19eb9b5ff6f5a40383d92350f1177a9413449c7b
[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 "main/core.h"
68 #include "glsl_symbol_table.h"
69 #include "ir.h"
70 #include "program.h"
71 #include "program/hash_table.h"
72 #include "linker.h"
73 #include "ir_optimization.h"
74
75 extern "C" {
76 #include "main/shaderobj.h"
77 }
78
79 /**
80 * Visitor that determines whether or not a variable is ever written.
81 */
82 class find_assignment_visitor : public ir_hierarchical_visitor {
83 public:
84 find_assignment_visitor(const char *name)
85 : name(name), found(false)
86 {
87 /* empty */
88 }
89
90 virtual ir_visitor_status visit_enter(ir_assignment *ir)
91 {
92 ir_variable *const var = ir->lhs->variable_referenced();
93
94 if (strcmp(name, var->name) == 0) {
95 found = true;
96 return visit_stop;
97 }
98
99 return visit_continue_with_parent;
100 }
101
102 virtual ir_visitor_status visit_enter(ir_call *ir)
103 {
104 exec_list_iterator sig_iter = ir->get_callee()->parameters.iterator();
105 foreach_iter(exec_list_iterator, iter, *ir) {
106 ir_rvalue *param_rval = (ir_rvalue *)iter.get();
107 ir_variable *sig_param = (ir_variable *)sig_iter.get();
108
109 if (sig_param->mode == ir_var_out ||
110 sig_param->mode == ir_var_inout) {
111 ir_variable *var = param_rval->variable_referenced();
112 if (var && strcmp(name, var->name) == 0) {
113 found = true;
114 return visit_stop;
115 }
116 }
117 sig_iter.next();
118 }
119
120 return visit_continue_with_parent;
121 }
122
123 bool variable_found()
124 {
125 return found;
126 }
127
128 private:
129 const char *name; /**< Find writes to a variable with this name. */
130 bool found; /**< Was a write to the variable found? */
131 };
132
133
134 /**
135 * Visitor that determines whether or not a variable is ever read.
136 */
137 class find_deref_visitor : public ir_hierarchical_visitor {
138 public:
139 find_deref_visitor(const char *name)
140 : name(name), found(false)
141 {
142 /* empty */
143 }
144
145 virtual ir_visitor_status visit(ir_dereference_variable *ir)
146 {
147 if (strcmp(this->name, ir->var->name) == 0) {
148 this->found = true;
149 return visit_stop;
150 }
151
152 return visit_continue;
153 }
154
155 bool variable_found() const
156 {
157 return this->found;
158 }
159
160 private:
161 const char *name; /**< Find writes to a variable with this name. */
162 bool found; /**< Was a write to the variable found? */
163 };
164
165
166 void
167 linker_error(gl_shader_program *prog, const char *fmt, ...)
168 {
169 va_list ap;
170
171 ralloc_strcat(&prog->InfoLog, "error: ");
172 va_start(ap, fmt);
173 ralloc_vasprintf_append(&prog->InfoLog, fmt, ap);
174 va_end(ap);
175
176 prog->LinkStatus = false;
177 }
178
179
180 void
181 linker_warning(gl_shader_program *prog, const char *fmt, ...)
182 {
183 va_list ap;
184
185 ralloc_strcat(&prog->InfoLog, "error: ");
186 va_start(ap, fmt);
187 ralloc_vasprintf_append(&prog->InfoLog, fmt, ap);
188 va_end(ap);
189
190 }
191
192
193 void
194 invalidate_variable_locations(gl_shader *sh, enum ir_variable_mode mode,
195 int generic_base)
196 {
197 foreach_list(node, sh->ir) {
198 ir_variable *const var = ((ir_instruction *) node)->as_variable();
199
200 if ((var == NULL) || (var->mode != (unsigned) mode))
201 continue;
202
203 /* Only assign locations for generic attributes / varyings / etc.
204 */
205 if ((var->location >= generic_base) && !var->explicit_location)
206 var->location = -1;
207 }
208 }
209
210
211 /**
212 * Determine the number of attribute slots required for a particular type
213 *
214 * This code is here because it implements the language rules of a specific
215 * GLSL version. Since it's a property of the language and not a property of
216 * types in general, it doesn't really belong in glsl_type.
217 */
218 unsigned
219 count_attribute_slots(const glsl_type *t)
220 {
221 /* From page 31 (page 37 of the PDF) of the GLSL 1.50 spec:
222 *
223 * "A scalar input counts the same amount against this limit as a vec4,
224 * so applications may want to consider packing groups of four
225 * unrelated float inputs together into a vector to better utilize the
226 * capabilities of the underlying hardware. A matrix input will use up
227 * multiple locations. The number of locations used will equal the
228 * number of columns in the matrix."
229 *
230 * The spec does not explicitly say how arrays are counted. However, it
231 * should be safe to assume the total number of slots consumed by an array
232 * is the number of entries in the array multiplied by the number of slots
233 * consumed by a single element of the array.
234 */
235
236 if (t->is_array())
237 return t->array_size() * count_attribute_slots(t->element_type());
238
239 if (t->is_matrix())
240 return t->matrix_columns;
241
242 return 1;
243 }
244
245
246 /**
247 * Verify that a vertex shader executable meets all semantic requirements
248 *
249 * \param shader Vertex shader executable to be verified
250 */
251 bool
252 validate_vertex_shader_executable(struct gl_shader_program *prog,
253 struct gl_shader *shader)
254 {
255 if (shader == NULL)
256 return true;
257
258 find_assignment_visitor find("gl_Position");
259 find.run(shader->ir);
260 if (!find.variable_found()) {
261 linker_error(prog, "vertex shader does not write to `gl_Position'\n");
262 return false;
263 }
264
265 return true;
266 }
267
268
269 /**
270 * Verify that a fragment shader executable meets all semantic requirements
271 *
272 * \param shader Fragment shader executable to be verified
273 */
274 bool
275 validate_fragment_shader_executable(struct gl_shader_program *prog,
276 struct gl_shader *shader)
277 {
278 if (shader == NULL)
279 return true;
280
281 find_assignment_visitor frag_color("gl_FragColor");
282 find_assignment_visitor frag_data("gl_FragData");
283
284 frag_color.run(shader->ir);
285 frag_data.run(shader->ir);
286
287 if (frag_color.variable_found() && frag_data.variable_found()) {
288 linker_error(prog, "fragment shader writes to both "
289 "`gl_FragColor' and `gl_FragData'\n");
290 return false;
291 }
292
293 return true;
294 }
295
296
297 /**
298 * Generate a string describing the mode of a variable
299 */
300 static const char *
301 mode_string(const ir_variable *var)
302 {
303 switch (var->mode) {
304 case ir_var_auto:
305 return (var->read_only) ? "global constant" : "global variable";
306
307 case ir_var_uniform: return "uniform";
308 case ir_var_in: return "shader input";
309 case ir_var_out: return "shader output";
310 case ir_var_inout: return "shader inout";
311
312 case ir_var_const_in:
313 case ir_var_temporary:
314 default:
315 assert(!"Should not get here.");
316 return "invalid variable";
317 }
318 }
319
320
321 /**
322 * Perform validation of global variables used across multiple shaders
323 */
324 bool
325 cross_validate_globals(struct gl_shader_program *prog,
326 struct gl_shader **shader_list,
327 unsigned num_shaders,
328 bool uniforms_only)
329 {
330 /* Examine all of the uniforms in all of the shaders and cross validate
331 * them.
