glsl/linker: Make separate allocations to track vertex and fragment shaders.
[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 "glsl_parser_extras.h"
70 #include "ir.h"
71 #include "program.h"
72 #include "program/hash_table.h"
73 #include "linker.h"
74 #include "link_varyings.h"
75 #include "ir_optimization.h"
76
77 extern "C" {
78 #include "main/shaderobj.h"
79 }
80
81 /**
82 * Visitor that determines whether or not a variable is ever written.
83 */
84 class find_assignment_visitor : public ir_hierarchical_visitor {
85 public:
86 find_assignment_visitor(const char *name)
87 : name(name), found(false)
88 {
89 /* empty */
90 }
91
92 virtual ir_visitor_status visit_enter(ir_assignment *ir)
93 {
94 ir_variable *const var = ir->lhs->variable_referenced();
95
96 if (strcmp(name, var->name) == 0) {
97 found = true;
98 return visit_stop;
99 }
100
101 return visit_continue_with_parent;
102 }
103
104 virtual ir_visitor_status visit_enter(ir_call *ir)
105 {
106 exec_list_iterator sig_iter = ir->callee->parameters.iterator();
107 foreach_iter(exec_list_iterator, iter, *ir) {
108 ir_rvalue *param_rval = (ir_rvalue *)iter.get();
109 ir_variable *sig_param = (ir_variable *)sig_iter.get();
110
111 if (sig_param->mode == ir_var_function_out ||
112 sig_param->mode == ir_var_function_inout) {
113 ir_variable *var = param_rval->variable_referenced();
114 if (var && strcmp(name, var->name) == 0) {
115 found = true;
116 return visit_stop;
117 }
118 }
119 sig_iter.next();
120 }
121
122 if (ir->return_deref != NULL) {
123 ir_variable *const var = ir->return_deref->variable_referenced();
124
125 if (strcmp(name, var->name) == 0) {
126 found = true;
127 return visit_stop;
128 }
129 }
130
131 return visit_continue_with_parent;
132 }
133
134 bool variable_found()
135 {
136 return found;
137 }
138
139 private:
140 const char *name; /**< Find writes to a variable with this name. */
141 bool found; /**< Was a write to the variable found? */
142 };
143
144
145 /**
146 * Visitor that determines whether or not a variable is ever read.
147 */
148 class find_deref_visitor : public ir_hierarchical_visitor {
149 public:
150 find_deref_visitor(const char *name)
151 : name(name), found(false)
152 {
153 /* empty */
154 }
155
156 virtual ir_visitor_status visit(ir_dereference_variable *ir)
157 {
158 if (strcmp(this->name, ir->var->name) == 0) {
159 this->found = true;
160 return visit_stop;
161 }
162
163 return visit_continue;
164 }
165
166 bool variable_found() const
167 {
168 return this->found;
169 }
170
171 private:
172 const char *name; /**< Find writes to a variable with this name. */
173 bool found; /**< Was a write to the variable found? */
174 };
175
176
177 void
178 linker_error(gl_shader_program *prog, const char *fmt, ...)
179 {
180 va_list ap;
181
182 ralloc_strcat(&prog->InfoLog, "error: ");
183 va_start(ap, fmt);
184 ralloc_vasprintf_append(&prog->InfoLog, fmt, ap);
185 va_end(ap);
186
187 prog->LinkStatus = false;
188 }
189
190
191 void
192 linker_warning(gl_shader_program *prog, const char *fmt, ...)
193 {
194 va_list ap;
195
196 ralloc_strcat(&prog->InfoLog, "error: ");
197 va_start(ap, fmt);
198 ralloc_vasprintf_append(&prog->InfoLog, fmt, ap);
199 va_end(ap);
200
201 }
202
203
204 /**
205 * Given a string identifying a program resource, break it into a base name
206 * and an optional array index in square brackets.
207 *
208 * If an array index is present, \c out_base_name_end is set to point to the
209 * "[" that precedes the array index, and the array index itself is returned
210 * as a long.
211 *
212 * If no array index is present (or if the array index is negative or
213 * mal-formed), \c out_base_name_end, is set to point to the null terminator
214 * at the end of the input string, and -1 is returned.
215 *
216 * Only the final array index is parsed; if the string contains other array
217 * indices (or structure field accesses), they are left in the base name.
218 *
219 * No attempt is made to check that the base name is properly formed;
220 * typically the caller will look up the base name in a hash table, so
221 * ill-formed base names simply turn into hash table lookup failures.
222 */
223 long
224 parse_program_resource_name(const GLchar *name,
225 const GLchar **out_base_name_end)
226 {
227 /* Section 7.3.1 ("Program Interfaces") of the OpenGL 4.3 spec says:
228 *
229 * "When an integer array element or block instance number is part of
230 * the name string, it will be specified in decimal form without a "+"
231 * or "-" sign or any extra leading zeroes. Additionally, the name
232 * string will not include white space anywhere in the string."
233 */
234
235 const size_t len = strlen(name);
236 *out_base_name_end = name + len;
237
238 if (len == 0 || name[len-1] != ']')
239 return -1;
240
241 /* Walk backwards over the string looking for a non-digit character. This
242 * had better be the opening bracket for an array index.
243 *
244 * Initially, i specifies the location of the ']'. Since the string may
245 * contain only the ']' charcater, walk backwards very carefully.
246 */
247 unsigned i;
248 for (i = len - 1; (i > 0) && isdigit(name[i-1]); --i)
249 /* empty */ ;
250
251 if ((i == 0) || name[i-1] != '[')
252 return -1;
253
254 long array_index = strtol(&name[i], NULL, 10);
255 if (array_index < 0)
256 return -1;
257
258 *out_base_name_end = name + (i - 1);
259 return array_index;
260 }
261
262
263 void
264 link_invalidate_variable_locations(gl_shader *sh, int input_base,
265 int output_base)
266 {
267 foreach_list(node, sh->ir) {
268 ir_variable *const var = ((ir_instruction *) node)->as_variable();
269
270 if (var == NULL)
271 continue;
272
273 int base;
274 switch (var->mode) {
275 case ir_var_shader_in:
276 base = input_base;
277 break;
278 case ir_var_shader_out:
279 base = output_base;
280 break;
281 default:
282 continue;
283 }
284
285 /* Only assign locations for generic attributes / varyings / etc.
286 */
287 if ((var->location >= base) && !var->explicit_location)
288 var->location = -1;
289
290 if ((var->location == -1) && !var->explicit_location) {
291 var->is_unmatched_generic_inout = 1;
292 var->location_frac = 0;
293 } else {
294 var->is_unmatched_generic_inout = 0;
295 }
296 }
297 }
298
299
300 /**
301 * Verify that a vertex shader executable meets all semantic requirements.
302 *
303 * Also sets prog->Vert.UsesClipDistance and prog->Vert.ClipDistanceArraySize
304 * as a side effect.
305 *
306 * \param shader Vertex shader executable to be verified
307 */
308 void
309 validate_vertex_shader_executable(struct gl_shader_program *prog,
310 struct gl_shader *shader)
311 {
312 if (shader == NULL)
313 return;
314
315 /* From the GLSL 1.10 spec, page 48:
316 *
317 * "The variable gl_Position is available only in the vertex
318 * language and is intended for writing the homogeneous vertex
319 * position. All executions of a well-formed vertex shader
320 * executable must write a value into this variable. [...] The
321 * variable gl_Position is available only in the vertex
322 * language and is intended for writing the homogeneous vertex
323 * position. All executions of a well-formed vertex shader
324 * executable must write a value into this variable."
325 *
326 * while in GLSL 1.40 this text is changed to:
327 *
328 * "The variable gl_Position is available only in the vertex
329 * language and is intended for writing the homogeneous vertex
330 * position. It can be written at any time during shader
331 * execution. It may also be read back by a vertex shader
332 * after being written. This value will be used by primitive
333 * assembly, clipping, culling, and other fixed functionality
334 * operations, if present, that operate on primitives after
335 * vertex processing has occurred. Its value is undefined if
336 * the vertex shader executable does not write gl_Position."
337 *
338 * GLSL ES 3.00 is similar to GLSL 1.40--failing to write to gl_Position is
339 * not an error.
