2 * Copyright © 2012 Intel Corporation
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
21 * DEALINGS IN THE SOFTWARE.
25 * \file link_varyings.cpp
27 * Linker functions related specifically to linking varyings between shader
32 #include "main/mtypes.h"
33 #include "glsl_symbol_table.h"
34 #include "glsl_parser_extras.h"
35 #include "ir_optimization.h"
37 #include "link_varyings.h"
38 #include "main/macros.h"
39 #include "program/hash_table.h"
44 * Get the varying type stripped of the outermost array if we're processing
45 * a stage whose varyings are arrays indexed by a vertex number (such as
46 * geometry shader inputs).
48 static const glsl_type
*
49 get_varying_type(const ir_variable
*var
, gl_shader_stage stage
)
51 const glsl_type
*type
= var
->type
;
53 if (!var
->data
.patch
&&
54 ((var
->data
.mode
== ir_var_shader_out
&&
55 stage
== MESA_SHADER_TESS_CTRL
) ||
56 (var
->data
.mode
== ir_var_shader_in
&&
57 (stage
== MESA_SHADER_TESS_CTRL
|| stage
== MESA_SHADER_TESS_EVAL
||
58 stage
== MESA_SHADER_GEOMETRY
)))) {
59 assert(type
->is_array());
60 type
= type
->fields
.array
;
67 create_xfb_varying_names(void *mem_ctx
, const glsl_type
*t
, char **name
,
68 size_t name_length
, unsigned *count
,
69 const char *ifc_member_name
,
70 const glsl_type
*ifc_member_t
, char ***varying_names
)
72 if (t
->is_interface()) {
73 size_t new_length
= name_length
;
75 assert(ifc_member_name
&& ifc_member_t
);
76 ralloc_asprintf_rewrite_tail(name
, &new_length
, ".%s", ifc_member_name
);
78 create_xfb_varying_names(mem_ctx
, ifc_member_t
, name
, new_length
, count
,
79 NULL
, NULL
, varying_names
);
80 } else if (t
->is_record()) {
81 for (unsigned i
= 0; i
< t
->length
; i
++) {
82 const char *field
= t
->fields
.structure
[i
].name
;
83 size_t new_length
= name_length
;
85 ralloc_asprintf_rewrite_tail(name
, &new_length
, ".%s", field
);
87 create_xfb_varying_names(mem_ctx
, t
->fields
.structure
[i
].type
, name
,
88 new_length
, count
, NULL
, NULL
,
91 } else if (t
->without_array()->is_record() ||
92 t
->without_array()->is_interface() ||
93 (t
->is_array() && t
->fields
.array
->is_array())) {
94 for (unsigned i
= 0; i
< t
->length
; i
++) {
95 size_t new_length
= name_length
;
97 /* Append the subscript to the current variable name */
98 ralloc_asprintf_rewrite_tail(name
, &new_length
, "[%u]", i
);
100 create_xfb_varying_names(mem_ctx
, t
->fields
.array
, name
, new_length
,
101 count
, ifc_member_name
, ifc_member_t
,
105 (*varying_names
)[(*count
)++] = ralloc_strdup(mem_ctx
, *name
);
110 process_xfb_layout_qualifiers(void *mem_ctx
, const gl_shader
*sh
,
111 unsigned *num_tfeedback_decls
,
112 char ***varying_names
)
114 bool has_xfb_qualifiers
= false;
116 /* We still need to enable transform feedback mode even if xfb_stride is
117 * only applied to a global out. Also we don't bother to propagate
118 * xfb_stride to interface block members so this will catch that case also.
120 for (unsigned j
= 0; j
< MAX_FEEDBACK_BUFFERS
; j
++) {
121 if (sh
->TransformFeedback
.BufferStride
[j
]) {
122 has_xfb_qualifiers
= true;
126 foreach_in_list(ir_instruction
, node
, sh
->ir
) {
127 ir_variable
*var
= node
->as_variable();
128 if (!var
|| var
->data
.mode
!= ir_var_shader_out
)
131 /* From the ARB_enhanced_layouts spec:
133 * "Any shader making any static use (after preprocessing) of any of
134 * these *xfb_* qualifiers will cause the shader to be in a
135 * transform feedback capturing mode and hence responsible for
136 * describing the transform feedback setup. This mode will capture
137 * any output selected by *xfb_offset*, directly or indirectly, to
138 * a transform feedback buffer."
140 if (var
->data
.explicit_xfb_buffer
|| var
->data
.explicit_xfb_stride
) {
141 has_xfb_qualifiers
= true;
144 if (var
->data
.explicit_xfb_offset
) {
145 *num_tfeedback_decls
+= var
->type
->varying_count();
146 has_xfb_qualifiers
= true;
150 if (*num_tfeedback_decls
== 0)
151 return has_xfb_qualifiers
;
154 *varying_names
= ralloc_array(mem_ctx
, char *, *num_tfeedback_decls
);
155 foreach_in_list(ir_instruction
, node
, sh
->ir
) {
156 ir_variable
*var
= node
->as_variable();
157 if (!var
|| var
->data
.mode
!= ir_var_shader_out
)
160 if (var
->data
.explicit_xfb_offset
) {
162 const glsl_type
*type
, *member_type
;
164 if (var
->data
.from_named_ifc_block
) {
165 type
= var
->get_interface_type();
166 /* Find the member type before it was altered by lowering */
168 type
->fields
.structure
[type
->field_index(var
->name
)].type
;
169 name
= ralloc_strdup(NULL
, type
->without_array()->name
);
173 name
= ralloc_strdup(NULL
, var
->name
);
175 create_xfb_varying_names(mem_ctx
, type
, &name
, strlen(name
), &i
,
176 var
->name
, member_type
, varying_names
);
181 assert(i
== *num_tfeedback_decls
);
182 return has_xfb_qualifiers
;
186 anonymous_struct_type_matches(const glsl_type
*output_type
,
187 const glsl_type
*to_match
)
189 while (output_type
->is_array() && to_match
->is_array()) {
190 /* if the lengths at each level don't match fail. */
191 if (output_type
->length
!= to_match
->length
)
193 output_type
= output_type
->fields
.array
;
194 to_match
= to_match
->fields
.array
;
197 if (output_type
->is_array() || to_match
->is_array())
199 return output_type
->is_anonymous() &&
200 to_match
->is_anonymous() &&
201 to_match
->record_compare(output_type
);
205 * Validate the types and qualifiers of an output from one stage against the
206 * matching input to another stage.
209 cross_validate_types_and_qualifiers(struct gl_shader_program
*prog
,
210 const ir_variable
*input
,
211 const ir_variable
*output
,
212 gl_shader_stage consumer_stage
,
213 gl_shader_stage producer_stage
)
215 /* Check that the types match between stages.
217 const glsl_type
*type_to_match
= input
->type
;
219 /* VS -> GS, VS -> TCS, VS -> TES, TES -> GS */
220 const bool extra_array_level
= (producer_stage
== MESA_SHADER_VERTEX
&&
221 consumer_stage
!= MESA_SHADER_FRAGMENT
) ||
222 consumer_stage
== MESA_SHADER_GEOMETRY
;
223 if (extra_array_level
) {
224 assert(type_to_match
->is_array());
225 type_to_match
= type_to_match
->fields
.array
;
228 if (type_to_match
!= output
->type
) {
229 /* There is a bit of a special case for gl_TexCoord. This
230 * built-in is unsized by default. Applications that variable
231 * access it must redeclare it with a size. There is some
232 * language in the GLSL spec that implies the fragment shader
233 * and vertex shader do not have to agree on this size. Other
234 * driver behave this way, and one or two applications seem to
237 * Neither declaration needs to be modified here because the array
238 * sizes are fixed later when update_array_sizes is called.
240 * From page 48 (page 54 of the PDF) of the GLSL 1.10 spec:
242 * "Unlike user-defined varying variables, the built-in
243 * varying variables don't have a strict one-to-one
244 * correspondence between the vertex language and the
245 * fragment language."
247 if (!output
->type
->is_array() || !is_gl_identifier(output
->name
)) {
248 bool anon_matches
= anonymous_struct_type_matches(output
->type
, type_to_match
);
252 "%s shader output `%s' declared as type `%s', "
253 "but %s shader input declared as type `%s'\n",
254 _mesa_shader_stage_to_string(producer_stage
),
257 _mesa_shader_stage_to_string(consumer_stage
),
264 /* Check that all of the qualifiers match between stages.
267 /* According to the OpenGL and OpenGLES GLSL specs, the centroid qualifier
268 * should match until OpenGL 4.3 and OpenGLES 3.1. The OpenGLES 3.0
269 * conformance test suite does not verify that the qualifiers must match.
270 * The deqp test suite expects the opposite (OpenGLES 3.1) behavior for
271 * OpenGLES 3.0 drivers, so we relax the checking in all cases.
273 if (false /* always skip the centroid check */ &&
274 prog
->Version
< (prog
->IsES
? 310 : 430) &&
275 input
->data
.centroid
!= output
->data
.centroid
) {
277 "%s shader output `%s' %s centroid qualifier, "
278 "but %s shader input %s centroid qualifier\n",
279 _mesa_shader_stage_to_string(producer_stage
),
281 (output
->data
.centroid
) ? "has" : "lacks",
282 _mesa_shader_stage_to_string(consumer_stage
),
283 (input
->data
.centroid
) ? "has" : "lacks");
287 if (input
->data
.sample
!= output
->data
.sample
) {
289 "%s shader output `%s' %s sample qualifier, "
290 "but %s shader input %s sample qualifier\n",
291 _mesa_shader_stage_to_string(producer_stage
),
293 (output
->data
.sample
) ? "has" : "lacks",
294 _mesa_shader_stage_to_string(consumer_stage
),
295 (input
->data
.sample
) ? "has" : "lacks");
299 if (input
->data
.patch
!= output
->data
.patch
) {
301 "%s shader output `%s' %s patch qualifier, "
302 "but %s shader input %s patch qualifier\n",
303 _mesa_shader_stage_to_string(producer_stage
),
305 (output
->data
.patch
) ? "has" : "lacks",
306 _mesa_shader_stage_to_string(consumer_stage
),
307 (input
->data
.patch
) ? "has" : "lacks");
311 if (!prog
->IsES
&& input
->data
.invariant
!= output
->data
.invariant
) {
313 "%s shader output `%s' %s invariant qualifier, "
314 "but %s shader input %s invariant qualifier\n",
315 _mesa_shader_stage_to_string(producer_stage
),
317 (output
->data
.invariant
) ? "has" : "lacks",
318 _mesa_shader_stage_to_string(consumer_stage
),
319 (input
->data
.invariant
) ? "has" : "lacks");
323 /* GLSL >= 4.40 removes text requiring interpolation qualifiers
324 * to match cross stage, they must only match within the same stage.
