linker.cpp revision 84a1273e7fe1216a4724ab13cd061a12b48893c2
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
75extern "C" {
76#include "main/shaderobj.h"
77}
78
79/**
80 * Visitor that determines whether or not a variable is ever written.
81 */
82class find_assignment_visitor : public ir_hierarchical_visitor {
83public:
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
128private:
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 */
137class find_deref_visitor : public ir_hierarchical_visitor {
138public:
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
160private:
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
166void
167linker_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
180void
181linker_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
193void
194link_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 */
218unsigned
219count_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 * Also sets prog->Vert.UsesClipDistance and prog->Vert.ClipDistanceArraySize
250 * as a side effect.
251 *
252 * \param shader  Vertex shader executable to be verified
253 */
254bool
255validate_vertex_shader_executable(struct gl_shader_program *prog,
256				  struct gl_shader *shader)
257{
258   if (shader == NULL)
259      return true;
260
261   find_assignment_visitor find("gl_Position");
262   find.run(shader->ir);
263   if (!find.variable_found()) {
264      linker_error(prog, "vertex shader does not write to `gl_Position'\n");
265      return false;
266   }
267
268   prog->Vert.ClipDistanceArraySize = 0;
269
270   if (prog->Version >= 130) {
271      /* From section 7.1 (Vertex Shader Special Variables) of the
272       * GLSL 1.30 spec:
273       *
274       *   "It is an error for a shader to statically write both
275       *   gl_ClipVertex and gl_ClipDistance."
276       */
277      find_assignment_visitor clip_vertex("gl_ClipVertex");
278      find_assignment_visitor clip_distance("gl_ClipDistance");
279
280      clip_vertex.run(shader->ir);
281      clip_distance.run(shader->ir);
282      if (clip_vertex.variable_found() && clip_distance.variable_found()) {
283         linker_error(prog, "vertex shader writes to both `gl_ClipVertex' "
284                      "and `gl_ClipDistance'\n");
285         return false;
286      }
287      prog->Vert.UsesClipDistance = clip_distance.variable_found();
288      ir_variable *clip_distance_var =
289         shader->symbols->get_variable("gl_ClipDistance");
290      if (clip_distance_var)
291         prog->Vert.ClipDistanceArraySize = clip_distance_var->type->length;
292   }
293
294   return true;
295}
296
297
298/**
299 * Verify that a fragment shader executable meets all semantic requirements
300 *
301 * \param shader  Fragment shader executable to be verified
302 */
303bool
304validate_fragment_shader_executable(struct gl_shader_program *prog,
305				    struct gl_shader *shader)
306{
307   if (shader == NULL)
308      return true;
309
310   find_assignment_visitor frag_color("gl_FragColor");
311   find_assignment_visitor frag_data("gl_FragData");
312
313   frag_color.run(shader->ir);
314   frag_data.run(shader->ir);
315
316   if (frag_color.variable_found() && frag_data.variable_found()) {
317      linker_error(prog,  "fragment shader writes to both "
318		   "`gl_FragColor' and `gl_FragData'\n");
319      return false;
320   }
321
322   return true;
323}
324
325
326/**
327 * Generate a string describing the mode of a variable
328 */
329static const char *
330mode_string(const ir_variable *var)
331{
332   switch (var->mode) {
333   case ir_var_auto:
334      return (var->read_only) ? "global constant" : "global variable";
335
336   case ir_var_uniform: return "uniform";
337   case ir_var_in:      return "shader input";
338   case ir_var_out:     return "shader output";
339   case ir_var_inout:   return "shader inout";
340
341   case ir_var_const_in:
342   case ir_var_temporary:
343   default:
344      assert(!"Should not get here.");
345      return "invalid variable";
346   }
347}
348
349
350/**
351 * Perform validation of global variables used across multiple shaders
352 */
353bool
354cross_validate_globals(struct gl_shader_program *prog,
355		       struct gl_shader **shader_list,
356		       unsigned num_shaders,
357		       bool uniforms_only)
358{
359   /* Examine all of the uniforms in all of the shaders and cross validate
360    * them.
361    */
362   glsl_symbol_table variables;
363   for (unsigned i = 0; i < num_shaders; i++) {
364      if (shader_list[i] == NULL)
365	 continue;
366
367      foreach_list(node, shader_list[i]->ir) {
368	 ir_variable *const var = ((ir_instruction *) node)->as_variable();
369
370	 if (var == NULL)
371	    continue;
372
373	 if (uniforms_only && (var->mode != ir_var_uniform))
374	    continue;
375
376	 /* Don't cross validate temporaries that are at global scope.  These
377	  * will eventually get pulled into the shaders 'main'.
378	  */
379	 if (var->mode == ir_var_temporary)
380	    continue;
381
382	 /* If a global with this name has already been seen, verify that the
383	  * new instance has the same type.  In addition, if the globals have
384	  * initializers, the values of the initializers must be the same.
385	  */
386	 ir_variable *const existing = variables.get_variable(var->name);
387	 if (existing != NULL) {
388	    if (var->type != existing->type) {
389	       /* Consider the types to be "the same" if both types are arrays
390		* of the same type and one of the arrays is implicitly sized.
391		* In addition, set the type of the linked variable to the
392		* explicitly sized array.
393		*/
394	       if (var->type->is_array()
395		   && existing->type->is_array()
396		   && (var->type->fields.array == existing->type->fields.array)
397		   && ((var->type->length == 0)
398		       || (existing->type->length == 0))) {
399		  if (var->type->length != 0) {
400		     existing->type = var->type;
401		  }
402	       } else {
403		  linker_error(prog, "%s `%s' declared as type "
404			       "`%s' and type `%s'\n",
405			       mode_string(var),
406			       var->name, var->type->name,
407			       existing->type->name);
408		  return false;
409	       }
410	    }
411
412	    if (var->explicit_location) {
413	       if (existing->explicit_location
414		   && (var->location != existing->location)) {
415		     linker_error(prog, "explicit locations for %s "
416				  "`%s' have differing values\n",
417				  mode_string(var), var->name);
418		     return false;
419	       }
420
421	       existing->location = var->location;
422	       existing->explicit_location = true;
423	    }
424
425	    /* Validate layout qualifiers for gl_FragDepth.
426	     *
427	     * From the AMD/ARB_conservative_depth specs:
428	     *
429	     *    "If gl_FragDepth is redeclared in any fragment shader in a
430	     *    program, it must be redeclared in all fragment shaders in
431	     *    that program that have static assignments to
432	     *    gl_FragDepth. All redeclarations of gl_FragDepth in all
433	     *    fragment shaders in a single program must have the same set
434	     *    of qualifiers."
435	     */
436	    if (strcmp(var->name, "gl_FragDepth") == 0) {
437	       bool layout_declared = var->depth_layout != ir_depth_layout_none;
438	       bool layout_differs =
439		  var->depth_layout != existing->depth_layout;
440
441	       if (layout_declared && layout_differs) {
442		  linker_error(prog,
443			       "All redeclarations of gl_FragDepth in all "
444			       "fragment shaders in a single program must have "
445			       "the same set of qualifiers.");
446	       }
447
448	       if (var->used && layout_differs) {
449		  linker_error(prog,
450			       "If gl_FragDepth is redeclared with a layout "
451			       "qualifier in any fragment shader, it must be "
452			       "redeclared with the same layout qualifier in "
453			       "all fragment shaders that have assignments to "
454			       "gl_FragDepth");
455	       }
456	    }
457
458	    /* Page 35 (page 41 of the PDF) of the GLSL 4.20 spec says:
459	     *
460	     *     "If a shared global has multiple initializers, the
461	     *     initializers must all be constant expressions, and they
462	     *     must all have the same value. Otherwise, a link error will
463	     *     result. (A shared global having only one initializer does
464	     *     not require that initializer to be a constant expression.)"
465	     *
466	     * Previous to 4.20 the GLSL spec simply said that initializers
467	     * must have the same value.  In this case of non-constant
468	     * initializers, this was impossible to determine.  As a result,
469	     * no vendor actually implemented that behavior.  The 4.20
470	     * behavior matches the implemented behavior of at least one other
471	     * vendor, so we'll implement that for all GLSL versions.
472	     */
473	    if (var->constant_initializer != NULL) {
474	       if (existing->constant_initializer != NULL) {
475		  if (!var->constant_initializer->has_value(existing->constant_initializer)) {
476		     linker_error(prog, "initializers for %s "
477				  "`%s' have differing values\n",
478				  mode_string(var), var->name);
479		     return false;
480		  }
481	       } else {
482		  /* If the first-seen instance of a particular uniform did not
483		   * have an initializer but a later instance does, copy the
484		   * initializer to the version stored in the symbol table.
485		   */
486		  /* FINISHME: This is wrong.  The constant_value field should
487		   * FINISHME: not be modified!  Imagine a case where a shader
488		   * FINISHME: without an initializer is linked in two different
489		   * FINISHME: programs with shaders that have differing
490		   * FINISHME: initializers.  Linking with the first will
491		   * FINISHME: modify the shader, and linking with the second
492		   * FINISHME: will fail.
493		   */
494		  existing->constant_initializer =
495		     var->constant_initializer->clone(ralloc_parent(existing),
496						      NULL);
497	       }
498	    }
499
500	    if (var->has_initializer) {
501	       if (existing->has_initializer
502		   && (var->constant_initializer == NULL
503		       || existing->constant_initializer == NULL)) {
504		  linker_error(prog,
505			       "shared global variable `%s' has multiple "
506			       "non-constant initializers.\n",
507			       var->name);
508		  return false;
509	       }
510
511	       /* Some instance had an initializer, so keep track of that.  In
512		* this location, all sorts of initializers (constant or
513		* otherwise) will propagate the existence to the variable
514		* stored in the symbol table.
