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/errors.h"
33 #include "main/mtypes.h"
34 #include "glsl_symbol_table.h"
35 #include "glsl_parser_extras.h"
36 #include "ir_optimization.h"
38 #include "link_varyings.h"
39 #include "main/macros.h"
40 #include "util/hash_table.h"
41 #include "util/u_math.h"
46 * Get the varying type stripped of the outermost array if we're processing
47 * a stage whose varyings are arrays indexed by a vertex number (such as
48 * geometry shader inputs).
50 static const glsl_type
*
51 get_varying_type(const ir_variable
*var
, gl_shader_stage stage
)
53 const glsl_type
*type
= var
->type
;
55 if (!var
->data
.patch
&&
56 ((var
->data
.mode
== ir_var_shader_out
&&
57 stage
== MESA_SHADER_TESS_CTRL
) ||
58 (var
->data
.mode
== ir_var_shader_in
&&
59 (stage
== MESA_SHADER_TESS_CTRL
|| stage
== MESA_SHADER_TESS_EVAL
||
60 stage
== MESA_SHADER_GEOMETRY
)))) {
61 assert(type
->is_array());
62 type
= type
->fields
.array
;
69 create_xfb_varying_names(void *mem_ctx
, const glsl_type
*t
, char **name
,
70 size_t name_length
, unsigned *count
,
71 const char *ifc_member_name
,
72 const glsl_type
*ifc_member_t
, char ***varying_names
)
74 if (t
->is_interface()) {
75 size_t new_length
= name_length
;
77 assert(ifc_member_name
&& ifc_member_t
);
78 ralloc_asprintf_rewrite_tail(name
, &new_length
, ".%s", ifc_member_name
);
80 create_xfb_varying_names(mem_ctx
, ifc_member_t
, name
, new_length
, count
,
81 NULL
, NULL
, varying_names
);
82 } else if (t
->is_struct()) {
83 for (unsigned i
= 0; i
< t
->length
; i
++) {
84 const char *field
= t
->fields
.structure
[i
].name
;
85 size_t new_length
= name_length
;
87 ralloc_asprintf_rewrite_tail(name
, &new_length
, ".%s", field
);
89 create_xfb_varying_names(mem_ctx
, t
->fields
.structure
[i
].type
, name
,
90 new_length
, count
, NULL
, NULL
,
93 } else if (t
->without_array()->is_struct() ||
94 t
->without_array()->is_interface() ||
95 (t
->is_array() && t
->fields
.array
->is_array())) {
96 for (unsigned i
= 0; i
< t
->length
; i
++) {
97 size_t new_length
= name_length
;
99 /* Append the subscript to the current variable name */
100 ralloc_asprintf_rewrite_tail(name
, &new_length
, "[%u]", i
);
102 create_xfb_varying_names(mem_ctx
, t
->fields
.array
, name
, new_length
,
103 count
, ifc_member_name
, ifc_member_t
,
107 (*varying_names
)[(*count
)++] = ralloc_strdup(mem_ctx
, *name
);
112 process_xfb_layout_qualifiers(void *mem_ctx
, const gl_linked_shader
*sh
,
113 struct gl_shader_program
*prog
,
114 unsigned *num_tfeedback_decls
,
115 char ***varying_names
)
117 bool has_xfb_qualifiers
= false;
119 /* We still need to enable transform feedback mode even if xfb_stride is
120 * only applied to a global out. Also we don't bother to propagate
121 * xfb_stride to interface block members so this will catch that case also.
123 for (unsigned j
= 0; j
< MAX_FEEDBACK_BUFFERS
; j
++) {
124 if (prog
->TransformFeedback
.BufferStride
[j
]) {
125 has_xfb_qualifiers
= true;
130 foreach_in_list(ir_instruction
, node
, sh
->ir
) {
131 ir_variable
*var
= node
->as_variable();
132 if (!var
|| var
->data
.mode
!= ir_var_shader_out
)
135 /* From the ARB_enhanced_layouts spec:
137 * "Any shader making any static use (after preprocessing) of any of
138 * these *xfb_* qualifiers will cause the shader to be in a
139 * transform feedback capturing mode and hence responsible for
140 * describing the transform feedback setup. This mode will capture
141 * any output selected by *xfb_offset*, directly or indirectly, to
142 * a transform feedback buffer."
144 if (var
->data
.explicit_xfb_buffer
|| var
->data
.explicit_xfb_stride
) {
145 has_xfb_qualifiers
= true;
148 if (var
->data
.explicit_xfb_offset
) {
149 *num_tfeedback_decls
+= var
->type
->varying_count();
150 has_xfb_qualifiers
= true;
154 if (*num_tfeedback_decls
== 0)
155 return has_xfb_qualifiers
;
158 *varying_names
= ralloc_array(mem_ctx
, char *, *num_tfeedback_decls
);
159 foreach_in_list(ir_instruction
, node
, sh
->ir
) {
160 ir_variable
*var
= node
->as_variable();
161 if (!var
|| var
->data
.mode
!= ir_var_shader_out
)
164 if (var
->data
.explicit_xfb_offset
) {
166 const glsl_type
*type
, *member_type
;
168 if (var
->data
.from_named_ifc_block
) {
169 type
= var
->get_interface_type();
171 /* Find the member type before it was altered by lowering */
172 const glsl_type
*type_wa
= type
->without_array();
174 type_wa
->fields
.structure
[type_wa
->field_index(var
->name
)].type
;
175 name
= ralloc_strdup(NULL
, type_wa
->name
);
179 name
= ralloc_strdup(NULL
, var
->name
);
181 create_xfb_varying_names(mem_ctx
, type
, &name
, strlen(name
), &i
,
182 var
->name
, member_type
, varying_names
);
187 assert(i
== *num_tfeedback_decls
);
188 return has_xfb_qualifiers
;
192 * Validate the types and qualifiers of an output from one stage against the
193 * matching input to another stage.
196 cross_validate_types_and_qualifiers(struct gl_context
*ctx
,
197 struct gl_shader_program
*prog
,
198 const ir_variable
*input
,
199 const ir_variable
*output
,
200 gl_shader_stage consumer_stage
,
201 gl_shader_stage producer_stage
)
203 /* Check that the types match between stages.