326 * From page 84 (page 90 of the PDF) of the GLSL 4.40 spec:
328 * "It is a link-time error if, within the same stage, the interpolation
329 * qualifiers of variables of the same name do not match.
332 if (input
->data
.interpolation
!= output
->data
.interpolation
&&
333 prog
->Version
< 440) {
335 "%s shader output `%s' specifies %s "
336 "interpolation qualifier, "
337 "but %s shader input specifies %s "
338 "interpolation qualifier\n",
339 _mesa_shader_stage_to_string(producer_stage
),
341 interpolation_string(output
->data
.interpolation
),
342 _mesa_shader_stage_to_string(consumer_stage
),
343 interpolation_string(input
->data
.interpolation
));
349 * Validate front and back color outputs against single color input
352 cross_validate_front_and_back_color(struct gl_shader_program
*prog
,
353 const ir_variable
*input
,
354 const ir_variable
*front_color
,
355 const ir_variable
*back_color
,
356 gl_shader_stage consumer_stage
,
357 gl_shader_stage producer_stage
)
359 if (front_color
!= NULL
&& front_color
->data
.assigned
)
360 cross_validate_types_and_qualifiers(prog
, input
, front_color
,
361 consumer_stage
, producer_stage
);
363 if (back_color
!= NULL
&& back_color
->data
.assigned
)
364 cross_validate_types_and_qualifiers(prog
, input
, back_color
,
365 consumer_stage
, producer_stage
);
369 * Validate that outputs from one stage match inputs of another
372 cross_validate_outputs_to_inputs(struct gl_shader_program
*prog
,
373 gl_shader
*producer
, gl_shader
*consumer
)
375 glsl_symbol_table parameters
;
376 ir_variable
*explicit_locations
[MAX_VARYINGS_INCL_PATCH
][4] =
379 /* Find all shader outputs in the "producer" stage.
381 foreach_in_list(ir_instruction
, node
, producer
->ir
) {
382 ir_variable
*const var
= node
->as_variable();
384 if (var
== NULL
|| var
->data
.mode
!= ir_var_shader_out
)
387 if (!var
->data
.explicit_location
388 || var
->data
.location
< VARYING_SLOT_VAR0
)
389 parameters
.add_variable(var
);
391 /* User-defined varyings with explicit locations are handled
392 * differently because they do not need to have matching names.
394 const glsl_type
*type
= get_varying_type(var
, producer
->Stage
);
395 unsigned num_elements
= type
->count_attribute_slots(false);
396 unsigned idx
= var
->data
.location
- VARYING_SLOT_VAR0
;
397 unsigned slot_limit
= idx
+ num_elements
;
400 if (var
->type
->without_array()->is_record()) {
401 /* The component qualifier can't be used on structs so just treat
402 * all component slots as used.
406 unsigned dmul
= var
->type
->is_double() ? 2 : 1;
407 last_comp
= var
->data
.location_frac
+
408 var
->type
->without_array()->vector_elements
* dmul
;
411 while (idx
< slot_limit
) {
412 unsigned i
= var
->data
.location_frac
;
413 while (i
< last_comp
) {
414 if (explicit_locations
[idx
][i
] != NULL
) {
416 "%s shader has multiple outputs explicitly "
417 "assigned to location %d and component %d\n",
418 _mesa_shader_stage_to_string(producer
->Stage
),
419 idx
, var
->data
.location_frac
);
423 /* Make sure all component at this location have the same type.
425 for (unsigned j
= 0; j
< 4; j
++) {
426 if (explicit_locations
[idx
][j
] &&
427 (explicit_locations
[idx
][j
]->type
->without_array()
428 ->base_type
!= var
->type
->without_array()->base_type
)) {
430 "Varyings sharing the same location must "
431 "have the same underlying numerical type. "
432 "Location %u component %u\n", idx
,
433 var
->data
.location_frac
);
438 explicit_locations
[idx
][i
] = var
;
441 /* We need to do some special handling for doubles as dvec3 and
442 * dvec4 consume two consecutive locations. We don't need to
443 * worry about components beginning at anything other than 0 as
444 * the spec does not allow this for dvec3 and dvec4.
446 if (i
== 3 && last_comp
> 4) {
447 last_comp
= last_comp
- 4;
448 /* Bump location index and reset the component index */
459 /* Find all shader inputs in the "consumer" stage. Any variables that have
460 * matching outputs already in the symbol table must have the same type and
463 * Exception: if the consumer is the geometry shader, then the inputs
464 * should be arrays and the type of the array element should match the type
465 * of the corresponding producer output.
467 foreach_in_list(ir_instruction
, node
, consumer
->ir
) {
468 ir_variable
*const input
= node
->as_variable();
470 if (input
== NULL
|| input
->data
.mode
!= ir_var_shader_in
)
473 if (strcmp(input
->name
, "gl_Color") == 0 && input
->data
.used
) {
474 const ir_variable
*const front_color
=
475 parameters
.get_variable("gl_FrontColor");
477 const ir_variable
*const back_color
=
478 parameters
.get_variable("gl_BackColor");
480 cross_validate_front_and_back_color(prog
, input
,
481 front_color
, back_color
,
482 consumer
->Stage
, producer
->Stage
);
483 } else if (strcmp(input
->name
, "gl_SecondaryColor") == 0 && input
->data
.used
) {
484 const ir_variable
*const front_color
=
485 parameters
.get_variable("gl_FrontSecondaryColor");
487 const ir_variable
*const back_color
=
488 parameters
.get_variable("gl_BackSecondaryColor");
490 cross_validate_front_and_back_color(prog
, input
,
491 front_color
, back_color
,
492 consumer
->Stage
, producer
->Stage
);
494 /* The rules for connecting inputs and outputs change in the presence
495 * of explicit locations. In this case, we no longer care about the
496 * names of the variables. Instead, we care only about the
497 * explicitly assigned location.
499 ir_variable
*output
= NULL
;
500 if (input
->data
.explicit_location
501 && input
->data
.location
>= VARYING_SLOT_VAR0
) {
503 const glsl_type
*type
= get_varying_type(input
, consumer
->Stage
);
504 unsigned num_elements
= type
->count_attribute_slots(false);
505 unsigned idx
= input
->data
.location
- VARYING_SLOT_VAR0
;
506 unsigned slot_limit
= idx
+ num_elements
;
508 while (idx
< slot_limit
) {
509 output
= explicit_locations
[idx
][input
->data
.location_frac
];
511 if (output
== NULL
||
512 input
->data
.location
!= output
->data
.location
) {
514 "%s shader input `%s' with explicit location "
515 "has no matching output\n",
516 _mesa_shader_stage_to_string(consumer
->Stage
),
523 output
= parameters
.get_variable(input
->name
);
526 if (output
!= NULL
) {
527 /* Interface blocks have their own validation elsewhere so don't
528 * try validating them here.
530 if (!(input
->get_interface_type() &&
531 output
->get_interface_type()))
532 cross_validate_types_and_qualifiers(prog
, input
, output
,
536 /* Check for input vars with unmatched output vars in prev stage
537 * taking into account that interface blocks could have a matching
538 * output but with different name, so we ignore them.
540 assert(!input
->data
.assigned
);
541 if (input
->data
.used
&& !input
->get_interface_type() &&
542 !input
->data
.explicit_location
&& !prog
->SeparateShader
)
544 "%s shader input `%s' "
545 "has no matching output in the previous stage\n",
546 _mesa_shader_stage_to_string(consumer
->Stage
),
554 * Demote shader inputs and outputs that are not used in other stages, and
555 * remove them via dead code elimination.
558 remove_unused_shader_inputs_and_outputs(bool is_separate_shader_object
,
560 enum ir_variable_mode mode
)
562 if (is_separate_shader_object
)
565 foreach_in_list(ir_instruction
, node
, sh
->ir
) {
566 ir_variable
*const var
= node
->as_variable();
568 if (var
== NULL
|| var
->data
.mode
!= int(mode
))
571 /* A shader 'in' or 'out' variable is only really an input or output if
572 * its value is used by other shader stages. This will cause the
573 * variable to have a location assigned.
575 if (var
->data
.is_unmatched_generic_inout
&& !var
->data
.is_xfb_only
) {
576 assert(var
->data
.mode
!= ir_var_temporary
);
577 var
->data
.mode
= ir_var_auto
;
581 /* Eliminate code that is now dead due to unused inputs/outputs being
584 while (do_dead_code(sh
->ir
, false))
590 * Initialize this object based on a string that was passed to
591 * glTransformFeedbackVaryings.
593 * If the input is mal-formed, this call still succeeds, but it sets
594 * this->var_name to a mal-formed input, so tfeedback_decl::find_output_var()
595 * will fail to find any matching variable.
598 tfeedback_decl::init(struct gl_context
*ctx
, const void *mem_ctx
,
601 /* We don't have to be pedantic about what is a valid GLSL variable name,
602 * because any variable with an invalid name can't exist in the IR anyway.