332 */
333 glsl_symbol_table variables;
334 for (unsigned i = 0; i < num_shaders; i++) {
335 if (shader_list[i] == NULL)
336 continue;
337
338 foreach_list(node, shader_list[i]->ir) {
339 ir_variable *const var = ((ir_instruction *) node)->as_variable();
340
341 if (var == NULL)
342 continue;
343
344 if (uniforms_only && (var->mode != ir_var_uniform))
345 continue;
346
347 /* Don't cross validate temporaries that are at global scope. These
348 * will eventually get pulled into the shaders 'main'.
349 */
350 if (var->mode == ir_var_temporary)
351 continue;
352
353 /* If a global with this name has already been seen, verify that the
354 * new instance has the same type. In addition, if the globals have
355 * initializers, the values of the initializers must be the same.
356 */
357 ir_variable *const existing = variables.get_variable(var->name);
358 if (existing != NULL) {
359 if (var->type != existing->type) {
360 /* Consider the types to be "the same" if both types are arrays
361 * of the same type and one of the arrays is implicitly sized.
362 * In addition, set the type of the linked variable to the
363 * explicitly sized array.
364 */
365 if (var->type->is_array()
366 && existing->type->is_array()
367 && (var->type->fields.array == existing->type->fields.array)
368 && ((var->type->length == 0)
369 || (existing->type->length == 0))) {
370 if (var->type->length != 0) {
371 existing->type = var->type;
372 }
373 } else {
374 linker_error(prog, "%s `%s' declared as type "
375 "`%s' and type `%s'\n",
376 mode_string(var),
377 var->name, var->type->name,
378 existing->type->name);
379 return false;
380 }
381 }
382
383 if (var->explicit_location) {
384 if (existing->explicit_location
385 && (var->location != existing->location)) {
386 linker_error(prog, "explicit locations for %s "
387 "`%s' have differing values\n",
388 mode_string(var), var->name);
389 return false;
390 }
391
392 existing->location = var->location;
393 existing->explicit_location = true;
394 }
395
396 /* Validate layout qualifiers for gl_FragDepth.
397 *
398 * From the AMD_conservative_depth spec:
399 * "If gl_FragDepth is redeclared in any fragment shader in
400 * a program, it must be redeclared in all fragment shaders in that
401 * program that have static assignments to gl_FragDepth. All
402 * redeclarations of gl_FragDepth in all fragment shaders in
403 * a single program must have the same set of qualifiers."
404 */
405 if (strcmp(var->name, "gl_FragDepth") == 0) {
406 bool layout_declared = var->depth_layout != ir_depth_layout_none;
407 bool layout_differs = var->depth_layout != existing->depth_layout;
408 if (layout_declared && layout_differs) {
409 linker_error(prog,
410 "All redeclarations of gl_FragDepth in all fragment shaders "
411 "in a single program must have the same set of qualifiers.");
412 }
413 if (var->used && layout_differs) {
414 linker_error(prog,
415 "If gl_FragDepth is redeclared with a layout qualifier in"
416 "any fragment shader, it must be redeclared with the same"
417 "layout qualifier in all fragment shaders that have"
418 "assignments to gl_FragDepth");
419 }
420 }
421
422 /* FINISHME: Handle non-constant initializers.
423 */
424 if (var->constant_value != NULL) {
425 if (existing->constant_value != NULL) {
426 if (!var->constant_value->has_value(existing->constant_value)) {
427 linker_error(prog, "initializers for %s "
428 "`%s' have differing values\n",
429 mode_string(var), var->name);
430 return false;
431 }
432 } else
433 /* If the first-seen instance of a particular uniform did not
434 * have an initializer but a later instance does, copy the
435 * initializer to the version stored in the symbol table.
436 */
437 /* FINISHME: This is wrong. The constant_value field should
438 * FINISHME: not be modified! Imagine a case where a shader
439 * FINISHME: without an initializer is linked in two different
440 * FINISHME: programs with shaders that have differing
441 * FINISHME: initializers. Linking with the first will
442 * FINISHME: modify the shader, and linking with the second
443 * FINISHME: will fail.
444 */
445 existing->constant_value =
446 var->constant_value->clone(ralloc_parent(existing), NULL);
447 }
448
449 if (existing->invariant != var->invariant) {
450 linker_error(prog, "declarations for %s `%s' have "
451 "mismatching invariant qualifiers\n",
452 mode_string(var), var->name);
453 return false;
454 }
455 if (existing->centroid != var->centroid) {
456 linker_error(prog, "declarations for %s `%s' have "
457 "mismatching centroid qualifiers\n",
458 mode_string(var), var->name);
459 return false;
460 }
461 } else
462 variables.add_variable(var);
463 }
464 }
465
466 return true;
467 }
468
469
470 /**
471 * Perform validation of uniforms used across multiple shader stages
472 */
473 bool
474 cross_validate_uniforms(struct gl_shader_program *prog)
475 {
476 return cross_validate_globals(prog, prog->_LinkedShaders,
477 MESA_SHADER_TYPES, true);
478 }
479
480
481 /**
482 * Validate that outputs from one stage match inputs of another
483 */
484 bool
485 cross_validate_outputs_to_inputs(struct gl_shader_program *prog,
486 gl_shader *producer, gl_shader *consumer)
487 {
488 glsl_symbol_table parameters;
489 /* FINISHME: Figure these out dynamically. */
490 const char *const producer_stage = "vertex";
491 const char *const consumer_stage = "fragment";
492
493 /* Find all shader outputs in the "producer" stage.
494 */
495 foreach_list(node, producer->ir) {
496 ir_variable *const var = ((ir_instruction *) node)->as_variable();
497
498 /* FINISHME: For geometry shaders, this should also look for inout
499 * FINISHME: variables.
500 */
501 if ((var == NULL) || (var->mode != ir_var_out))
502 continue;
503
504 parameters.add_variable(var);
505 }
506
507
508 /* Find all shader inputs in the "consumer" stage. Any variables that have
509 * matching outputs already in the symbol table must have the same type and
510 * qualifiers.
511 */
512 foreach_list(node, consumer->ir) {
513 ir_variable *const input = ((ir_instruction *) node)->as_variable();
514
515 /* FINISHME: For geometry shaders, this should also look for inout
516 * FINISHME: variables.
517 */
518 if ((input == NULL) || (input->mode != ir_var_in))
519 continue;
520
521 ir_variable *const output = parameters.get_variable(input->name);
522 if (output != NULL) {
523 /* Check that the types match between stages.
524 */
525 if (input->type != output->type) {
526 /* There is a bit of a special case for gl_TexCoord. This
527 * built-in is unsized by default. Applications that variable
528 * access it must redeclare it with a size. There is some
529 * language in the GLSL spec that implies the fragment shader
530 * and vertex shader do not have to agree on this size. Other
531 * driver behave this way, and one or two applications seem to
532 * rely on it.
533 *
534 * Neither declaration needs to be modified here because the array
535 * sizes are fixed later when update_array_sizes is called.
536 *
537 * From page 48 (page 54 of the PDF) of the GLSL 1.10 spec:
538 *
539 * "Unlike user-defined varying variables, the built-in
540 * varying variables don't have a strict one-to-one
541 * correspondence between the vertex language and the
542 * fragment language."