515		*/
516	       existing->has_initializer = true;
517	    }
518
519	    if (existing->invariant != var->invariant) {
520	       linker_error(prog, "declarations for %s `%s' have "
521			    "mismatching invariant qualifiers\n",
522			    mode_string(var), var->name);
523	       return false;
524	    }
525            if (existing->centroid != var->centroid) {
526               linker_error(prog, "declarations for %s `%s' have "
527			    "mismatching centroid qualifiers\n",
528			    mode_string(var), var->name);
529               return false;
530            }
531	 } else
532	    variables.add_variable(var);
533      }
534   }
535
536   return true;
537}
538
539
540/**
541 * Perform validation of uniforms used across multiple shader stages
542 */
543bool
544cross_validate_uniforms(struct gl_shader_program *prog)
545{
546   return cross_validate_globals(prog, prog->_LinkedShaders,
547				 MESA_SHADER_TYPES, true);
548}
549
550
551/**
552 * Validate that outputs from one stage match inputs of another
553 */
554bool
555cross_validate_outputs_to_inputs(struct gl_shader_program *prog,
556				 gl_shader *producer, gl_shader *consumer)
557{
558   glsl_symbol_table parameters;
559   /* FINISHME: Figure these out dynamically. */
560   const char *const producer_stage = "vertex";
561   const char *const consumer_stage = "fragment";
562
563   /* Find all shader outputs in the "producer" stage.
564    */
565   foreach_list(node, producer->ir) {
566      ir_variable *const var = ((ir_instruction *) node)->as_variable();
567
568      /* FINISHME: For geometry shaders, this should also look for inout
569       * FINISHME: variables.
570       */
571      if ((var == NULL) || (var->mode != ir_var_out))
572	 continue;
573
574      parameters.add_variable(var);
575   }
576
577
578   /* Find all shader inputs in the "consumer" stage.  Any variables that have
579    * matching outputs already in the symbol table must have the same type and
580    * qualifiers.
581    */
582   foreach_list(node, consumer->ir) {
583      ir_variable *const input = ((ir_instruction *) node)->as_variable();
584
585      /* FINISHME: For geometry shaders, this should also look for inout
586       * FINISHME: variables.
587       */
588      if ((input == NULL) || (input->mode != ir_var_in))
589	 continue;
590
591      ir_variable *const output = parameters.get_variable(input->name);
592      if (output != NULL) {
593	 /* Check that the types match between stages.
594	  */
595	 if (input->type != output->type) {
596	    /* There is a bit of a special case for gl_TexCoord.  This
597	     * built-in is unsized by default.  Applications that variable
598	     * access it must redeclare it with a size.  There is some
599	     * language in the GLSL spec that implies the fragment shader
600	     * and vertex shader do not have to agree on this size.  Other
601	     * driver behave this way, and one or two applications seem to
602	     * rely on it.
603	     *
604	     * Neither declaration needs to be modified here because the array
605	     * sizes are fixed later when update_array_sizes is called.
606	     *
607	     * From page 48 (page 54 of the PDF) of the GLSL 1.10 spec:
608	     *
609	     *     "Unlike user-defined varying variables, the built-in
610	     *     varying variables don't have a strict one-to-one
611	     *     correspondence between the vertex language and the
612	     *     fragment language."
613	     */
614	    if (!output->type->is_array()
615		|| (strncmp("gl_", output->name, 3) != 0)) {
616	       linker_error(prog,
617			    "%s shader output `%s' declared as type `%s', "
618			    "but %s shader input declared as type `%s'\n",
619			    producer_stage, output->name,
620			    output->type->name,
621			    consumer_stage, input->type->name);
622	       return false;
623	    }
624	 }
625
626	 /* Check that all of the qualifiers match between stages.
627	  */
628	 if (input->centroid != output->centroid) {
629	    linker_error(prog,
630			 "%s shader output `%s' %s centroid qualifier, "
631			 "but %s shader input %s centroid qualifier\n",
632			 producer_stage,
633			 output->name,
634			 (output->centroid) ? "has" : "lacks",
635			 consumer_stage,
636			 (input->centroid) ? "has" : "lacks");
637	    return false;
638	 }
639
640	 if (input->invariant != output->invariant) {
641	    linker_error(prog,
642			 "%s shader output `%s' %s invariant qualifier, "
643			 "but %s shader input %s invariant qualifier\n",
644			 producer_stage,
645			 output->name,
646			 (output->invariant) ? "has" : "lacks",
647			 consumer_stage,
648			 (input->invariant) ? "has" : "lacks");
649	    return false;
650	 }
651
652	 if (input->interpolation != output->interpolation) {
653	    linker_error(prog,
654			 "%s shader output `%s' specifies %s "
655			 "interpolation qualifier, "
656			 "but %s shader input specifies %s "
657			 "interpolation qualifier\n",
658			 producer_stage,
659			 output->name,
660			 output->interpolation_string(),
661			 consumer_stage,
662			 input->interpolation_string());
663	    return false;
664	 }
665      }
666   }
667
668   return true;
669}
670
671
672/**
673 * Populates a shaders symbol table with all global declarations
674 */
675static void
676populate_symbol_table(gl_shader *sh)
677{
678   sh->symbols = new(sh) glsl_symbol_table;
679
680   foreach_list(node, sh->ir) {
681      ir_instruction *const inst = (ir_instruction *) node;
682      ir_variable *var;
683      ir_function *func;
684
685      if ((func = inst->as_function()) != NULL) {
686	 sh->symbols->add_function(func);
687      } else if ((var = inst->as_variable()) != NULL) {
688	 sh->symbols->add_variable(var);
689      }
690   }
691}
692
693
694/**
695 * Remap variables referenced in an instruction tree
696 *
697 * This is used when instruction trees are cloned from one shader and placed in
698 * another.  These trees will contain references to \c ir_variable nodes that
699 * do not exist in the target shader.  This function finds these \c ir_variable
700 * references and replaces the references with matching variables in the target
701 * shader.
702 *
703 * If there is no matching variable in the target shader, a clone of the
704 * \c ir_variable is made and added to the target shader.  The new variable is
705 * added to \b both the instruction stream and the symbol table.
706 *
707 * \param inst         IR tree that is to be processed.
708 * \param symbols      Symbol table containing global scope symbols in the
709 *                     linked shader.
710 * \param instructions Instruction stream where new variable declarations
711 *                     should be added.
712 */
713void
714remap_variables(ir_instruction *inst, struct gl_shader *target,
715		hash_table *temps)
716{
717   class remap_visitor : public ir_hierarchical_visitor {
718   public:
719	 remap_visitor(struct gl_shader *target,
720		    hash_table *temps)
721      {
722	 this->target = target;
723	 this->symbols = target->symbols;
724	 this->instructions = target->ir;
725	 this->temps = temps;
726      }
727
728      virtual ir_visitor_status visit(ir_dereference_variable *ir)
729      {
730	 if (ir->var->mode == ir_var_temporary) {
731	    ir_variable *var = (ir_variable *) hash_table_find(temps, ir->var);
732
733	    assert(var != NULL);
734	    ir->var = var;
735	    return visit_continue;
736	 }
737
738	 ir_variable *const existing =
739	    this->symbols->get_variable(ir->var->name);
740	 if (existing != NULL)
741	    ir->var = existing;
742	 else {
743	    ir_variable *copy = ir->var->clone(this->target, NULL);
744
745	    this->symbols->add_variable(copy);
746	    this->instructions->push_head(copy);
747	    ir->var = copy;
748	 }
749
750	 return visit_continue;
751      }
752
753   private:
754      struct gl_shader *target;
755      glsl_symbol_table *symbols;
756      exec_list *instructions;
757      hash_table *temps;
758   };
759
760   remap_visitor v(target, temps);
761
762   inst->accept(&v);
763}
764
765
766/**
767 * Move non-declarations from one instruction stream to another
768 *
769 * The intended usage pattern of this function is to pass the pointer to the
770 * head sentinel of a list (i.e., a pointer to the list cast to an \c exec_node
771 * pointer) for \c last and \c false for \c make_copies on the first
772 * call.  Successive calls pass the return value of the previous call for
773 * \c last and \c true for \c make_copies.
774 *
775 * \param instructions Source instruction stream
776 * \param last         Instruction after which new instructions should be
777 *                     inserted in the target instruction stream
778 * \param make_copies  Flag selecting whether instructions in \c instructions
779 *                     should be copied (via \c ir_instruction::clone) into the
780 *                     target list or moved.
781 *
782 * \return
783 * The new "last" instruction in the target instruction stream.  This pointer
784 * is suitable for use as the \c last parameter of a later call to this
785 * function.
786 */
787exec_node *
788move_non_declarations(exec_list *instructions, exec_node *last,
789		      bool make_copies, gl_shader *target)
790{
791   hash_table *temps = NULL;
792
793   if (make_copies)
794      temps = hash_table_ctor(0, hash_table_pointer_hash,
795			      hash_table_pointer_compare);
796
797   foreach_list_safe(node, instructions) {
798      ir_instruction *inst = (ir_instruction *) node;
799
800      if (inst->as_function())
801	 continue;
802
803      ir_variable *var = inst->as_variable();
804      if ((var != NULL) && (var->mode != ir_var_temporary))
805	 continue;
806
807      assert(inst->as_assignment()
808	     || ((var != NULL) && (var->mode == ir_var_temporary)));
809
810      if (make_copies) {
811	 inst = inst->clone(target, NULL);
812
813	 if (var != NULL)
814	    hash_table_insert(temps, inst, var);
815	 else
816	    remap_variables(inst, target, temps);
817      } else {
818	 inst->remove();
819      }
820
821      last->insert_after(inst);
822      last = inst;
823   }
824
825   if (make_copies)
826      hash_table_dtor(temps);
827
828   return last;
829}
830
831/**
832 * Get the function signature for main from a shader
833 */
834static ir_function_signature *
835get_main_function_signature(gl_shader *sh)
836{
837   ir_function *const f = sh->symbols->get_function("main");
838   if (f != NULL) {
839      exec_list void_parameters;
840
841      /* Look for the 'void main()' signature and ensure that it's defined.
842       * This keeps the linker from accidentally pick a shader that just
843       * contains a prototype for main.