340 */
341 if (prog->Version < (prog->IsES ? 300 : 140)) {
342 find_assignment_visitor find("gl_Position");
343 find.run(shader->ir);
344 if (!find.variable_found()) {
345 linker_error(prog, "vertex shader does not write to `gl_Position'\n");
346 return;
347 }
348 }
349
350 prog->Vert.ClipDistanceArraySize = 0;
351
352 if (!prog->IsES && prog->Version >= 130) {
353 /* From section 7.1 (Vertex Shader Special Variables) of the
354 * GLSL 1.30 spec:
355 *
356 * "It is an error for a shader to statically write both
357 * gl_ClipVertex and gl_ClipDistance."
358 *
359 * This does not apply to GLSL ES shaders, since GLSL ES defines neither
360 * gl_ClipVertex nor gl_ClipDistance.
361 */
362 find_assignment_visitor clip_vertex("gl_ClipVertex");
363 find_assignment_visitor clip_distance("gl_ClipDistance");
364
365 clip_vertex.run(shader->ir);
366 clip_distance.run(shader->ir);
367 if (clip_vertex.variable_found() && clip_distance.variable_found()) {
368 linker_error(prog, "vertex shader writes to both `gl_ClipVertex' "
369 "and `gl_ClipDistance'\n");
370 return;
371 }
372 prog->Vert.UsesClipDistance = clip_distance.variable_found();
373 ir_variable *clip_distance_var =
374 shader->symbols->get_variable("gl_ClipDistance");
375 if (clip_distance_var)
376 prog->Vert.ClipDistanceArraySize = clip_distance_var->type->length;
377 }
378 }
379
380
381 /**
382 * Verify that a fragment shader executable meets all semantic requirements
383 *
384 * \param shader Fragment shader executable to be verified
385 */
386 void
387 validate_fragment_shader_executable(struct gl_shader_program *prog,
388 struct gl_shader *shader)
389 {
390 if (shader == NULL)
391 return;
392
393 find_assignment_visitor frag_color("gl_FragColor");
394 find_assignment_visitor frag_data("gl_FragData");
395
396 frag_color.run(shader->ir);
397 frag_data.run(shader->ir);
398
399 if (frag_color.variable_found() && frag_data.variable_found()) {
400 linker_error(prog, "fragment shader writes to both "
401 "`gl_FragColor' and `gl_FragData'\n");
402 }
403 }
404
405
406 /**
407 * Generate a string describing the mode of a variable
408 */
409 static const char *
410 mode_string(const ir_variable *var)
411 {
412 switch (var->mode) {
413 case ir_var_auto:
414 return (var->read_only) ? "global constant" : "global variable";
415
416 case ir_var_uniform: return "uniform";
417 case ir_var_shader_in: return "shader input";
418 case ir_var_shader_out: return "shader output";
419
420 case ir_var_const_in:
421 case ir_var_temporary:
422 default:
423 assert(!"Should not get here.");
424 return "invalid variable";
425 }
426 }
427
428
429 /**
430 * Perform validation of global variables used across multiple shaders
431 */
432 void
433 cross_validate_globals(struct gl_shader_program *prog,
434 struct gl_shader **shader_list,
435 unsigned num_shaders,
436 bool uniforms_only)
437 {
438 /* Examine all of the uniforms in all of the shaders and cross validate
439 * them.
440 */
441 glsl_symbol_table variables;
442 for (unsigned i = 0; i < num_shaders; i++) {
443 if (shader_list[i] == NULL)
444 continue;
445
446 foreach_list(node, shader_list[i]->ir) {
447 ir_variable *const var = ((ir_instruction *) node)->as_variable();
448
449 if (var == NULL)
450 continue;
451
452 if (uniforms_only && (var->mode != ir_var_uniform))
453 continue;
454
455 /* Don't cross validate temporaries that are at global scope. These
456 * will eventually get pulled into the shaders 'main'.
457 */
458 if (var->mode == ir_var_temporary)
459 continue;
460
461 /* If a global with this name has already been seen, verify that the
462 * new instance has the same type. In addition, if the globals have
463 * initializers, the values of the initializers must be the same.
464 */
465 ir_variable *const existing = variables.get_variable(var->name);
466 if (existing != NULL) {
467 if (var->type != existing->type) {
468 /* Consider the types to be "the same" if both types are arrays
469 * of the same type and one of the arrays is implicitly sized.
470 * In addition, set the type of the linked variable to the
471 * explicitly sized array.
472 */
473 if (var->type->is_array()
474 && existing->type->is_array()
475 && (var->type->fields.array == existing->type->fields.array)
476 && ((var->type->length == 0)
477 || (existing->type->length == 0))) {
478 if (var->type->length != 0) {
479 existing->type = var->type;
480 }
481 } else {
482 linker_error(prog, "%s `%s' declared as type "
483 "`%s' and type `%s'\n",
484 mode_string(var),
485 var->name, var->type->name,
486 existing->type->name);
487 return;
488 }
489 }
490
491 if (var->explicit_location) {
492 if (existing->explicit_location
493 && (var->location != existing->location)) {
494 linker_error(prog, "explicit locations for %s "
495 "`%s' have differing values\n",
496 mode_string(var), var->name);
497 return;
498 }
499
500 existing->location = var->location;
501 existing->explicit_location = true;
502 }
503
504 /* From the GLSL 4.20 specification:
505 * "A link error will result if two compilation units in a program
506 * specify different integer-constant bindings for the same
507 * opaque-uniform name. However, it is not an error to specify a
508 * binding on some but not all declarations for the same name"
509 */
510 if (var->explicit_binding) {
511 if (existing->explicit_binding &&
512 var->binding != existing->binding) {
513 linker_error(prog, "explicit bindings for %s "
514 "`%s' have differing values\n",
515 mode_string(var), var->name);
516 return;
517 }
518
519 existing->binding = var->binding;
520 existing->explicit_binding = true;
521 }
522
523 /* Validate layout qualifiers for gl_FragDepth.
524 *
525 * From the AMD/ARB_conservative_depth specs:
526 *
527 * "If gl_FragDepth is redeclared in any fragment shader in a
528 * program, it must be redeclared in all fragment shaders in
529 * that program that have static assignments to
530 * gl_FragDepth. All redeclarations of gl_FragDepth in all
531 * fragment shaders in a single program must have the same set
532 * of qualifiers."
533 */
534 if (strcmp(var->name, "gl_FragDepth") == 0) {
535 bool layout_declared = var->depth_layout != ir_depth_layout_none;
536 bool layout_differs =
537 var->depth_layout != existing->depth_layout;
538
539 if (layout_declared && layout_differs) {
540 linker_error(prog,
541 "All redeclarations of gl_FragDepth in all "
542 "fragment shaders in a single program must have "
543 "the same set of qualifiers.");
544 }
545
546 if (var->used && layout_differs) {
547 linker_error(prog,
548 "If gl_FragDepth is redeclared with a layout "
549 "qualifier in any fragment shader, it must be "
550 "redeclared with the same layout qualifier in "
551 "all fragment shaders that have assignments to "
552 "gl_FragDepth");
553 }
554 }
555
556 /* Page 35 (page 41 of the PDF) of the GLSL 4.20 spec says:
557 *
558 * "If a shared global has multiple initializers, the
559 * initializers must all be constant expressions, and they
560 * must all have the same value. Otherwise, a link error will
561 * result. (A shared global having only one initializer does
562 * not require that initializer to be a constant expression.)"
563 *
564 * Previous to 4.20 the GLSL spec simply said that initializers
565 * must have the same value. In this case of non-constant
566 * initializers, this was impossible to determine. As a result,
567 * no vendor actually implemented that behavior. The 4.20
568 * behavior matches the implemented behavior of at least one other
569 * vendor, so we'll implement that for all GLSL versions.
570 */
571 if (var->constant_initializer != NULL) {
572 if (existing->constant_initializer != NULL) {
573 if (!var->constant_initializer->has_value(existing->constant_initializer)) {
574 linker_error(prog, "initializers for %s "
575 "`%s' have differing values\n",
576 mode_string(var), var->name);
577 return;
578 }
579 } else {
580 /* If the first-seen instance of a particular uniform did not
581 * have an initializer but a later instance does, copy the
582 * initializer to the version stored in the symbol table.
583 */
584 /* FINISHME: This is wrong. The constant_value field should
585 * FINISHME: not be modified! Imagine a case where a shader
586 * FINISHME: without an initializer is linked in two different
587 * FINISHME: programs with shaders that have differing
588 * FINISHME: initializers. Linking with the first will
589 * FINISHME: modify the shader, and linking with the second
590 * FINISHME: will fail.