606 this->orig_name
= input
;
607 this->lowered_builtin_array_variable
= none
;
608 this->skip_components
= 0;
609 this->next_buffer_separator
= false;
610 this->matched_candidate
= NULL
;
615 if (ctx
->Extensions
.ARB_transform_feedback3
) {
616 /* Parse gl_NextBuffer. */
617 if (strcmp(input
, "gl_NextBuffer") == 0) {
618 this->next_buffer_separator
= true;
622 /* Parse gl_SkipComponents. */
623 if (strcmp(input
, "gl_SkipComponents1") == 0)
624 this->skip_components
= 1;
625 else if (strcmp(input
, "gl_SkipComponents2") == 0)
626 this->skip_components
= 2;
627 else if (strcmp(input
, "gl_SkipComponents3") == 0)
628 this->skip_components
= 3;
629 else if (strcmp(input
, "gl_SkipComponents4") == 0)
630 this->skip_components
= 4;
632 if (this->skip_components
)
636 /* Parse a declaration. */
637 const char *base_name_end
;
638 long subscript
= parse_program_resource_name(input
, &base_name_end
);
639 this->var_name
= ralloc_strndup(mem_ctx
, input
, base_name_end
- input
);
640 if (this->var_name
== NULL
) {
641 _mesa_error_no_memory(__func__
);
645 if (subscript
>= 0) {
646 this->array_subscript
= subscript
;
647 this->is_subscripted
= true;
649 this->is_subscripted
= false;
652 /* For drivers that lower gl_ClipDistance to gl_ClipDistanceMESA, this
653 * class must behave specially to account for the fact that gl_ClipDistance
654 * is converted from a float[8] to a vec4[2].
656 if (ctx
->Const
.ShaderCompilerOptions
[MESA_SHADER_VERTEX
].LowerCombinedClipCullDistance
&&
657 strcmp(this->var_name
, "gl_ClipDistance") == 0) {
658 this->lowered_builtin_array_variable
= clip_distance
;
660 if (ctx
->Const
.ShaderCompilerOptions
[MESA_SHADER_VERTEX
].LowerCombinedClipCullDistance
&&
661 strcmp(this->var_name
, "gl_CullDistance") == 0) {
662 this->lowered_builtin_array_variable
= cull_distance
;
665 if (ctx
->Const
.LowerTessLevel
&&
666 (strcmp(this->var_name
, "gl_TessLevelOuter") == 0))
667 this->lowered_builtin_array_variable
= tess_level_outer
;
668 if (ctx
->Const
.LowerTessLevel
&&
669 (strcmp(this->var_name
, "gl_TessLevelInner") == 0))
670 this->lowered_builtin_array_variable
= tess_level_inner
;
675 * Determine whether two tfeedback_decl objects refer to the same variable and
676 * array index (if applicable).
679 tfeedback_decl::is_same(const tfeedback_decl
&x
, const tfeedback_decl
&y
)
681 assert(x
.is_varying() && y
.is_varying());
683 if (strcmp(x
.var_name
, y
.var_name
) != 0)
685 if (x
.is_subscripted
!= y
.is_subscripted
)
687 if (x
.is_subscripted
&& x
.array_subscript
!= y
.array_subscript
)
694 * Assign a location and stream ID for this tfeedback_decl object based on the
695 * transform feedback candidate found by find_candidate.
697 * If an error occurs, the error is reported through linker_error() and false
701 tfeedback_decl::assign_location(struct gl_context
*ctx
,
702 struct gl_shader_program
*prog
)
704 assert(this->is_varying());
706 unsigned fine_location
707 = this->matched_candidate
->toplevel_var
->data
.location
* 4
708 + this->matched_candidate
->toplevel_var
->data
.location_frac
709 + this->matched_candidate
->offset
;
710 const unsigned dmul
=
711 this->matched_candidate
->type
->without_array()->is_double() ? 2 : 1;
713 if (this->matched_candidate
->type
->is_array()) {
715 const unsigned matrix_cols
=
716 this->matched_candidate
->type
->fields
.array
->matrix_columns
;
717 const unsigned vector_elements
=
718 this->matched_candidate
->type
->fields
.array
->vector_elements
;
719 unsigned actual_array_size
;
720 switch (this->lowered_builtin_array_variable
) {
722 actual_array_size
= prog
->LastClipDistanceArraySize
;
725 actual_array_size
= prog
->LastCullDistanceArraySize
;
727 case tess_level_outer
:
728 actual_array_size
= 4;
730 case tess_level_inner
:
731 actual_array_size
= 2;
735 actual_array_size
= this->matched_candidate
->type
->array_size();
739 if (this->is_subscripted
) {
740 /* Check array bounds. */
741 if (this->array_subscript
>= actual_array_size
) {
742 linker_error(prog
, "Transform feedback varying %s has index "
743 "%i, but the array size is %u.",
744 this->orig_name
, this->array_subscript
,
748 unsigned array_elem_size
= this->lowered_builtin_array_variable
?
749 1 : vector_elements
* matrix_cols
* dmul
;
750 fine_location
+= array_elem_size
* this->array_subscript
;
753 this->size
= actual_array_size
;
755 this->vector_elements
= vector_elements
;
756 this->matrix_columns
= matrix_cols
;
757 if (this->lowered_builtin_array_variable
)
758 this->type
= GL_FLOAT
;
760 this->type
= this->matched_candidate
->type
->fields
.array
->gl_type
;
762 /* Regular variable (scalar, vector, or matrix) */
763 if (this->is_subscripted
) {
764 linker_error(prog
, "Transform feedback varying %s requested, "
765 "but %s is not an array.",
766 this->orig_name
, this->var_name
);
770 this->vector_elements
= this->matched_candidate
->type
->vector_elements
;
771 this->matrix_columns
= this->matched_candidate
->type
->matrix_columns
;
772 this->type
= this->matched_candidate
->type
->gl_type
;
774 this->location
= fine_location
/ 4;
775 this->location_frac
= fine_location
% 4;
777 /* From GL_EXT_transform_feedback:
778 * A program will fail to link if:
780 * * the total number of components to capture in any varying
781 * variable in <varyings> is greater than the constant
782 * MAX_TRANSFORM_FEEDBACK_SEPARATE_COMPONENTS_EXT and the
783 * buffer mode is SEPARATE_ATTRIBS_EXT;
785 if (prog
->TransformFeedback
.BufferMode
== GL_SEPARATE_ATTRIBS
&&
786 this->num_components() >
787 ctx
->Const
.MaxTransformFeedbackSeparateComponents
) {
788 linker_error(prog
, "Transform feedback varying %s exceeds "
789 "MAX_TRANSFORM_FEEDBACK_SEPARATE_COMPONENTS.",
794 /* Only transform feedback varyings can be assigned to non-zero streams,
795 * so assign the stream id here.
797 this->stream_id
= this->matched_candidate
->toplevel_var
->data
.stream
;
799 unsigned array_offset
= this->array_subscript
* 4 * dmul
;
800 unsigned struct_offset
= this->matched_candidate
->offset
* 4 * dmul
;
801 this->buffer
= this->matched_candidate
->toplevel_var
->data
.xfb_buffer
;
802 this->offset
= this->matched_candidate
->toplevel_var
->data
.offset
+
803 array_offset
+ struct_offset
;
810 tfeedback_decl::get_num_outputs() const
812 if (!this->is_varying()) {
815 return (this->num_components() + this->location_frac
+ 3)/4;
820 * Update gl_transform_feedback_info to reflect this tfeedback_decl.
822 * If an error occurs, the error is reported through linker_error() and false
826 tfeedback_decl::store(struct gl_context
*ctx
, struct gl_shader_program
*prog
,
827 struct gl_transform_feedback_info
*info
,
828 unsigned buffer
, unsigned buffer_index
,
829 const unsigned max_outputs
, bool *explicit_stride
,
830 bool has_xfb_qualifiers
) const
832 unsigned xfb_offset
= 0;
833 unsigned size
= this->size
;
834 /* Handle gl_SkipComponents. */
835 if (this->skip_components
) {
836 info
->Buffers
[buffer
].Stride
+= this->skip_components
;
837 size
= this->skip_components
;
841 if (this->next_buffer_separator
) {
846 if (has_xfb_qualifiers
) {
847 xfb_offset
= this->offset
/ 4;
849 xfb_offset
= info
->Buffers
[buffer
].Stride
;
851 info
->Varyings
[info
->NumVarying
].Offset
= xfb_offset
* 4;
854 unsigned location
= this->location
;
855 unsigned location_frac
= this->location_frac
;
856 unsigned num_components
= this->num_components();
857 while (num_components
> 0) {
858 unsigned output_size
= MIN2(num_components
, 4 - location_frac
);
859 assert((info
->NumOutputs
== 0 && max_outputs
== 0) ||
860 info
->NumOutputs
< max_outputs
);
862 /* From the ARB_enhanced_layouts spec:
864 * "If such a block member or variable is not written during a shader
865 * invocation, the buffer contents at the assigned offset will be
866 * undefined. Even if there are no static writes to a variable or
867 * member that is assigned a transform feedback offset, the space is
868 * still allocated in the buffer and still affects the stride."
870 if (this->is_varying_written()) {
871 info
->Outputs
[info
->NumOutputs
].ComponentOffset
= location_frac
;
872 info
->Outputs
[info
->NumOutputs
].OutputRegister
= location
;
873 info
->Outputs
[info
->NumOutputs
].NumComponents
= output_size
;
874 info
->Outputs
[info
->NumOutputs
].StreamId
= stream_id
;
875 info
->Outputs
[info
->NumOutputs
].OutputBuffer
= buffer
;
876 info
->Outputs
[info
->NumOutputs
].DstOffset
= xfb_offset
;
879 info
->Buffers
[buffer
].Stream
= this->stream_id
;
880 xfb_offset
+= output_size
;
882 num_components
-= output_size
;
888 if (explicit_stride
&& explicit_stride
[buffer
]) {
889 if (this->is_double() && info
->Buffers
[buffer
].Stride
% 2) {
890 linker_error(prog
, "invalid qualifier xfb_stride=%d must be a "
891 "multiple of 8 as its applied to a type that is or "
892 "contains a double.",
893 info
->Buffers
[buffer
].Stride
* 4);
897 if ((this->offset
/ 4) / info
->Buffers
[buffer
].Stride
!=
898 (xfb_offset
- 1) / info
->Buffers
[buffer
].Stride
) {
899 linker_error(prog
, "xfb_offset (%d) overflows xfb_stride (%d) for "
900 "buffer (%d)", xfb_offset
* 4,
901 info
->Buffers
[buffer
].Stride
* 4, buffer
);
905 info
->Buffers
[buffer
].Stride
= xfb_offset
;
908 /* From GL_EXT_transform_feedback:
909 * A program will fail to link if:
911 * * the total number of components to capture is greater than
912 * the constant MAX_TRANSFORM_FEEDBACK_INTERLEAVED_COMPONENTS_EXT
913 * and the buffer mode is INTERLEAVED_ATTRIBS_EXT.