844       *
845       * We don't have to check for multiple definitions of main (in multiple
846       * shaders) because that would have already been caught above.
847       */
848      ir_function_signature *sig = f->matching_signature(&void_parameters);
849      if ((sig != NULL) && sig->is_defined) {
850	 return sig;
851      }
852   }
853
854   return NULL;
855}
856
857
858/**
859 * This class is only used in link_intrastage_shaders() below but declaring
860 * it inside that function leads to compiler warnings with some versions of
861 * gcc.
862 */
863class array_sizing_visitor : public ir_hierarchical_visitor {
864public:
865   virtual ir_visitor_status visit(ir_variable *var)
866   {
867      if (var->type->is_array() && (var->type->length == 0)) {
868         const glsl_type *type =
869            glsl_type::get_array_instance(var->type->fields.array,
870                                          var->max_array_access + 1);
871         assert(type != NULL);
872         var->type = type;
873      }
874      return visit_continue;
875   }
876};
877
878
879/**
880 * Combine a group of shaders for a single stage to generate a linked shader
881 *
882 * \note
883 * If this function is supplied a single shader, it is cloned, and the new
884 * shader is returned.
885 */
886static struct gl_shader *
887link_intrastage_shaders(void *mem_ctx,
888			struct gl_context *ctx,
889			struct gl_shader_program *prog,
890			struct gl_shader **shader_list,
891			unsigned num_shaders)
892{
893   /* Check that global variables defined in multiple shaders are consistent.
894    */
895   if (!cross_validate_globals(prog, shader_list, num_shaders, false))
896      return NULL;
897
898   /* Check that there is only a single definition of each function signature
899    * across all shaders.
900    */
901   for (unsigned i = 0; i < (num_shaders - 1); i++) {
902      foreach_list(node, shader_list[i]->ir) {
903	 ir_function *const f = ((ir_instruction *) node)->as_function();
904
905	 if (f == NULL)
906	    continue;
907
908	 for (unsigned j = i + 1; j < num_shaders; j++) {
909	    ir_function *const other =
910	       shader_list[j]->symbols->get_function(f->name);
911
912	    /* If the other shader has no function (and therefore no function
913	     * signatures) with the same name, skip to the next shader.
914	     */
915	    if (other == NULL)
916	       continue;
917
918	    foreach_iter (exec_list_iterator, iter, *f) {
919	       ir_function_signature *sig =
920		  (ir_function_signature *) iter.get();
921
922	       if (!sig->is_defined || sig->is_builtin)
923		  continue;
924
925	       ir_function_signature *other_sig =
926		  other->exact_matching_signature(& sig->parameters);
927
928	       if ((other_sig != NULL) && other_sig->is_defined
929		   && !other_sig->is_builtin) {
930		  linker_error(prog, "function `%s' is multiply defined",
931			       f->name);
932		  return NULL;
933	       }
934	    }
935	 }
936      }
937   }
938
939   /* Find the shader that defines main, and make a clone of it.
940    *
941    * Starting with the clone, search for undefined references.  If one is
942    * found, find the shader that defines it.  Clone the reference and add
943    * it to the shader.  Repeat until there are no undefined references or
944    * until a reference cannot be resolved.
945    */
946   gl_shader *main = NULL;
947   for (unsigned i = 0; i < num_shaders; i++) {
948      if (get_main_function_signature(shader_list[i]) != NULL) {
949	 main = shader_list[i];
950	 break;
951      }
952   }
953
954   if (main == NULL) {
955      linker_error(prog, "%s shader lacks `main'\n",
956		   (shader_list[0]->Type == GL_VERTEX_SHADER)
957		   ? "vertex" : "fragment");
958      return NULL;
959   }
960
961   gl_shader *linked = ctx->Driver.NewShader(NULL, 0, main->Type);
962   linked->ir = new(linked) exec_list;
963   clone_ir_list(mem_ctx, linked->ir, main->ir);
964
965   populate_symbol_table(linked);
966
967   /* The a pointer to the main function in the final linked shader (i.e., the
968    * copy of the original shader that contained the main function).
969    */
970   ir_function_signature *const main_sig = get_main_function_signature(linked);
971
972   /* Move any instructions other than variable declarations or function
973    * declarations into main.
974    */
975   exec_node *insertion_point =
976      move_non_declarations(linked->ir, (exec_node *) &main_sig->body, false,
977			    linked);
978
979   for (unsigned i = 0; i < num_shaders; i++) {
980      if (shader_list[i] == main)
981	 continue;
982
983      insertion_point = move_non_declarations(shader_list[i]->ir,
984					      insertion_point, true, linked);
985   }
986
987   /* Resolve initializers for global variables in the linked shader.
988    */
989   unsigned num_linking_shaders = num_shaders;
990   for (unsigned i = 0; i < num_shaders; i++)
991      num_linking_shaders += shader_list[i]->num_builtins_to_link;
992
993   gl_shader **linking_shaders =
994      (gl_shader **) calloc(num_linking_shaders, sizeof(gl_shader *));
995
996   memcpy(linking_shaders, shader_list,
997	  sizeof(linking_shaders[0]) * num_shaders);
998
999   unsigned idx = num_shaders;
1000   for (unsigned i = 0; i < num_shaders; i++) {
1001      memcpy(&linking_shaders[idx], shader_list[i]->builtins_to_link,
1002	     sizeof(linking_shaders[0]) * shader_list[i]->num_builtins_to_link);
1003      idx += shader_list[i]->num_builtins_to_link;
1004   }
1005
1006   assert(idx == num_linking_shaders);
1007
1008   if (!link_function_calls(prog, linked, linking_shaders,
1009			    num_linking_shaders)) {
1010      ctx->Driver.DeleteShader(ctx, linked);
1011      linked = NULL;
1012   }
1013
1014   free(linking_shaders);
1015
1016#ifdef DEBUG
1017   /* At this point linked should contain all of the linked IR, so
1018    * validate it to make sure nothing went wrong.
1019    */
1020   if (linked)
1021      validate_ir_tree(linked->ir);
1022#endif
1023
1024   /* Make a pass over all variable declarations to ensure that arrays with
1025    * unspecified sizes have a size specified.  The size is inferred from the
1026    * max_array_access field.
1027    */
1028   if (linked != NULL) {
1029      array_sizing_visitor v;
1030
1031      v.run(linked->ir);
1032   }
1033
1034   return linked;
1035}
1036
1037/**
1038 * Update the sizes of linked shader uniform arrays to the maximum
1039 * array index used.
1040 *
1041 * From page 81 (page 95 of the PDF) of the OpenGL 2.1 spec:
1042 *
1043 *     If one or more elements of an array are active,
1044 *     GetActiveUniform will return the name of the array in name,
1045 *     subject to the restrictions listed above. The type of the array
1046 *     is returned in type. The size parameter contains the highest
1047 *     array element index used, plus one. The compiler or linker
1048 *     determines the highest index used.  There will be only one
1049 *     active uniform reported by the GL per uniform array.
1050
1051 */
1052static void
1053update_array_sizes(struct gl_shader_program *prog)
1054{
1055   for (unsigned i = 0; i < MESA_SHADER_TYPES; i++) {
1056	 if (prog->_LinkedShaders[i] == NULL)
1057	    continue;
1058
1059      foreach_list(node, prog->_LinkedShaders[i]->ir) {
1060	 ir_variable *const var = ((ir_instruction *) node)->as_variable();
1061
1062	 if ((var == NULL) || (var->mode != ir_var_uniform &&
1063			       var->mode != ir_var_in &&
1064			       var->mode != ir_var_out) ||
1065	     !var->type->is_array())
1066	    continue;
1067
1068	 unsigned int size = var->max_array_access;
1069	 for (unsigned j = 0; j < MESA_SHADER_TYPES; j++) {
1070	       if (prog->_LinkedShaders[j] == NULL)
1071		  continue;
1072
1073	    foreach_list(node2, prog->_LinkedShaders[j]->ir) {
1074	       ir_variable *other_var = ((ir_instruction *) node2)->as_variable();
1075	       if (!other_var)
1076		  continue;
1077
1078	       if (strcmp(var->name, other_var->name) == 0 &&
1079		   other_var->max_array_access > size) {
1080		  size = other_var->max_array_access;
1081	       }
1082	    }
1083	 }
1084
1085	 if (size + 1 != var->type->fields.array->length) {
1086	    /* If this is a built-in uniform (i.e., it's backed by some
1087	     * fixed-function state), adjust the number of state slots to
1088	     * match the new array size.  The number of slots per array entry
1089	     * is not known.  It seems safe to assume that the total number of
1090	     * slots is an integer multiple of the number of array elements.
1091	     * Determine the number of slots per array element by dividing by
1092	     * the old (total) size.
1093	     */
1094	    if (var->num_state_slots > 0) {
1095	       var->num_state_slots = (size + 1)
1096		  * (var->num_state_slots / var->type->length);
1097	    }
1098
1099	    var->type = glsl_type::get_array_instance(var->type->fields.array,
1100						      size + 1);
1101	    /* FINISHME: We should update the types of array
1102	     * dereferences of this variable now.
1103	     */
1104	 }
1105      }
1106   }
1107}
1108
1109/**
1110 * Find a contiguous set of available bits in a bitmask.
1111 *
1112 * \param used_mask     Bits representing used (1) and unused (0) locations
1113 * \param needed_count  Number of contiguous bits needed.
1114 *
1115 * \return
1116 * Base location of the available bits on success or -1 on failure.
1117 */
1118int
1119find_available_slots(unsigned used_mask, unsigned needed_count)
1120{
1121   unsigned needed_mask = (1 << needed_count) - 1;
1122   const int max_bit_to_test = (8 * sizeof(used_mask)) - needed_count;
1123
1124   /* The comparison to 32 is redundant, but without it GCC emits "warning:
1125    * cannot optimize possibly infinite loops" for the loop below.