591 */
592 existing->constant_initializer =
593 var->constant_initializer->clone(ralloc_parent(existing),
594 NULL);
595 }
596 }
597
598 if (var->has_initializer) {
599 if (existing->has_initializer
600 && (var->constant_initializer == NULL
601 || existing->constant_initializer == NULL)) {
602 linker_error(prog,
603 "shared global variable `%s' has multiple "
604 "non-constant initializers.\n",
605 var->name);
606 return;
607 }
608
609 /* Some instance had an initializer, so keep track of that. In
610 * this location, all sorts of initializers (constant or
611 * otherwise) will propagate the existence to the variable
612 * stored in the symbol table.
613 */
614 existing->has_initializer = true;
615 }
616
617 if (existing->invariant != var->invariant) {
618 linker_error(prog, "declarations for %s `%s' have "
619 "mismatching invariant qualifiers\n",
620 mode_string(var), var->name);
621 return;
622 }
623 if (existing->centroid != var->centroid) {
624 linker_error(prog, "declarations for %s `%s' have "
625 "mismatching centroid qualifiers\n",
626 mode_string(var), var->name);
627 return;
628 }
629 } else
630 variables.add_variable(var);
631 }
632 }
633 }
634
635
636 /**
637 * Perform validation of uniforms used across multiple shader stages
638 */
639 void
640 cross_validate_uniforms(struct gl_shader_program *prog)
641 {
642 cross_validate_globals(prog, prog->_LinkedShaders,
643 MESA_SHADER_TYPES, true);
644 }
645
646 /**
647 * Accumulates the array of prog->UniformBlocks and checks that all
648 * definitons of blocks agree on their contents.
649 */
650 static bool
651 interstage_cross_validate_uniform_blocks(struct gl_shader_program *prog)
652 {
653 unsigned max_num_uniform_blocks = 0;
654 for (unsigned i = 0; i < MESA_SHADER_TYPES; i++) {
655 if (prog->_LinkedShaders[i])
656 max_num_uniform_blocks += prog->_LinkedShaders[i]->NumUniformBlocks;
657 }
658
659 for (unsigned i = 0; i < MESA_SHADER_TYPES; i++) {
660 struct gl_shader *sh = prog->_LinkedShaders[i];
661
662 prog->UniformBlockStageIndex[i] = ralloc_array(prog, int,
663 max_num_uniform_blocks);
664 for (unsigned int j = 0; j < max_num_uniform_blocks; j++)
665 prog->UniformBlockStageIndex[i][j] = -1;
666
667 if (sh == NULL)
668 continue;
669
670 for (unsigned int j = 0; j < sh->NumUniformBlocks; j++) {
671 int index = link_cross_validate_uniform_block(prog,
672 &prog->UniformBlocks,
673 &prog->NumUniformBlocks,
674 &sh->UniformBlocks[j]);
675
676 if (index == -1) {
677 linker_error(prog, "uniform block `%s' has mismatching definitions",
678 sh->UniformBlocks[j].Name);
679 return false;
680 }
681
682 prog->UniformBlockStageIndex[i][index] = j;
683 }
684 }
685
686 return true;
687 }
688
689
690 /**
691 * Populates a shaders symbol table with all global declarations
692 */
693 static void
694 populate_symbol_table(gl_shader *sh)
695 {
696 sh->symbols = new(sh) glsl_symbol_table;
697
698 foreach_list(node, sh->ir) {
699 ir_instruction *const inst = (ir_instruction *) node;
700 ir_variable *var;
701 ir_function *func;
702
703 if ((func = inst->as_function()) != NULL) {
704 sh->symbols->add_function(func);
705 } else if ((var = inst->as_variable()) != NULL) {
706 sh->symbols->add_variable(var);
707 }
708 }
709 }
710
711
712 /**
713 * Remap variables referenced in an instruction tree
714 *
715 * This is used when instruction trees are cloned from one shader and placed in
716 * another. These trees will contain references to \c ir_variable nodes that
717 * do not exist in the target shader. This function finds these \c ir_variable
718 * references and replaces the references with matching variables in the target
719 * shader.
720 *
721 * If there is no matching variable in the target shader, a clone of the
722 * \c ir_variable is made and added to the target shader. The new variable is
723 * added to \b both the instruction stream and the symbol table.
724 *
725 * \param inst IR tree that is to be processed.
726 * \param symbols Symbol table containing global scope symbols in the
727 * linked shader.
728 * \param instructions Instruction stream where new variable declarations
729 * should be added.
730 */
731 void
732 remap_variables(ir_instruction *inst, struct gl_shader *target,
733 hash_table *temps)
734 {
735 class remap_visitor : public ir_hierarchical_visitor {
736 public:
737 remap_visitor(struct gl_shader *target,
738 hash_table *temps)
739 {
740 this->target = target;
741 this->symbols = target->symbols;
742 this->instructions = target->ir;
743 this->temps = temps;
744 }
745
746 virtual ir_visitor_status visit(ir_dereference_variable *ir)
747 {
748 if (ir->var->mode == ir_var_temporary) {
749 ir_variable *var = (ir_variable *) hash_table_find(temps, ir->var);
750
751 assert(var != NULL);
752 ir->var = var;
753 return visit_continue;
754 }
755
756 ir_variable *const existing =
757 this->symbols->get_variable(ir->var->name);
758 if (existing != NULL)
759 ir->var = existing;
760 else {
761 ir_variable *copy = ir->var->clone(this->target, NULL);
762
763 this->symbols->add_variable(copy);
764 this->instructions->push_head(copy);
765 ir->var = copy;
766 }
767
768 return visit_continue;
769 }
770
771 private:
772 struct gl_shader *target;
773 glsl_symbol_table *symbols;
774 exec_list *instructions;
775 hash_table *temps;
776 };
777
778 remap_visitor v(target, temps);
779
780 inst->accept(&v);
781 }
782
783
784 /**
785 * Move non-declarations from one instruction stream to another
786 *
787 * The intended usage pattern of this function is to pass the pointer to the
788 * head sentinel of a list (i.e., a pointer to the list cast to an \c exec_node
789 * pointer) for \c last and \c false for \c make_copies on the first
790 * call. Successive calls pass the return value of the previous call for
791 * \c last and \c true for \c make_copies.
792 *
793 * \param instructions Source instruction stream
794 * \param last Instruction after which new instructions should be
795 * inserted in the target instruction stream
796 * \param make_copies Flag selecting whether instructions in \c instructions
797 * should be copied (via \c ir_instruction::clone) into the
798 * target list or moved.
799 *
800 * \return
801 * The new "last" instruction in the target instruction stream. This pointer
802 * is suitable for use as the \c last parameter of a later call to this
803 * function.
804 */
805 exec_node *
806 move_non_declarations(exec_list *instructions, exec_node *last,
807 bool make_copies, gl_shader *target)
808 {
809 hash_table *temps = NULL;
810
811 if (make_copies)
812 temps = hash_table_ctor(0, hash_table_pointer_hash,
813 hash_table_pointer_compare);
814
815 foreach_list_safe(node, instructions) {
816 ir_instruction *inst = (ir_instruction *) node;
817
818 if (inst->as_function())
819 continue;
820
821 ir_variable *var = inst->as_variable();
822 if ((var != NULL) && (var->mode != ir_var_temporary))
823 continue;
824
825 assert(inst->as_assignment()
826 || inst->as_call()
827 || inst->as_if() /* for initializers with the ?: operator */
828 || ((var != NULL) && (var->mode == ir_var_temporary)));
829
830 if (make_copies) {
831 inst = inst->clone(target, NULL);
832
833 if (var != NULL)
834 hash_table_insert(temps, inst, var);
835 else
836 remap_variables(inst, target, temps);
837 } else {
838 inst->remove();
839 }
840
841 last->insert_after(inst);
842 last = inst;
843 }
844
845 if (make_copies)
846 hash_table_dtor(temps);
847
848 return last;
849 }
850
851 /**
852 * Get the function signature for main from a shader
853 */
854 static ir_function_signature *
855 get_main_function_signature(gl_shader *sh)
856 {
857 ir_function *const f = sh->symbols->get_function("main");
858 if (f != NULL) {
859 exec_list void_parameters;
860
861 /* Look for the 'void main()' signature and ensure that it's defined.
862 * This keeps the linker from accidentally pick a shader that just
863 * contains a prototype for main.
864 *
865 * We don't have to check for multiple definitions of main (in multiple
866 * shaders) because that would have already been caught above.
867 */
868 ir_function_signature *sig = f->matching_signature(&void_parameters);
869 if ((sig != NULL) && sig->is_defined) {
870 return sig;
871 }
872 }
873
874 return NULL;
875 }
876
877
878 /**
879 * This class is only used in link_intrastage_shaders() below but declaring
880 * it inside that function leads to compiler warnings with some versions of
881 * gcc.