205 const glsl_type
*type_to_match
= input
->type
;
207 /* VS -> GS, VS -> TCS, VS -> TES, TES -> GS */
208 const bool extra_array_level
= (producer_stage
== MESA_SHADER_VERTEX
&&
209 consumer_stage
!= MESA_SHADER_FRAGMENT
) ||
210 consumer_stage
== MESA_SHADER_GEOMETRY
;
211 if (extra_array_level
) {
212 assert(type_to_match
->is_array());
213 type_to_match
= type_to_match
->fields
.array
;
216 if (type_to_match
!= output
->type
) {
217 /* There is a bit of a special case for gl_TexCoord. This
218 * built-in is unsized by default. Applications that variable
219 * access it must redeclare it with a size. There is some
220 * language in the GLSL spec that implies the fragment shader
221 * and vertex shader do not have to agree on this size. Other
222 * driver behave this way, and one or two applications seem to
225 * Neither declaration needs to be modified here because the array
226 * sizes are fixed later when update_array_sizes is called.
228 * From page 48 (page 54 of the PDF) of the GLSL 1.10 spec:
230 * "Unlike user-defined varying variables, the built-in
231 * varying variables don't have a strict one-to-one
232 * correspondence between the vertex language and the
233 * fragment language."
235 if (!output
->type
->is_array() || !is_gl_identifier(output
->name
)) {
237 "%s shader output `%s' declared as type `%s', "
238 "but %s shader input declared as type `%s'\n",
239 _mesa_shader_stage_to_string(producer_stage
),
242 _mesa_shader_stage_to_string(consumer_stage
),
248 /* Check that all of the qualifiers match between stages.
251 /* According to the OpenGL and OpenGLES GLSL specs, the centroid qualifier
252 * should match until OpenGL 4.3 and OpenGLES 3.1. The OpenGLES 3.0
253 * conformance test suite does not verify that the qualifiers must match.
254 * The deqp test suite expects the opposite (OpenGLES 3.1) behavior for
255 * OpenGLES 3.0 drivers, so we relax the checking in all cases.
257 if (false /* always skip the centroid check */ &&
258 prog
->data
->Version
< (prog
->IsES
? 310 : 430) &&
259 input
->data
.centroid
!= output
->data
.centroid
) {
261 "%s shader output `%s' %s centroid qualifier, "
262 "but %s shader input %s centroid qualifier\n",
263 _mesa_shader_stage_to_string(producer_stage
),
265 (output
->data
.centroid
) ? "has" : "lacks",
266 _mesa_shader_stage_to_string(consumer_stage
),
267 (input
->data
.centroid
) ? "has" : "lacks");
271 if (input
->data
.sample
!= output
->data
.sample
) {
273 "%s shader output `%s' %s sample qualifier, "
274 "but %s shader input %s sample qualifier\n",
275 _mesa_shader_stage_to_string(producer_stage
),
277 (output
->data
.sample
) ? "has" : "lacks",
278 _mesa_shader_stage_to_string(consumer_stage
),
279 (input
->data
.sample
) ? "has" : "lacks");
283 if (input
->data
.patch
!= output
->data
.patch
) {
285 "%s shader output `%s' %s patch qualifier, "
286 "but %s shader input %s patch qualifier\n",
287 _mesa_shader_stage_to_string(producer_stage
),
289 (output
->data
.patch
) ? "has" : "lacks",
290 _mesa_shader_stage_to_string(consumer_stage
),
291 (input
->data
.patch
) ? "has" : "lacks");
295 /* The GLSL 4.30 and GLSL ES 3.00 specifications say:
297 * "As only outputs need be declared with invariant, an output from
298 * one shader stage will still match an input of a subsequent stage
299 * without the input being declared as invariant."
301 * while GLSL 4.20 says:
303 * "For variables leaving one shader and coming into another shader,
304 * the invariant keyword has to be used in both shaders, or a link
305 * error will result."
307 * and GLSL ES 1.00 section 4.6.4 "Invariance and Linking" says:
309 * "The invariance of varyings that are declared in both the vertex
310 * and fragment shaders must match."
312 if (input
->data
.invariant
!= output
->data
.invariant
&&
313 prog
->data
->Version
< (prog
->IsES
? 300 : 430)) {
315 "%s shader output `%s' %s invariant qualifier, "
316 "but %s shader input %s invariant qualifier\n",
317 _mesa_shader_stage_to_string(producer_stage
),
319 (output
->data
.invariant
) ? "has" : "lacks",
320 _mesa_shader_stage_to_string(consumer_stage
),
321 (input
->data
.invariant
) ? "has" : "lacks");
325 /* GLSL >= 4.40 removes text requiring interpolation qualifiers
326 * to match cross stage, they must only match within the same stage.
328 * From page 84 (page 90 of the PDF) of the GLSL 4.40 spec:
330 * "It is a link-time error if, within the same stage, the interpolation
331 * qualifiers of variables of the same name do not match.