1126    */
1127   if ((needed_count == 0) || (max_bit_to_test < 0) || (max_bit_to_test > 32))
1128      return -1;
1129
1130   for (int i = 0; i <= max_bit_to_test; i++) {
1131      if ((needed_mask & ~used_mask) == needed_mask)
1132	 return i;
1133
1134      needed_mask <<= 1;
1135   }
1136
1137   return -1;
1138}
1139
1140
1141/**
1142 * Assign locations for either VS inputs for FS outputs
1143 *
1144 * \param prog          Shader program whose variables need locations assigned
1145 * \param target_index  Selector for the program target to receive location
1146 *                      assignmnets.  Must be either \c MESA_SHADER_VERTEX or
1147 *                      \c MESA_SHADER_FRAGMENT.
1148 * \param max_index     Maximum number of generic locations.  This corresponds
1149 *                      to either the maximum number of draw buffers or the
1150 *                      maximum number of generic attributes.
1151 *
1152 * \return
1153 * If locations are successfully assigned, true is returned.  Otherwise an
1154 * error is emitted to the shader link log and false is returned.
1155 */
1156bool
1157assign_attribute_or_color_locations(gl_shader_program *prog,
1158				    unsigned target_index,
1159				    unsigned max_index)
1160{
1161   /* Mark invalid locations as being used.
1162    */
1163   unsigned used_locations = (max_index >= 32)
1164      ? ~0 : ~((1 << max_index) - 1);
1165
1166   assert((target_index == MESA_SHADER_VERTEX)
1167	  || (target_index == MESA_SHADER_FRAGMENT));
1168
1169   gl_shader *const sh = prog->_LinkedShaders[target_index];
1170   if (sh == NULL)
1171      return true;
1172
1173   /* Operate in a total of four passes.
1174    *
1175    * 1. Invalidate the location assignments for all vertex shader inputs.
1176    *
1177    * 2. Assign locations for inputs that have user-defined (via
1178    *    glBindVertexAttribLocation) locations and outputs that have
1179    *    user-defined locations (via glBindFragDataLocation).
1180    *
1181    * 3. Sort the attributes without assigned locations by number of slots
1182    *    required in decreasing order.  Fragmentation caused by attribute
1183    *    locations assigned by the application may prevent large attributes
1184    *    from having enough contiguous space.
1185    *
1186    * 4. Assign locations to any inputs without assigned locations.
1187    */
1188
1189   const int generic_base = (target_index == MESA_SHADER_VERTEX)
1190      ? (int) VERT_ATTRIB_GENERIC0 : (int) FRAG_RESULT_DATA0;
1191
1192   const enum ir_variable_mode direction =
1193      (target_index == MESA_SHADER_VERTEX) ? ir_var_in : ir_var_out;
1194
1195
1196   link_invalidate_variable_locations(sh, direction, generic_base);
1197
1198   /* Temporary storage for the set of attributes that need locations assigned.
1199    */
1200   struct temp_attr {
1201      unsigned slots;
1202      ir_variable *var;
1203
1204      /* Used below in the call to qsort. */
1205      static int compare(const void *a, const void *b)
1206      {
1207	 const temp_attr *const l = (const temp_attr *) a;
1208	 const temp_attr *const r = (const temp_attr *) b;
1209
1210	 /* Reversed because we want a descending order sort below. */
1211	 return r->slots - l->slots;
1212      }
1213   } to_assign[16];
1214
1215   unsigned num_attr = 0;
1216
1217   foreach_list(node, sh->ir) {
1218      ir_variable *const var = ((ir_instruction *) node)->as_variable();
1219
1220      if ((var == NULL) || (var->mode != (unsigned) direction))
1221	 continue;
1222
1223      if (var->explicit_location) {
1224	 if ((var->location >= (int)(max_index + generic_base))
1225	     || (var->location < 0)) {
1226	    linker_error(prog,
1227			 "invalid explicit location %d specified for `%s'\n",
1228			 (var->location < 0)
1229			 ? var->location : var->location - generic_base,
1230			 var->name);
1231	    return false;
1232	 }
1233      } else if (target_index == MESA_SHADER_VERTEX) {
1234	 unsigned binding;
1235
1236	 if (prog->AttributeBindings->get(binding, var->name)) {
1237	    assert(binding >= VERT_ATTRIB_GENERIC0);
1238	    var->location = binding;
1239	 }
1240      } else if (target_index == MESA_SHADER_FRAGMENT) {
1241	 unsigned binding;
1242
1243	 if (prog->FragDataBindings->get(binding, var->name)) {
1244	    assert(binding >= FRAG_RESULT_DATA0);
1245	    var->location = binding;
1246	 }
1247      }
1248
1249      /* If the variable is not a built-in and has a location statically
1250       * assigned in the shader (presumably via a layout qualifier), make sure
1251       * that it doesn't collide with other assigned locations.  Otherwise,
1252       * add it to the list of variables that need linker-assigned locations.
1253       */
1254      const unsigned slots = count_attribute_slots(var->type);
1255      if (var->location != -1) {
1256	 if (var->location >= generic_base) {
1257	    /* From page 61 of the OpenGL 4.0 spec:
1258	     *
1259	     *     "LinkProgram will fail if the attribute bindings assigned
1260	     *     by BindAttribLocation do not leave not enough space to
1261	     *     assign a location for an active matrix attribute or an
1262	     *     active attribute array, both of which require multiple
1263	     *     contiguous generic attributes."
1264	     *
1265	     * Previous versions of the spec contain similar language but omit
1266	     * the bit about attribute arrays.
1267	     *
1268	     * Page 61 of the OpenGL 4.0 spec also says:
1269	     *
1270	     *     "It is possible for an application to bind more than one
1271	     *     attribute name to the same location. This is referred to as
1272	     *     aliasing. This will only work if only one of the aliased
1273	     *     attributes is active in the executable program, or if no
1274	     *     path through the shader consumes more than one attribute of
1275	     *     a set of attributes aliased to the same location. A link
1276	     *     error can occur if the linker determines that every path
1277	     *     through the shader consumes multiple aliased attributes,
1278	     *     but implementations are not required to generate an error
1279	     *     in this case."
1280	     *
1281	     * These two paragraphs are either somewhat contradictory, or I
1282	     * don't fully understand one or both of them.
1283	     */
1284	    /* FINISHME: The code as currently written does not support
1285	     * FINISHME: attribute location aliasing (see comment above).
1286	     */
1287	    /* Mask representing the contiguous slots that will be used by
1288	     * this attribute.
1289	     */
1290	    const unsigned attr = var->location - generic_base;
1291	    const unsigned use_mask = (1 << slots) - 1;
1292
1293	    /* Generate a link error if the set of bits requested for this
1294	     * attribute overlaps any previously allocated bits.
1295	     */
1296	    if ((~(use_mask << attr) & used_locations) != used_locations) {
1297	       linker_error(prog,
1298			    "insufficient contiguous attribute locations "
1299			    "available for vertex shader input `%s'",
1300			    var->name);
1301	       return false;
1302	    }
1303
1304	    used_locations |= (use_mask << attr);
1305	 }
1306
1307	 continue;
1308      }
1309
1310      to_assign[num_attr].slots = slots;
1311      to_assign[num_attr].var = var;
1312      num_attr++;
1313   }
1314
1315   /* If all of the attributes were assigned locations by the application (or
1316    * are built-in attributes with fixed locations), return early.  This should
1317    * be the common case.
1318    */
1319   if (num_attr == 0)
1320      return true;
1321
1322   qsort(to_assign, num_attr, sizeof(to_assign[0]), temp_attr::compare);
1323
1324   if (target_index == MESA_SHADER_VERTEX) {
1325      /* VERT_ATTRIB_GENERIC0 is a pseudo-alias for VERT_ATTRIB_POS.  It can
1326       * only be explicitly assigned by via glBindAttribLocation.  Mark it as
1327       * reserved to prevent it from being automatically allocated below.
1328       */
1329      find_deref_visitor find("gl_Vertex");
1330      find.run(sh->ir);
1331      if (find.variable_found())
1332	 used_locations |= (1 << 0);
1333   }
1334
1335   for (unsigned i = 0; i < num_attr; i++) {
1336      /* Mask representing the contiguous slots that will be used by this
1337       * attribute.
1338       */
1339      const unsigned use_mask = (1 << to_assign[i].slots) - 1;
1340
1341      int location = find_available_slots(used_locations, to_assign[i].slots);
1342
1343      if (location < 0) {
1344	 const char *const string = (target_index == MESA_SHADER_VERTEX)
1345	    ? "vertex shader input" : "fragment shader output";
1346
1347	 linker_error(prog,
1348		      "insufficient contiguous attribute locations "
1349		      "available for %s `%s'",
1350		      string, to_assign[i].var->name);
1351	 return false;
1352      }
1353
1354      to_assign[i].var->location = generic_base + location;
1355      used_locations |= (use_mask << location);
1356   }
1357
1358   return true;
1359}
1360
1361
1362/**
1363 * Demote shader inputs and outputs that are not used in other stages
1364 */
1365void
1366demote_shader_inputs_and_outputs(gl_shader *sh, enum ir_variable_mode mode)
1367{
1368   foreach_list(node, sh->ir) {
1369      ir_variable *const var = ((ir_instruction *) node)->as_variable();
1370
1371      if ((var == NULL) || (var->mode != int(mode)))
1372	 continue;
1373
1374      /* A shader 'in' or 'out' variable is only really an input or output if
1375       * its value is used by other shader stages.  This will cause the variable
1376       * to have a location assigned.
1377       */
1378      if (var->location == -1) {
1379	 var->mode = ir_var_auto;
1380      }
1381   }
1382}
1383
1384
1385/**
1386 * Data structure tracking information about a transform feedback declaration
1387 * during linking.