543 */
544 if (!output->type->is_array()
545 || (strncmp("gl_", output->name, 3) != 0)) {
546 linker_error(prog,
547 "%s shader output `%s' declared as type `%s', "
548 "but %s shader input declared as type `%s'\n",
549 producer_stage, output->name,
550 output->type->name,
551 consumer_stage, input->type->name);
552 return false;
553 }
554 }
555
556 /* Check that all of the qualifiers match between stages.
557 */
558 if (input->centroid != output->centroid) {
559 linker_error(prog,
560 "%s shader output `%s' %s centroid qualifier, "
561 "but %s shader input %s centroid qualifier\n",
562 producer_stage,
563 output->name,
564 (output->centroid) ? "has" : "lacks",
565 consumer_stage,
566 (input->centroid) ? "has" : "lacks");
567 return false;
568 }
569
570 if (input->invariant != output->invariant) {
571 linker_error(prog,
572 "%s shader output `%s' %s invariant qualifier, "
573 "but %s shader input %s invariant qualifier\n",
574 producer_stage,
575 output->name,
576 (output->invariant) ? "has" : "lacks",
577 consumer_stage,
578 (input->invariant) ? "has" : "lacks");
579 return false;
580 }
581
582 if (input->interpolation != output->interpolation) {
583 linker_error(prog,
584 "%s shader output `%s' specifies %s "
585 "interpolation qualifier, "
586 "but %s shader input specifies %s "
587 "interpolation qualifier\n",
588 producer_stage,
589 output->name,
590 output->interpolation_string(),
591 consumer_stage,
592 input->interpolation_string());
593 return false;
594 }
595 }
596 }
597
598 return true;
599 }
600
601
602 /**
603 * Populates a shaders symbol table with all global declarations
604 */
605 static void
606 populate_symbol_table(gl_shader *sh)
607 {
608 sh->symbols = new(sh) glsl_symbol_table;
609
610 foreach_list(node, sh->ir) {
611 ir_instruction *const inst = (ir_instruction *) node;
612 ir_variable *var;
613 ir_function *func;
614
615 if ((func = inst->as_function()) != NULL) {
616 sh->symbols->add_function(func);
617 } else if ((var = inst->as_variable()) != NULL) {
618 sh->symbols->add_variable(var);
619 }
620 }
621 }
622
623
624 /**
625 * Remap variables referenced in an instruction tree
626 *
627 * This is used when instruction trees are cloned from one shader and placed in
628 * another. These trees will contain references to \c ir_variable nodes that
629 * do not exist in the target shader. This function finds these \c ir_variable
630 * references and replaces the references with matching variables in the target
631 * shader.
632 *
633 * If there is no matching variable in the target shader, a clone of the
634 * \c ir_variable is made and added to the target shader. The new variable is
635 * added to \b both the instruction stream and the symbol table.
636 *
637 * \param inst IR tree that is to be processed.
638 * \param symbols Symbol table containing global scope symbols in the
639 * linked shader.
640 * \param instructions Instruction stream where new variable declarations
641 * should be added.
642 */
643 void
644 remap_variables(ir_instruction *inst, struct gl_shader *target,
645 hash_table *temps)
646 {
647 class remap_visitor : public ir_hierarchical_visitor {
648 public:
649 remap_visitor(struct gl_shader *target,
650 hash_table *temps)
651 {
652 this->target = target;
653 this->symbols = target->symbols;
654 this->instructions = target->ir;
655 this->temps = temps;
656 }
657
658 virtual ir_visitor_status visit(ir_dereference_variable *ir)
659 {
660 if (ir->var->mode == ir_var_temporary) {
661 ir_variable *var = (ir_variable *) hash_table_find(temps, ir->var);
662
663 assert(var != NULL);
664 ir->var = var;
665 return visit_continue;
666 }
667
668 ir_variable *const existing =
669 this->symbols->get_variable(ir->var->name);
670 if (existing != NULL)
671 ir->var = existing;
672 else {
673 ir_variable *copy = ir->var->clone(this->target, NULL);
674
675 this->symbols->add_variable(copy);
676 this->instructions->push_head(copy);
677 ir->var = copy;
678 }
679
680 return visit_continue;
681 }
682
683 private:
684 struct gl_shader *target;
685 glsl_symbol_table *symbols;
686 exec_list *instructions;
687 hash_table *temps;
688 };
689
690 remap_visitor v(target, temps);
691
692 inst->accept(&v);
693 }
694
695
696 /**
697 * Move non-declarations from one instruction stream to another
698 *
699 * The intended usage pattern of this function is to pass the pointer to the
700 * head sentinel of a list (i.e., a pointer to the list cast to an \c exec_node
701 * pointer) for \c last and \c false for \c make_copies on the first
702 * call. Successive calls pass the return value of the previous call for
703 * \c last and \c true for \c make_copies.
704 *
705 * \param instructions Source instruction stream
706 * \param last Instruction after which new instructions should be
707 * inserted in the target instruction stream
708 * \param make_copies Flag selecting whether instructions in \c instructions
709 * should be copied (via \c ir_instruction::clone) into the
710 * target list or moved.
711 *
712 * \return
713 * The new "last" instruction in the target instruction stream. This pointer
714 * is suitable for use as the \c last parameter of a later call to this
715 * function.
716 */
717 exec_node *
718 move_non_declarations(exec_list *instructions, exec_node *last,
719 bool make_copies, gl_shader *target)
720 {
721 hash_table *temps = NULL;
722
723 if (make_copies)
724 temps = hash_table_ctor(0, hash_table_pointer_hash,
725 hash_table_pointer_compare);
726
727 foreach_list_safe(node, instructions) {
728 ir_instruction *inst = (ir_instruction *) node;
729
730 if (inst->as_function())
731 continue;
732
733 ir_variable *var = inst->as_variable();
734 if ((var != NULL) && (var->mode != ir_var_temporary))
735 continue;
736
737 assert(inst->as_assignment()
738 || ((var != NULL) && (var->mode == ir_var_temporary)));
739
740 if (make_copies) {
741 inst = inst->clone(target, NULL);
742
743 if (var != NULL)
744 hash_table_insert(temps, inst, var);
745 else
746 remap_variables(inst, target, temps);
747 } else {
748 inst->remove();
749 }
750
751 last->insert_after(inst);
752 last = inst;
753 }
754
755 if (make_copies)
756 hash_table_dtor(temps);
757
758 return last;
759 }
760
761 /**
762 * Get the function signature for main from a shader
763 */
764 static ir_function_signature *
765 get_main_function_signature(gl_shader *sh)
766 {
767 ir_function *const f = sh->symbols->get_function("main");
768 if (f != NULL) {
769 exec_list void_parameters;
770
771 /* Look for the 'void main()' signature and ensure that it's defined.
772 * This keeps the linker from accidentally pick a shader that just
773 * contains a prototype for main.
774 *
775 * We don't have to check for multiple definitions of main (in multiple
776 * shaders) because that would have already been caught above.
777 */
778 ir_function_signature *sig = f->matching_signature(&void_parameters);
779 if ((sig != NULL) && sig->is_defined) {
780 return sig;
781 }
782 }
783
784 return NULL;
785 }
786
787
788 /**
789 * Combine a group of shaders for a single stage to generate a linked shader
790 *
791 * \note
792 * If this function is supplied a single shader, it is cloned, and the new
793 * shader is returned.
794 */
795 static struct gl_shader *
796 link_intrastage_shaders(void *mem_ctx,
797 struct gl_context *ctx,
798 struct gl_shader_program *prog,
799 struct gl_shader **shader_list,
800 unsigned num_shaders)
801 {
802 /* Check that global variables defined in multiple shaders are consistent.