915 * From GL_ARB_enhanced_layouts:
917 * "The resulting stride (implicit or explicit) must be less than or
918 * equal to the implementation-dependent constant
919 * gl_MaxTransformFeedbackInterleavedComponents."
921 if ((prog
->TransformFeedback
.BufferMode
== GL_INTERLEAVED_ATTRIBS
||
922 has_xfb_qualifiers
) &&
923 info
->Buffers
[buffer
].Stride
>
924 ctx
->Const
.MaxTransformFeedbackInterleavedComponents
) {
925 linker_error(prog
, "The MAX_TRANSFORM_FEEDBACK_INTERLEAVED_COMPONENTS "
926 "limit has been exceeded.");
931 info
->Varyings
[info
->NumVarying
].Name
= ralloc_strdup(prog
,
933 info
->Varyings
[info
->NumVarying
].Type
= this->type
;
934 info
->Varyings
[info
->NumVarying
].Size
= size
;
935 info
->Varyings
[info
->NumVarying
].BufferIndex
= buffer_index
;
937 info
->Buffers
[buffer
].NumVaryings
++;
943 const tfeedback_candidate
*
944 tfeedback_decl::find_candidate(gl_shader_program
*prog
,
945 hash_table
*tfeedback_candidates
)
947 const char *name
= this->var_name
;
948 switch (this->lowered_builtin_array_variable
) {
950 name
= this->var_name
;
953 name
= "gl_ClipDistanceMESA";
956 name
= "gl_CullDistanceMESA";
958 case tess_level_outer
:
959 name
= "gl_TessLevelOuterMESA";
961 case tess_level_inner
:
962 name
= "gl_TessLevelInnerMESA";
965 this->matched_candidate
= (const tfeedback_candidate
*)
966 hash_table_find(tfeedback_candidates
, name
);
967 if (!this->matched_candidate
) {
968 /* From GL_EXT_transform_feedback:
969 * A program will fail to link if:
971 * * any variable name specified in the <varyings> array is not
972 * declared as an output in the geometry shader (if present) or
973 * the vertex shader (if no geometry shader is present);
975 linker_error(prog
, "Transform feedback varying %s undeclared.",
978 return this->matched_candidate
;
983 * Parse all the transform feedback declarations that were passed to
984 * glTransformFeedbackVaryings() and store them in tfeedback_decl objects.
986 * If an error occurs, the error is reported through linker_error() and false
990 parse_tfeedback_decls(struct gl_context
*ctx
, struct gl_shader_program
*prog
,
991 const void *mem_ctx
, unsigned num_names
,
992 char **varying_names
, tfeedback_decl
*decls
)
994 for (unsigned i
= 0; i
< num_names
; ++i
) {
995 decls
[i
].init(ctx
, mem_ctx
, varying_names
[i
]);
997 if (!decls
[i
].is_varying())
1000 /* From GL_EXT_transform_feedback:
1001 * A program will fail to link if:
1003 * * any two entries in the <varyings> array specify the same varying
1006 * We interpret this to mean "any two entries in the <varyings> array
1007 * specify the same varying variable and array index", since transform
1008 * feedback of arrays would be useless otherwise.
1010 for (unsigned j
= 0; j
< i
; ++j
) {
1011 if (!decls
[j
].is_varying())
1014 if (tfeedback_decl::is_same(decls
[i
], decls
[j
])) {
1015 linker_error(prog
, "Transform feedback varying %s specified "
1016 "more than once.", varying_names
[i
]);
1026 cmp_xfb_offset(const void * x_generic
, const void * y_generic
)
1028 tfeedback_decl
*x
= (tfeedback_decl
*) x_generic
;
1029 tfeedback_decl
*y
= (tfeedback_decl
*) y_generic
;
1031 if (x
->get_buffer() != y
->get_buffer())
1032 return x
->get_buffer() - y
->get_buffer();
1033 return x
->get_offset() - y
->get_offset();
1037 * Store transform feedback location assignments into
1038 * prog->LinkedTransformFeedback based on the data stored in tfeedback_decls.
1040 * If an error occurs, the error is reported through linker_error() and false
1044 store_tfeedback_info(struct gl_context
*ctx
, struct gl_shader_program
*prog
,
1045 unsigned num_tfeedback_decls
,
1046 tfeedback_decl
*tfeedback_decls
, bool has_xfb_qualifiers
)
1048 /* Make sure MaxTransformFeedbackBuffers is less than 32 so the bitmask for
1049 * tracking the number of buffers doesn't overflow.
1051 assert(ctx
->Const
.MaxTransformFeedbackBuffers
< 32);
1053 bool separate_attribs_mode
=
1054 prog
->TransformFeedback
.BufferMode
== GL_SEPARATE_ATTRIBS
;
1056 ralloc_free(prog
->LinkedTransformFeedback
.Varyings
);
1057 ralloc_free(prog
->LinkedTransformFeedback
.Outputs
);
1059 memset(&prog
->LinkedTransformFeedback
, 0,
1060 sizeof(prog
->LinkedTransformFeedback
));
1062 /* The xfb_offset qualifier does not have to be used in increasing order
1063 * however some drivers expect to receive the list of transform feedback
1064 * declarations in order so sort it now for convenience.
1066 if (has_xfb_qualifiers
)
1067 qsort(tfeedback_decls
, num_tfeedback_decls
, sizeof(*tfeedback_decls
),
1070 prog
->LinkedTransformFeedback
.Varyings
=
1072 struct gl_transform_feedback_varying_info
,
1073 num_tfeedback_decls
);
1075 unsigned num_outputs
= 0;
1076 for (unsigned i
= 0; i
< num_tfeedback_decls
; ++i
) {
1077 if (tfeedback_decls
[i
].is_varying_written())
1078 num_outputs
+= tfeedback_decls
[i
].get_num_outputs();
1081 prog
->LinkedTransformFeedback
.Outputs
=
1083 struct gl_transform_feedback_output
,
1086 unsigned num_buffers
= 0;
1087 unsigned buffers
= 0;
1089 if (!has_xfb_qualifiers
&& separate_attribs_mode
) {
1090 /* GL_SEPARATE_ATTRIBS */
1091 for (unsigned i
= 0; i
< num_tfeedback_decls
; ++i
) {
1092 if (!tfeedback_decls
[i
].store(ctx
, prog
, &prog
->LinkedTransformFeedback
,
1093 num_buffers
, num_buffers
, num_outputs
,
1094 NULL
, has_xfb_qualifiers
))
1097 buffers
|= 1 << num_buffers
;
1102 /* GL_INVERLEAVED_ATTRIBS */
1103 int buffer_stream_id
= -1;
1105 num_tfeedback_decls
? tfeedback_decls
[0].get_buffer() : 0;
1106 bool explicit_stride
[MAX_FEEDBACK_BUFFERS
] = { false };
1108 /* Apply any xfb_stride global qualifiers */
1109 if (has_xfb_qualifiers
) {
1110 for (unsigned j
= 0; j
< MAX_FEEDBACK_BUFFERS
; j
++) {
1111 if (prog
->TransformFeedback
.BufferStride
[j
]) {
1113 explicit_stride
[j
] = true;
1114 prog
->LinkedTransformFeedback
.Buffers
[j
].Stride
=
1115 prog
->TransformFeedback
.BufferStride
[j
] / 4;
1120 for (unsigned i
= 0; i
< num_tfeedback_decls
; ++i
) {
1121 if (has_xfb_qualifiers
&&
1122 buffer
!= tfeedback_decls
[i
].get_buffer()) {
1123 /* we have moved to the next buffer so reset stream id */
1124 buffer_stream_id
= -1;
1128 if (tfeedback_decls
[i
].is_next_buffer_separator()) {
1129 if (!tfeedback_decls
[i
].store(ctx
, prog
,
1130 &prog
->LinkedTransformFeedback
,
1131 buffer
, num_buffers
, num_outputs
,
1132 explicit_stride
, has_xfb_qualifiers
))
1135 buffer_stream_id
= -1;
1137 } else if (tfeedback_decls
[i
].is_varying()) {
1138 if (buffer_stream_id
== -1) {
1139 /* First varying writing to this buffer: remember its stream */
1140 buffer_stream_id
= (int) tfeedback_decls
[i
].get_stream_id();
1141 } else if (buffer_stream_id
!=
1142 (int) tfeedback_decls
[i
].get_stream_id()) {
1143 /* Varying writes to the same buffer from a different stream */
1145 "Transform feedback can't capture varyings belonging "
1146 "to different vertex streams in a single buffer. "
1147 "Varying %s writes to buffer from stream %u, other "
1148 "varyings in the same buffer write from stream %u.",
1149 tfeedback_decls
[i
].name(),
1150 tfeedback_decls
[i
].get_stream_id(),
1156 if (has_xfb_qualifiers
) {
1157 buffer
= tfeedback_decls
[i
].get_buffer();
1159 buffer
= num_buffers
;
1161 buffers
|= 1 << buffer
;
1163 if (!tfeedback_decls
[i
].store(ctx
, prog
,
1164 &prog
->LinkedTransformFeedback
,
1165 buffer
, num_buffers
, num_outputs
,
1166 explicit_stride
, has_xfb_qualifiers
))
1171 assert(prog
->LinkedTransformFeedback
.NumOutputs
== num_outputs
);
1173 prog
->LinkedTransformFeedback
.ActiveBuffers
= buffers
;
1180 * Data structure recording the relationship between outputs of one shader
1181 * stage (the "producer") and inputs of another (the "consumer").