1388 */
1389class tfeedback_decl
1390{
1391public:
1392   bool init(struct gl_context *ctx, struct gl_shader_program *prog,
1393             const void *mem_ctx, const char *input);
1394   static bool is_same(const tfeedback_decl &x, const tfeedback_decl &y);
1395   bool assign_location(struct gl_context *ctx, struct gl_shader_program *prog,
1396                        ir_variable *output_var);
1397   bool accumulate_num_outputs(struct gl_shader_program *prog, unsigned *count);
1398   bool store(struct gl_context *ctx, struct gl_shader_program *prog,
1399              struct gl_transform_feedback_info *info, unsigned buffer,
1400              unsigned varying, const unsigned max_outputs) const;
1401
1402
1403   /**
1404    * True if assign_location() has been called for this object.
1405    */
1406   bool is_assigned() const
1407   {
1408      return this->location != -1;
1409   }
1410
1411   /**
1412    * Determine whether this object refers to the variable var.
1413    */
1414   bool matches_var(ir_variable *var) const
1415   {
1416      if (this->is_clip_distance_mesa)
1417         return strcmp(var->name, "gl_ClipDistanceMESA") == 0;
1418      else
1419         return strcmp(var->name, this->var_name) == 0;
1420   }
1421
1422   /**
1423    * The total number of varying components taken up by this variable.  Only
1424    * valid if is_assigned() is true.
1425    */
1426   unsigned num_components() const
1427   {
1428      if (this->is_clip_distance_mesa)
1429         return this->size;
1430      else
1431         return this->vector_elements * this->matrix_columns * this->size;
1432   }
1433
1434private:
1435   /**
1436    * The name that was supplied to glTransformFeedbackVaryings.  Used for
1437    * error reporting and glGetTransformFeedbackVarying().
1438    */
1439   const char *orig_name;
1440
1441   /**
1442    * The name of the variable, parsed from orig_name.
1443    */
1444   const char *var_name;
1445
1446   /**
1447    * True if the declaration in orig_name represents an array.
1448    */
1449   bool is_subscripted;
1450
1451   /**
1452    * If is_subscripted is true, the subscript that was specified in orig_name.
1453    */
1454   unsigned array_subscript;
1455
1456   /**
1457    * True if the variable is gl_ClipDistance and the driver lowers
1458    * gl_ClipDistance to gl_ClipDistanceMESA.
1459    */
1460   bool is_clip_distance_mesa;
1461
1462   /**
1463    * The vertex shader output location that the linker assigned for this
1464    * variable.  -1 if a location hasn't been assigned yet.
1465    */
1466   int location;
1467
1468   /**
1469    * If location != -1, the number of vector elements in this variable, or 1
1470    * if this variable is a scalar.
1471    */
1472   unsigned vector_elements;
1473
1474   /**
1475    * If location != -1, the number of matrix columns in this variable, or 1
1476    * if this variable is not a matrix.
1477    */
1478   unsigned matrix_columns;
1479
1480   /** Type of the varying returned by glGetTransformFeedbackVarying() */
1481   GLenum type;
1482
1483   /**
1484    * If location != -1, the size that should be returned by
1485    * glGetTransformFeedbackVarying().
1486    */
1487   unsigned size;
1488};
1489
1490
1491/**
1492 * Initialize this object based on a string that was passed to
1493 * glTransformFeedbackVaryings.  If there is a parse error, the error is
1494 * reported using linker_error(), and false is returned.
1495 */
1496bool
1497tfeedback_decl::init(struct gl_context *ctx, struct gl_shader_program *prog,
1498                     const void *mem_ctx, const char *input)
1499{
1500   /* We don't have to be pedantic about what is a valid GLSL variable name,
1501    * because any variable with an invalid name can't exist in the IR anyway.
1502    */
1503
1504   this->location = -1;
1505   this->orig_name = input;
1506   this->is_clip_distance_mesa = false;
1507
1508   const char *bracket = strrchr(input, '[');
1509
1510   if (bracket) {
1511      this->var_name = ralloc_strndup(mem_ctx, input, bracket - input);
1512      if (sscanf(bracket, "[%u]", &this->array_subscript) != 1) {
1513         linker_error(prog, "Cannot parse transform feedback varying %s", input);
1514         return false;
1515      }
1516      this->is_subscripted = true;
1517   } else {
1518      this->var_name = ralloc_strdup(mem_ctx, input);
1519      this->is_subscripted = false;
1520   }
1521
1522   /* For drivers that lower gl_ClipDistance to gl_ClipDistanceMESA, this
1523    * class must behave specially to account for the fact that gl_ClipDistance
1524    * is converted from a float[8] to a vec4[2].
1525    */
1526   if (ctx->ShaderCompilerOptions[MESA_SHADER_VERTEX].LowerClipDistance &&
1527       strcmp(this->var_name, "gl_ClipDistance") == 0) {
1528      this->is_clip_distance_mesa = true;
1529   }
1530
1531   return true;
1532}
1533
1534
1535/**
1536 * Determine whether two tfeedback_decl objects refer to the same variable and
1537 * array index (if applicable).
1538 */
1539bool
1540tfeedback_decl::is_same(const tfeedback_decl &x, const tfeedback_decl &y)
1541{
1542   if (strcmp(x.var_name, y.var_name) != 0)
1543      return false;
1544   if (x.is_subscripted != y.is_subscripted)
1545      return false;
1546   if (x.is_subscripted && x.array_subscript != y.array_subscript)
1547      return false;
1548   return true;
1549}
1550
1551
1552/**
1553 * Assign a location for this tfeedback_decl object based on the location
1554 * assignment in output_var.
1555 *
1556 * If an error occurs, the error is reported through linker_error() and false
1557 * is returned.
1558 */
1559bool
1560tfeedback_decl::assign_location(struct gl_context *ctx,
1561                                struct gl_shader_program *prog,
1562                                ir_variable *output_var)
1563{
1564   if (output_var->type->is_array()) {
1565      /* Array variable */
1566      const unsigned matrix_cols =
1567         output_var->type->fields.array->matrix_columns;
1568      unsigned actual_array_size = this->is_clip_distance_mesa ?
1569         prog->Vert.ClipDistanceArraySize : output_var->type->array_size();
1570
1571      if (this->is_subscripted) {
1572         /* Check array bounds. */
1573         if (this->array_subscript >= actual_array_size) {
1574            linker_error(prog, "Transform feedback varying %s has index "
1575                         "%i, but the array size is %u.",
1576                         this->orig_name, this->array_subscript,
1577                         actual_array_size);
1578            return false;
1579         }
1580         if (this->is_clip_distance_mesa) {
1581            this->location =
1582               output_var->location + this->array_subscript / 4;
1583         } else {
1584            this->location =
1585               output_var->location + this->array_subscript * matrix_cols;
1586         }
1587         this->size = 1;
1588      } else {
1589         this->location = output_var->location;
1590         this->size = actual_array_size;
1591      }
1592      this->vector_elements = output_var->type->fields.array->vector_elements;
1593      this->matrix_columns = matrix_cols;
1594      if (this->is_clip_distance_mesa)
1595         this->type = GL_FLOAT;
1596      else
1597         this->type = output_var->type->fields.array->gl_type;
1598   } else {
1599      /* Regular variable (scalar, vector, or matrix) */
1600      if (this->is_subscripted) {
1601         linker_error(prog, "Transform feedback varying %s requested, "
1602                      "but %s is not an array.",
1603                      this->orig_name, this->var_name);
1604         return false;
1605      }
1606      this->location = output_var->location;
1607      this->size = 1;
1608      this->vector_elements = output_var->type->vector_elements;
1609      this->matrix_columns = output_var->type->matrix_columns;
1610      this->type = output_var->type->gl_type;
1611   }
1612
1613   /* From GL_EXT_transform_feedback:
1614    *   A program will fail to link if:
1615    *
1616    *   * the total number of components to capture in any varying
1617    *     variable in <varyings> is greater than the constant
1618    *     MAX_TRANSFORM_FEEDBACK_SEPARATE_COMPONENTS_EXT and the
1619    *     buffer mode is SEPARATE_ATTRIBS_EXT;
1620    */
1621   if (prog->TransformFeedback.BufferMode == GL_SEPARATE_ATTRIBS &&
1622       this->num_components() >
1623       ctx->Const.MaxTransformFeedbackSeparateComponents) {
1624      linker_error(prog, "Transform feedback varying %s exceeds "
1625                   "MAX_TRANSFORM_FEEDBACK_SEPARATE_COMPONENTS.",
1626                   this->orig_name);
1627      return false;
1628   }
1629
1630   return true;
1631}
1632
1633
1634bool
1635tfeedback_decl::accumulate_num_outputs(struct gl_shader_program *prog,
1636                                       unsigned *count)
1637{
1638   if (!this->is_assigned()) {
1639      /* From GL_EXT_transform_feedback:
1640       *   A program will fail to link if:
1641       *
1642       *   * any variable name specified in the <varyings> array is not
1643       *     declared as an output in the geometry shader (if present) or
1644       *     the vertex shader (if no geometry shader is present);
1645       */
1646      linker_error(prog, "Transform feedback varying %s undeclared.",
1647                   this->orig_name);
1648      return false;
1649   }
1650
1651   unsigned translated_size = this->size;
1652   if (this->is_clip_distance_mesa)
1653      translated_size = (translated_size + 3) / 4;
1654
1655   *count += translated_size * this->matrix_columns;
1656
1657   return true;
1658}
1659
1660
1661/**
1662 * Update gl_transform_feedback_info to reflect this tfeedback_decl.
1663 *
1664 * If an error occurs, the error is reported through linker_error() and false
1665 * is returned.
1666 */
1667bool
1668tfeedback_decl::store(struct gl_context *ctx, struct gl_shader_program *prog,
1669                      struct gl_transform_feedback_info *info,
1670                      unsigned buffer,
1671                      unsigned varying, const unsigned max_outputs) const
1672{
1673   /* From GL_EXT_transform_feedback:
1674    *   A program will fail to link if:
1675    *
1676    *     * the total number of components to capture is greater than
1677    *       the constant MAX_TRANSFORM_FEEDBACK_INTERLEAVED_COMPONENTS_EXT
1678    *       and the buffer mode is INTERLEAVED_ATTRIBS_EXT.
