 &&
2755 !prog->SeparateShader) {
2756 linker_error(prog, "Geometry shader must be linked with "
2757 "vertex shader\n");
2758 goto done;
2759 }
2760
2761 /* Compute shaders have additional restrictions. */
2762 if (num_shaders[MESA_SHADER_COMPUTE] > 0 &&
2763 num_shaders[MESA_SHADER_COMPUTE] != prog->NumShaders) {
2764 linker_error(prog, "Compute shaders may not be linked with any other "
2765 "type of shader\n");
2766 }
2767
2768 for (unsigned int i = 0; i < MESA_SHADER_STAGES; i++) {
2769 if (prog->_LinkedShaders[i] != NULL)
2770 ctx->Driver.DeleteShader(ctx, prog->_LinkedShaders[i]);
2771
2772 prog->_LinkedShaders[i] = NULL;
2773 }
2774
2775 /* Link all shaders for a particular stage and validate the result.
2776 */
2777 for (int stage = 0; stage < MESA_SHADER_STAGES; stage++) {
2778 if (num_shaders[stage] > 0) {
2779 gl_shader *const sh =
2780 link_intrastage_shaders(mem_ctx, ctx, prog, shader_list[stage],
2781 num_shaders[stage]);
2782
2783 if (!prog->LinkStatus)
2784 goto done;
2785
2786 switch (stage) {
2787 case MESA_SHADER_VERTEX:
2788 validate_vertex_shader_executable(prog, sh);
2789 break;
2790 case MESA_SHADER_GEOMETRY:
2791 validate_geometry_shader_executable(prog, sh);
2792 break;
2793 case MESA_SHADER_FRAGMENT:
2794 validate_fragment_shader_executable(prog, sh);
2795 break;
2796 }
2797 if (!prog->LinkStatus)
2798 goto done;
2799
2800 _mesa_reference_shader(ctx, &prog->_LinkedShaders[stage], sh);
2801 }
2802 }
2803
2804 if (num_shaders[MESA_SHADER_GEOMETRY] > 0)
2805 prog->LastClipDistanceArraySize = prog->Geom.ClipDistanceArraySize;
2806 else if (num_shaders[MESA_SHADER_VERTEX] > 0)
2807 prog->LastClipDistanceArraySize = prog->Vert.ClipDistanceArraySize;
2808 else
2809 prog->LastClipDistanceArraySize = 0; /* Not used */
2810
2811 /* Here begins the inter-stage linking phase. Some initial validation is
2812 * performed, then locations are assigned for uniforms, attributes, and
2813 * varyings.
2814 */
2815 cross_validate_uniforms(prog);
2816 if (!prog->LinkStatus)
2817 goto done;
2818
2819 unsigned prev;
2820
2821 for (prev = 0; prev <= MESA_SHADER_FRAGMENT; prev++) {
2822 if (prog->_LinkedShaders[prev] != NULL)
2823 break;
2824 }
2825
2826 check_explicit_uniform_locations(ctx, prog);
2827 if (!prog->LinkStatus)
2828 goto done;
2829
2830 /* Validate the inputs of each stage with the output of the preceding
2831 * stage.
2832 */
2833 for (unsigned i = prev + 1; i <= MESA_SHADER_FRAGMENT; i++) {
2834 if (prog->_LinkedShaders[i] == NULL)
2835 continue;
2836
2837 validate_interstage_inout_blocks(prog, prog->_LinkedShaders[prev],
2838 prog->_LinkedShaders[i]);
2839 if (!prog->LinkStatus)
2840 goto done;
2841
2842 cross_validate_outputs_to_inputs(prog,
2843 prog->_LinkedShaders[prev],
2844 prog->_LinkedShaders[i]);
2845 if (!prog->LinkStatus)
2846 goto done;
2847
2848 prev = i;
2849 }
2850
2851 /* Cross-validate uniform blocks between shader stages */
2852 validate_interstage_uniform_blocks(prog, prog->_LinkedShaders,
2853 MESA_SHADER_STAGES);
2854 if (!prog->LinkStatus)
2855 goto done;
2856
2857 for (unsigned int i = 0; i < MESA_SHADER_STAGES; i++) {
2858 if (prog->_LinkedShaders[i] != NULL)
2859 lower_named_interface_blocks(mem_ctx, prog->_LinkedShaders[i]);
2860 }
2861
2862 /* Implement the GLSL 1.30+ rule for discard vs infinite loops Do
2863 * it before optimization because we want most of the checks to get
2864 * dropped thanks to constant propagation.