882 */
883 class array_sizing_visitor : public ir_hierarchical_visitor {
884 public:
885 virtual ir_visitor_status visit(ir_variable *var)
886 {
887 if (var->type->is_array() && (var->type->length == 0)) {
888 const glsl_type *type =
889 glsl_type::get_array_instance(var->type->fields.array,
890 var->max_array_access + 1);
891 assert(type != NULL);
892 var->type = type;
893 }
894 return visit_continue;
895 }
896 };
897
898 /**
899 * Combine a group of shaders for a single stage to generate a linked shader
900 *
901 * \note
902 * If this function is supplied a single shader, it is cloned, and the new
903 * shader is returned.
904 */
905 static struct gl_shader *
906 link_intrastage_shaders(void *mem_ctx,
907 struct gl_context *ctx,
908 struct gl_shader_program *prog,
909 struct gl_shader **shader_list,
910 unsigned num_shaders)
911 {
912 struct gl_uniform_block *uniform_blocks = NULL;
913
914 /* Check that global variables defined in multiple shaders are consistent.
915 */
916 cross_validate_globals(prog, shader_list, num_shaders, false);
917 if (!prog->LinkStatus)
918 return NULL;
919
920 /* Check that interface blocks defined in multiple shaders are consistent.
921 */
922 validate_intrastage_interface_blocks(prog, (const gl_shader **)shader_list,
923 num_shaders);
924 if (!prog->LinkStatus)
925 return NULL;
926
927 /* Link up uniform blocks defined within this stage. */
928 const unsigned num_uniform_blocks =
929 link_uniform_blocks(mem_ctx, prog, shader_list, num_shaders,
930 &uniform_blocks);
931
932 /* Check that there is only a single definition of each function signature
933 * across all shaders.
934 */
935 for (unsigned i = 0; i < (num_shaders - 1); i++) {
936 foreach_list(node, shader_list[i]->ir) {
937 ir_function *const f = ((ir_instruction *) node)->as_function();
938
939 if (f == NULL)
940 continue;
941
942 for (unsigned j = i + 1; j < num_shaders; j++) {
943 ir_function *const other =
944 shader_list[j]->symbols->get_function(f->name);
945
946 /* If the other shader has no function (and therefore no function
947 * signatures) with the same name, skip to the next shader.
948 */
949 if (other == NULL)
950 continue;
951
952 foreach_iter (exec_list_iterator, iter, *f) {
953 ir_function_signature *sig =
954 (ir_function_signature *) iter.get();
955
956 if (!sig->is_defined || sig->is_builtin)
957 continue;
958
959 ir_function_signature *other_sig =
960 other->exact_matching_signature(& sig->parameters);
961
962 if ((other_sig != NULL) && other_sig->is_defined
963 && !other_sig->is_builtin) {
964 linker_error(prog, "function `%s' is multiply defined",
965 f->name);
966 return NULL;
967 }
968 }
969 }
970 }
971 }
972
973 /* Find the shader that defines main, and make a clone of it.
974 *
975 * Starting with the clone, search for undefined references. If one is
976 * found, find the shader that defines it. Clone the reference and add
977 * it to the shader. Repeat until there are no undefined references or
978 * until a reference cannot be resolved.
979 */
980 gl_shader *main = NULL;
981 for (unsigned i = 0; i < num_shaders; i++) {
982 if (get_main_function_signature(shader_list[i]) != NULL) {
983 main = shader_list[i];
984 break;
985 }
986 }
987
988 if (main == NULL) {
989 linker_error(prog, "%s shader lacks `main'\n",
990 _mesa_glsl_shader_target_name(shader_list[0]->Type));
991 return NULL;
992 }
993
994 gl_shader *linked = ctx->Driver.NewShader(NULL, 0, main->Type);
995 linked->ir = new(linked) exec_list;
996 clone_ir_list(mem_ctx, linked->ir, main->ir);
997
998 linked->UniformBlocks = uniform_blocks;
999 linked->NumUniformBlocks = num_uniform_blocks;
1000 ralloc_steal(linked, linked->UniformBlocks);
1001
1002 populate_symbol_table(linked);
1003
1004 /* The a pointer to the main function in the final linked shader (i.e., the
1005 * copy of the original shader that contained the main function).
1006 */
1007 ir_function_signature *const main_sig = get_main_function_signature(linked);
1008
1009 /* Move any instructions other than variable declarations or function
1010 * declarations into main.
1011 */
1012 exec_node *insertion_point =
1013 move_non_declarations(linked->ir, (exec_node *) &main_sig->body, false,
1014 linked);
1015
1016 for (unsigned i = 0; i < num_shaders; i++) {
1017 if (shader_list[i] == main)
1018 continue;
1019
1020 insertion_point = move_non_declarations(shader_list[i]->ir,
1021 insertion_point, true, linked);
1022 }
1023
1024 /* Resolve initializers for global variables in the linked shader.
1025 */
1026 unsigned num_linking_shaders = num_shaders;
1027 for (unsigned i = 0; i < num_shaders; i++)
1028 num_linking_shaders += shader_list[i]->num_builtins_to_link;
1029
1030 gl_shader **linking_shaders =
1031 (gl_shader **) calloc(num_linking_shaders, sizeof(gl_shader *));
1032
1033 memcpy(linking_shaders, shader_list,
1034 sizeof(linking_shaders[0]) * num_shaders);
1035
1036 unsigned idx = num_shaders;
1037 for (unsigned i = 0; i < num_shaders; i++) {
1038 memcpy(&linking_shaders[idx], shader_list[i]->builtins_to_link,
1039 sizeof(linking_shaders[0]) * shader_list[i]->num_builtins_to_link);
1040 idx += shader_list[i]->num_builtins_to_link;
1041 }
1042
1043 assert(idx == num_linking_shaders);
1044
1045 if (!link_function_calls(prog, linked, linking_shaders,
1046 num_linking_shaders)) {
1047 ctx->Driver.DeleteShader(ctx, linked);
1048 linked = NULL;
1049 }
1050
1051 free(linking_shaders);
1052
1053 /* At this point linked should contain all of the linked IR, so
1054 * validate it to make sure nothing went wrong.
1055 */
1056 if (linked)
1057 validate_ir_tree(linked->ir);
1058
1059 /* Make a pass over all variable declarations to ensure that arrays with
1060 * unspecified sizes have a size specified. The size is inferred from the
1061 * max_array_access field.
1062 */
1063 if (linked != NULL) {
1064 array_sizing_visitor v;
1065
1066 v.run(linked->ir);
1067 }
1068
1069 return linked;
1070 }
1071
1072 /**
1073 * Update the sizes of linked shader uniform arrays to the maximum
1074 * array index used.
1075 *
1076 * From page 81 (page 95 of the PDF) of the OpenGL 2.1 spec:
1077 *
1078 * If one or more elements of an array are active,
1079 * GetActiveUniform will return the name of the array in name,
1080 * subject to the restrictions listed above. The type of the array
1081 * is returned in type. The size parameter contains the highest
1082 * array element index used, plus one. The compiler or linker
1083 * determines the highest index used. There will be only one
1084 * active uniform reported by the GL per uniform array.
1085
1086 */
1087 static void
1088 update_array_sizes(struct gl_shader_program *prog)
1089 {
1090 for (unsigned i = 0; i < MESA_SHADER_TYPES; i++) {
1091 if (prog->_LinkedShaders[i] == NULL)
1092 continue;
1093
1094 foreach_list(node, prog->_LinkedShaders[i]->ir) {
1095 ir_variable *const var = ((ir_instruction *) node)->as_variable();
1096
1097 if ((var == NULL) || (var->mode != ir_var_uniform) ||
1098 !var->type->is_array())
1099 continue;
1100
1101 /* GL_ARB_uniform_buffer_object says that std140 uniforms
1102 * will not be eliminated. Since we always do std140, just
1103 * don't resize arrays in UBOs.