333 * Section 4.3.9 (Interpolation) of the GLSL ES 3.00 spec says:
335 * "When no interpolation qualifier is present, smooth interpolation
338 * So we match variables where one is smooth and the other has no explicit
341 unsigned input_interpolation
= input
->data
.interpolation
;
342 unsigned output_interpolation
= output
->data
.interpolation
;
344 if (input_interpolation
== INTERP_MODE_NONE
)
345 input_interpolation
= INTERP_MODE_SMOOTH
;
346 if (output_interpolation
== INTERP_MODE_NONE
)
347 output_interpolation
= INTERP_MODE_SMOOTH
;
349 if (input_interpolation
!= output_interpolation
&&
350 prog
->data
->Version
< 440) {
351 if (!ctx
->Const
.AllowGLSLCrossStageInterpolationMismatch
) {
353 "%s shader output `%s' specifies %s "
354 "interpolation qualifier, "
355 "but %s shader input specifies %s "
356 "interpolation qualifier\n",
357 _mesa_shader_stage_to_string(producer_stage
),
359 interpolation_string(output
->data
.interpolation
),
360 _mesa_shader_stage_to_string(consumer_stage
),
361 interpolation_string(input
->data
.interpolation
));
365 "%s shader output `%s' specifies %s "
366 "interpolation qualifier, "
367 "but %s shader input specifies %s "
368 "interpolation qualifier\n",
369 _mesa_shader_stage_to_string(producer_stage
),
371 interpolation_string(output
->data
.interpolation
),
372 _mesa_shader_stage_to_string(consumer_stage
),
373 interpolation_string(input
->data
.interpolation
));
379 * Validate front and back color outputs against single color input
382 cross_validate_front_and_back_color(struct gl_context
*ctx
,
383 struct gl_shader_program
*prog
,
384 const ir_variable
*input
,
385 const ir_variable
*front_color
,
386 const ir_variable
*back_color
,
387 gl_shader_stage consumer_stage
,
388 gl_shader_stage producer_stage
)
390 if (front_color
!= NULL
&& front_color
->data
.assigned
)
391 cross_validate_types_and_qualifiers(ctx
, prog
, input
, front_color
,
392 consumer_stage
, producer_stage
);
394 if (back_color
!= NULL
&& back_color
->data
.assigned
)
395 cross_validate_types_and_qualifiers(ctx
, prog
, input
, back_color
,
396 consumer_stage
, producer_stage
);
400 compute_variable_location_slot(ir_variable
*var
, gl_shader_stage stage
)
402 unsigned location_start
= VARYING_SLOT_VAR0
;
405 case MESA_SHADER_VERTEX
:
406 if (var
->data
.mode
== ir_var_shader_in
)
407 location_start
= VERT_ATTRIB_GENERIC0
;
409 case MESA_SHADER_TESS_CTRL
:
410 case MESA_SHADER_TESS_EVAL
:
412 location_start
= VARYING_SLOT_PATCH0
;
414 case MESA_SHADER_FRAGMENT
:
415 if (var
->data
.mode
== ir_var_shader_out
)
416 location_start
= FRAG_RESULT_DATA0
;
422 return var
->data
.location
- location_start
;
425 struct explicit_location_info
{
427 unsigned numerical_type
;
428 unsigned interpolation
;
434 static inline unsigned
435 get_numerical_type(const glsl_type
*type
)
437 /* From the OpenGL 4.6 spec, section 4.4.1 Input Layout Qualifiers, Page 68,
438 * (Location aliasing):
440 * "Further, when location aliasing, the aliases sharing the location
441 * must have the same underlying numerical type (floating-point or
444 if (type
->is_float() || type
->is_double())
445 return GLSL_TYPE_FLOAT
;
446 return GLSL_TYPE_INT
;
450 check_location_aliasing(struct explicit_location_info explicit_locations
[][4],
454 unsigned location_limit
,
455 const glsl_type
*type
,
456 unsigned interpolation
,
460 gl_shader_program
*prog
,
461 gl_shader_stage stage
)
464 if (type
->without_array()->is_struct()) {
465 /* The component qualifier can't be used on structs so just treat
466 * all component slots as used.
470 unsigned dmul
= type
->without_array()->is_64bit() ? 2 : 1;
471 last_comp
= component
+ type
->without_array()->vector_elements
* dmul
;
474 while (location
< location_limit
) {
477 struct explicit_location_info
*info
=
478 &explicit_locations
[location
][comp
];
481 /* Component aliasing is not alloed */
482 if (comp
>= component
&& comp
< last_comp
) {
484 "%s shader has multiple %sputs explicitly "
485 "assigned to location %d and component %d\n",
486 _mesa_shader_stage_to_string(stage
),
487 var
->data
.mode
== ir_var_shader_in
? "in" : "out",
491 /* For all other used components we need to have matching
492 * types, interpolation and auxiliary storage
494 if (info
->numerical_type
!=
495 get_numerical_type(type
->without_array())) {
497 "Varyings sharing the same location must "
498 "have the same underlying numerical type. "
499 "Location %u component %u\n",
504 if (info
->interpolation
!= interpolation
) {
506 "%s shader has multiple %sputs at explicit "
507 "location %u with different interpolation "
509 _mesa_shader_stage_to_string(stage
),
510 var
->data
.mode
== ir_var_shader_in
?
511 "in" : "out", location
);
515 if (info
->centroid
!= centroid
||
516 info
->sample
!= sample
||
517 info
->patch
!= patch
) {
519 "%s shader has multiple %sputs at explicit "
520 "location %u with different aux storage\n",
521 _mesa_shader_stage_to_string(stage
),
522 var
->data
.mode
== ir_var_shader_in
?
523 "in" : "out", location
);
527 } else if (comp
>= component
&& comp
< last_comp
) {
529 info
->numerical_type
= get_numerical_type(type
->without_array());
530 info
->interpolation
= interpolation
;
531 info
->centroid
= centroid
;
532 info
->sample
= sample
;
538 /* We need to do some special handling for doubles as dvec3 and
539 * dvec4 consume two consecutive locations. We don't need to
540 * worry about components beginning at anything other than 0 as
541 * the spec does not allow this for dvec3 and dvec4.
543 if (comp
== 4 && last_comp
> 4) {
544 last_comp
= last_comp
- 4;
545 /* Bump location index and reset the component index */
559 validate_explicit_variable_location(struct gl_context
*ctx
,
560 struct explicit_location_info explicit_locations
[][4],
562 gl_shader_program
*prog
,
563 gl_linked_shader
*sh
)
565 const glsl_type
*type
= get_varying_type(var
, sh
->Stage
);
566 unsigned num_elements
= type
->count_attribute_slots(false);
567 unsigned idx
= compute_variable_location_slot(var
, sh
->Stage
);
568 unsigned slot_limit
= idx
+ num_elements
;
570 /* Vertex shader inputs and fragment shader outputs are validated in
571 * assign_attribute_or_color_locations() so we should not attempt to
572 * validate them again here.