803 */
804 if (!cross_validate_globals(prog, shader_list, num_shaders, false))
805 return NULL;
806
807 /* Check that there is only a single definition of each function signature
808 * across all shaders.
809 */
810 for (unsigned i = 0; i < (num_shaders - 1); i++) {
811 foreach_list(node, shader_list[i]->ir) {
812 ir_function *const f = ((ir_instruction *) node)->as_function();
813
814 if (f == NULL)
815 continue;
816
817 for (unsigned j = i + 1; j < num_shaders; j++) {
818 ir_function *const other =
819 shader_list[j]->symbols->get_function(f->name);
820
821 /* If the other shader has no function (and therefore no function
822 * signatures) with the same name, skip to the next shader.
823 */
824 if (other == NULL)
825 continue;
826
827 foreach_iter (exec_list_iterator, iter, *f) {
828 ir_function_signature *sig =
829 (ir_function_signature *) iter.get();
830
831 if (!sig->is_defined || sig->is_builtin)
832 continue;
833
834 ir_function_signature *other_sig =
835 other->exact_matching_signature(& sig->parameters);
836
837 if ((other_sig != NULL) && other_sig->is_defined
838 && !other_sig->is_builtin) {
839 linker_error(prog, "function `%s' is multiply defined",
840 f->name);
841 return NULL;
842 }
843 }
844 }
845 }
846 }
847
848 /* Find the shader that defines main, and make a clone of it.
849 *
850 * Starting with the clone, search for undefined references. If one is
851 * found, find the shader that defines it. Clone the reference and add
852 * it to the shader. Repeat until there are no undefined references or
853 * until a reference cannot be resolved.
854 */
855 gl_shader *main = NULL;
856 for (unsigned i = 0; i < num_shaders; i++) {
857 if (get_main_function_signature(shader_list[i]) != NULL) {
858 main = shader_list[i];
859 break;
860 }
861 }
862
863 if (main == NULL) {
864 linker_error(prog, "%s shader lacks `main'\n",
865 (shader_list[0]->Type == GL_VERTEX_SHADER)
866 ? "vertex" : "fragment");
867 return NULL;
868 }
869
870 gl_shader *linked = ctx->Driver.NewShader(NULL, 0, main->Type);
871 linked->ir = new(linked) exec_list;
872 clone_ir_list(mem_ctx, linked->ir, main->ir);
873
874 populate_symbol_table(linked);
875
876 /* The a pointer to the main function in the final linked shader (i.e., the
877 * copy of the original shader that contained the main function).
878 */
879 ir_function_signature *const main_sig = get_main_function_signature(linked);
880
881 /* Move any instructions other than variable declarations or function
882 * declarations into main.
883 */
884 exec_node *insertion_point =
885 move_non_declarations(linked->ir, (exec_node *) &main_sig->body, false,
886 linked);
887
888 for (unsigned i = 0; i < num_shaders; i++) {
889 if (shader_list[i] == main)
890 continue;
891
892 insertion_point = move_non_declarations(shader_list[i]->ir,
893 insertion_point, true, linked);
894 }
895
896 /* Resolve initializers for global variables in the linked shader.
897 */
898 unsigned num_linking_shaders = num_shaders;
899 for (unsigned i = 0; i < num_shaders; i++)
900 num_linking_shaders += shader_list[i]->num_builtins_to_link;
901
902 gl_shader **linking_shaders =
903 (gl_shader **) calloc(num_linking_shaders, sizeof(gl_shader *));
904
905 memcpy(linking_shaders, shader_list,
906 sizeof(linking_shaders[0]) * num_shaders);
907
908 unsigned idx = num_shaders;
909 for (unsigned i = 0; i < num_shaders; i++) {
910 memcpy(&linking_shaders[idx], shader_list[i]->builtins_to_link,
911 sizeof(linking_shaders[0]) * shader_list[i]->num_builtins_to_link);
912 idx += shader_list[i]->num_builtins_to_link;
913 }
914
915 assert(idx == num_linking_shaders);
916
917 if (!link_function_calls(prog, linked, linking_shaders,
918 num_linking_shaders)) {
919 ctx->Driver.DeleteShader(ctx, linked);
920 linked = NULL;
921 }
922
923 free(linking_shaders);
924
925 /* Make a pass over all variable declarations to ensure that arrays with
926 * unspecified sizes have a size specified. The size is inferred from the
927 * max_array_access field.
928 */
929 if (linked != NULL) {
930 class array_sizing_visitor : public ir_hierarchical_visitor {
931 public:
932 virtual ir_visitor_status visit(ir_variable *var)
933 {
934 if (var->type->is_array() && (var->type->length == 0)) {
935 const glsl_type *type =
936 glsl_type::get_array_instance(var->type->fields.array,
937 var->max_array_access + 1);
938
939 assert(type != NULL);
940 var->type = type;
941 }
942
943 return visit_continue;
944 }
945 } v;
946
947 v.run(linked->ir);
948 }
949
950 return linked;
951 }
952
953
954 struct uniform_node {
955 exec_node link;
956 struct gl_uniform *u;
957 unsigned slots;
958 };
959
960 /**
961 * Update the sizes of linked shader uniform arrays to the maximum
962 * array index used.
963 *
964 * From page 81 (page 95 of the PDF) of the OpenGL 2.1 spec:
965 *
966 * If one or more elements of an array are active,
967 * GetActiveUniform will return the name of the array in name,
968 * subject to the restrictions listed above. The type of the array
969 * is returned in type. The size parameter contains the highest
970 * array element index used, plus one. The compiler or linker
971 * determines the highest index used. There will be only one
972 * active uniform reported by the GL per uniform array.
973
974 */
975 static void
976 update_array_sizes(struct gl_shader_program *prog)
977 {
978 for (unsigned i = 0; i < MESA_SHADER_TYPES; i++) {
979 if (prog->_LinkedShaders[i] == NULL)
980 continue;
981
982 foreach_list(node, prog->_LinkedShaders[i]->ir) {
983 ir_variable *const var = ((ir_instruction *) node)->as_variable();
984
985 if ((var == NULL) || (var->mode != ir_var_uniform &&
986 var->mode != ir_var_in &&
987 var->mode != ir_var_out) ||
988 !var->type->is_array())
989 continue;
990
991 unsigned int size = var->max_array_access;
992 for (unsigned j = 0; j < MESA_SHADER_TYPES; j++) {
993 if (prog->_LinkedShaders[j] == NULL)
994 continue;
995
996 foreach_list(node2, prog->_LinkedShaders[j]->ir) {
997 ir_variable *other_var = ((ir_instruction *) node2)->as_variable();
998 if (!other_var)
999 continue;
1000
1001 if (strcmp(var->name, other_var->name) == 0 &&
1002 other_var->max_array_access > size) {
1003 size = other_var->max_array_access;
1004 }
1005 }
1006 }
1007
1008 if (size + 1 != var->type->fields.array->length) {
1009 /* If this is a built-in uniform (i.e., it's backed by some
1010 * fixed-function state), adjust the number of state slots to
1011 * match the new array size. The number of slots per array entry
1012 * is not known. It seems safe to assume that the total number of
1013 * slots is an integer multiple of the number of array elements.
1014 * Determine the number of slots per array element by dividing by
1015 * the old (total) size.