1104 */
1105 if (var->is_in_uniform_block())
1106 continue;
1107
1108 unsigned int size = var->max_array_access;
1109 for (unsigned j = 0; j < MESA_SHADER_TYPES; j++) {
1110 if (prog->_LinkedShaders[j] == NULL)
1111 continue;
1112
1113 foreach_list(node2, prog->_LinkedShaders[j]->ir) {
1114 ir_variable *other_var = ((ir_instruction *) node2)->as_variable();
1115 if (!other_var)
1116 continue;
1117
1118 if (strcmp(var->name, other_var->name) == 0 &&
1119 other_var->max_array_access > size) {
1120 size = other_var->max_array_access;
1121 }
1122 }
1123 }
1124
1125 if (size + 1 != var->type->length) {
1126 /* If this is a built-in uniform (i.e., it's backed by some
1127 * fixed-function state), adjust the number of state slots to
1128 * match the new array size. The number of slots per array entry
1129 * is not known. It seems safe to assume that the total number of
1130 * slots is an integer multiple of the number of array elements.
1131 * Determine the number of slots per array element by dividing by
1132 * the old (total) size.
1133 */
1134 if (var->num_state_slots > 0) {
1135 var->num_state_slots = (size + 1)
1136 * (var->num_state_slots / var->type->length);
1137 }
1138
1139 var->type = glsl_type::get_array_instance(var->type->fields.array,
1140 size + 1);
1141 /* FINISHME: We should update the types of array
1142 * dereferences of this variable now.
1143 */
1144 }
1145 }
1146 }
1147 }
1148
1149 /**
1150 * Find a contiguous set of available bits in a bitmask.
1151 *
1152 * \param used_mask Bits representing used (1) and unused (0) locations
1153 * \param needed_count Number of contiguous bits needed.
1154 *
1155 * \return
1156 * Base location of the available bits on success or -1 on failure.
1157 */
1158 int
1159 find_available_slots(unsigned used_mask, unsigned needed_count)
1160 {
1161 unsigned needed_mask = (1 << needed_count) - 1;
1162 const int max_bit_to_test = (8 * sizeof(used_mask)) - needed_count;
1163
1164 /* The comparison to 32 is redundant, but without it GCC emits "warning:
1165 * cannot optimize possibly infinite loops" for the loop below.
1166 */
1167 if ((needed_count == 0) || (max_bit_to_test < 0) || (max_bit_to_test > 32))
1168 return -1;
1169
1170 for (int i = 0; i <= max_bit_to_test; i++) {
1171 if ((needed_mask & ~used_mask) == needed_mask)
1172 return i;
1173
1174 needed_mask <<= 1;
1175 }
1176
1177 return -1;
1178 }
1179
1180
1181 /**
1182 * Assign locations for either VS inputs for FS outputs
1183 *
1184 * \param prog Shader program whose variables need locations assigned
1185 * \param target_index Selector for the program target to receive location
1186 * assignmnets. Must be either \c MESA_SHADER_VERTEX or
1187 * \c MESA_SHADER_FRAGMENT.
1188 * \param max_index Maximum number of generic locations. This corresponds
1189 * to either the maximum number of draw buffers or the
1190 * maximum number of generic attributes.
1191 *
1192 * \return
1193 * If locations are successfully assigned, true is returned. Otherwise an
1194 * error is emitted to the shader link log and false is returned.
1195 */
1196 bool
1197 assign_attribute_or_color_locations(gl_shader_program *prog,
1198 unsigned target_index,
1199 unsigned max_index)
1200 {
1201 /* Mark invalid locations as being used.
1202 */
1203 unsigned used_locations = (max_index >= 32)
1204 ? ~0 : ~((1 << max_index) - 1);
1205
1206 assert((target_index == MESA_SHADER_VERTEX)
1207 || (target_index == MESA_SHADER_FRAGMENT));
1208
1209 gl_shader *const sh = prog->_LinkedShaders[target_index];
1210 if (sh == NULL)
1211 return true;
1212
1213 /* Operate in a total of four passes.
1214 *
1215 * 1. Invalidate the location assignments for all vertex shader inputs.
1216 *
1217 * 2. Assign locations for inputs that have user-defined (via
1218 * glBindVertexAttribLocation) locations and outputs that have
1219 * user-defined locations (via glBindFragDataLocation).
1220 *
1221 * 3. Sort the attributes without assigned locations by number of slots
1222 * required in decreasing order. Fragmentation caused by attribute
1223 * locations assigned by the application may prevent large attributes
1224 * from having enough contiguous space.
1225 *
1226 * 4. Assign locations to any inputs without assigned locations.
1227 */
1228
1229 const int generic_base = (target_index == MESA_SHADER_VERTEX)
1230 ? (int) VERT_ATTRIB_GENERIC0 : (int) FRAG_RESULT_DATA0;
1231
1232 const enum ir_variable_mode direction =
1233 (target_index == MESA_SHADER_VERTEX)
1234 ? ir_var_shader_in : ir_var_shader_out;
1235
1236
1237 /* Temporary storage for the set of attributes that need locations assigned.
1238 */
1239 struct temp_attr {
1240 unsigned slots;
1241 ir_variable *var;
1242
1243 /* Used below in the call to qsort. */
1244 static int compare(const void *a, const void *b)
1245 {
1246 const temp_attr *const l = (const temp_attr *) a;
1247 const temp_attr *const r = (const temp_attr *) b;
1248
1249 /* Reversed because we want a descending order sort below. */
1250 return r->slots - l->slots;
1251 }
1252 } to_assign[16];
1253
1254 unsigned num_attr = 0;
1255
1256 foreach_list(node, sh->ir) {
1257 ir_variable *const var = ((ir_instruction *) node)->as_variable();
1258
1259 if ((var == NULL) || (var->mode != (unsigned) direction))
1260 continue;
1261
1262 if (var->explicit_location) {
1263 if ((var->location >= (int)(max_index + generic_base))
1264 || (var->location < 0)) {
1265 linker_error(prog,
1266 "invalid explicit location %d specified for `%s'\n",
1267 (var->location < 0)
1268 ? var->location : var->location - generic_base,
1269 var->name);
1270 return false;
1271 }
1272 } else if (target_index == MESA_SHADER_VERTEX) {
1273 unsigned binding;
1274
1275 if (prog->AttributeBindings->get(binding, var->name)) {
1276 assert(binding >= VERT_ATTRIB_GENERIC0);
1277 var->location = binding;
1278 var->is_unmatched_generic_inout = 0;
1279 }
1280 } else if (target_index == MESA_SHADER_FRAGMENT) {
1281 unsigned binding;
1282 unsigned index;
1283
1284 if (prog->FragDataBindings->get(binding, var->name)) {
1285 assert(binding >= FRAG_RESULT_DATA0);
1286 var->location = binding;
1287 var->is_unmatched_generic_inout = 0;
1288
1289 if (prog->FragDataIndexBindings->get(index, var->name)) {
1290 var->index = index;
1291 }
1292 }
1293 }
1294
1295 /* If the variable is not a built-in and has a location statically
1296 * assigned in the shader (presumably via a layout qualifier), make sure
1297 * that it doesn't collide with other assigned locations. Otherwise,
1298 * add it to the list of variables that need linker-assigned locations.
1299 */
1300 const unsigned slots = var->type->count_attribute_slots();
1301 if (var->location != -1) {
1302 if (var->location >= generic_base && var->index < 1) {
1303 /* From page 61 of the OpenGL 4.0 spec:
1304 *
1305 * "LinkProgram will fail if the attribute bindings assigned
1306 * by BindAttribLocation do not leave not enough space to
1307 * assign a location for an active matrix attribute or an
1308 * active attribute array, both of which require multiple
1309 * contiguous generic attributes."
1310 *
1311 * Previous versions of the spec contain similar language but omit
1312 * the bit about attribute arrays.
1313 *
1314 * Page 61 of the OpenGL 4.0 spec also says:
1315 *
1316 * "It is possible for an application to bind more than one
1317 * attribute name to the same location. This is referred to as
1318 * aliasing. This will only work if only one of the aliased
1319 * attributes is active in the executable program, or if no
1320 * path through the shader consumes more than one attribute of
1321 * a set of attributes aliased to the same location. A link
1322 * error can occur if the linker determines that every path
1323 * through the shader consumes multiple aliased attributes,
1324 * but implementations are not required to generate an error
1325 * in this case."
1326 *
1327 * These two paragraphs are either somewhat contradictory, or I
1328 * don't fully understand one or both of them.
1329 */
1330 /* FINISHME: The code as currently written does not support
1331 * FINISHME: attribute location aliasing (see comment above).
1332 */
1333 /* Mask representing the contiguous slots that will be used by
1334 * this attribute.
1335 */
1336 const unsigned attr = var->location - generic_base;
1337 const unsigned use_mask = (1 << slots) - 1;
1338
1339 /* Generate a link error if the set of bits requested for this
1340 * attribute overlaps any previously allocated bits.