2865 *
2866 * This rule also applies to GLSL ES 3.00.
2867 */
2868 if (max_version >= (is_es_prog ? 300 : 130)) {
2869 struct gl_shader *sh = prog->_LinkedShaders[MESA_SHADER_FRAGMENT];
2870 if (sh) {
2871 lower_discard_flow(sh->ir);
2872 }
2873 }
2874
2875 if (!interstage_cross_validate_uniform_blocks(prog))
2876 goto done;
2877
2878 /* Do common optimization before assigning storage for attributes,
2879 * uniforms, and varyings. Later optimization could possibly make
2880 * some of that unused.
2881 */
2882 for (unsigned i = 0; i < MESA_SHADER_STAGES; i++) {
2883 if (prog->_LinkedShaders[i] == NULL)
2884 continue;
2885
2886 detect_recursion_linked(prog, prog->_LinkedShaders[i]->ir);
2887 if (!prog->LinkStatus)
2888 goto done;
2889
2890 if (ctx->Const.ShaderCompilerOptions[i].LowerClipDistance) {
2891 lower_clip_distance(prog->_LinkedShaders[i]);
2892 }
2893
2894 while (do_common_optimization(prog->_LinkedShaders[i]->ir, true, false,
2895 &ctx->Const.ShaderCompilerOptions[i],
2896 ctx->Const.NativeIntegers))
2897 ;
2898
2899 lower_const_arrays_to_uniforms(prog->_LinkedShaders[i]->ir);
2900 }
2901
2902 /* Check and validate stream emissions in geometry shaders */
2903 validate_geometry_shader_emissions(ctx, prog);
2904
2905 /* Mark all generic shader inputs and outputs as unpaired. */
2906 for (unsigned i = MESA_SHADER_VERTEX; i <= MESA_SHADER_FRAGMENT; i++) {
2907 if (prog->_LinkedShaders[i] != NULL) {
2908 link_invalidate_variable_locations(prog->_LinkedShaders[i]->ir);
2909 }
2910 }
2911
2912 /* FINISHME: The value of the max_attribute_index parameter is
2913 * FINISHME: implementation dependent based on the value of
2914 * FINISHME: GL_MAX_VERTEX_ATTRIBS. GL_MAX_VERTEX_ATTRIBS must be
2915 * FINISHME: at least 16, so hardcode 16 for now.
2916 */
2917 if (!assign_attribute_or_color_locations(prog, MESA_SHADER_VERTEX, 16)) {
2918 goto done;
2919 }
2920
2921 if (!assign_attribute_or_color_locations(prog, MESA_SHADER_FRAGMENT, MAX2(ctx->Const.MaxDrawBuffers, ctx->Const.MaxDualSourceDrawBuffers))) {
2922 goto done;
2923 }
2924
2925 unsigned first, last;
2926
2927 first = MESA_SHADER_STAGES;
2928 last = 0;
2929
2930 /* Determine first and last stage. */
2931 for (unsigned i = 0; i < MESA_SHADER_STAGES; i++) {
2932 if (!prog->_LinkedShaders[i])
2933 continue;
2934 if (first == MESA_SHADER_STAGES)
2935 first = i;
2936 last = i;
2937 }
2938
2939 if (num_tfeedback_decls != 0) {
2940 /* From GL_EXT_transform_feedback:
2941 * A program will fail to link if:
2942 *
2943 * * the <count> specified by TransformFeedbackVaryingsEXT is
2944 * non-zero, but the program object has no vertex or geometry
2945 * shader;
2946 */
2947 if (first == MESA_SHADER_FRAGMENT) {
2948 linker_error(prog, "Transform feedback varyings specified, but "
2949 "no vertex or geometry shader is present.\n");
2950 goto done;
2951 }
2952
2953 tfeedback_decls = ralloc_array(mem_ctx, tfeedback_decl,
2954 prog->TransformFeedback.NumVarying);
2955 if (!parse_tfeedback_decls(ctx, prog, mem_ctx, num_tfeedback_decls,
2956 prog->TransformFeedback.VaryingNames,
2957 tfeedback_decls))
2958 goto done;
2959 }
2960
2961 /* Linking the stages in the opposite order (from fragment to vertex)
2962 * ensures that inter-shader outputs written to in an earlier stage are
2963 * eliminated if they are (transitively) not used in a later stage.