1016 */
1017 if (var->num_state_slots > 0) {
1018 var->num_state_slots = (size + 1)
1019 * (var->num_state_slots / var->type->length);
1020 }
1021
1022 var->type = glsl_type::get_array_instance(var->type->fields.array,
1023 size + 1);
1024 /* FINISHME: We should update the types of array
1025 * dereferences of this variable now.
1026 */
1027 }
1028 }
1029 }
1030 }
1031
1032 static void
1033 add_uniform(void *mem_ctx, exec_list *uniforms, struct hash_table *ht,
1034 const char *name, const glsl_type *type, GLenum shader_type,
1035 unsigned *next_shader_pos, unsigned *total_uniforms)
1036 {
1037 if (type->is_record()) {
1038 for (unsigned int i = 0; i < type->length; i++) {
1039 const glsl_type *field_type = type->fields.structure[i].type;
1040 char *field_name = ralloc_asprintf(mem_ctx, "%s.%s", name,
1041 type->fields.structure[i].name);
1042
1043 add_uniform(mem_ctx, uniforms, ht, field_name, field_type,
1044 shader_type, next_shader_pos, total_uniforms);
1045 }
1046 } else {
1047 uniform_node *n = (uniform_node *) hash_table_find(ht, name);
1048 unsigned int vec4_slots;
1049 const glsl_type *array_elem_type = NULL;
1050
1051 if (type->is_array()) {
1052 array_elem_type = type->fields.array;
1053 /* Array of structures. */
1054 if (array_elem_type->is_record()) {
1055 for (unsigned int i = 0; i < type->length; i++) {
1056 char *elem_name = ralloc_asprintf(mem_ctx, "%s[%d]", name, i);
1057 add_uniform(mem_ctx, uniforms, ht, elem_name, array_elem_type,
1058 shader_type, next_shader_pos, total_uniforms);
1059 }
1060 return;
1061 }
1062 }
1063
1064 /* Fix the storage size of samplers at 1 vec4 each. Be sure to pad out
1065 * vectors to vec4 slots.
1066 */
1067 if (type->is_array()) {
1068 if (array_elem_type->is_sampler())
1069 vec4_slots = type->length;
1070 else
1071 vec4_slots = type->length * array_elem_type->matrix_columns;
1072 } else if (type->is_sampler()) {
1073 vec4_slots = 1;
1074 } else {
1075 vec4_slots = type->matrix_columns;
1076 }
1077
1078 if (n == NULL) {
1079 n = (uniform_node *) calloc(1, sizeof(struct uniform_node));
1080 n->u = (gl_uniform *) calloc(1, sizeof(struct gl_uniform));
1081 n->slots = vec4_slots;
1082
1083 n->u->Name = strdup(name);
1084 n->u->Type = type;
1085 n->u->VertPos = -1;
1086 n->u->FragPos = -1;
1087 n->u->GeomPos = -1;
1088 (*total_uniforms)++;
1089
1090 hash_table_insert(ht, n, name);
1091 uniforms->push_tail(& n->link);
1092 }
1093
1094 switch (shader_type) {
1095 case GL_VERTEX_SHADER:
1096 n->u->VertPos = *next_shader_pos;
1097 break;
1098 case GL_FRAGMENT_SHADER:
1099 n->u->FragPos = *next_shader_pos;
1100 break;
1101 case GL_GEOMETRY_SHADER:
1102 n->u->GeomPos = *next_shader_pos;
1103 break;
1104 }
1105
1106 (*next_shader_pos) += vec4_slots;
1107 }
1108 }
1109
1110 void
1111 assign_uniform_locations(struct gl_shader_program *prog)
1112 {
1113 /* */
1114 exec_list uniforms;
1115 unsigned total_uniforms = 0;
1116 hash_table *ht = hash_table_ctor(32, hash_table_string_hash,
1117 hash_table_string_compare);
1118 void *mem_ctx = ralloc_context(NULL);
1119
1120 for (unsigned i = 0; i < MESA_SHADER_TYPES; i++) {
1121 if (prog->_LinkedShaders[i] == NULL)
1122 continue;
1123
1124 unsigned next_position = 0;
1125
1126 foreach_list(node, prog->_LinkedShaders[i]->ir) {
1127 ir_variable *const var = ((ir_instruction *) node)->as_variable();
1128
1129 if ((var == NULL) || (var->mode != ir_var_uniform))
1130 continue;
1131
1132 if (strncmp(var->name, "gl_", 3) == 0) {
1133 /* At the moment, we don't allocate uniform locations for
1134 * builtin uniforms. It's permitted by spec, and we'll
1135 * likely switch to doing that at some point, but not yet.
1136 */
1137 continue;
1138 }
1139
1140 var->location = next_position;
1141 add_uniform(mem_ctx, &uniforms, ht, var->name, var->type,
1142 prog->_LinkedShaders[i]->Type,
1143 &next_position, &total_uniforms);
1144 }
1145 }
1146
1147 ralloc_free(mem_ctx);
1148
1149 gl_uniform_list *ul = (gl_uniform_list *)
1150 calloc(1, sizeof(gl_uniform_list));
1151
1152 ul->Size = total_uniforms;
1153 ul->NumUniforms = total_uniforms;
1154 ul->Uniforms = (gl_uniform *) calloc(total_uniforms, sizeof(gl_uniform));
1155
1156 unsigned idx = 0;
1157 uniform_node *next;
1158 for (uniform_node *node = (uniform_node *) uniforms.head
1159 ; node->link.next != NULL
1160 ; node = next) {
1161 next = (uniform_node *) node->link.next;
1162
1163 node->link.remove();
1164 memcpy(&ul->Uniforms[idx], node->u, sizeof(gl_uniform));
1165 idx++;
1166
1167 free(node->u);
1168 free(node);
1169 }
1170
1171 hash_table_dtor(ht);
1172
1173 prog->Uniforms = ul;
1174 }
1175
1176
1177 /**
1178 * Find a contiguous set of available bits in a bitmask.
1179 *
1180 * \param used_mask Bits representing used (1) and unused (0) locations
1181 * \param needed_count Number of contiguous bits needed.
1182 *
1183 * \return
1184 * Base location of the available bits on success or -1 on failure.
1185 */
1186 int
1187 find_available_slots(unsigned used_mask, unsigned needed_count)
1188 {
1189 unsigned needed_mask = (1 << needed_count) - 1;
1190 const int max_bit_to_test = (8 * sizeof(used_mask)) - needed_count;
1191
1192 /* The comparison to 32 is redundant, but without it GCC emits "warning:
1193 * cannot optimize possibly infinite loops" for the loop below.
1194 */
1195 if ((needed_count == 0) || (max_bit_to_test < 0) || (max_bit_to_test > 32))
1196 return -1;
1197
1198 for (int i = 0; i <= max_bit_to_test; i++) {
1199 if ((needed_mask & ~used_mask) == needed_mask)
1200 return i;
1201
1202 needed_mask <<= 1;
1203 }
1204
1205 return -1;
1206 }
1207
1208
1209 /**
1210 * Assign locations for either VS inputs for FS outputs
1211 *
1212 * \param prog Shader program whose variables need locations assigned
1213 * \param target_index Selector for the program target to receive location
1214 * assignmnets. Must be either \c MESA_SHADER_VERTEX or
1215 * \c MESA_SHADER_FRAGMENT.
1216 * \param max_index Maximum number of generic locations. This corresponds
1217 * to either the maximum number of draw buffers or the
1218 * maximum number of generic attributes.