1183 class varying_matches
1186 varying_matches(bool disable_varying_packing
, bool xfb_enabled
,
1187 gl_shader_stage producer_stage
,
1188 gl_shader_stage consumer_stage
);
1190 void record(ir_variable
*producer_var
, ir_variable
*consumer_var
);
1191 unsigned assign_locations(struct gl_shader_program
*prog
,
1192 uint64_t reserved_slots
);
1193 void store_locations() const;
1196 bool is_varying_packing_safe(const glsl_type
*type
,
1197 const ir_variable
*var
);
1200 * If true, this driver disables varying packing, so all varyings need to
1201 * be aligned on slot boundaries, and take up a number of slots equal to
1202 * their number of matrix columns times their array size.
1204 * Packing may also be disabled because our current packing method is not
1205 * safe in SSO or versions of OpenGL where interpolation qualifiers are not
1206 * guaranteed to match across stages.
1208 const bool disable_varying_packing
;
1211 * If true, this driver has transform feedback enabled. The transform
1212 * feedback code requires at least some packing be done even when varying
1213 * packing is disabled, fortunately where transform feedback requires
1214 * packing it's safe to override the disabled setting. See
1215 * is_varying_packing_safe().
1217 const bool xfb_enabled
;
1220 * Enum representing the order in which varyings are packed within a
1223 * Currently we pack vec4's first, then vec2's, then scalar values, then
1224 * vec3's. This order ensures that the only vectors that are at risk of
1225 * having to be "double parked" (split between two adjacent varying slots)
1228 enum packing_order_enum
{
1231 PACKING_ORDER_SCALAR
,
1235 static unsigned compute_packing_class(const ir_variable
*var
);
1236 static packing_order_enum
compute_packing_order(const ir_variable
*var
);
1237 static int match_comparator(const void *x_generic
, const void *y_generic
);
1238 static int xfb_comparator(const void *x_generic
, const void *y_generic
);
1241 * Structure recording the relationship between a single producer output
1242 * and a single consumer input.
1246 * Packing class for this varying, computed by compute_packing_class().
1248 unsigned packing_class
;
1251 * Packing order for this varying, computed by compute_packing_order().
1253 packing_order_enum packing_order
;
1254 unsigned num_components
;
1257 * The output variable in the producer stage.
1259 ir_variable
*producer_var
;
1262 * The input variable in the consumer stage.
1264 ir_variable
*consumer_var
;
1267 * The location which has been assigned for this varying. This is
1268 * expressed in multiples of a float, with the first generic varying
1269 * (i.e. the one referred to by VARYING_SLOT_VAR0) represented by the
1272 unsigned generic_location
;
1276 * The number of elements in the \c matches array that are currently in
1279 unsigned num_matches
;
1282 * The number of elements that were set aside for the \c matches array when
1285 unsigned matches_capacity
;
1287 gl_shader_stage producer_stage
;
1288 gl_shader_stage consumer_stage
;
1291 } /* anonymous namespace */
1293 varying_matches::varying_matches(bool disable_varying_packing
,
1295 gl_shader_stage producer_stage
,
1296 gl_shader_stage consumer_stage
)
1297 : disable_varying_packing(disable_varying_packing
),
1298 xfb_enabled(xfb_enabled
),
1299 producer_stage(producer_stage
),
1300 consumer_stage(consumer_stage
)
1302 /* Note: this initial capacity is rather arbitrarily chosen to be large
1303 * enough for many cases without wasting an unreasonable amount of space.
1304 * varying_matches::record() will resize the array if there are more than
1305 * this number of varyings.
1307 this->matches_capacity
= 8;
1308 this->matches
= (match
*)
1309 malloc(sizeof(*this->matches
) * this->matches_capacity
);
1310 this->num_matches
= 0;
1314 varying_matches::~varying_matches()
1316 free(this->matches
);
1321 * Packing is always safe on individual arrays, structures, and matrices. It
1322 * is also safe if the varying is only used for transform feedback.
1325 varying_matches::is_varying_packing_safe(const glsl_type
*type
,
1326 const ir_variable
*var
)
1328 if (consumer_stage
== MESA_SHADER_TESS_EVAL
||
1329 consumer_stage
== MESA_SHADER_TESS_CTRL
||
1330 producer_stage
== MESA_SHADER_TESS_CTRL
)
1333 return xfb_enabled
&& (type
->is_array() || type
->is_record() ||
1334 type
->is_matrix() || var
->data
.is_xfb_only
);
1339 * Record the given producer/consumer variable pair in the list of variables
1340 * that should later be assigned locations.
1342 * It is permissible for \c consumer_var to be NULL (this happens if a
1343 * variable is output by the producer and consumed by transform feedback, but
1344 * not consumed by the consumer).
1346 * If \c producer_var has already been paired up with a consumer_var, or
1347 * producer_var is part of fixed pipeline functionality (and hence already has
1348 * a location assigned), this function has no effect.
1350 * Note: as a side effect this function may change the interpolation type of
1351 * \c producer_var, but only when the change couldn't possibly affect
1355 varying_matches::record(ir_variable
*producer_var
, ir_variable
*consumer_var
)
1357 assert(producer_var
!= NULL
|| consumer_var
!= NULL
);
1359 if ((producer_var
&& (!producer_var
->data
.is_unmatched_generic_inout
||
1360 producer_var
->data
.explicit_location
)) ||
1361 (consumer_var
&& (!consumer_var
->data
.is_unmatched_generic_inout
||
1362 consumer_var
->data
.explicit_location
))) {
1363 /* Either a location already exists for this variable (since it is part
1364 * of fixed functionality), or it has already been recorded as part of a
1370 bool needs_flat_qualifier
= consumer_var
== NULL
&&
1371 (producer_var
->type
->contains_integer() ||
1372 producer_var
->type
->contains_double());
1374 if (needs_flat_qualifier
||
1375 (consumer_stage
!= -1 && consumer_stage
!= MESA_SHADER_FRAGMENT
)) {
1376 /* Since this varying is not being consumed by the fragment shader, its
1377 * interpolation type varying cannot possibly affect rendering.
1378 * Also, this variable is non-flat and is (or contains) an integer
1380 * If the consumer stage is unknown, don't modify the interpolation
1381 * type as it could affect rendering later with separate shaders.
1383 * lower_packed_varyings requires all integer varyings to flat,
1384 * regardless of where they appear. We can trivially satisfy that
1385 * requirement by changing the interpolation type to flat here.
1388 producer_var
->data
.centroid
= false;
1389 producer_var
->data
.sample
= false;
1390 producer_var
->data
.interpolation
= INTERP_QUALIFIER_FLAT
;
1394 consumer_var
->data
.centroid
= false;
1395 consumer_var
->data
.sample
= false;
1396 consumer_var
->data
.interpolation
= INTERP_QUALIFIER_FLAT
;
1400 if (this->num_matches
== this->matches_capacity
) {
1401 this->matches_capacity
*= 2;
1402 this->matches
= (match
*)
1403 realloc(this->matches
,
1404 sizeof(*this->matches
) * this->matches_capacity
);
1407 const ir_variable
*const var
= (producer_var
!= NULL
)
1408 ? producer_var
: consumer_var
;
1409 const gl_shader_stage stage
= (producer_var
!= NULL
)
1410 ? producer_stage
: consumer_stage
;
1411 const glsl_type
*type
= get_varying_type(var
, stage
);
1413 this->matches
[this->num_matches
].packing_class
1414 = this->compute_packing_class(var
);
1415 this->matches
[this->num_matches
].packing_order
1416 = this->compute_packing_order(var
);
1417 if (this->disable_varying_packing
&& !is_varying_packing_safe(type
, var
)) {
1418 unsigned slots
= type
->count_attribute_slots(false);
1419 this->matches
[this->num_matches
].num_components
= slots
* 4;
1421 this->matches
[this->num_matches
].num_components
1422 = type
->component_slots();
1424 this->matches
[this->num_matches
].producer_var
= producer_var
;
1425 this->matches
[this->num_matches
].consumer_var
= consumer_var
;
1426 this->num_matches
++;
1428 producer_var
->data
.is_unmatched_generic_inout
= 0;
1430 consumer_var
->data
.is_unmatched_generic_inout
= 0;
1435 * Choose locations for all of the variable matches that were previously
1436 * passed to varying_matches::record().
1439 varying_matches::assign_locations(struct gl_shader_program
*prog
,
1440 uint64_t reserved_slots
)
1442 /* If packing has been disabled then we cannot safely sort the varyings by
1443 * class as it may mean we are using a version of OpenGL where
1444 * interpolation qualifiers are not guaranteed to be matching across
1445 * shaders, sorting in this case could result in mismatching shader
1447 * When packing is disabled the sort orders varyings used by transform
1448 * feedback first, but also depends on *undefined behaviour* of qsort to
1449 * reverse the order of the varyings. See: xfb_comparator().