2964 */
2965 int next;
2966
2967 if (first < MESA_SHADER_FRAGMENT) {
2968 gl_shader *const sh = prog->_LinkedShaders[last];
2969
2970 if (first == MESA_SHADER_GEOMETRY) {
2971 /* There was no vertex shader, but we still have to assign varying
2972 * locations for use by geometry shader inputs in SSO.
2973 *
2974 * If the shader is not separable (i.e., prog->SeparateShader is
2975 * false), linking will have already failed when first is
2976 * MESA_SHADER_GEOMETRY.
2977 */
2978 if (!assign_varying_locations(ctx, mem_ctx, prog,
2979 NULL, prog->_LinkedShaders[first],
2980 num_tfeedback_decls, tfeedback_decls,
2981 prog->Geom.VerticesIn))
2982 goto done;
2983 }
2984
2985 if (last != MESA_SHADER_FRAGMENT &&
2986 (num_tfeedback_decls != 0 || prog->SeparateShader)) {
2987 /* There was no fragment shader, but we still have to assign varying
2988 * locations for use by transform feedback.
2989 */
2990 if (!assign_varying_locations(ctx, mem_ctx, prog,
2991 sh, NULL,
2992 num_tfeedback_decls, tfeedback_decls,
2993 0))
2994 goto done;
2995 }
2996
2997 do_dead_builtin_varyings(ctx, sh, NULL,
2998 num_tfeedback_decls, tfeedback_decls);
2999
3000 if (!prog->SeparateShader)
3001 demote_shader_inputs_and_outputs(sh, ir_var_shader_out);
3002
3003 /* Eliminate code that is now dead due to unused outputs being demoted.
3004 */
3005 while (do_dead_code(sh->ir, false))
3006 ;
3007 }
3008 else if (first == MESA_SHADER_FRAGMENT) {
3009 /* If the program only contains a fragment shader...
3010 */
3011 gl_shader *const sh = prog->_LinkedShaders[first];
3012
3013 do_dead_builtin_varyings(ctx, NULL, sh,
3014 num_tfeedback_decls, tfeedback_decls);
3015
3016 if (prog->SeparateShader) {
3017 if (!assign_varying_locations(ctx, mem_ctx, prog,
3018 NULL /* producer */,
3019 sh /* consumer */,
3020 0 /* num_tfeedback_decls */,
3021 NULL /* tfeedback_decls */,
3022 0 /* gs_input_vertices */))
3023 goto done;
3024 } else
3025 demote_shader_inputs_and_outputs(sh, ir_var_shader_in);
3026
3027 while (do_dead_code(sh->ir, false))
3028 ;
3029 }
3030
3031 next = last;
3032 for (int i = next - 1; i >= 0; i--) {
3033 if (prog->_LinkedShaders[i] == NULL)
3034 continue;
3035
3036 gl_shader *const sh_i = prog->_LinkedShaders[i];
3037 gl_shader *const sh_next = prog->_LinkedShaders[next];
3038 unsigned gs_input_vertices =
3039 next == MESA_SHADER_GEOMETRY ? prog->Geom.VerticesIn : 0;
3040
3041 if (!assign_varying_locations(ctx, mem_ctx, prog, sh_i, sh_next,
3042 next == MESA_SHADER_FRAGMENT ? num_tfeedback_decls : 0,
3043 tfeedback_decls, gs_input_vertices))
3044 goto done;
3045
3046 do_dead_builtin_varyings(ctx, sh_i, sh_next,
3047 next == MESA_SHADER_FRAGMENT ? num_tfeedback_decls : 0,
3048 tfeedback_decls);
3049
3050 demote_shader_inputs_and_outputs(sh_i, ir_var_shader_out);
3051 demote_shader_inputs_and_outputs(sh_next, ir_var_shader_in);
3052
3053 /* Eliminate code that is now dead due to unused outputs being demoted.