1341 */
1342 if ((~(use_mask << attr) & used_locations) != used_locations) {
1343 const char *const string = (target_index == MESA_SHADER_VERTEX)
1344 ? "vertex shader input" : "fragment shader output";
1345 linker_error(prog,
1346 "insufficient contiguous locations "
1347 "available for %s `%s' %d %d %d", string,
1348 var->name, used_locations, use_mask, attr);
1349 return false;
1350 }
1351
1352 used_locations |= (use_mask << attr);
1353 }
1354
1355 continue;
1356 }
1357
1358 to_assign[num_attr].slots = slots;
1359 to_assign[num_attr].var = var;
1360 num_attr++;
1361 }
1362
1363 /* If all of the attributes were assigned locations by the application (or
1364 * are built-in attributes with fixed locations), return early. This should
1365 * be the common case.
1366 */
1367 if (num_attr == 0)
1368 return true;
1369
1370 qsort(to_assign, num_attr, sizeof(to_assign[0]), temp_attr::compare);
1371
1372 if (target_index == MESA_SHADER_VERTEX) {
1373 /* VERT_ATTRIB_GENERIC0 is a pseudo-alias for VERT_ATTRIB_POS. It can
1374 * only be explicitly assigned by via glBindAttribLocation. Mark it as
1375 * reserved to prevent it from being automatically allocated below.
1376 */
1377 find_deref_visitor find("gl_Vertex");
1378 find.run(sh->ir);
1379 if (find.variable_found())
1380 used_locations |= (1 << 0);
1381 }
1382
1383 for (unsigned i = 0; i < num_attr; i++) {
1384 /* Mask representing the contiguous slots that will be used by this
1385 * attribute.
1386 */
1387 const unsigned use_mask = (1 << to_assign[i].slots) - 1;
1388
1389 int location = find_available_slots(used_locations, to_assign[i].slots);
1390
1391 if (location < 0) {
1392 const char *const string = (target_index == MESA_SHADER_VERTEX)
1393 ? "vertex shader input" : "fragment shader output";
1394
1395 linker_error(prog,
1396 "insufficient contiguous locations "
1397 "available for %s `%s'",
1398 string, to_assign[i].var->name);
1399 return false;
1400 }
1401
1402 to_assign[i].var->location = generic_base + location;
1403 to_assign[i].var->is_unmatched_generic_inout = 0;
1404 used_locations |= (use_mask << location);
1405 }
1406
1407 return true;
1408 }
1409
1410
1411 /**
1412 * Demote shader inputs and outputs that are not used in other stages
1413 */
1414 void
1415 demote_shader_inputs_and_outputs(gl_shader *sh, enum ir_variable_mode mode)
1416 {
1417 foreach_list(node, sh->ir) {
1418 ir_variable *const var = ((ir_instruction *) node)->as_variable();
1419
1420 if ((var == NULL) || (var->mode != int(mode)))
1421 continue;
1422
1423 /* A shader 'in' or 'out' variable is only really an input or output if
1424 * its value is used by other shader stages. This will cause the variable
1425 * to have a location assigned.
1426 */
1427 if (var->is_unmatched_generic_inout) {
1428 var->mode = ir_var_auto;
1429 }
1430 }
1431 }
1432
1433
1434 /**
1435 * Store the gl_FragDepth layout in the gl_shader_program struct.
1436 */
1437 static void
1438 store_fragdepth_layout(struct gl_shader_program *prog)
1439 {
1440 if (prog->_LinkedShaders[MESA_SHADER_FRAGMENT] == NULL) {
1441 return;
1442 }
1443
1444 struct exec_list *ir = prog->_LinkedShaders[MESA_SHADER_FRAGMENT]->ir;
1445
1446 /* We don't look up the gl_FragDepth symbol directly because if
1447 * gl_FragDepth is not used in the shader, it's removed from the IR.
1448 * However, the symbol won't be removed from the symbol table.
1449 *
1450 * We're only interested in the cases where the variable is NOT removed
1451 * from the IR.
1452 */
1453 foreach_list(node, ir) {
1454 ir_variable *const var = ((ir_instruction *) node)->as_variable();
1455
1456 if (var == NULL || var->mode != ir_var_shader_out) {
1457 continue;
1458 }
1459
1460 if (strcmp(var->name, "gl_FragDepth") == 0) {
1461 switch (var->depth_layout) {
1462 case ir_depth_layout_none:
1463 prog->FragDepthLayout = FRAG_DEPTH_LAYOUT_NONE;
1464 return;
1465 case ir_depth_layout_any:
1466 prog->FragDepthLayout = FRAG_DEPTH_LAYOUT_ANY;
1467 return;
1468 case ir_depth_layout_greater:
1469 prog->FragDepthLayout = FRAG_DEPTH_LAYOUT_GREATER;
1470 return;
1471 case ir_depth_layout_less:
1472 prog->FragDepthLayout = FRAG_DEPTH_LAYOUT_LESS;
1473 return;
1474 case ir_depth_layout_unchanged:
1475 prog->FragDepthLayout = FRAG_DEPTH_LAYOUT_UNCHANGED;
1476 return;
1477 default:
1478 assert(0);
1479 return;
1480 }
1481 }
1482 }
1483 }
1484
1485 /**
1486 * Validate the resources used by a program versus the implementation limits
1487 */
1488 static void
1489 check_resources(struct gl_context *ctx, struct gl_shader_program *prog)
1490 {
1491 static const char *const shader_names[MESA_SHADER_TYPES] = {
1492 "vertex", "geometry", "fragment"
1493 };
1494
1495 const unsigned max_samplers[MESA_SHADER_TYPES] = {
1496 ctx->Const.VertexProgram.MaxTextureImageUnits,
1497 ctx->Const.GeometryProgram.MaxTextureImageUnits,
1498 ctx->Const.FragmentProgram.MaxTextureImageUnits
1499 };
1500
1501 const unsigned max_default_uniform_components[MESA_SHADER_TYPES] = {
1502 ctx->Const.VertexProgram.MaxUniformComponents,
1503 ctx->Const.GeometryProgram.MaxUniformComponents,
1504 ctx->Const.FragmentProgram.MaxUniformComponents
1505 };
1506
1507 const unsigned max_combined_uniform_components[MESA_SHADER_TYPES] = {
1508 ctx->Const.VertexProgram.MaxCombinedUniformComponents,
1509 ctx->Const.GeometryProgram.MaxCombinedUniformComponents,
1510 ctx->Const.FragmentProgram.MaxCombinedUniformComponents
1511 };
1512
1513 const unsigned max_uniform_blocks[MESA_SHADER_TYPES] = {
1514 ctx->Const.VertexProgram.MaxUniformBlocks,
1515 ctx->Const.GeometryProgram.MaxUniformBlocks,
1516 ctx->Const.FragmentProgram.MaxUniformBlocks
1517 };
1518
1519 for (unsigned i = 0; i < MESA_SHADER_TYPES; i++) {
1520 struct gl_shader *sh = prog->_LinkedShaders[i];
1521
1522 if (sh == NULL)
1523 continue;
1524
1525 if (sh->num_samplers > max_samplers[i]) {
1526 linker_error(prog, "Too many %s shader texture samplers",
1527 shader_names[i]);
1528 }
1529
1530 if (sh->num_uniform_components > max_default_uniform_components[i]) {
1531 if (ctx->Const.GLSLSkipStrictMaxUniformLimitCheck) {
1532 linker_warning(prog, "Too many %s shader default uniform block "
1533 "components, but the driver will try to optimize "
1534 "them out; this is non-portable out-of-spec "
1535 "behavior\n",
1536 shader_names[i]);
1537 } else {
1538 linker_error(prog, "Too many %s shader default uniform block "
1539 "components",
1540 shader_names[i]);
1541 }
1542 }
1543
1544 if (sh->num_combined_uniform_components >
1545 max_combined_uniform_components[i]) {
1546 if (ctx->Const.GLSLSkipStrictMaxUniformLimitCheck) {
1547 linker_warning(prog, "Too many %s shader uniform components, "
1548 "but the driver will try to optimize them out; "
1549 "this is non-portable out-of-spec behavior\n",
1550 shader_names[i]);
1551 } else {
1552 linker_error(prog, "Too many %s shader uniform components",
1553 shader_names[i]);
1554 }
1555 }
1556 }
1557
1558 unsigned blocks[MESA_SHADER_TYPES] = {0};
1559 unsigned total_uniform_blocks = 0;
1560
1561 for (unsigned i = 0; i < prog->NumUniformBlocks; i++) {
1562 for (unsigned j = 0; j < MESA_SHADER_TYPES; j++) {
1563 if (prog->UniformBlockStageIndex[j][i] != -1) {
1564 blocks[j]++;
1565 total_uniform_blocks++;
1566 }
1567 }
1568
1569 if (total_uniform_blocks > ctx->Const.MaxCombinedUniformBlocks) {
1570 linker_error(prog, "Too many combined uniform blocks (%d/%d)",
1571 prog->NumUniformBlocks,
1572 ctx->Const.MaxCombinedUniformBlocks);
1573 } else {
1574 for (unsigned i = 0; i < MESA_SHADER_TYPES; i++) {
1575 if (blocks[i] > max_uniform_blocks[i]) {
1576 linker_error(prog, "Too many %s uniform blocks (%d/%d)",
1577 shader_names[i],
1578 blocks[i],
1579 max_uniform_blocks[i]);
1580 break;
1581 }
1582 }
1583 }
1584 }
1585 }
1586
1587 void
1588 link_shaders(struct gl_context *ctx, struct gl_shader_program *prog)
1589 {
1590 tfeedback_decl *tfeedback_decls = NULL;
1591 unsigned num_tfeedback_decls = prog->TransformFeedback.NumVarying;
1592
1593 void *mem_ctx = ralloc_context(NULL); // temporary linker context
1594
1595 prog->LinkStatus = true; /* All error paths will set this to false */
1596 prog->Validated = false;
1597 prog->_Used = false;
1598
1599 ralloc_free(prog->InfoLog);
1600 prog->InfoLog = ralloc_strdup(NULL, "");
1601
1602 ralloc_free(prog->UniformBlocks);
1603 prog->UniformBlocks = NULL;
1604 prog->NumUniformBlocks = 0;
1605 for (int i = 0; i < MESA_SHADER_TYPES; i++) {
1606 ralloc_free(prog->UniformBlockStageIndex[i]);
1607 prog->UniformBlockStageIndex[i] = NULL;
1608 }
1609
1610 /* Separate the shaders into groups based on their type.