575 if (var
->data
.mode
== ir_var_shader_out
) {
576 assert(sh
->Stage
!= MESA_SHADER_FRAGMENT
);
578 ctx
->Const
.Program
[sh
->Stage
].MaxOutputComponents
/ 4;
580 assert(var
->data
.mode
== ir_var_shader_in
);
581 assert(sh
->Stage
!= MESA_SHADER_VERTEX
);
583 ctx
->Const
.Program
[sh
->Stage
].MaxInputComponents
/ 4;
586 if (slot_limit
> slot_max
) {
588 "Invalid location %u in %s shader\n",
589 idx
, _mesa_shader_stage_to_string(sh
->Stage
));
593 const glsl_type
*type_without_array
= type
->without_array();
594 if (type_without_array
->is_interface()) {
595 for (unsigned i
= 0; i
< type_without_array
->length
; i
++) {
596 glsl_struct_field
*field
= &type_without_array
->fields
.structure
[i
];
597 unsigned field_location
= field
->location
-
598 (field
->patch
? VARYING_SLOT_PATCH0
: VARYING_SLOT_VAR0
);
599 if (!check_location_aliasing(explicit_locations
, var
,
601 0, field_location
+ 1,
603 field
->interpolation
,
611 } else if (!check_location_aliasing(explicit_locations
, var
,
612 idx
, var
->data
.location_frac
,
614 var
->data
.interpolation
,
626 * Validate explicit locations for the inputs to the first stage and the
627 * outputs of the last stage in an SSO program (everything in between is
628 * validated in cross_validate_outputs_to_inputs).
631 validate_sso_explicit_locations(struct gl_context
*ctx
,
632 struct gl_shader_program
*prog
,
633 gl_shader_stage first_stage
,
634 gl_shader_stage last_stage
)
636 assert(prog
->SeparateShader
);
638 /* VS inputs and FS outputs are validated in
639 * assign_attribute_or_color_locations()
641 bool validate_first_stage
= first_stage
!= MESA_SHADER_VERTEX
;
642 bool validate_last_stage
= last_stage
!= MESA_SHADER_FRAGMENT
;
643 if (!validate_first_stage
&& !validate_last_stage
)
646 struct explicit_location_info explicit_locations
[MAX_VARYING
][4];
648 gl_shader_stage stages
[2] = { first_stage
, last_stage
};
649 bool validate_stage
[2] = { validate_first_stage
, validate_last_stage
};
650 ir_variable_mode var_direction
[2] = { ir_var_shader_in
, ir_var_shader_out
};
652 for (unsigned i
= 0; i
< 2; i
++) {
653 if (!validate_stage
[i
])
656 gl_shader_stage stage
= stages
[i
];
658 gl_linked_shader
*sh
= prog
->_LinkedShaders
[stage
];
661 memset(explicit_locations
, 0, sizeof(explicit_locations
));
663 foreach_in_list(ir_instruction
, node
, sh
->ir
) {
664 ir_variable
*const var
= node
->as_variable();
667 !var
->data
.explicit_location
||
668 var
->data
.location
< VARYING_SLOT_VAR0
||
669 var
->data
.mode
!= var_direction
[i
])
672 if (!validate_explicit_variable_location(
673 ctx
, explicit_locations
, var
, prog
, sh
)) {
681 * Validate that outputs from one stage match inputs of another
684 cross_validate_outputs_to_inputs(struct gl_context
*ctx
,
685 struct gl_shader_program
*prog
,
686 gl_linked_shader
*producer
,
687 gl_linked_shader
*consumer
)
689 glsl_symbol_table parameters
;
690 struct explicit_location_info output_explicit_locations
[MAX_VARYING
][4] = { 0 };
691 struct explicit_location_info input_explicit_locations
[MAX_VARYING
][4] = { 0 };
693 /* Find all shader outputs in the "producer" stage.
695 foreach_in_list(ir_instruction
, node
, producer
->ir
) {
696 ir_variable
*const var
= node
->as_variable();
698 if (var
== NULL
|| var
->data
.mode
!= ir_var_shader_out
)
701 if (!var
->data
.explicit_location
702 || var
->data
.location
< VARYING_SLOT_VAR0
)
703 parameters
.add_variable(var
);
705 /* User-defined varyings with explicit locations are handled
706 * differently because they do not need to have matching names.
708 if (!validate_explicit_variable_location(ctx
,
709 output_explicit_locations
,
710 var
, prog
, producer
)) {
717 /* Find all shader inputs in the "consumer" stage. Any variables that have
718 * matching outputs already in the symbol table must have the same type and
721 * Exception: if the consumer is the geometry shader, then the inputs
722 * should be arrays and the type of the array element should match the type
723 * of the corresponding producer output.
725 foreach_in_list(ir_instruction
, node
, consumer
->ir
) {
726 ir_variable
*const input
= node
->as_variable();
728 if (input
== NULL
|| input
->data
.mode
!= ir_var_shader_in
)
731 if (strcmp(input
->name
, "gl_Color") == 0 && input
->data
.used
) {
732 const ir_variable
*const front_color
=
733 parameters
.get_variable("gl_FrontColor");
735 const ir_variable
*const back_color
=
736 parameters
.get_variable("gl_BackColor");
738 cross_validate_front_and_back_color(ctx
, prog
, input
,
739 front_color
, back_color
,
740 consumer
->Stage
, producer
->Stage
);
741 } else if (strcmp(input
->name
, "gl_SecondaryColor") == 0 && input
->data
.used
) {
742 const ir_variable
*const front_color
=
743 parameters
.get_variable("gl_FrontSecondaryColor");
745 const ir_variable
*const back_color
=
746 parameters
.get_variable("gl_BackSecondaryColor");
748 cross_validate_front_and_back_color(ctx
, prog
, input
,
749 front_color
, back_color
,
750 consumer
->Stage
, producer
->Stage
);
752 /* The rules for connecting inputs and outputs change in the presence
753 * of explicit locations. In this case, we no longer care about the
754 * names of the variables. Instead, we care only about the
755 * explicitly assigned location.
757 ir_variable
*output
= NULL
;
758 if (input
->data
.explicit_location
759 && input
->data
.location
>= VARYING_SLOT_VAR0
) {
761 const glsl_type
*type
= get_varying_type(input
, consumer
->Stage
);
762 unsigned num_elements
= type
->count_attribute_slots(false);
764 compute_variable_location_slot(input
, consumer
->Stage
);
765 unsigned slot_limit
= idx
+ num_elements
;
767 if (!validate_explicit_variable_location(ctx
,
768 input_explicit_locations
,
769 input
, prog
, consumer
)) {
773 while (idx
< slot_limit
) {
774 if (idx
>= MAX_VARYING
) {
776 "Invalid location %u in %s shader\n", idx
,
777 _mesa_shader_stage_to_string(consumer
->Stage
));
781 output
= output_explicit_locations
[idx
][input
->data
.location_frac
].var
;
783 if (output
== NULL
) {
784 /* A linker failure should only happen when there is no
785 * output declaration and there is Static Use of the
788 if (input
->data
.used
) {
790 "%s shader input `%s' with explicit location "
791 "has no matching output\n",
792 _mesa_shader_stage_to_string(consumer
->Stage
),
796 } else if (input
->data
.location
!= output
->data
.location
) {
798 "%s shader input `%s' with explicit location "
799 "has no matching output\n",
800 _mesa_shader_stage_to_string(consumer
->Stage
),
807 output
= parameters
.get_variable(input
->name
);
810 if (output
!= NULL
) {
811 /* Interface blocks have their own validation elsewhere so don't
812 * try validating them here.