1679    */
1680   if (prog->TransformFeedback.BufferMode == GL_INTERLEAVED_ATTRIBS &&
1681       info->BufferStride[buffer] + this->num_components() >
1682       ctx->Const.MaxTransformFeedbackInterleavedComponents) {
1683      linker_error(prog, "The MAX_TRANSFORM_FEEDBACK_INTERLEAVED_COMPONENTS "
1684                   "limit has been exceeded.");
1685      return false;
1686   }
1687
1688   unsigned translated_size = this->size;
1689   if (this->is_clip_distance_mesa)
1690      translated_size = (translated_size + 3) / 4;
1691   unsigned components_so_far = 0;
1692   for (unsigned index = 0; index < translated_size; ++index) {
1693      for (unsigned v = 0; v < this->matrix_columns; ++v) {
1694         unsigned num_components = this->vector_elements;
1695         assert(info->NumOutputs < max_outputs);
1696         info->Outputs[info->NumOutputs].ComponentOffset = 0;
1697         if (this->is_clip_distance_mesa) {
1698            if (this->is_subscripted) {
1699               num_components = 1;
1700               info->Outputs[info->NumOutputs].ComponentOffset =
1701                  this->array_subscript % 4;
1702            } else {
1703               num_components = MIN2(4, this->size - components_so_far);
1704            }
1705         }
1706         info->Outputs[info->NumOutputs].OutputRegister =
1707            this->location + v + index * this->matrix_columns;
1708         info->Outputs[info->NumOutputs].NumComponents = num_components;
1709         info->Outputs[info->NumOutputs].OutputBuffer = buffer;
1710         info->Outputs[info->NumOutputs].DstOffset = info->BufferStride[buffer];
1711         ++info->NumOutputs;
1712         info->BufferStride[buffer] += num_components;
1713         components_so_far += num_components;
1714      }
1715   }
1716   assert(components_so_far == this->num_components());
1717
1718   info->Varyings[varying].Name = ralloc_strdup(prog, this->orig_name);
1719   info->Varyings[varying].Type = this->type;
1720   info->Varyings[varying].Size = this->size;
1721   info->NumVarying++;
1722
1723   return true;
1724}
1725
1726
1727/**
1728 * Parse all the transform feedback declarations that were passed to
1729 * glTransformFeedbackVaryings() and store them in tfeedback_decl objects.
1730 *
1731 * If an error occurs, the error is reported through linker_error() and false
1732 * is returned.
1733 */
1734static bool
1735parse_tfeedback_decls(struct gl_context *ctx, struct gl_shader_program *prog,
1736                      const void *mem_ctx, unsigned num_names,
1737                      char **varying_names, tfeedback_decl *decls)
1738{
1739   for (unsigned i = 0; i < num_names; ++i) {
1740      if (!decls[i].init(ctx, prog, mem_ctx, varying_names[i]))
1741         return false;
1742      /* From GL_EXT_transform_feedback:
1743       *   A program will fail to link if:
1744       *
1745       *   * any two entries in the <varyings> array specify the same varying
1746       *     variable;
1747       *
1748       * We interpret this to mean "any two entries in the <varyings> array
1749       * specify the same varying variable and array index", since transform
1750       * feedback of arrays would be useless otherwise.
1751       */
1752      for (unsigned j = 0; j < i; ++j) {
1753         if (tfeedback_decl::is_same(decls[i], decls[j])) {
1754            linker_error(prog, "Transform feedback varying %s specified "
1755                         "more than once.", varying_names[i]);
1756            return false;
1757         }
1758      }
1759   }
1760   return true;
1761}
1762
1763
1764/**
1765 * Assign a location for a variable that is produced in one pipeline stage
1766 * (the "producer") and consumed in the next stage (the "consumer").
1767 *
1768 * \param input_var is the input variable declaration in the consumer.
1769 *
1770 * \param output_var is the output variable declaration in the producer.
1771 *
1772 * \param input_index is the counter that keeps track of assigned input
1773 *        locations in the consumer.
1774 *
1775 * \param output_index is the counter that keeps track of assigned output
1776 *        locations in the producer.
1777 *
1778 * It is permissible for \c input_var to be NULL (this happens if a variable
1779 * is output by the producer and consumed by transform feedback, but not
1780 * consumed by the consumer).
1781 *
1782 * If the variable has already been assigned a location, this function has no
1783 * effect.
1784 */
1785void
1786assign_varying_location(ir_variable *input_var, ir_variable *output_var,
1787                        unsigned *input_index, unsigned *output_index)
1788{
1789   if (output_var->location != -1) {
1790      /* Location already assigned. */
1791      return;
1792   }
1793
1794   if (input_var) {
1795      assert(input_var->location == -1);
1796      input_var->location = *input_index;
1797   }
1798
1799   output_var->location = *output_index;
1800
1801   /* FINISHME: Support for "varying" records in GLSL 1.50. */
1802   assert(!output_var->type->is_record());
1803
1804   if (output_var->type->is_array()) {
1805      const unsigned slots = output_var->type->length
1806         * output_var->type->fields.array->matrix_columns;
1807
1808      *output_index += slots;
1809      *input_index += slots;
1810   } else {
1811      const unsigned slots = output_var->type->matrix_columns;
1812
1813      *output_index += slots;
1814      *input_index += slots;
1815   }
1816}
1817
1818
1819/**
1820 * Assign locations for all variables that are produced in one pipeline stage
1821 * (the "producer") and consumed in the next stage (the "consumer").
1822 *
1823 * Variables produced by the producer may also be consumed by transform
1824 * feedback.
1825 *
1826 * \param num_tfeedback_decls is the number of declarations indicating
1827 *        variables that may be consumed by transform feedback.
1828 *
1829 * \param tfeedback_decls is a pointer to an array of tfeedback_decl objects
1830 *        representing the result of parsing the strings passed to
1831 *        glTransformFeedbackVaryings().  assign_location() will be called for
1832 *        each of these objects that matches one of the outputs of the
1833 *        producer.
1834 *
1835 * When num_tfeedback_decls is nonzero, it is permissible for the consumer to
1836 * be NULL.  In this case, varying locations are assigned solely based on the
1837 * requirements of transform feedback.
1838 */
1839bool
1840assign_varying_locations(struct gl_context *ctx,
1841			 struct gl_shader_program *prog,
1842			 gl_shader *producer, gl_shader *consumer,
1843                         unsigned num_tfeedback_decls,
1844                         tfeedback_decl *tfeedback_decls)
1845{
1846   /* FINISHME: Set dynamically when geometry shader support is added. */
1847   unsigned output_index = VERT_RESULT_VAR0;
1848   unsigned input_index = FRAG_ATTRIB_VAR0;
1849
1850   /* Operate in a total of three passes.
1851    *
1852    * 1. Assign locations for any matching inputs and outputs.
1853    *
1854    * 2. Mark output variables in the producer that do not have locations as
1855    *    not being outputs.  This lets the optimizer eliminate them.
1856    *
1857    * 3. Mark input variables in the consumer that do not have locations as
1858    *    not being inputs.  This lets the optimizer eliminate them.
1859    */
1860
1861   link_invalidate_variable_locations(producer, ir_var_out, VERT_RESULT_VAR0);
1862   if (consumer)
1863      link_invalidate_variable_locations(consumer, ir_var_in, FRAG_ATTRIB_VAR0);
1864
1865   foreach_list(node, producer->ir) {
1866      ir_variable *const output_var = ((ir_instruction *) node)->as_variable();
1867
1868      if ((output_var == NULL) || (output_var->mode != ir_var_out))
1869	 continue;
1870
1871      ir_variable *input_var =
1872	 consumer ? consumer->symbols->get_variable(output_var->name) : NULL;
1873
1874      if (input_var && input_var->mode != ir_var_in)
1875         input_var = NULL;
1876
1877      if (input_var) {
1878         assign_varying_location(input_var, output_var, &input_index,
1879                                 &output_index);
1880      }
1881
1882      for (unsigned i = 0; i < num_tfeedback_decls; ++i) {
1883         if (!tfeedback_decls[i].is_assigned() &&
1884             tfeedback_decls[i].matches_var(output_var)) {
1885            if (output_var->location == -1) {
1886               assign_varying_location(input_var, output_var, &input_index,
1887                                       &output_index);
1888            }
1889            if (!tfeedback_decls[i].assign_location(ctx, prog, output_var))
1890               return false;
1891         }
1892      }
1893   }
1894
1895   unsigned varying_vectors = 0;
1896
1897   if (consumer) {
1898      foreach_list(node, consumer->ir) {
1899         ir_variable *const var = ((ir_instruction *) node)->as_variable();
1900
1901         if ((var == NULL) || (var->mode != ir_var_in))
1902            continue;
1903
1904         if (var->location == -1) {
1905            if (prog->Version <= 120) {
1906               /* On page 25 (page 31 of the PDF) of the GLSL 1.20 spec:
1907                *
1908                *     Only those varying variables used (i.e. read) in
1909                *     the fragment shader executable must be written to
1910                *     by the vertex shader executable; declaring
1911                *     superfluous varying variables in a vertex shader is
1912                *     permissible.
1913                *
1914                * We interpret this text as meaning that the VS must
1915                * write the variable for the FS to read it.  See
1916                * "glsl1-varying read but not written" in piglit.
1917                */
1918
1919               linker_error(prog, "fragment shader varying %s not written "
1920                            "by vertex shader\n.", var->name);
1921            }
1922
1923            /* An 'in' variable is only really a shader input if its
1924             * value is written by the previous stage.
1925             */
1926            var->mode = ir_var_auto;
1927         } else {
1928            /* The packing rules are used for vertex shader inputs are also
1929             * used for fragment shader inputs.