1219 *
1220 * \return
1221 * If locations are successfully assigned, true is returned. Otherwise an
1222 * error is emitted to the shader link log and false is returned.
1223 *
1224 * \bug
1225 * Locations set via \c glBindFragDataLocation are not currently supported.
1226 * Only locations assigned automatically by the linker, explicitly set by a
1227 * layout qualifier, or explicitly set by a built-in variable (e.g., \c
1228 * gl_FragColor) are supported for fragment shaders.
1229 */
1230 bool
1231 assign_attribute_or_color_locations(gl_shader_program *prog,
1232 unsigned target_index,
1233 unsigned max_index)
1234 {
1235 /* Mark invalid locations as being used.
1236 */
1237 unsigned used_locations = (max_index >= 32)
1238 ? ~0 : ~((1 << max_index) - 1);
1239
1240 assert((target_index == MESA_SHADER_VERTEX)
1241 || (target_index == MESA_SHADER_FRAGMENT));
1242
1243 gl_shader *const sh = prog->_LinkedShaders[target_index];
1244 if (sh == NULL)
1245 return true;
1246
1247 /* Operate in a total of four passes.
1248 *
1249 * 1. Invalidate the location assignments for all vertex shader inputs.
1250 *
1251 * 2. Assign locations for inputs that have user-defined (via
1252 * glBindVertexAttribLocation) locations.
1253 *
1254 * 3. Sort the attributes without assigned locations by number of slots
1255 * required in decreasing order. Fragmentation caused by attribute
1256 * locations assigned by the application may prevent large attributes
1257 * from having enough contiguous space.
1258 *
1259 * 4. Assign locations to any inputs without assigned locations.
1260 */
1261
1262 const int generic_base = (target_index == MESA_SHADER_VERTEX)
1263 ? (int) VERT_ATTRIB_GENERIC0 : (int) FRAG_RESULT_DATA0;
1264
1265 const enum ir_variable_mode direction =
1266 (target_index == MESA_SHADER_VERTEX) ? ir_var_in : ir_var_out;
1267
1268
1269 invalidate_variable_locations(sh, direction, generic_base);
1270
1271 if ((target_index == MESA_SHADER_VERTEX) && (prog->Attributes != NULL)) {
1272 for (unsigned i = 0; i < prog->Attributes->NumParameters; i++) {
1273 ir_variable *const var =
1274 sh->symbols->get_variable(prog->Attributes->Parameters[i].Name);
1275
1276 /* Note: attributes that occupy multiple slots, such as arrays or
1277 * matrices, may appear in the attrib array multiple times.
1278 */
1279 if ((var == NULL) || (var->location != -1))
1280 continue;
1281
1282 /* From page 61 of the OpenGL 4.0 spec:
1283 *
1284 * "LinkProgram will fail if the attribute bindings assigned by
1285 * BindAttribLocation do not leave not enough space to assign a
1286 * location for an active matrix attribute or an active attribute
1287 * array, both of which require multiple contiguous generic
1288 * attributes."
1289 *
1290 * Previous versions of the spec contain similar language but omit the
1291 * bit about attribute arrays.
1292 *
1293 * Page 61 of the OpenGL 4.0 spec also says:
1294 *
1295 * "It is possible for an application to bind more than one
1296 * attribute name to the same location. This is referred to as
1297 * aliasing. This will only work if only one of the aliased
1298 * attributes is active in the executable program, or if no path
1299 * through the shader consumes more than one attribute of a set
1300 * of attributes aliased to the same location. A link error can
1301 * occur if the linker determines that every path through the
1302 * shader consumes multiple aliased attributes, but
1303 * implementations are not required to generate an error in this
1304 * case."
1305 *
1306 * These two paragraphs are either somewhat contradictory, or I don't
1307 * fully understand one or both of them.
1308 */
1309 /* FINISHME: The code as currently written does not support attribute
1310 * FINISHME: location aliasing (see comment above).
1311 */
1312 const int attr = prog->Attributes->Parameters[i].StateIndexes[0];
1313 const unsigned slots = count_attribute_slots(var->type);
1314
1315 /* Mask representing the contiguous slots that will be used by this
1316 * attribute.
1317 */
1318 const unsigned use_mask = (1 << slots) - 1;
1319
1320 /* Generate a link error if the set of bits requested for this
1321 * attribute overlaps any previously allocated bits.
1322 */
1323 if ((~(use_mask << attr) & used_locations) != used_locations) {
1324 linker_error(prog,
1325 "insufficient contiguous attribute locations "
1326 "available for vertex shader input `%s'",
1327 var->name);
1328 return false;
1329 }
1330
1331 var->location = VERT_ATTRIB_GENERIC0 + attr;
1332 used_locations |= (use_mask << attr);
1333 }
1334 }
1335
1336 /* Temporary storage for the set of attributes that need locations assigned.
1337 */
1338 struct temp_attr {
1339 unsigned slots;
1340 ir_variable *var;
1341
1342 /* Used below in the call to qsort. */
1343 static int compare(const void *a, const void *b)
1344 {
1345 const temp_attr *const l = (const temp_attr *) a;
1346 const temp_attr *const r = (const temp_attr *) b;
1347
1348 /* Reversed because we want a descending order sort below. */
1349 return r->slots - l->slots;
1350 }
1351 } to_assign[16];
1352
1353 unsigned num_attr = 0;
1354
1355 foreach_list(node, sh->ir) {
1356 ir_variable *const var = ((ir_instruction *) node)->as_variable();
1357
1358 if ((var == NULL) || (var->mode != (unsigned) direction))
1359 continue;
1360
1361 if (var->explicit_location) {
1362 const unsigned slots = count_attribute_slots(var->type);
1363 const unsigned use_mask = (1 << slots) - 1;
1364 const int attr = var->location - generic_base;
1365
1366 if ((var->location >= (int)(max_index + generic_base))
1367 || (var->location < 0)) {
1368 linker_error(prog,
1369 "invalid explicit location %d specified for `%s'\n",
1370 (var->location < 0) ? var->location : attr,
1371 var->name);
1372 return false;
1373 } else if (var->location >= generic_base) {
1374 used_locations |= (use_mask << attr);
1375 }
1376 }
1377
1378 /* The location was explicitly assigned, nothing to do here.
1379 */
1380 if (var->location != -1)
1381 continue;
1382
1383 to_assign[num_attr].slots = count_attribute_slots(var->type);
1384 to_assign[num_attr].var = var;
1385 num_attr++;
1386 }
1387
1388 /* If all of the attributes were assigned locations by the application (or
1389 * are built-in attributes with fixed locations), return early. This should
1390 * be the common case.
1391 */
1392 if (num_attr == 0)
1393 return true;
1394
1395 qsort(to_assign, num_attr, sizeof(to_assign[0]), temp_attr::compare);
1396
1397 if (target_index == MESA_SHADER_VERTEX) {
1398 /* VERT_ATTRIB_GENERIC0 is a pseudo-alias for VERT_ATTRIB_POS. It can
1399 * only be explicitly assigned by via glBindAttribLocation. Mark it as
1400 * reserved to prevent it from being automatically allocated below.
1401 */
1402 find_deref_visitor find("gl_Vertex");
1403 find.run(sh->ir);
1404 if (find.variable_found())
1405 used_locations |= (1 << 0);
1406 }
1407
1408 for (unsigned i = 0; i < num_attr; i++) {
1409 /* Mask representing the contiguous slots that will be used by this
1410 * attribute.