814 if (!(input
->get_interface_type() &&
815 output
->get_interface_type()))
816 cross_validate_types_and_qualifiers(ctx
, prog
, input
, output
,
820 /* Check for input vars with unmatched output vars in prev stage
821 * taking into account that interface blocks could have a matching
822 * output but with different name, so we ignore them.
824 assert(!input
->data
.assigned
);
825 if (input
->data
.used
&& !input
->get_interface_type() &&
826 !input
->data
.explicit_location
&& !prog
->SeparateShader
)
828 "%s shader input `%s' "
829 "has no matching output in the previous stage\n",
830 _mesa_shader_stage_to_string(consumer
->Stage
),
838 * Demote shader inputs and outputs that are not used in other stages, and
839 * remove them via dead code elimination.
842 remove_unused_shader_inputs_and_outputs(bool is_separate_shader_object
,
843 gl_linked_shader
*sh
,
844 enum ir_variable_mode mode
)
846 if (is_separate_shader_object
)
849 foreach_in_list(ir_instruction
, node
, sh
->ir
) {
850 ir_variable
*const var
= node
->as_variable();
852 if (var
== NULL
|| var
->data
.mode
!= int(mode
))
855 /* A shader 'in' or 'out' variable is only really an input or output if
856 * its value is used by other shader stages. This will cause the
857 * variable to have a location assigned.
859 if (var
->data
.is_unmatched_generic_inout
&& !var
->data
.is_xfb_only
) {
860 assert(var
->data
.mode
!= ir_var_temporary
);
862 /* Assign zeros to demoted inputs to allow more optimizations. */
863 if (var
->data
.mode
== ir_var_shader_in
&& !var
->constant_value
)
864 var
->constant_value
= ir_constant::zero(var
, var
->type
);
866 var
->data
.mode
= ir_var_auto
;
870 /* Eliminate code that is now dead due to unused inputs/outputs being
873 while (do_dead_code(sh
->ir
, false))
879 * Initialize this object based on a string that was passed to
880 * glTransformFeedbackVaryings.
882 * If the input is mal-formed, this call still succeeds, but it sets
883 * this->var_name to a mal-formed input, so tfeedback_decl::find_output_var()
884 * will fail to find any matching variable.
887 tfeedback_decl::init(struct gl_context
*ctx
, const void *mem_ctx
,
890 /* We don't have to be pedantic about what is a valid GLSL variable name,
891 * because any variable with an invalid name can't exist in the IR anyway.
895 this->orig_name
= input
;
896 this->lowered_builtin_array_variable
= none
;
897 this->skip_components
= 0;
898 this->next_buffer_separator
= false;
899 this->matched_candidate
= NULL
;
904 if (ctx
->Extensions
.ARB_transform_feedback3
) {
905 /* Parse gl_NextBuffer. */
906 if (strcmp(input
, "gl_NextBuffer") == 0) {
907 this->next_buffer_separator
= true;
911 /* Parse gl_SkipComponents. */
912 if (strcmp(input
, "gl_SkipComponents1") == 0)
913 this->skip_components
= 1;
914 else if (strcmp(input
, "gl_SkipComponents2") == 0)
915 this->skip_components
= 2;
916 else if (strcmp(input
, "gl_SkipComponents3") == 0)
917 this->skip_components
= 3;
918 else if (strcmp(input
, "gl_SkipComponents4") == 0)
919 this->skip_components
= 4;
921 if (this->skip_components
)
925 /* Parse a declaration. */
926 const char *base_name_end
;
927 long subscript
= parse_program_resource_name(input
, &base_name_end
);
928 this->var_name
= ralloc_strndup(mem_ctx
, input
, base_name_end
- input
);
929 if (this->var_name
== NULL
) {
930 _mesa_error_no_memory(__func__
);
934 if (subscript
>= 0) {
935 this->array_subscript
= subscript
;
936 this->is_subscripted
= true;
938 this->is_subscripted
= false;
941 /* For drivers that lower gl_ClipDistance to gl_ClipDistanceMESA, this
942 * class must behave specially to account for the fact that gl_ClipDistance
943 * is converted from a float[8] to a vec4[2].
945 if (ctx
->Const
.ShaderCompilerOptions
[MESA_SHADER_VERTEX
].LowerCombinedClipCullDistance
&&
946 strcmp(this->var_name
, "gl_ClipDistance") == 0) {
947 this->lowered_builtin_array_variable
= clip_distance
;
949 if (ctx
->Const
.ShaderCompilerOptions
[MESA_SHADER_VERTEX
].LowerCombinedClipCullDistance
&&
950 strcmp(this->var_name
, "gl_CullDistance") == 0) {
951 this->lowered_builtin_array_variable
= cull_distance
;
954 if (ctx
->Const
.LowerTessLevel
&&
955 (strcmp(this->var_name
, "gl_TessLevelOuter") == 0))
956 this->lowered_builtin_array_variable
= tess_level_outer
;
957 if (ctx
->Const
.LowerTessLevel
&&
958 (strcmp(this->var_name
, "gl_TessLevelInner") == 0))
959 this->lowered_builtin_array_variable
= tess_level_inner
;
964 * Determine whether two tfeedback_decl objects refer to the same variable and
965 * array index (if applicable).
968 tfeedback_decl::is_same(const tfeedback_decl
&x
, const tfeedback_decl
&y
)
970 assert(x
.is_varying() && y
.is_varying());
972 if (strcmp(x
.var_name
, y
.var_name
) != 0)
974 if (x
.is_subscripted
!= y
.is_subscripted
)
976 if (x
.is_subscripted
&& x
.array_subscript
!= y
.array_subscript
)
983 * Assign a location and stream ID for this tfeedback_decl object based on the
984 * transform feedback candidate found by find_candidate.