1451 if (!this->disable_varying_packing
) {
1452 /* Sort varying matches into an order that makes them easy to pack. */
1453 qsort(this->matches
, this->num_matches
, sizeof(*this->matches
),
1454 &varying_matches::match_comparator
);
1456 /* Only sort varyings that are only used by transform feedback. */
1457 qsort(this->matches
, this->num_matches
, sizeof(*this->matches
),
1458 &varying_matches::xfb_comparator
);
1461 unsigned generic_location
= 0;
1462 unsigned generic_patch_location
= MAX_VARYING
*4;
1463 bool previous_var_xfb_only
= false;
1465 for (unsigned i
= 0; i
< this->num_matches
; i
++) {
1466 unsigned *location
= &generic_location
;
1468 const ir_variable
*var
;
1469 const glsl_type
*type
;
1470 bool is_vertex_input
= false;
1471 if (matches
[i
].consumer_var
) {
1472 var
= matches
[i
].consumer_var
;
1473 type
= get_varying_type(var
, consumer_stage
);
1474 if (consumer_stage
== MESA_SHADER_VERTEX
)
1475 is_vertex_input
= true;
1477 var
= matches
[i
].producer_var
;
1478 type
= get_varying_type(var
, producer_stage
);
1481 if (var
->data
.patch
)
1482 location
= &generic_patch_location
;
1484 /* Advance to the next slot if this varying has a different packing
1485 * class than the previous one, and we're not already on a slot
1488 * Also advance to the next slot if packing is disabled. This makes sure
1489 * we don't assign varyings the same locations which is possible
1490 * because we still pack individual arrays, records and matrices even
1491 * when packing is disabled. Note we don't advance to the next slot if
1492 * we can pack varyings together that are only used for transform
1495 if ((this->disable_varying_packing
&&
1496 !(previous_var_xfb_only
&& var
->data
.is_xfb_only
)) ||
1497 (i
> 0 && this->matches
[i
- 1].packing_class
1498 != this->matches
[i
].packing_class
)) {
1499 *location
= ALIGN(*location
, 4);
1502 previous_var_xfb_only
= var
->data
.is_xfb_only
;
1504 unsigned num_elements
= type
->count_attribute_slots(is_vertex_input
);
1506 if (this->disable_varying_packing
&&
1507 !is_varying_packing_safe(type
, var
))
1510 slot_end
= type
->without_array()->vector_elements
;
1511 slot_end
+= *location
- 1;
1513 /* FIXME: We could be smarter in the below code and loop back over
1514 * trying to fill any locations that we skipped because we couldn't pack
1515 * the varying between an explicit location. For now just let the user
1516 * hit the linking error if we run out of room and suggest they use
1517 * explicit locations.
1519 for (unsigned j
= 0; j
< num_elements
; j
++) {
1520 while ((slot_end
< MAX_VARYING
* 4u) &&
1521 ((reserved_slots
& (UINT64_C(1) << *location
/ 4u) ||
1522 (reserved_slots
& (UINT64_C(1) << slot_end
/ 4u))))) {
1524 *location
= ALIGN(*location
+ 1, 4);
1525 slot_end
= *location
;
1527 /* reset the counter and try again */
1531 /* Increase the slot to make sure there is enough room for next
1534 if (this->disable_varying_packing
&&
1535 !is_varying_packing_safe(type
, var
))
1538 slot_end
+= type
->without_array()->vector_elements
;
1541 if (!var
->data
.patch
&& *location
>= MAX_VARYING
* 4u) {
1542 linker_error(prog
, "insufficient contiguous locations available for "
1543 "%s it is possible an array or struct could not be "
1544 "packed between varyings with explicit locations. Try "
1545 "using an explicit location for arrays and structs.",
1549 this->matches
[i
].generic_location
= *location
;
1551 *location
+= this->matches
[i
].num_components
;
1554 return (generic_location
+ 3) / 4;
1559 * Update the producer and consumer shaders to reflect the locations
1560 * assignments that were made by varying_matches::assign_locations().
1563 varying_matches::store_locations() const
1565 for (unsigned i
= 0; i
< this->num_matches
; i
++) {
1566 ir_variable
*producer_var
= this->matches
[i
].producer_var
;
1567 ir_variable
*consumer_var
= this->matches
[i
].consumer_var
;
1568 unsigned generic_location
= this->matches
[i
].generic_location
;
1569 unsigned slot
= generic_location
/ 4;
1570 unsigned offset
= generic_location
% 4;
1573 producer_var
->data
.location
= VARYING_SLOT_VAR0
+ slot
;
1574 producer_var
->data
.location_frac
= offset
;
1578 assert(consumer_var
->data
.location
== -1);
1579 consumer_var
->data
.location
= VARYING_SLOT_VAR0
+ slot
;
1580 consumer_var
->data
.location_frac
= offset
;
1587 * Compute the "packing class" of the given varying. This is an unsigned
1588 * integer with the property that two variables in the same packing class can
1589 * be safely backed into the same vec4.
1592 varying_matches::compute_packing_class(const ir_variable
*var
)
1594 /* Without help from the back-end, there is no way to pack together
1595 * variables with different interpolation types, because
1596 * lower_packed_varyings must choose exactly one interpolation type for
1597 * each packed varying it creates.
1599 * However, we can safely pack together floats, ints, and uints, because:
1601 * - varyings of base type "int" and "uint" must use the "flat"
1602 * interpolation type, which can only occur in GLSL 1.30 and above.
1604 * - On platforms that support GLSL 1.30 and above, lower_packed_varyings
1605 * can store flat floats as ints without losing any information (using
1606 * the ir_unop_bitcast_* opcodes).
1608 * Therefore, the packing class depends only on the interpolation type.
1610 unsigned packing_class
= var
->data
.centroid
| (var
->data
.sample
<< 1) |
1611 (var
->data
.patch
<< 2);
1613 packing_class
+= var
->data
.interpolation
;
1614 return packing_class
;
1619 * Compute the "packing order" of the given varying. This is a sort key we
1620 * use to determine when to attempt to pack the given varying relative to
1621 * other varyings in the same packing class.
1623 varying_matches::packing_order_enum
1624 varying_matches::compute_packing_order(const ir_variable
*var
)
1626 const glsl_type
*element_type
= var
->type
;
1628 while (element_type
->base_type
== GLSL_TYPE_ARRAY
) {
1629 element_type
= element_type
->fields
.array
;
1632 switch (element_type
->component_slots() % 4) {
1633 case 1: return PACKING_ORDER_SCALAR
;
1634 case 2: return PACKING_ORDER_VEC2
;
1635 case 3: return PACKING_ORDER_VEC3
;
1636 case 0: return PACKING_ORDER_VEC4
;
1638 assert(!"Unexpected value of vector_elements");
1639 return PACKING_ORDER_VEC4
;
1645 * Comparison function passed to qsort() to sort varyings by packing_class and
1646 * then by packing_order.
1649 varying_matches::match_comparator(const void *x_generic
, const void *y_generic
)
1651 const match
*x
= (const match
*) x_generic
;
1652 const match
*y
= (const match
*) y_generic
;
1654 if (x
->packing_class
!= y
->packing_class
)
1655 return x
->packing_class
- y
->packing_class
;
1656 return x
->packing_order
- y
->packing_order
;
1661 * Comparison function passed to qsort() to sort varyings used only by
1662 * transform feedback when packing of other varyings is disabled.
1665 varying_matches::xfb_comparator(const void *x_generic
, const void *y_generic
)
1667 const match
*x
= (const match
*) x_generic
;
1669 if (x
->producer_var
!= NULL
&& x
->producer_var
->data
.is_xfb_only
)
1670 return match_comparator(x_generic
, y_generic
);
1672 /* FIXME: When the comparator returns 0 it means the elements being
1673 * compared are equivalent. However the qsort documentation says:
1675 * "The order of equivalent elements is undefined."
1677 * In practice the sort ends up reversing the order of the varyings which
1678 * means locations are also assigned in this reversed order and happens to
1679 * be what we want. This is also whats happening in
1680 * varying_matches::match_comparator().
1687 * Is the given variable a varying variable to be counted against the
1688 * limit in ctx->Const.MaxVarying?
1689 * This includes variables such as texcoords, colors and generic
1690 * varyings, but excludes variables such as gl_FrontFacing and gl_FragCoord.
1693 var_counts_against_varying_limit(gl_shader_stage stage
, const ir_variable
*var
)
1695 /* Only fragment shaders will take a varying variable as an input */
1696 if (stage
== MESA_SHADER_FRAGMENT
&&
1697 var
->data
.mode
== ir_var_shader_in
) {
1698 switch (var
->data
.location
) {
1699 case VARYING_SLOT_POS
:
1700 case VARYING_SLOT_FACE
:
1701 case VARYING_SLOT_PNTC
:
1712 * Visitor class that generates tfeedback_candidate structs describing all
1713 * possible targets of transform feedback.
1715 * tfeedback_candidate structs are stored in the hash table
1716 * tfeedback_candidates, which is passed to the constructor. This hash table
1717 * maps varying names to instances of the tfeedback_candidate struct.
1719 class tfeedback_candidate_generator
: public program_resource_visitor
1722 tfeedback_candidate_generator(void *mem_ctx
,
1723 hash_table
*tfeedback_candidates
)
1725 tfeedback_candidates(tfeedback_candidates
),
1731 void process(ir_variable
*var
)
1733 /* All named varying interface blocks should be flattened by now */
1734 assert(!var
->is_interface_instance());
1736 this->toplevel_var
= var
;
1737 this->varying_floats
= 0;
1738 program_resource_visitor::process(var
);
1742 virtual void visit_field(const glsl_type
*type
, const char *name
,
1745 assert(!type
->without_array()->is_record());
1746 assert(!type
->without_array()->is_interface());
1750 tfeedback_candidate
*candidate
1751 = rzalloc(this->mem_ctx
, tfeedback_candidate
);
1752 candidate
->toplevel_var
= this->toplevel_var
;
1753 candidate
->type
= type
;
1754 candidate
->offset
= this->varying_floats
;
1755 hash_table_insert(this->tfeedback_candidates
, candidate
,
1756 ralloc_strdup(this->mem_ctx
, name
));
1757 this->varying_floats
+= type
->component_slots();
1761 * Memory context used to allocate hash table keys and values.
1763 void * const mem_ctx
;
1766 * Hash table in which tfeedback_candidate objects should be stored.
1768 hash_table
* const tfeedback_candidates
;
1771 * Pointer to the toplevel variable that is being traversed.
1773 ir_variable
*toplevel_var
;
1776 * Total number of varying floats that have been visited so far. This is
1777 * used to determine the offset to each varying within the toplevel
1780 unsigned varying_floats
;
1787 populate_consumer_input_sets(void *mem_ctx
, exec_list
*ir
,
1788 hash_table
*consumer_inputs
,
1789 hash_table
*consumer_interface_inputs
,
1790 ir_variable
*consumer_inputs_with_locations
[VARYING_SLOT_TESS_MAX
])
1792 memset(consumer_inputs_with_locations
,
1794 sizeof(consumer_inputs_with_locations
[0]) * VARYING_SLOT_TESS_MAX
);
1796 foreach_in_list(ir_instruction
, node
, ir
) {
1797 ir_variable
*const input_var
= node
->as_variable();
1799 if (input_var
!= NULL
&& input_var
->data
.mode
== ir_var_shader_in
) {
1800 /* All interface blocks should have been lowered by this point */
1801 assert(!input_var
->type
->is_interface());
1803 if (input_var
->data
.explicit_location
) {
1804 /* assign_varying_locations only cares about finding the
1805 * ir_variable at the start of a contiguous location block.