1411 */
1412 const unsigned use_mask = (1 << to_assign[i].slots) - 1;
1413
1414 int location = find_available_slots(used_locations, to_assign[i].slots);
1415
1416 if (location < 0) {
1417 const char *const string = (target_index == MESA_SHADER_VERTEX)
1418 ? "vertex shader input" : "fragment shader output";
1419
1420 linker_error(prog,
1421 "insufficient contiguous attribute locations "
1422 "available for %s `%s'",
1423 string, to_assign[i].var->name);
1424 return false;
1425 }
1426
1427 to_assign[i].var->location = generic_base + location;
1428 used_locations |= (use_mask << location);
1429 }
1430
1431 return true;
1432 }
1433
1434
1435 /**
1436 * Demote shader inputs and outputs that are not used in other stages
1437 */
1438 void
1439 demote_shader_inputs_and_outputs(gl_shader *sh, enum ir_variable_mode mode)
1440 {
1441 foreach_list(node, sh->ir) {
1442 ir_variable *const var = ((ir_instruction *) node)->as_variable();
1443
1444 if ((var == NULL) || (var->mode != int(mode)))
1445 continue;
1446
1447 /* A shader 'in' or 'out' variable is only really an input or output if
1448 * its value is used by other shader stages. This will cause the variable
1449 * to have a location assigned.
1450 */
1451 if (var->location == -1) {
1452 var->mode = ir_var_auto;
1453 }
1454 }
1455 }
1456
1457
1458 bool
1459 assign_varying_locations(struct gl_context *ctx,
1460 struct gl_shader_program *prog,
1461 gl_shader *producer, gl_shader *consumer)
1462 {
1463 /* FINISHME: Set dynamically when geometry shader support is added. */
1464 unsigned output_index = VERT_RESULT_VAR0;
1465 unsigned input_index = FRAG_ATTRIB_VAR0;
1466
1467 /* Operate in a total of three passes.
1468 *
1469 * 1. Assign locations for any matching inputs and outputs.
1470 *
1471 * 2. Mark output variables in the producer that do not have locations as
1472 * not being outputs. This lets the optimizer eliminate them.
1473 *
1474 * 3. Mark input variables in the consumer that do not have locations as
1475 * not being inputs. This lets the optimizer eliminate them.
1476 */
1477
1478 invalidate_variable_locations(producer, ir_var_out, VERT_RESULT_VAR0);
1479 invalidate_variable_locations(consumer, ir_var_in, FRAG_ATTRIB_VAR0);
1480
1481 foreach_list(node, producer->ir) {
1482 ir_variable *const output_var = ((ir_instruction *) node)->as_variable();
1483
1484 if ((output_var == NULL) || (output_var->mode != ir_var_out)
1485 || (output_var->location != -1))
1486 continue;
1487
1488 ir_variable *const input_var =
1489 consumer->symbols->get_variable(output_var->name);
1490
1491 if ((input_var == NULL) || (input_var->mode != ir_var_in))
1492 continue;
1493
1494 assert(input_var->location == -1);
1495
1496 output_var->location = output_index;
1497 input_var->location = input_index;
1498
1499 /* FINISHME: Support for "varying" records in GLSL 1.50. */
1500 assert(!output_var->type->is_record());
1501
1502 if (output_var->type->is_array()) {
1503 const unsigned slots = output_var->type->length
1504 * output_var->type->fields.array->matrix_columns;
1505
1506 output_index += slots;
1507 input_index += slots;
1508 } else {
1509 const unsigned slots = output_var->type->matrix_columns;
1510
1511 output_index += slots;
1512 input_index += slots;
1513 }
1514 }
1515
1516 unsigned varying_vectors = 0;
1517
1518 foreach_list(node, consumer->ir) {
1519 ir_variable *const var = ((ir_instruction *) node)->as_variable();
1520
1521 if ((var == NULL) || (var->mode != ir_var_in))
1522 continue;
1523
1524 if (var->location == -1) {
1525 if (prog->Version <= 120) {
1526 /* On page 25 (page 31 of the PDF) of the GLSL 1.20 spec:
1527 *
1528 * Only those varying variables used (i.e. read) in
1529 * the fragment shader executable must be written to
1530 * by the vertex shader executable; declaring
1531 * superfluous varying variables in a vertex shader is
1532 * permissible.
1533 *
1534 * We interpret this text as meaning that the VS must
1535 * write the variable for the FS to read it. See
1536 * "glsl1-varying read but not written" in piglit.
1537 */
1538
1539 linker_error(prog, "fragment shader varying %s not written "
1540 "by vertex shader\n.", var->name);
1541 }
1542
1543 /* An 'in' variable is only really a shader input if its
1544 * value is written by the previous stage.
1545 */
1546 var->mode = ir_var_auto;
1547 } else {
1548 /* The packing rules are used for vertex shader inputs are also used
1549 * for fragment shader inputs.
1550 */
1551 varying_vectors += count_attribute_slots(var->type);
1552 }
1553 }
1554
1555 if (ctx->API == API_OPENGLES2 || prog->Version == 100) {
1556 if (varying_vectors > ctx->Const.MaxVarying) {
1557 linker_error(prog, "shader uses too many varying vectors "
1558 "(%u > %u)\n",
1559 varying_vectors, ctx->Const.MaxVarying);
1560 return false;
1561 }
1562 } else {
1563 const unsigned float_components = varying_vectors * 4;
1564 if (float_components > ctx->Const.MaxVarying * 4) {
1565 linker_error(prog, "shader uses too many varying components "
1566 "(%u > %u)\n",
1567 float_components, ctx->Const.MaxVarying * 4);
1568 return false;
1569 }
1570 }
1571
1572 return true;
1573 }
1574
1575
1576 void
1577 link_shaders(struct gl_context *ctx, struct gl_shader_program *prog)
1578 {
1579 void *mem_ctx = ralloc_context(NULL); // temporary linker context
1580
1581 prog->LinkStatus = false;
1582 prog->Validated = false;
1583 prog->_Used = false;
1584
1585 if (prog->InfoLog != NULL)
1586 ralloc_free(prog->InfoLog);
1587
1588 prog->InfoLog = ralloc_strdup(NULL, "");
1589
1590 /* Separate the shaders into groups based on their type.
1591 */
1592 struct gl_shader **vert_shader_list;
1593 unsigned num_vert_shaders = 0;
1594 struct gl_shader **frag_shader_list;
1595 unsigned num_frag_shaders = 0;
1596
1597 vert_shader_list = (struct gl_shader **)
1598 calloc(2 * prog->NumShaders, sizeof(struct gl_shader *));
1599 frag_shader_list = &vert_shader_list[prog->NumShaders];
1600
1601 unsigned min_version = UINT_MAX;
1602 unsigned max_version = 0;
1603 for (unsigned i = 0; i < prog->NumShaders; i++) {
1604 min_version = MIN2(min_version, prog->Shaders[i]->Version);
1605 max_version = MAX2(max_version, prog->Shaders[i]->Version);
1606
1607 switch (prog->Shaders[i]->Type) {
1608 case GL_VERTEX_SHADER:
1609 vert_shader_list[num_vert_shaders] = prog->Shaders[i];
1610 num_vert_shaders++;
1611 break;
1612 case GL_FRAGMENT_SHADER:
1613 frag_shader_list[num_frag_shaders] = prog->Shaders[i];
1614 num_frag_shaders++;
1615 break;
1616 case GL_GEOMETRY_SHADER:
1617 /* FINISHME: Support geometry shaders. */
1618 assert(prog->Shaders[i]->Type != GL_GEOMETRY_SHADER);
1619 break;
1620 }
1621 }
1622
1623 /* Previous to GLSL version 1.30, different compilation units could mix and
1624 * match shading language versions. With GLSL 1.30 and later, the versions
1625 * of all shaders must match.