986 * If an error occurs, the error is reported through linker_error() and false
990 tfeedback_decl::assign_location(struct gl_context
*ctx
,
991 struct gl_shader_program
*prog
)
993 assert(this->is_varying());
995 unsigned fine_location
996 = this->matched_candidate
->toplevel_var
->data
.location
* 4
997 + this->matched_candidate
->toplevel_var
->data
.location_frac
998 + this->matched_candidate
->offset
;
999 const unsigned dmul
=
1000 this->matched_candidate
->type
->without_array()->is_64bit() ? 2 : 1;
1002 if (this->matched_candidate
->type
->is_array()) {
1003 /* Array variable */
1004 const unsigned matrix_cols
=
1005 this->matched_candidate
->type
->fields
.array
->matrix_columns
;
1006 const unsigned vector_elements
=
1007 this->matched_candidate
->type
->fields
.array
->vector_elements
;
1008 unsigned actual_array_size
;
1009 switch (this->lowered_builtin_array_variable
) {
1011 actual_array_size
= prog
->last_vert_prog
?
1012 prog
->last_vert_prog
->info
.clip_distance_array_size
: 0;
1015 actual_array_size
= prog
->last_vert_prog
?
1016 prog
->last_vert_prog
->info
.cull_distance_array_size
: 0;
1018 case tess_level_outer
:
1019 actual_array_size
= 4;
1021 case tess_level_inner
:
1022 actual_array_size
= 2;
1026 actual_array_size
= this->matched_candidate
->type
->array_size();
1030 if (this->is_subscripted
) {
1031 /* Check array bounds. */
1032 if (this->array_subscript
>= actual_array_size
) {
1033 linker_error(prog
, "Transform feedback varying %s has index "
1034 "%i, but the array size is %u.",
1035 this->orig_name
, this->array_subscript
,
1039 unsigned array_elem_size
= this->lowered_builtin_array_variable
?
1040 1 : vector_elements
* matrix_cols
* dmul
;
1041 fine_location
+= array_elem_size
* this->array_subscript
;
1044 this->size
= actual_array_size
;
1046 this->vector_elements
= vector_elements
;
1047 this->matrix_columns
= matrix_cols
;
1048 if (this->lowered_builtin_array_variable
)
1049 this->type
= GL_FLOAT
;
1051 this->type
= this->matched_candidate
->type
->fields
.array
->gl_type
;
1053 /* Regular variable (scalar, vector, or matrix) */
1054 if (this->is_subscripted
) {
1055 linker_error(prog
, "Transform feedback varying %s requested, "
1056 "but %s is not an array.",
1057 this->orig_name
, this->var_name
);
1061 this->vector_elements
= this->matched_candidate
->type
->vector_elements
;
1062 this->matrix_columns
= this->matched_candidate
->type
->matrix_columns
;
1063 this->type
= this->matched_candidate
->type
->gl_type
;
1065 this->location
= fine_location
/ 4;
1066 this->location_frac
= fine_location
% 4;
1068 /* From GL_EXT_transform_feedback:
1069 * A program will fail to link if:
1071 * * the total number of components to capture in any varying
1072 * variable in <varyings> is greater than the constant
1073 * MAX_TRANSFORM_FEEDBACK_SEPARATE_COMPONENTS_EXT and the
1074 * buffer mode is SEPARATE_ATTRIBS_EXT;
1076 if (prog
->TransformFeedback
.BufferMode
== GL_SEPARATE_ATTRIBS
&&
1077 this->num_components() >
1078 ctx
->Const
.MaxTransformFeedbackSeparateComponents
) {
1079 linker_error(prog
, "Transform feedback varying %s exceeds "
1080 "MAX_TRANSFORM_FEEDBACK_SEPARATE_COMPONENTS.",
1085 /* Only transform feedback varyings can be assigned to non-zero streams,
1086 * so assign the stream id here.
1088 this->stream_id
= this->matched_candidate
->toplevel_var
->data
.stream
;
1090 unsigned array_offset
= this->array_subscript
* 4 * dmul
;
1091 unsigned struct_offset
= this->matched_candidate
->offset
* 4 * dmul
;
1092 this->buffer
= this->matched_candidate
->toplevel_var
->data
.xfb_buffer
;
1093 this->offset
= this->matched_candidate
->toplevel_var
->data
.offset
+
1094 array_offset
+ struct_offset
;
1101 tfeedback_decl::get_num_outputs() const
1103 if (!this->is_varying()) {
1106 return (this->num_components() + this->location_frac
+ 3)/4;
1111 * Update gl_transform_feedback_info to reflect this tfeedback_decl.
1113 * If an error occurs, the error is reported through linker_error() and false
1117 tfeedback_decl::store(struct gl_context
*ctx
, struct gl_shader_program
*prog
,
1118 struct gl_transform_feedback_info
*info
,
1119 unsigned buffer
, unsigned buffer_index
,
1120 const unsigned max_outputs
, bool *explicit_stride
,
1121 bool has_xfb_qualifiers
) const
1123 unsigned xfb_offset
= 0;
1124 unsigned size
= this->size
;
1125 /* Handle gl_SkipComponents. */
1126 if (this->skip_components
) {
1127 info
->Buffers
[buffer
].Stride
+= this->skip_components
;
1128 size
= this->skip_components
;
1132 if (this->next_buffer_separator
) {
1137 if (has_xfb_qualifiers
) {
1138 xfb_offset
= this->offset
/ 4;
1140 xfb_offset
= info
->Buffers
[buffer
].Stride
;
1142 info
->Varyings
[info
->NumVarying
].Offset
= xfb_offset
* 4;
1145 unsigned location
= this->location
;
1146 unsigned location_frac
= this->location_frac
;
1147 unsigned num_components
= this->num_components();
1148 while (num_components
> 0) {
1149 unsigned output_size
= MIN2(num_components
, 4 - location_frac
);
1150 assert((info
->NumOutputs
== 0 && max_outputs
== 0) ||
1151 info
->NumOutputs
< max_outputs
);
1153 /* From the ARB_enhanced_layouts spec:
1155 * "If such a block member or variable is not written during a shader
1156 * invocation, the buffer contents at the assigned offset will be
1157 * undefined. Even if there are no static writes to a variable or
1158 * member that is assigned a transform feedback offset, the space is
1159 * still allocated in the buffer and still affects the stride."