1807 * - For !producer, consumer_inputs_with_locations isn't used.
1809 * - For !consumer, consumer_inputs_with_locations is empty.
1811 * For consumer && producer, if you were trying to set some
1812 * ir_variable to the middle of a location block on the other side
1813 * of producer/consumer, cross_validate_outputs_to_inputs() should
1814 * be link-erroring due to either type mismatch or location
1815 * overlaps. If the variables do match up, then they've got a
1816 * matching data.location and you only looked at
1817 * consumer_inputs_with_locations[var->data.location], not any
1818 * following entries for the array/structure.
1820 consumer_inputs_with_locations
[input_var
->data
.location
] =
1822 } else if (input_var
->get_interface_type() != NULL
) {
1823 char *const iface_field_name
=
1824 ralloc_asprintf(mem_ctx
, "%s.%s",
1825 input_var
->get_interface_type()->without_array()->name
,
1827 hash_table_insert(consumer_interface_inputs
, input_var
,
1830 hash_table_insert(consumer_inputs
, input_var
,
1831 ralloc_strdup(mem_ctx
, input_var
->name
));
1838 * Find a variable from the consumer that "matches" the specified variable
1840 * This function only finds inputs with names that match. There is no
1841 * validation (here) that the types, etc. are compatible.
1844 get_matching_input(void *mem_ctx
,
1845 const ir_variable
*output_var
,
1846 hash_table
*consumer_inputs
,
1847 hash_table
*consumer_interface_inputs
,
1848 ir_variable
*consumer_inputs_with_locations
[VARYING_SLOT_TESS_MAX
])
1850 ir_variable
*input_var
;
1852 if (output_var
->data
.explicit_location
) {
1853 input_var
= consumer_inputs_with_locations
[output_var
->data
.location
];
1854 } else if (output_var
->get_interface_type() != NULL
) {
1855 char *const iface_field_name
=
1856 ralloc_asprintf(mem_ctx
, "%s.%s",
1857 output_var
->get_interface_type()->without_array()->name
,
1860 (ir_variable
*) hash_table_find(consumer_interface_inputs
,
1864 (ir_variable
*) hash_table_find(consumer_inputs
, output_var
->name
);
1867 return (input_var
== NULL
|| input_var
->data
.mode
!= ir_var_shader_in
)
1874 io_variable_cmp(const void *_a
, const void *_b
)
1876 const ir_variable
*const a
= *(const ir_variable
**) _a
;
1877 const ir_variable
*const b
= *(const ir_variable
**) _b
;
1879 if (a
->data
.explicit_location
&& b
->data
.explicit_location
)
1880 return b
->data
.location
- a
->data
.location
;
1882 if (a
->data
.explicit_location
&& !b
->data
.explicit_location
)
1885 if (!a
->data
.explicit_location
&& b
->data
.explicit_location
)
1888 return -strcmp(a
->name
, b
->name
);
1892 * Sort the shader IO variables into canonical order
1895 canonicalize_shader_io(exec_list
*ir
, enum ir_variable_mode io_mode
)
1897 ir_variable
*var_table
[MAX_PROGRAM_OUTPUTS
* 4];
1898 unsigned num_variables
= 0;
1900 foreach_in_list(ir_instruction
, node
, ir
) {
1901 ir_variable
*const var
= node
->as_variable();
1903 if (var
== NULL
|| var
->data
.mode
!= io_mode
)
1906 /* If we have already encountered more I/O variables that could
1907 * successfully link, bail.
1909 if (num_variables
== ARRAY_SIZE(var_table
))
1912 var_table
[num_variables
++] = var
;
1915 if (num_variables
== 0)
1918 /* Sort the list in reverse order (io_variable_cmp handles this). Later
1919 * we're going to push the variables on to the IR list as a stack, so we
1920 * want the last variable (in canonical order) to be first in the list.
1922 qsort(var_table
, num_variables
, sizeof(var_table
[0]), io_variable_cmp
);
1924 /* Remove the variable from it's current location in the IR, and put it at
1927 for (unsigned i
= 0; i
< num_variables
; i
++) {
1928 var_table
[i
]->remove();
1929 ir
->push_head(var_table
[i
]);
1934 * Generate a bitfield map of the explicit locations for shader varyings.
1936 * Note: For Tessellation shaders we are sitting right on the limits of the
1937 * 64 bit map. Per-vertex and per-patch both have separate location domains
1938 * with a max of MAX_VARYING.
1941 reserved_varying_slot(struct gl_shader
*stage
, ir_variable_mode io_mode
)
1943 assert(io_mode
== ir_var_shader_in
|| io_mode
== ir_var_shader_out
);
1944 /* Avoid an overflow of the returned value */
1945 assert(MAX_VARYINGS_INCL_PATCH
<= 64);
1953 foreach_in_list(ir_instruction
, node
, stage
->ir
) {
1954 ir_variable
*const var
= node
->as_variable();
1956 if (var
== NULL
|| var
->data
.mode
!= io_mode
||
1957 !var
->data
.explicit_location
||
1958 var
->data
.location
< VARYING_SLOT_VAR0
)
1961 var_slot
= var
->data
.location
- VARYING_SLOT_VAR0
;
1963 unsigned num_elements
= get_varying_type(var
, stage
->Stage
)
1964 ->count_attribute_slots(stage
->Stage
== MESA_SHADER_VERTEX
);
1965 for (unsigned i
= 0; i
< num_elements
; i
++) {
1966 if (var_slot
>= 0 && var_slot
< MAX_VARYINGS_INCL_PATCH
)
1967 slots
|= UINT64_C(1) << var_slot
;
1977 * Assign locations for all variables that are produced in one pipeline stage
1978 * (the "producer") and consumed in the next stage (the "consumer").
1980 * Variables produced by the producer may also be consumed by transform
1983 * \param num_tfeedback_decls is the number of declarations indicating
1984 * variables that may be consumed by transform feedback.
1986 * \param tfeedback_decls is a pointer to an array of tfeedback_decl objects
1987 * representing the result of parsing the strings passed to
1988 * glTransformFeedbackVaryings(). assign_location() will be called for
1989 * each of these objects that matches one of the outputs of the
1992 * When num_tfeedback_decls is nonzero, it is permissible for the consumer to
1993 * be NULL. In this case, varying locations are assigned solely based on the
1994 * requirements of transform feedback.
1997 assign_varying_locations(struct gl_context
*ctx
,
1999 struct gl_shader_program
*prog
,
2000 gl_shader
*producer
, gl_shader
*consumer
,
2001 unsigned num_tfeedback_decls
,
2002 tfeedback_decl
*tfeedback_decls
)
2004 /* Tessellation shaders treat inputs and outputs as shared memory and can
2005 * access inputs and outputs of other invocations.
2006 * Therefore, they can't be lowered to temps easily (and definitely not
2009 bool unpackable_tess
=
2010 (consumer
&& consumer
->Stage
== MESA_SHADER_TESS_EVAL
) ||
2011 (consumer
&& consumer
->Stage
== MESA_SHADER_TESS_CTRL
) ||
2012 (producer
&& producer
->Stage
== MESA_SHADER_TESS_CTRL
);
2014 /* Transform feedback code assumes varying arrays are packed, so if the
2015 * driver has disabled varying packing, make sure to at least enable
2016 * packing required by transform feedback.
2019 ctx
->Extensions
.EXT_transform_feedback
&& !unpackable_tess
;
2021 /* Disable varying packing for GL 4.4+ as there is no guarantee
2022 * that interpolation qualifiers will match between shaders in these
2023 * versions. We also disable packing on outward facing interfaces for
2024 * SSO because in ES we need to retain the unpacked varying information
2025 * for draw time validation. For desktop GL we could allow packing for
2026 * versions < 4.4 but it's just safer not to do packing.
2028 * Packing is still enabled on individual arrays, structs, and matrices as
2029 * these are required by the transform feedback code and it is still safe
2030 * to do so. We also enable packing when a varying is only used for
2031 * transform feedback and its not a SSO.
2033 * Varying packing currently only packs together varyings with matching
2034 * interpolation qualifiers as the backends assume all packed components
2035 * are to be processed in the same way. Therefore we cannot do packing in
2036 * these versions of GL without the risk of mismatching interfaces.
2038 * From Section 4.5 (Interpolation Qualifiers) of the GLSL 4.30 spec:
2040 * "The type and presence of interpolation qualifiers of variables with
2041 * the same name declared in all linked shaders for the same cross-stage
2042 * interface must match, otherwise the link command will fail.
2044 * When comparing an output from one stage to an input of a subsequent
2045 * stage, the input and output don't match if their interpolation
2046 * qualifiers (or lack thereof) are not the same."
2048 * This text was also in at least revison 7 of the 4.40 spec but is no
2049 * longer in revision 9 and not in the 4.50 spec.
2051 bool disable_varying_packing
=
2052 ctx
->Const
.DisableVaryingPacking
|| unpackable_tess
;
2053 if ((ctx
->API
== API_OPENGL_CORE
&& ctx
->Version
>= 44) ||
2054 (prog
->SeparateShader
&& (producer
== NULL
|| consumer
== NULL
)))
2055 disable_varying_packing
= true;
2057 varying_matches
matches(disable_varying_packing
, xfb_enabled
,
2058 producer
? producer
->Stage
: (gl_shader_stage
)-1,
2059 consumer
? consumer
->Stage
: (gl_shader_stage
)-1);
2060 hash_table
*tfeedback_candidates
2061 = hash_table_ctor(0, hash_table_string_hash
, hash_table_string_compare
);
2062 hash_table
*consumer_inputs
2063 = hash_table_ctor(0, hash_table_string_hash
, hash_table_string_compare
);
2064 hash_table
*consumer_interface_inputs
2065 = hash_table_ctor(0, hash_table_string_hash
, hash_table_string_compare
);
2066 ir_variable
*consumer_inputs_with_locations
[VARYING_SLOT_TESS_MAX
] = {
2070 unsigned consumer_vertices
= 0;
2071 if (consumer
&& consumer
->Stage
== MESA_SHADER_GEOMETRY
)
2072 consumer_vertices
= prog
->Geom
.VerticesIn
;
2074 /* Operate in a total of four passes.