1626 */
1627 assert(min_version >= 100);
1628 assert(max_version <= 130);
1629 if ((max_version >= 130 || min_version == 100)
1630 && min_version != max_version) {
1631 linker_error(prog, "all shaders must use same shading "
1632 "language version\n");
1633 goto done;
1634 }
1635
1636 prog->Version = max_version;
1637
1638 for (unsigned int i = 0; i < MESA_SHADER_TYPES; i++) {
1639 if (prog->_LinkedShaders[i] != NULL)
1640 ctx->Driver.DeleteShader(ctx, prog->_LinkedShaders[i]);
1641
1642 prog->_LinkedShaders[i] = NULL;
1643 }
1644
1645 /* Link all shaders for a particular stage and validate the result.
1646 */
1647 if (num_vert_shaders > 0) {
1648 gl_shader *const sh =
1649 link_intrastage_shaders(mem_ctx, ctx, prog, vert_shader_list,
1650 num_vert_shaders);
1651
1652 if (sh == NULL)
1653 goto done;
1654
1655 if (!validate_vertex_shader_executable(prog, sh))
1656 goto done;
1657
1658 _mesa_reference_shader(ctx, &prog->_LinkedShaders[MESA_SHADER_VERTEX],
1659 sh);
1660 }
1661
1662 if (num_frag_shaders > 0) {
1663 gl_shader *const sh =
1664 link_intrastage_shaders(mem_ctx, ctx, prog, frag_shader_list,
1665 num_frag_shaders);
1666
1667 if (sh == NULL)
1668 goto done;
1669
1670 if (!validate_fragment_shader_executable(prog, sh))
1671 goto done;
1672
1673 _mesa_reference_shader(ctx, &prog->_LinkedShaders[MESA_SHADER_FRAGMENT],
1674 sh);
1675 }
1676
1677 /* Here begins the inter-stage linking phase. Some initial validation is
1678 * performed, then locations are assigned for uniforms, attributes, and
1679 * varyings.
1680 */
1681 if (cross_validate_uniforms(prog)) {
1682 unsigned prev;
1683
1684 for (prev = 0; prev < MESA_SHADER_TYPES; prev++) {
1685 if (prog->_LinkedShaders[prev] != NULL)
1686 break;
1687 }
1688
1689 /* Validate the inputs of each stage with the output of the preceding
1690 * stage.
1691 */
1692 for (unsigned i = prev + 1; i < MESA_SHADER_TYPES; i++) {
1693 if (prog->_LinkedShaders[i] == NULL)
1694 continue;
1695
1696 if (!cross_validate_outputs_to_inputs(prog,
1697 prog->_LinkedShaders[prev],
1698 prog->_LinkedShaders[i]))
1699 goto done;
1700
1701 prev = i;
1702 }
1703
1704 prog->LinkStatus = true;
1705 }
1706
1707 /* Do common optimization before assigning storage for attributes,
1708 * uniforms, and varyings. Later optimization could possibly make
1709 * some of that unused.
1710 */
1711 for (unsigned i = 0; i < MESA_SHADER_TYPES; i++) {
1712 if (prog->_LinkedShaders[i] == NULL)
1713 continue;
1714
1715 detect_recursion_linked(prog, prog->_LinkedShaders[i]->ir);
1716 if (!prog->LinkStatus)
1717 goto done;
1718
1719 while (do_common_optimization(prog->_LinkedShaders[i]->ir, true, 32))
1720 ;
1721 }
1722
1723 update_array_sizes(prog);
1724
1725 assign_uniform_locations(prog);
1726
1727 /* FINISHME: The value of the max_attribute_index parameter is
1728 * FINISHME: implementation dependent based on the value of
1729 * FINISHME: GL_MAX_VERTEX_ATTRIBS. GL_MAX_VERTEX_ATTRIBS must be
1730 * FINISHME: at least 16, so hardcode 16 for now.
1731 */
1732 if (!assign_attribute_or_color_locations(prog, MESA_SHADER_VERTEX, 16)) {
1733 goto done;
1734 }
1735
1736 if (!assign_attribute_or_color_locations(prog, MESA_SHADER_FRAGMENT, ctx->Const.MaxDrawBuffers)) {
1737 goto done;
1738 }
1739
1740 unsigned prev;
1741 for (prev = 0; prev < MESA_SHADER_TYPES; prev++) {
1742 if (prog->_LinkedShaders[prev] != NULL)
1743 break;
1744 }
1745
1746 for (unsigned i = prev + 1; i < MESA_SHADER_TYPES; i++) {
1747 if (prog->_LinkedShaders[i] == NULL)
1748 continue;
1749
1750 if (!assign_varying_locations(ctx, prog,
1751 prog->_LinkedShaders[prev],
1752 prog->_LinkedShaders[i])) {
1753 goto done;
1754 }
1755
1756 prev = i;
1757 }
1758
1759 if (prog->_LinkedShaders[MESA_SHADER_VERTEX] != NULL) {
1760 demote_shader_inputs_and_outputs(prog->_LinkedShaders[MESA_SHADER_VERTEX],
1761 ir_var_out);
1762 }
1763
1764 if (prog->_LinkedShaders[MESA_SHADER_GEOMETRY] != NULL) {
1765 gl_shader *const sh = prog->_LinkedShaders[MESA_SHADER_GEOMETRY];
1766
1767 demote_shader_inputs_and_outputs(sh, ir_var_in);
1768 demote_shader_inputs_and_outputs(sh, ir_var_inout);
1769 demote_shader_inputs_and_outputs(sh, ir_var_out);
1770 }
1771
1772 if (prog->_LinkedShaders[MESA_SHADER_FRAGMENT] != NULL) {
1773 gl_shader *const sh = prog->_LinkedShaders[MESA_SHADER_FRAGMENT];
1774
1775 demote_shader_inputs_and_outputs(sh, ir_var_in);
1776 }
1777
1778 /* OpenGL ES requires that a vertex shader and a fragment shader both be
1779 * present in a linked program. By checking for use of shading language
1780 * version 1.00, we also catch the GL_ARB_ES2_compatibility case.
1781 */
1782 if (ctx->API == API_OPENGLES2 || prog->Version == 100) {
1783 if (prog->_LinkedShaders[MESA_SHADER_VERTEX] == NULL) {
1784 linker_error(prog, "program lacks a vertex shader\n");
1785 } else if (prog->_LinkedShaders[MESA_SHADER_FRAGMENT] == NULL) {
1786 linker_error(prog, "program lacks a fragment shader\n");
1787 }
1788 }
1789
1790 /* FINISHME: Assign fragment shader output locations. */
1791
1792 done:
1793 free(vert_shader_list);
1794
1795 for (unsigned i = 0; i < MESA_SHADER_TYPES; i++) {
1796 if (prog->_LinkedShaders[i] == NULL)
1797 continue;
1798
1799 /* Retain any live IR, but trash the rest. */
1800 reparent_ir(prog->_LinkedShaders[i]->ir, prog->_LinkedShaders[i]->ir);
1801 }
1802
1803 ralloc_free(mem_ctx);
1804 }