1930             */
1931            varying_vectors += count_attribute_slots(var->type);
1932         }
1933      }
1934   }
1935
1936   if (ctx->API == API_OPENGLES2 || prog->Version == 100) {
1937      if (varying_vectors > ctx->Const.MaxVarying) {
1938         if (ctx->Const.GLSLSkipStrictMaxVaryingLimitCheck) {
1939            linker_warning(prog, "shader uses too many varying vectors "
1940                           "(%u > %u), but the driver will try to optimize "
1941                           "them out; this is non-portable out-of-spec "
1942                           "behavior\n",
1943                           varying_vectors, ctx->Const.MaxVarying);
1944         } else {
1945            linker_error(prog, "shader uses too many varying vectors "
1946                         "(%u > %u)\n",
1947                         varying_vectors, ctx->Const.MaxVarying);
1948            return false;
1949         }
1950      }
1951   } else {
1952      const unsigned float_components = varying_vectors * 4;
1953      if (float_components > ctx->Const.MaxVarying * 4) {
1954         if (ctx->Const.GLSLSkipStrictMaxVaryingLimitCheck) {
1955            linker_warning(prog, "shader uses too many varying components "
1956                           "(%u > %u), but the driver will try to optimize "
1957                           "them out; this is non-portable out-of-spec "
1958                           "behavior\n",
1959                           float_components, ctx->Const.MaxVarying * 4);
1960         } else {
1961            linker_error(prog, "shader uses too many varying components "
1962                         "(%u > %u)\n",
1963                         float_components, ctx->Const.MaxVarying * 4);
1964            return false;
1965         }
1966      }
1967   }
1968
1969   return true;
1970}
1971
1972
1973/**
1974 * Store transform feedback location assignments into
1975 * prog->LinkedTransformFeedback based on the data stored in tfeedback_decls.
1976 *
1977 * If an error occurs, the error is reported through linker_error() and false
1978 * is returned.
1979 */
1980static bool
1981store_tfeedback_info(struct gl_context *ctx, struct gl_shader_program *prog,
1982                     unsigned num_tfeedback_decls,
1983                     tfeedback_decl *tfeedback_decls)
1984{
1985   bool separate_attribs_mode =
1986      prog->TransformFeedback.BufferMode == GL_SEPARATE_ATTRIBS;
1987
1988   ralloc_free(prog->LinkedTransformFeedback.Varyings);
1989   ralloc_free(prog->LinkedTransformFeedback.Outputs);
1990
1991   memset(&prog->LinkedTransformFeedback, 0,
1992          sizeof(prog->LinkedTransformFeedback));
1993
1994   prog->LinkedTransformFeedback.NumBuffers =
1995      separate_attribs_mode ? num_tfeedback_decls : 1;
1996
1997   prog->LinkedTransformFeedback.Varyings =
1998      rzalloc_array(prog,
1999		    struct gl_transform_feedback_varying_info,
2000		    num_tfeedback_decls);
2001
2002   unsigned num_outputs = 0;
2003   for (unsigned i = 0; i < num_tfeedback_decls; ++i)
2004      if (!tfeedback_decls[i].accumulate_num_outputs(prog, &num_outputs))
2005         return false;
2006
2007   prog->LinkedTransformFeedback.Outputs =
2008      rzalloc_array(prog,
2009                    struct gl_transform_feedback_output,
2010                    num_outputs);
2011
2012   for (unsigned i = 0; i < num_tfeedback_decls; ++i) {
2013      unsigned buffer = separate_attribs_mode ? i : 0;
2014      if (!tfeedback_decls[i].store(ctx, prog, &prog->LinkedTransformFeedback,
2015                                    buffer, i, num_outputs))
2016         return false;
2017   }
2018   assert(prog->LinkedTransformFeedback.NumOutputs == num_outputs);
2019
2020   return true;
2021}
2022
2023/**
2024 * Store the gl_FragDepth layout in the gl_shader_program struct.
2025 */
2026static void
2027store_fragdepth_layout(struct gl_shader_program *prog)
2028{
2029   if (prog->_LinkedShaders[MESA_SHADER_FRAGMENT] == NULL) {
2030      return;
2031   }
2032
2033   struct exec_list *ir = prog->_LinkedShaders[MESA_SHADER_FRAGMENT]->ir;
2034
2035   /* We don't look up the gl_FragDepth symbol directly because if
2036    * gl_FragDepth is not used in the shader, it's removed from the IR.
2037    * However, the symbol won't be removed from the symbol table.
2038    *
2039    * We're only interested in the cases where the variable is NOT removed
2040    * from the IR.
2041    */
2042   foreach_list(node, ir) {
2043      ir_variable *const var = ((ir_instruction *) node)->as_variable();
2044
2045      if (var == NULL || var->mode != ir_var_out) {
2046         continue;
2047      }
2048
2049      if (strcmp(var->name, "gl_FragDepth") == 0) {
2050         switch (var->depth_layout) {
2051         case ir_depth_layout_none:
2052            prog->FragDepthLayout = FRAG_DEPTH_LAYOUT_NONE;
2053            return;
2054         case ir_depth_layout_any:
2055            prog->FragDepthLayout = FRAG_DEPTH_LAYOUT_ANY;
2056            return;
2057         case ir_depth_layout_greater:
2058            prog->FragDepthLayout = FRAG_DEPTH_LAYOUT_GREATER;
2059            return;
2060         case ir_depth_layout_less:
2061            prog->FragDepthLayout = FRAG_DEPTH_LAYOUT_LESS;
2062            return;
2063         case ir_depth_layout_unchanged:
2064            prog->FragDepthLayout = FRAG_DEPTH_LAYOUT_UNCHANGED;
2065            return;
2066         default:
2067            assert(0);
2068            return;
2069         }
2070      }
2071   }
2072}
2073
2074/**
2075 * Validate the resources used by a program versus the implementation limits
2076 */
2077static bool
2078check_resources(struct gl_context *ctx, struct gl_shader_program *prog)
2079{
2080   static const char *const shader_names[MESA_SHADER_TYPES] = {
2081      "vertex", "fragment", "geometry"
2082   };
2083
2084   const unsigned max_samplers[MESA_SHADER_TYPES] = {
2085      ctx->Const.MaxVertexTextureImageUnits,
2086      ctx->Const.MaxTextureImageUnits,
2087      ctx->Const.MaxGeometryTextureImageUnits
2088   };
2089
2090   const unsigned max_uniform_components[MESA_SHADER_TYPES] = {
2091      ctx->Const.VertexProgram.MaxUniformComponents,
2092      ctx->Const.FragmentProgram.MaxUniformComponents,
2093      0          /* FINISHME: Geometry shaders. */
2094   };
2095
2096   for (unsigned i = 0; i < MESA_SHADER_TYPES; i++) {
2097      struct gl_shader *sh = prog->_LinkedShaders[i];
2098
2099      if (sh == NULL)
2100	 continue;
2101
2102      if (sh->num_samplers > max_samplers[i]) {
2103	 linker_error(prog, "Too many %s shader texture samplers",
2104		      shader_names[i]);
2105      }
2106
2107      if (sh->num_uniform_components > max_uniform_components[i]) {
2108         if (ctx->Const.GLSLSkipStrictMaxUniformLimitCheck) {
2109            linker_warning(prog, "Too many %s shader uniform components, "
2110                           "but the driver will try to optimize them out; "
2111                           "this is non-portable out-of-spec behavior\n",
2112                           shader_names[i]);
2113         } else {
2114            linker_error(prog, "Too many %s shader uniform components",
2115                         shader_names[i]);
2116         }
2117      }
2118   }
2119
2120   return prog->LinkStatus;
2121}
2122
2123void
2124link_shaders(struct gl_context *ctx, struct gl_shader_program *prog)
2125{
2126   tfeedback_decl *tfeedback_decls = NULL;
2127   unsigned num_tfeedback_decls = prog->TransformFeedback.NumVarying;
2128
2129   void *mem_ctx = ralloc_context(NULL); // temporary linker context
2130
2131   prog->LinkStatus = false;
2132   prog->Validated = false;
2133   prog->_Used = false;
2134
2135   if (prog->InfoLog != NULL)
2136      ralloc_free(prog->InfoLog);
2137
2138   prog->InfoLog = ralloc_strdup(NULL, "");
2139
2140   /* Separate the shaders into groups based on their type.
2141    */
2142   struct gl_shader **vert_shader_list;
2143   unsigned num_vert_shaders = 0;
2144   struct gl_shader **frag_shader_list;
2145   unsigned num_frag_shaders = 0;
2146
2147   vert_shader_list = (struct gl_shader **)
2148      calloc(2 * prog->NumShaders, sizeof(struct gl_shader *));
2149   frag_shader_list =  &vert_shader_list[prog->NumShaders];
2150
2151   unsigned min_version = UINT_MAX;
2152   unsigned max_version = 0;
2153   for (unsigned i = 0; i < prog->NumShaders; i++) {
2154      min_version = MIN2(min_version, prog->Shaders[i]->Version);
2155      max_version = MAX2(max_version, prog->Shaders[i]->Version);
2156
2157      switch (prog->Shaders[i]->Type) {
2158      case GL_VERTEX_SHADER:
2159	 vert_shader_list[num_vert_shaders] = prog->Shaders[i];
2160	 num_vert_shaders++;
2161	 break;
2162      case GL_FRAGMENT_SHADER:
2163	 frag_shader_list[num_frag_shaders] = prog->Shaders[i];
2164	 num_frag_shaders++;
2165	 break;
2166      case GL_GEOMETRY_SHADER:
2167	 /* FINISHME: Support geometry shaders. */
2168	 assert(prog->Shaders[i]->Type != GL_GEOMETRY_SHADER);
2169	 break;
2170      }
2171   }
2172
2173   /* Previous to GLSL version 1.30, different compilation units could mix and
2174    * match shading language versions.  With GLSL 1.30 and later, the versions
2175    * of all shaders must match.