3054 */
3055 while (do_dead_code(sh_i->ir, false))
3056 ;
3057 while (do_dead_code(sh_next->ir, false))
3058 ;
3059
3060 /* This must be done after all dead varyings are eliminated. */
3061 if (!check_against_output_limit(ctx, prog, sh_i))
3062 goto done;
3063 if (!check_against_input_limit(ctx, prog, sh_next))
3064 goto done;
3065
3066 next = i;
3067 }
3068
3069 if (!store_tfeedback_info(ctx, prog, num_tfeedback_decls, tfeedback_decls))
3070 goto done;
3071
3072 update_array_sizes(prog);
3073 link_assign_uniform_locations(prog, ctx->Const.UniformBooleanTrue);
3074 link_assign_atomic_counter_resources(ctx, prog);
3075 store_fragdepth_layout(prog);
3076
3077 check_resources(ctx, prog);
3078 check_image_resources(ctx, prog);
3079 link_check_atomic_counter_resources(ctx, prog);
3080
3081 if (!prog->LinkStatus)
3082 goto done;
3083
3084 /* OpenGL ES requires that a vertex shader and a fragment shader both be
3085 * present in a linked program. GL_ARB_ES2_compatibility doesn't say
3086 * anything about shader linking when one of the shaders (vertex or
3087 * fragment shader) is absent. So, the extension shouldn't change the
3088 * behavior specified in GLSL specification.
3089 */
3090 if (!prog->SeparateShader && ctx->API == API_OPENGLES2) {
3091 if (prog->_LinkedShaders[MESA_SHADER_VERTEX] == NULL) {
3092 linker_error(prog, "program lacks a vertex shader\n");
3093 } else if (prog->_LinkedShaders[MESA_SHADER_FRAGMENT] == NULL) {
3094 linker_error(prog, "program lacks a fragment shader\n");
3095 }
3096 }
3097
3098 build_program_resource_list(ctx, prog);
3099 if (!prog->LinkStatus)
3100 goto done;
3101
3102 /* FINISHME: Assign fragment shader output locations. */
3103
3104 done:
3105 for (unsigned i = 0; i < MESA_SHADER_STAGES; i++) {
3106 free(shader_list[i]);
3107 if (prog->_LinkedShaders[i] == NULL)
3108 continue;
3109
3110 /* Do a final validation step to make sure that the IR wasn't
3111 * invalidated by any modifications performed after intrastage linking.
3112 */
3113 validate_ir_tree(prog->_LinkedShaders[i]->ir);
3114
3115 /* Retain any live IR, but trash the rest. */
3116 reparent_ir(prog->_LinkedShaders[i]->ir, prog->_LinkedShaders[i]->ir);
3117
3118 /* The symbol table in the linked shaders may contain references to
3119 * variables that were removed (e.g., unused uniforms). Since it may
3120 * contain junk, there is no possible valid use. Delete it and set the
3121 * pointer to NULL.
3122 */
3123 delete prog->_LinkedShaders[i]->symbols;
3124 prog->_LinkedShaders[i]->symbols = NULL;
3125 }
3126
3127 ralloc_free(mem_ctx);
3128 }