1161 if (this->is_varying_written()) {
1162 info
->Outputs
[info
->NumOutputs
].ComponentOffset
= location_frac
;
1163 info
->Outputs
[info
->NumOutputs
].OutputRegister
= location
;
1164 info
->Outputs
[info
->NumOutputs
].NumComponents
= output_size
;
1165 info
->Outputs
[info
->NumOutputs
].StreamId
= stream_id
;
1166 info
->Outputs
[info
->NumOutputs
].OutputBuffer
= buffer
;
1167 info
->Outputs
[info
->NumOutputs
].DstOffset
= xfb_offset
;
1170 info
->Buffers
[buffer
].Stream
= this->stream_id
;
1171 xfb_offset
+= output_size
;
1173 num_components
-= output_size
;
1179 if (explicit_stride
&& explicit_stride
[buffer
]) {
1180 if (this->is_64bit() && info
->Buffers
[buffer
].Stride
% 2) {
1181 linker_error(prog
, "invalid qualifier xfb_stride=%d must be a "
1182 "multiple of 8 as its applied to a type that is or "
1183 "contains a double.",
1184 info
->Buffers
[buffer
].Stride
* 4);
1188 if ((this->offset
/ 4) / info
->Buffers
[buffer
].Stride
!=
1189 (xfb_offset
- 1) / info
->Buffers
[buffer
].Stride
) {
1190 linker_error(prog
, "xfb_offset (%d) overflows xfb_stride (%d) for "
1191 "buffer (%d)", xfb_offset
* 4,
1192 info
->Buffers
[buffer
].Stride
* 4, buffer
);
1196 info
->Buffers
[buffer
].Stride
= xfb_offset
;
1199 /* From GL_EXT_transform_feedback:
1200 * A program will fail to link if:
1202 * * the total number of components to capture is greater than
1203 * the constant MAX_TRANSFORM_FEEDBACK_INTERLEAVED_COMPONENTS_EXT
1204 * and the buffer mode is INTERLEAVED_ATTRIBS_EXT.
1206 * From GL_ARB_enhanced_layouts:
1208 * "The resulting stride (implicit or explicit) must be less than or
1209 * equal to the implementation-dependent constant
1210 * gl_MaxTransformFeedbackInterleavedComponents."
1212 if ((prog
->TransformFeedback
.BufferMode
== GL_INTERLEAVED_ATTRIBS
||
1213 has_xfb_qualifiers
) &&
1214 info
->Buffers
[buffer
].Stride
>
1215 ctx
->Const
.MaxTransformFeedbackInterleavedComponents
) {
1216 linker_error(prog
, "The MAX_TRANSFORM_FEEDBACK_INTERLEAVED_COMPONENTS "
1217 "limit has been exceeded.");
1222 info
->Varyings
[info
->NumVarying
].Name
= ralloc_strdup(prog
,
1224 info
->Varyings
[info
->NumVarying
].Type
= this->type
;
1225 info
->Varyings
[info
->NumVarying
].Size
= size
;
1226 info
->Varyings
[info
->NumVarying
].BufferIndex
= buffer_index
;
1228 info
->Buffers
[buffer
].NumVaryings
++;
1234 const tfeedback_candidate
*
1235 tfeedback_decl::find_candidate(gl_shader_program
*prog
,
1236 hash_table
*tfeedback_candidates
)
1238 const char *name
= this->var_name
;
1239 switch (this->lowered_builtin_array_variable
) {
1241 name
= this->var_name
;
1244 name
= "gl_ClipDistanceMESA";
1247 name
= "gl_CullDistanceMESA";
1249 case tess_level_outer
:
1250 name
= "gl_TessLevelOuterMESA";
1252 case tess_level_inner
:
1253 name
= "gl_TessLevelInnerMESA";
1256 hash_entry
*entry
= _mesa_hash_table_search(tfeedback_candidates
, name
);
1258 this->matched_candidate
= entry
?
1259 (const tfeedback_candidate
*) entry
->data
: NULL
;
1261 if (!this->matched_candidate
) {
1262 /* From GL_EXT_transform_feedback:
1263 * A program will fail to link if:
1265 * * any variable name specified in the <varyings> array is not
1266 * declared as an output in the geometry shader (if present) or
1267 * the vertex shader (if no geometry shader is present);
1269 linker_error(prog
, "Transform feedback varying %s undeclared.",
1273 return this->matched_candidate
;
1278 * Parse all the transform feedback declarations that were passed to
1279 * glTransformFeedbackVaryings() and store them in tfeedback_decl objects.
1281 * If an error occurs, the error is reported through linker_error() and false
1285 parse_tfeedback_decls(struct gl_context
*ctx
, struct gl_shader_program
*prog
,
1286 const void *mem_ctx
, unsigned num_names
,
1287 char **varying_names
, tfeedback_decl
*decls
)
1289 for (unsigned i
= 0; i
< num_names
; ++i
) {
1290 decls
[i
].init(ctx
, mem_ctx
, varying_names
[i
]);
1292 if (!decls
[i
].is_varying())
1295 /* From GL_EXT_transform_feedback:
1296 * A program will fail to link if:
1298 * * any two entries in the <varyings> array specify the same varying
1301 * We interpret this to mean "any two entries in the <varyings> array
1302 * specify the same varying variable and array index", since transform
1303 * feedback of arrays would be useless otherwise.
1305 for (unsigned j
= 0; j
< i
; ++j
) {
1306 if (decls
[j
].is_varying()) {
1307 if (tfeedback_decl::is_same(decls
[i
], decls
[j
])) {
1308 linker_error(prog
, "Transform feedback varying %s specified "
1309 "more than once.", varying_names
[i
]);
1320 cmp_xfb_offset(const void * x_generic
, const void * y_generic
)
1322 tfeedback_decl
*x
= (tfeedback_decl
*) x_generic
;
1323 tfeedback_decl
*y
= (tfeedback_decl
*) y_generic
;
1325 if (x
->get_buffer() != y
->get_buffer())
1326 return x
->get_buffer() - y
->get_buffer();
1327 return x
->get_offset() - y
->get_offset();
1331 * Store transform feedback location assignments into
1332 * prog->sh.LinkedTransformFeedback based on the data stored in
1335 * If an error occurs, the error is reported through linker_error() and false
1339 store_tfeedback_info(struct gl_context
*ctx
, struct gl_shader_program
*prog
,
1340 unsigned num_tfeedback_decls
,
1341 tfeedback_decl
*tfeedback_decls
, bool has_xfb_qualifiers
)
1343 if (!prog
->last_vert_prog
)
1346 /* Make sure MaxTransformFeedbackBuffers is less than 32 so the bitmask for
1347 * tracking the number of buffers doesn't overflow.