2176    */
2177   assert(min_version >= 100);
2178   assert(max_version <= 130);
2179   if ((max_version >= 130 || min_version == 100)
2180       && min_version != max_version) {
2181      linker_error(prog, "all shaders must use same shading "
2182		   "language version\n");
2183      goto done;
2184   }
2185
2186   prog->Version = max_version;
2187
2188   for (unsigned int i = 0; i < MESA_SHADER_TYPES; i++) {
2189      if (prog->_LinkedShaders[i] != NULL)
2190	 ctx->Driver.DeleteShader(ctx, prog->_LinkedShaders[i]);
2191
2192      prog->_LinkedShaders[i] = NULL;
2193   }
2194
2195   /* Link all shaders for a particular stage and validate the result.
2196    */
2197   if (num_vert_shaders > 0) {
2198      gl_shader *const sh =
2199	 link_intrastage_shaders(mem_ctx, ctx, prog, vert_shader_list,
2200				 num_vert_shaders);
2201
2202      if (sh == NULL)
2203	 goto done;
2204
2205      if (!validate_vertex_shader_executable(prog, sh))
2206	 goto done;
2207
2208      _mesa_reference_shader(ctx, &prog->_LinkedShaders[MESA_SHADER_VERTEX],
2209			     sh);
2210   }
2211
2212   if (num_frag_shaders > 0) {
2213      gl_shader *const sh =
2214	 link_intrastage_shaders(mem_ctx, ctx, prog, frag_shader_list,
2215				 num_frag_shaders);
2216
2217      if (sh == NULL)
2218	 goto done;
2219
2220      if (!validate_fragment_shader_executable(prog, sh))
2221	 goto done;
2222
2223      _mesa_reference_shader(ctx, &prog->_LinkedShaders[MESA_SHADER_FRAGMENT],
2224			     sh);
2225   }
2226
2227   /* Here begins the inter-stage linking phase.  Some initial validation is
2228    * performed, then locations are assigned for uniforms, attributes, and
2229    * varyings.
2230    */
2231   if (cross_validate_uniforms(prog)) {
2232      unsigned prev;
2233
2234      for (prev = 0; prev < MESA_SHADER_TYPES; prev++) {
2235	 if (prog->_LinkedShaders[prev] != NULL)
2236	    break;
2237      }
2238
2239      /* Validate the inputs of each stage with the output of the preceding
2240       * stage.
2241       */
2242      for (unsigned i = prev + 1; i < MESA_SHADER_TYPES; i++) {
2243	 if (prog->_LinkedShaders[i] == NULL)
2244	    continue;
2245
2246	 if (!cross_validate_outputs_to_inputs(prog,
2247					       prog->_LinkedShaders[prev],
2248					       prog->_LinkedShaders[i]))
2249	    goto done;
2250
2251	 prev = i;
2252      }
2253
2254      prog->LinkStatus = true;
2255   }
2256
2257   /* Do common optimization before assigning storage for attributes,
2258    * uniforms, and varyings.  Later optimization could possibly make
2259    * some of that unused.
2260    */
2261   for (unsigned i = 0; i < MESA_SHADER_TYPES; i++) {
2262      if (prog->_LinkedShaders[i] == NULL)
2263	 continue;
2264
2265      detect_recursion_linked(prog, prog->_LinkedShaders[i]->ir);
2266      if (!prog->LinkStatus)
2267	 goto done;
2268
2269      if (ctx->ShaderCompilerOptions[i].LowerClipDistance)
2270         lower_clip_distance(prog->_LinkedShaders[i]->ir);
2271
2272      while (do_common_optimization(prog->_LinkedShaders[i]->ir, true, false, 32))
2273	 ;
2274   }
2275
2276   /* FINISHME: The value of the max_attribute_index parameter is
2277    * FINISHME: implementation dependent based on the value of
2278    * FINISHME: GL_MAX_VERTEX_ATTRIBS.  GL_MAX_VERTEX_ATTRIBS must be
2279    * FINISHME: at least 16, so hardcode 16 for now.
2280    */
2281   if (!assign_attribute_or_color_locations(prog, MESA_SHADER_VERTEX, 16)) {
2282      goto done;
2283   }
2284
2285   if (!assign_attribute_or_color_locations(prog, MESA_SHADER_FRAGMENT, ctx->Const.MaxDrawBuffers)) {
2286      goto done;
2287   }
2288
2289   unsigned prev;
2290   for (prev = 0; prev < MESA_SHADER_TYPES; prev++) {
2291      if (prog->_LinkedShaders[prev] != NULL)
2292	 break;
2293   }
2294
2295   if (num_tfeedback_decls != 0) {
2296      /* From GL_EXT_transform_feedback:
2297       *   A program will fail to link if:
2298       *
2299       *   * the <count> specified by TransformFeedbackVaryingsEXT is
2300       *     non-zero, but the program object has no vertex or geometry
2301       *     shader;
2302       */
2303      if (prev >= MESA_SHADER_FRAGMENT) {
2304         linker_error(prog, "Transform feedback varyings specified, but "
2305                      "no vertex or geometry shader is present.");
2306         goto done;
2307      }
2308
2309      tfeedback_decls = ralloc_array(mem_ctx, tfeedback_decl,
2310                                     prog->TransformFeedback.NumVarying);
2311      if (!parse_tfeedback_decls(ctx, prog, mem_ctx, num_tfeedback_decls,
2312                                 prog->TransformFeedback.VaryingNames,
2313                                 tfeedback_decls))
2314         goto done;
2315   }
2316
2317   for (unsigned i = prev + 1; i < MESA_SHADER_TYPES; i++) {
2318      if (prog->_LinkedShaders[i] == NULL)
2319	 continue;
2320
2321      if (!assign_varying_locations(
2322             ctx, prog, prog->_LinkedShaders[prev], prog->_LinkedShaders[i],
2323             i == MESA_SHADER_FRAGMENT ? num_tfeedback_decls : 0,
2324             tfeedback_decls))
2325	 goto done;
2326
2327      prev = i;
2328   }
2329
2330   if (prev != MESA_SHADER_FRAGMENT && num_tfeedback_decls != 0) {
2331      /* There was no fragment shader, but we still have to assign varying
2332       * locations for use by transform feedback.
2333       */
2334      if (!assign_varying_locations(
2335             ctx, prog, prog->_LinkedShaders[prev], NULL, num_tfeedback_decls,
2336             tfeedback_decls))
2337         goto done;
2338   }
2339
2340   if (!store_tfeedback_info(ctx, prog, num_tfeedback_decls, tfeedback_decls))
2341      goto done;
2342
2343   if (prog->_LinkedShaders[MESA_SHADER_VERTEX] != NULL) {
2344      demote_shader_inputs_and_outputs(prog->_LinkedShaders[MESA_SHADER_VERTEX],
2345				       ir_var_out);
2346
2347      /* Eliminate code that is now dead due to unused vertex outputs being
2348       * demoted.
2349       */
2350      while (do_dead_code(prog->_LinkedShaders[MESA_SHADER_VERTEX]->ir, false))
2351	 ;
2352   }
2353
2354   if (prog->_LinkedShaders[MESA_SHADER_GEOMETRY] != NULL) {
2355      gl_shader *const sh = prog->_LinkedShaders[MESA_SHADER_GEOMETRY];
2356
2357      demote_shader_inputs_and_outputs(sh, ir_var_in);
2358      demote_shader_inputs_and_outputs(sh, ir_var_inout);
2359      demote_shader_inputs_and_outputs(sh, ir_var_out);
2360
2361      /* Eliminate code that is now dead due to unused geometry outputs being
2362       * demoted.
2363       */
2364      while (do_dead_code(prog->_LinkedShaders[MESA_SHADER_GEOMETRY]->ir, false))
2365	 ;
2366   }
2367
2368   if (prog->_LinkedShaders[MESA_SHADER_FRAGMENT] != NULL) {
2369      gl_shader *const sh = prog->_LinkedShaders[MESA_SHADER_FRAGMENT];
2370
2371      demote_shader_inputs_and_outputs(sh, ir_var_in);
2372
2373      /* Eliminate code that is now dead due to unused fragment inputs being
2374       * demoted.  This shouldn't actually do anything other than remove
2375       * declarations of the (now unused) global variables.
2376       */
2377      while (do_dead_code(prog->_LinkedShaders[MESA_SHADER_FRAGMENT]->ir, false))
2378	 ;
2379   }
2380
2381   update_array_sizes(prog);
2382   link_assign_uniform_locations(prog);
2383   store_fragdepth_layout(prog);
2384
2385   if (!check_resources(ctx, prog))
2386      goto done;
2387
2388   /* OpenGL ES requires that a vertex shader and a fragment shader both be
2389    * present in a linked program.  By checking for use of shading language
2390    * version 1.00, we also catch the GL_ARB_ES2_compatibility case.
2391    */
2392   if (!prog->InternalSeparateShader &&
2393       (ctx->API == API_OPENGLES2 || prog->Version == 100)) {
2394      if (prog->_LinkedShaders[MESA_SHADER_VERTEX] == NULL) {
2395	 linker_error(prog, "program lacks a vertex shader\n");
2396      } else if (prog->_LinkedShaders[MESA_SHADER_FRAGMENT] == NULL) {
2397	 linker_error(prog, "program lacks a fragment shader\n");
2398      }
2399   }
2400
2401   /* FINISHME: Assign fragment shader output locations. */
2402
2403done:
2404   free(vert_shader_list);
2405
2406   for (unsigned i = 0; i < MESA_SHADER_TYPES; i++) {
2407      if (prog->_LinkedShaders[i] == NULL)
2408	 continue;
2409
2410      /* Retain any live IR, but trash the rest. */
2411      reparent_ir(prog->_LinkedShaders[i]->ir, prog->_LinkedShaders[i]->ir);
2412
2413      /* The symbol table in the linked shaders may contain references to
2414       * variables that were removed (e.g., unused uniforms).  Since it may
2415       * contain junk, there is no possible valid use.  Delete it and set the
2416       * pointer to NULL.
2417       */
2418      delete prog->_LinkedShaders[i]->symbols;
2419      prog->_LinkedShaders[i]->symbols = NULL;
2420   }
2421
2422   ralloc_free(mem_ctx);
2423}
2424