1611 */
1612 struct gl_shader **vert_shader_list;
1613 unsigned num_vert_shaders = 0;
1614 struct gl_shader **frag_shader_list;
1615 unsigned num_frag_shaders = 0;
1616
1617 vert_shader_list = (struct gl_shader **)
1618 calloc(prog->NumShaders, sizeof(struct gl_shader *));
1619 frag_shader_list = (struct gl_shader **)
1620 calloc(prog->NumShaders, sizeof(struct gl_shader *));
1621
1622 unsigned min_version = UINT_MAX;
1623 unsigned max_version = 0;
1624 const bool is_es_prog =
1625 (prog->NumShaders > 0 && prog->Shaders[0]->IsES) ? true : false;
1626 for (unsigned i = 0; i < prog->NumShaders; i++) {
1627 min_version = MIN2(min_version, prog->Shaders[i]->Version);
1628 max_version = MAX2(max_version, prog->Shaders[i]->Version);
1629
1630 if (prog->Shaders[i]->IsES != is_es_prog) {
1631 linker_error(prog, "all shaders must use same shading "
1632 "language version\n");
1633 goto done;
1634 }
1635
1636 switch (prog->Shaders[i]->Type) {
1637 case GL_VERTEX_SHADER:
1638 vert_shader_list[num_vert_shaders] = prog->Shaders[i];
1639 num_vert_shaders++;
1640 break;
1641 case GL_FRAGMENT_SHADER:
1642 frag_shader_list[num_frag_shaders] = prog->Shaders[i];
1643 num_frag_shaders++;
1644 break;
1645 case GL_GEOMETRY_SHADER:
1646 /* FINISHME: Support geometry shaders. */
1647 assert(prog->Shaders[i]->Type != GL_GEOMETRY_SHADER);
1648 break;
1649 }
1650 }
1651
1652 /* Previous to GLSL version 1.30, different compilation units could mix and
1653 * match shading language versions. With GLSL 1.30 and later, the versions
1654 * of all shaders must match.
1655 *
1656 * GLSL ES has never allowed mixing of shading language versions.
1657 */
1658 if ((is_es_prog || max_version >= 130)
1659 && min_version != max_version) {
1660 linker_error(prog, "all shaders must use same shading "
1661 "language version\n");
1662 goto done;
1663 }
1664
1665 prog->Version = max_version;
1666 prog->IsES = is_es_prog;
1667
1668 for (unsigned int i = 0; i < MESA_SHADER_TYPES; i++) {
1669 if (prog->_LinkedShaders[i] != NULL)
1670 ctx->Driver.DeleteShader(ctx, prog->_LinkedShaders[i]);
1671
1672 prog->_LinkedShaders[i] = NULL;
1673 }
1674
1675 /* Link all shaders for a particular stage and validate the result.
1676 */
1677 if (num_vert_shaders > 0) {
1678 gl_shader *const sh =
1679 link_intrastage_shaders(mem_ctx, ctx, prog, vert_shader_list,
1680 num_vert_shaders);
1681
1682 if (!prog->LinkStatus)
1683 goto done;
1684
1685 validate_vertex_shader_executable(prog, sh);
1686 if (!prog->LinkStatus)
1687 goto done;
1688
1689 _mesa_reference_shader(ctx, &prog->_LinkedShaders[MESA_SHADER_VERTEX],
1690 sh);
1691 }
1692
1693 if (num_frag_shaders > 0) {
1694 gl_shader *const sh =
1695 link_intrastage_shaders(mem_ctx, ctx, prog, frag_shader_list,
1696 num_frag_shaders);
1697
1698 if (!prog->LinkStatus)
1699 goto done;
1700
1701 validate_fragment_shader_executable(prog, sh);
1702 if (!prog->LinkStatus)
1703 goto done;
1704
1705 _mesa_reference_shader(ctx, &prog->_LinkedShaders[MESA_SHADER_FRAGMENT],
1706 sh);
1707 }
1708
1709 /* Here begins the inter-stage linking phase. Some initial validation is
1710 * performed, then locations are assigned for uniforms, attributes, and
1711 * varyings.
1712 */
1713 cross_validate_uniforms(prog);
1714 if (!prog->LinkStatus)
1715 goto done;
1716
1717 unsigned prev;
1718
1719 for (prev = 0; prev < MESA_SHADER_TYPES; prev++) {
1720 if (prog->_LinkedShaders[prev] != NULL)
1721 break;
1722 }
1723
1724 /* Validate the inputs of each stage with the output of the preceding
1725 * stage.
1726 */
1727 for (unsigned i = prev + 1; i < MESA_SHADER_TYPES; i++) {
1728 if (prog->_LinkedShaders[i] == NULL)
1729 continue;
1730
1731 validate_interstage_interface_blocks(prog, prog->_LinkedShaders[prev],
1732 prog->_LinkedShaders[i]);
1733 if (!prog->LinkStatus)
1734 goto done;
1735
1736 cross_validate_outputs_to_inputs(prog,
1737 prog->_LinkedShaders[prev],
1738 prog->_LinkedShaders[i]);
1739 if (!prog->LinkStatus)
1740 goto done;
1741
1742 prev = i;
1743 }
1744
1745
1746 for (unsigned int i = 0; i < MESA_SHADER_TYPES; i++) {
1747 if (prog->_LinkedShaders[i] != NULL)
1748 lower_named_interface_blocks(mem_ctx, prog->_LinkedShaders[i]);
1749 }
1750
1751 /* Implement the GLSL 1.30+ rule for discard vs infinite loops Do
1752 * it before optimization because we want most of the checks to get
1753 * dropped thanks to constant propagation.
1754 *
1755 * This rule also applies to GLSL ES 3.00.
1756 */
1757 if (max_version >= (is_es_prog ? 300 : 130)) {
1758 struct gl_shader *sh = prog->_LinkedShaders[MESA_SHADER_FRAGMENT];
1759 if (sh) {
1760 lower_discard_flow(sh->ir);
1761 }
1762 }
1763
1764 if (!interstage_cross_validate_uniform_blocks(prog))
1765 goto done;
1766
1767 /* Do common optimization before assigning storage for attributes,
1768 * uniforms, and varyings. Later optimization could possibly make
1769 * some of that unused.