2076 * 1. Sort inputs / outputs into a canonical order. This is necessary so
2077 * that inputs / outputs of separable shaders will be assigned
2078 * predictable locations regardless of the order in which declarations
2079 * appeared in the shader source.
2081 * 2. Assign locations for any matching inputs and outputs.
2083 * 3. Mark output variables in the producer that do not have locations as
2084 * not being outputs. This lets the optimizer eliminate them.
2086 * 4. Mark input variables in the consumer that do not have locations as
2087 * not being inputs. This lets the optimizer eliminate them.
2090 canonicalize_shader_io(consumer
->ir
, ir_var_shader_in
);
2093 canonicalize_shader_io(producer
->ir
, ir_var_shader_out
);
2096 linker::populate_consumer_input_sets(mem_ctx
, consumer
->ir
,
2098 consumer_interface_inputs
,
2099 consumer_inputs_with_locations
);
2102 foreach_in_list(ir_instruction
, node
, producer
->ir
) {
2103 ir_variable
*const output_var
= node
->as_variable();
2105 if (output_var
== NULL
|| output_var
->data
.mode
!= ir_var_shader_out
)
2108 /* Only geometry shaders can use non-zero streams */
2109 assert(output_var
->data
.stream
== 0 ||
2110 (output_var
->data
.stream
< MAX_VERTEX_STREAMS
&&
2111 producer
->Stage
== MESA_SHADER_GEOMETRY
));
2113 if (num_tfeedback_decls
> 0) {
2114 tfeedback_candidate_generator
g(mem_ctx
, tfeedback_candidates
);
2115 g
.process(output_var
);
2118 ir_variable
*const input_var
=
2119 linker::get_matching_input(mem_ctx
, output_var
, consumer_inputs
,
2120 consumer_interface_inputs
,
2121 consumer_inputs_with_locations
);
2123 /* If a matching input variable was found, add this output (and the
2124 * input) to the set. If this is a separable program and there is no
2125 * consumer stage, add the output.
2127 * Always add TCS outputs. They are shared by all invocations
2128 * within a patch and can be used as shared memory.
2130 if (input_var
|| (prog
->SeparateShader
&& consumer
== NULL
) ||
2131 producer
->Type
== GL_TESS_CONTROL_SHADER
) {
2132 matches
.record(output_var
, input_var
);
2135 /* Only stream 0 outputs can be consumed in the next stage */
2136 if (input_var
&& output_var
->data
.stream
!= 0) {
2137 linker_error(prog
, "output %s is assigned to stream=%d but "
2138 "is linked to an input, which requires stream=0",
2139 output_var
->name
, output_var
->data
.stream
);
2144 /* If there's no producer stage, then this must be a separable program.
2145 * For example, we may have a program that has just a fragment shader.
2146 * Later this program will be used with some arbitrary vertex (or
2147 * geometry) shader program. This means that locations must be assigned
2148 * for all the inputs.
2150 foreach_in_list(ir_instruction
, node
, consumer
->ir
) {
2151 ir_variable
*const input_var
= node
->as_variable();
2153 if (input_var
== NULL
|| input_var
->data
.mode
!= ir_var_shader_in
)
2156 matches
.record(NULL
, input_var
);
2160 for (unsigned i
= 0; i
< num_tfeedback_decls
; ++i
) {
2161 if (!tfeedback_decls
[i
].is_varying())
2164 const tfeedback_candidate
*matched_candidate
2165 = tfeedback_decls
[i
].find_candidate(prog
, tfeedback_candidates
);
2167 if (matched_candidate
== NULL
) {
2168 hash_table_dtor(tfeedback_candidates
);
2169 hash_table_dtor(consumer_inputs
);
2170 hash_table_dtor(consumer_interface_inputs
);
2174 if (matched_candidate
->toplevel_var
->data
.is_unmatched_generic_inout
) {
2175 matched_candidate
->toplevel_var
->data
.is_xfb_only
= 1;
2176 matches
.record(matched_candidate
->toplevel_var
, NULL
);
2180 const uint64_t reserved_slots
=
2181 reserved_varying_slot(producer
, ir_var_shader_out
) |
2182 reserved_varying_slot(consumer
, ir_var_shader_in
);
2184 const unsigned slots_used
= matches
.assign_locations(prog
, reserved_slots
);
2185 matches
.store_locations();
2187 for (unsigned i
= 0; i
< num_tfeedback_decls
; ++i
) {
2188 if (!tfeedback_decls
[i
].is_varying())
2191 if (!tfeedback_decls
[i
].assign_location(ctx
, prog
)) {
2192 hash_table_dtor(tfeedback_candidates
);
2193 hash_table_dtor(consumer_inputs
);
2194 hash_table_dtor(consumer_interface_inputs
);
2199 hash_table_dtor(tfeedback_candidates
);
2200 hash_table_dtor(consumer_inputs
);
2201 hash_table_dtor(consumer_interface_inputs
);
2203 if (consumer
&& producer
) {
2204 foreach_in_list(ir_instruction
, node
, consumer
->ir
) {
2205 ir_variable
*const var
= node
->as_variable();
2207 if (var
&& var
->data
.mode
== ir_var_shader_in
&&
2208 var
->data
.is_unmatched_generic_inout
) {
2209 if (!prog
->IsES
&& prog
->Version
<= 120) {
2210 /* On page 25 (page 31 of the PDF) of the GLSL 1.20 spec:
2212 * Only those varying variables used (i.e. read) in
2213 * the fragment shader executable must be written to
2214 * by the vertex shader executable; declaring
2215 * superfluous varying variables in a vertex shader is
2218 * We interpret this text as meaning that the VS must
2219 * write the variable for the FS to read it. See
2220 * "glsl1-varying read but not written" in piglit.
2222 linker_error(prog
, "%s shader varying %s not written "
2224 _mesa_shader_stage_to_string(consumer
->Stage
),
2226 _mesa_shader_stage_to_string(producer
->Stage
));
2228 linker_warning(prog
, "%s shader varying %s not written "
2230 _mesa_shader_stage_to_string(consumer
->Stage
),
2232 _mesa_shader_stage_to_string(producer
->Stage
));
2237 /* Now that validation is done its safe to remove unused varyings. As
2238 * we have both a producer and consumer its safe to remove unused
2239 * varyings even if the program is a SSO because the stages are being
2240 * linked together i.e. we have a multi-stage SSO.
2242 remove_unused_shader_inputs_and_outputs(false, producer
,
2244 remove_unused_shader_inputs_and_outputs(false, consumer
,
2249 lower_packed_varyings(mem_ctx
, slots_used
, ir_var_shader_out
,
2250 0, producer
, disable_varying_packing
,
2255 lower_packed_varyings(mem_ctx
, slots_used
, ir_var_shader_in
,
2256 consumer_vertices
, consumer
,
2257 disable_varying_packing
, xfb_enabled
);
2264 check_against_output_limit(struct gl_context
*ctx
,
2265 struct gl_shader_program
*prog
,
2266 gl_shader
*producer
)
2268 unsigned output_vectors
= 0;
2270 foreach_in_list(ir_instruction
, node
, producer
->ir
) {
2271 ir_variable
*const var
= node
->as_variable();
2273 if (var
&& var
->data
.mode
== ir_var_shader_out
&&
2274 var_counts_against_varying_limit(producer
->Stage
, var
)) {
2275 /* outputs for fragment shader can't be doubles */
2276 output_vectors
+= var
->type
->count_attribute_slots(false);
2280 assert(producer
->Stage
!= MESA_SHADER_FRAGMENT
);
2281 unsigned max_output_components
=
2282 ctx
->Const
.Program
[producer
->Stage
].MaxOutputComponents
;
2284 const unsigned output_components
= output_vectors
* 4;
2285 if (output_components
> max_output_components
) {
2286 if (ctx
->API
== API_OPENGLES2
|| prog
->IsES
)
2287 linker_error(prog
, "%s shader uses too many output vectors "
2289 _mesa_shader_stage_to_string(producer
->Stage
),
2291 max_output_components
/ 4);
2293 linker_error(prog
, "%s shader uses too many output components "
2295 _mesa_shader_stage_to_string(producer
->Stage
),
2297 max_output_components
);
2306 check_against_input_limit(struct gl_context
*ctx
,
2307 struct gl_shader_program
*prog
,
2308 gl_shader
*consumer
)
2310 unsigned input_vectors
= 0;
2312 foreach_in_list(ir_instruction
, node
, consumer
->ir
) {
2313 ir_variable
*const var
= node
->as_variable();
2315 if (var
&& var
->data
.mode
== ir_var_shader_in
&&
2316 var_counts_against_varying_limit(consumer
->Stage
, var
)) {
2317 /* vertex inputs aren't varying counted */
2318 input_vectors
+= var
->type
->count_attribute_slots(false);
2322 assert(consumer
->Stage
!= MESA_SHADER_VERTEX
);
2323 unsigned max_input_components
=
2324 ctx
->Const
.Program
[consumer
->Stage
].MaxInputComponents
;
2326 const unsigned input_components
= input_vectors
* 4;
2327 if (input_components
> max_input_components
) {
2328 if (ctx
->API
== API_OPENGLES2
|| prog
->IsES
)
2329 linker_error(prog
, "%s shader uses too many input vectors "
2331 _mesa_shader_stage_to_string(consumer
->Stage
),
2333 max_input_components
/ 4);
2335 linker_error(prog
, "%s shader uses too many input components "
2337 _mesa_shader_stage_to_string(consumer
->Stage
),
2339 max_input_components
);