1770 */
1771 for (unsigned i = 0; i < MESA_SHADER_TYPES; i++) {
1772 if (prog->_LinkedShaders[i] == NULL)
1773 continue;
1774
1775 detect_recursion_linked(prog, prog->_LinkedShaders[i]->ir);
1776 if (!prog->LinkStatus)
1777 goto done;
1778
1779 if (ctx->ShaderCompilerOptions[i].LowerClipDistance) {
1780 lower_clip_distance(prog->_LinkedShaders[i]);
1781 }
1782
1783 unsigned max_unroll = ctx->ShaderCompilerOptions[i].MaxUnrollIterations;
1784
1785 while (do_common_optimization(prog->_LinkedShaders[i]->ir, true, false, max_unroll, &ctx->ShaderCompilerOptions[i]))
1786 ;
1787 }
1788
1789 /* Mark all generic shader inputs and outputs as unpaired. */
1790 if (prog->_LinkedShaders[MESA_SHADER_VERTEX] != NULL) {
1791 link_invalidate_variable_locations(
1792 prog->_LinkedShaders[MESA_SHADER_VERTEX],
1793 VERT_ATTRIB_GENERIC0, VARYING_SLOT_VAR0);
1794 }
1795 /* FINISHME: Geometry shaders not implemented yet */
1796 if (prog->_LinkedShaders[MESA_SHADER_FRAGMENT] != NULL) {
1797 link_invalidate_variable_locations(
1798 prog->_LinkedShaders[MESA_SHADER_FRAGMENT],
1799 VARYING_SLOT_VAR0, FRAG_RESULT_DATA0);
1800 }
1801
1802 /* FINISHME: The value of the max_attribute_index parameter is
1803 * FINISHME: implementation dependent based on the value of
1804 * FINISHME: GL_MAX_VERTEX_ATTRIBS. GL_MAX_VERTEX_ATTRIBS must be
1805 * FINISHME: at least 16, so hardcode 16 for now.
1806 */
1807 if (!assign_attribute_or_color_locations(prog, MESA_SHADER_VERTEX, 16)) {
1808 goto done;
1809 }
1810
1811 if (!assign_attribute_or_color_locations(prog, MESA_SHADER_FRAGMENT, MAX2(ctx->Const.MaxDrawBuffers, ctx->Const.MaxDualSourceDrawBuffers))) {
1812 goto done;
1813 }
1814
1815 unsigned first;
1816 for (first = 0; first < MESA_SHADER_TYPES; first++) {
1817 if (prog->_LinkedShaders[first] != NULL)
1818 break;
1819 }
1820
1821 if (num_tfeedback_decls != 0) {
1822 /* From GL_EXT_transform_feedback:
1823 * A program will fail to link if:
1824 *
1825 * * the <count> specified by TransformFeedbackVaryingsEXT is
1826 * non-zero, but the program object has no vertex or geometry
1827 * shader;
1828 */
1829 if (first >= MESA_SHADER_FRAGMENT) {
1830 linker_error(prog, "Transform feedback varyings specified, but "
1831 "no vertex or geometry shader is present.");
1832 goto done;
1833 }
1834
1835 tfeedback_decls = ralloc_array(mem_ctx, tfeedback_decl,
1836 prog->TransformFeedback.NumVarying);
1837 if (!parse_tfeedback_decls(ctx, prog, mem_ctx, num_tfeedback_decls,
1838 prog->TransformFeedback.VaryingNames,
1839 tfeedback_decls))
1840 goto done;
1841 }
1842
1843 /* Linking the stages in the opposite order (from fragment to vertex)
1844 * ensures that inter-shader outputs written to in an earlier stage are
1845 * eliminated if they are (transitively) not used in a later stage.
1846 */
1847 int last, next;
1848 for (last = MESA_SHADER_TYPES-1; last >= 0; last--) {
1849 if (prog->_LinkedShaders[last] != NULL)
1850 break;
1851 }
1852
1853 if (last >= 0 && last < MESA_SHADER_FRAGMENT) {
1854 gl_shader *const sh = prog->_LinkedShaders[last];
1855
1856 if (num_tfeedback_decls != 0) {
1857 /* There was no fragment shader, but we still have to assign varying
1858 * locations for use by transform feedback.
1859 */
1860 if (!assign_varying_locations(ctx, mem_ctx, prog,
1861 sh, NULL,
1862 num_tfeedback_decls, tfeedback_decls))
1863 goto done;
1864 }
1865
1866 do_dead_builtin_varyings(ctx, sh->ir, NULL,
1867 num_tfeedback_decls, tfeedback_decls);
1868
1869 demote_shader_inputs_and_outputs(sh, ir_var_shader_out);
1870
1871 /* Eliminate code that is now dead due to unused outputs being demoted.
1872 */
1873 while (do_dead_code(sh->ir, false))
1874 ;
1875 }
1876 else if (first == MESA_SHADER_FRAGMENT) {
1877 /* If the program only contains a fragment shader...
1878 */
1879 gl_shader *const sh = prog->_LinkedShaders[first];
1880
1881 do_dead_builtin_varyings(ctx, NULL, sh->ir,
1882 num_tfeedback_decls, tfeedback_decls);
1883
1884 demote_shader_inputs_and_outputs(sh, ir_var_shader_in);
1885
1886 while (do_dead_code(sh->ir, false))
1887 ;
1888 }
1889
1890 next = last;
1891 for (int i = next - 1; i >= 0; i--) {
1892 if (prog->_LinkedShaders[i] == NULL)
1893 continue;
1894
1895 gl_shader *const sh_i = prog->_LinkedShaders[i];
1896 gl_shader *const sh_next = prog->_LinkedShaders[next];
1897
1898 if (!assign_varying_locations(ctx, mem_ctx, prog, sh_i, sh_next,
1899 next == MESA_SHADER_FRAGMENT ? num_tfeedback_decls : 0,
1900 tfeedback_decls))
1901 goto done;
1902
1903 do_dead_builtin_varyings(ctx, sh_i->ir, sh_next->ir,
1904 next == MESA_SHADER_FRAGMENT ? num_tfeedback_decls : 0,
1905 tfeedback_decls);
1906
1907 demote_shader_inputs_and_outputs(sh_i, ir_var_shader_out);
1908 demote_shader_inputs_and_outputs(sh_next, ir_var_shader_in);
1909
1910 /* Eliminate code that is now dead due to unused outputs being demoted.
1911 */
1912 while (do_dead_code(sh_i->ir, false))
1913 ;
1914 while (do_dead_code(sh_next->ir, false))
1915 ;
1916
1917 /* This must be done after all dead varyings are eliminated. */
1918 if (!check_against_varying_limit(ctx, prog, sh_next))
1919 goto done;
1920
1921 next = i;
1922 }
1923
1924 if (!store_tfeedback_info(ctx, prog, num_tfeedback_decls, tfeedback_decls))
1925 goto done;
1926
1927 update_array_sizes(prog);
1928 link_assign_uniform_locations(prog);
1929 store_fragdepth_layout(prog);
1930
1931 check_resources(ctx, prog);
1932 if (!prog->LinkStatus)
1933 goto done;
1934
1935 /* OpenGL ES requires that a vertex shader and a fragment shader both be
1936 * present in a linked program. By checking prog->IsES, we also
1937 * catch the GL_ARB_ES2_compatibility case.
1938 */
1939 if (!prog->InternalSeparateShader &&
1940 (ctx->API == API_OPENGLES2 || prog->IsES)) {
1941 if (prog->_LinkedShaders[MESA_SHADER_VERTEX] == NULL) {
1942 linker_error(prog, "program lacks a vertex shader\n");
1943 } else if (prog->_LinkedShaders[MESA_SHADER_FRAGMENT] == NULL) {
1944 linker_error(prog, "program lacks a fragment shader\n");
1945 }
1946 }
1947
1948 /* FINISHME: Assign fragment shader output locations. */
1949
1950 done:
1951 free(vert_shader_list);
1952 free(frag_shader_list);
1953
1954 for (unsigned i = 0; i < MESA_SHADER_TYPES; i++) {
1955 if (prog->_LinkedShaders[i] == NULL)
1956 continue;
1957
1958 /* Retain any live IR, but trash the rest. */
1959 reparent_ir(prog->_LinkedShaders[i]->ir, prog->_LinkedShaders[i]->ir);
1960
1961 /* The symbol table in the linked shaders may contain references to
1962 * variables that were removed (e.g., unused uniforms). Since it may
1963 * contain junk, there is no possible valid use. Delete it and set the
1964 * pointer to NULL.
1965 */
1966 delete prog->_LinkedShaders[i]->symbols;
1967 prog->_LinkedShaders[i]->symbols = NULL;
1968 }
1969
1970 ralloc_free(mem_ctx);
1971 }