1349 assert(ctx
->Const
.MaxTransformFeedbackBuffers
< 32);
1351 bool separate_attribs_mode
=
1352 prog
->TransformFeedback
.BufferMode
== GL_SEPARATE_ATTRIBS
;
1354 struct gl_program
*xfb_prog
= prog
->last_vert_prog
;
1355 xfb_prog
->sh
.LinkedTransformFeedback
=
1356 rzalloc(xfb_prog
, struct gl_transform_feedback_info
);
1358 /* The xfb_offset qualifier does not have to be used in increasing order
1359 * however some drivers expect to receive the list of transform feedback
1360 * declarations in order so sort it now for convenience.
1362 if (has_xfb_qualifiers
) {
1363 qsort(tfeedback_decls
, num_tfeedback_decls
, sizeof(*tfeedback_decls
),
1367 xfb_prog
->sh
.LinkedTransformFeedback
->Varyings
=
1368 rzalloc_array(xfb_prog
, struct gl_transform_feedback_varying_info
,
1369 num_tfeedback_decls
);
1371 unsigned num_outputs
= 0;
1372 for (unsigned i
= 0; i
< num_tfeedback_decls
; ++i
) {
1373 if (tfeedback_decls
[i
].is_varying_written())
1374 num_outputs
+= tfeedback_decls
[i
].get_num_outputs();
1377 xfb_prog
->sh
.LinkedTransformFeedback
->Outputs
=
1378 rzalloc_array(xfb_prog
, struct gl_transform_feedback_output
,
1381 unsigned num_buffers
= 0;
1382 unsigned buffers
= 0;
1384 if (!has_xfb_qualifiers
&& separate_attribs_mode
) {
1385 /* GL_SEPARATE_ATTRIBS */
1386 for (unsigned i
= 0; i
< num_tfeedback_decls
; ++i
) {
1387 if (!tfeedback_decls
[i
].store(ctx
, prog
,
1388 xfb_prog
->sh
.LinkedTransformFeedback
,
1389 num_buffers
, num_buffers
, num_outputs
,
1390 NULL
, has_xfb_qualifiers
))
1393 buffers
|= 1 << num_buffers
;
1398 /* GL_INVERLEAVED_ATTRIBS */
1399 int buffer_stream_id
= -1;
1401 num_tfeedback_decls
? tfeedback_decls
[0].get_buffer() : 0;
1402 bool explicit_stride
[MAX_FEEDBACK_BUFFERS
] = { false };
1404 /* Apply any xfb_stride global qualifiers */
1405 if (has_xfb_qualifiers
) {
1406 for (unsigned j
= 0; j
< MAX_FEEDBACK_BUFFERS
; j
++) {
1407 if (prog
->TransformFeedback
.BufferStride
[j
]) {
1408 explicit_stride
[j
] = true;
1409 xfb_prog
->sh
.LinkedTransformFeedback
->Buffers
[j
].Stride
=
1410 prog
->TransformFeedback
.BufferStride
[j
] / 4;
1415 for (unsigned i
= 0; i
< num_tfeedback_decls
; ++i
) {
1416 if (has_xfb_qualifiers
&&
1417 buffer
!= tfeedback_decls
[i
].get_buffer()) {
1418 /* we have moved to the next buffer so reset stream id */
1419 buffer_stream_id
= -1;
1423 if (tfeedback_decls
[i
].is_next_buffer_separator()) {
1424 if (!tfeedback_decls
[i
].store(ctx
, prog
,
1425 xfb_prog
->sh
.LinkedTransformFeedback
,
1426 buffer
, num_buffers
, num_outputs
,
1427 explicit_stride
, has_xfb_qualifiers
))
1430 buffer_stream_id
= -1;
1434 if (has_xfb_qualifiers
) {
1435 buffer
= tfeedback_decls
[i
].get_buffer();
1437 buffer
= num_buffers
;
1440 if (tfeedback_decls
[i
].is_varying()) {
1441 if (buffer_stream_id
== -1) {
1442 /* First varying writing to this buffer: remember its stream */
1443 buffer_stream_id
= (int) tfeedback_decls
[i
].get_stream_id();
1445 /* Only mark a buffer as active when there is a varying
1446 * attached to it. This behaviour is based on a revised version
1447 * of section 13.2.2 of the GL 4.6 spec.
1449 buffers
|= 1 << buffer
;
1450 } else if (buffer_stream_id
!=
1451 (int) tfeedback_decls
[i
].get_stream_id()) {
1452 /* Varying writes to the same buffer from a different stream */
1454 "Transform feedback can't capture varyings belonging "
1455 "to different vertex streams in a single buffer. "
1456 "Varying %s writes to buffer from stream %u, other "
1457 "varyings in the same buffer write from stream %u.",
1458 tfeedback_decls
[i
].name(),
1459 tfeedback_decls
[i
].get_stream_id(),
1465 if (!tfeedback_decls
[i
].store(ctx
, prog
,
1466 xfb_prog
->sh
.LinkedTransformFeedback
,
1467 buffer
, num_buffers
, num_outputs
,
1468 explicit_stride
, has_xfb_qualifiers
))
1473 assert(xfb_prog
->sh
.LinkedTransformFeedback
->NumOutputs
== num_outputs
);
1475 xfb_prog
->sh
.LinkedTransformFeedback
->ActiveBuffers
= buffers
;
1482 * Data structure recording the relationship between outputs of one shader
1483 * stage (the "producer") and inputs of another (the "consumer").
1485 class varying_matches
1488 varying_matches(bool disable_varying_packing
, bool xfb_enabled
,
1489 bool enhanced_layouts_enabled
,
1490 gl_shader_stage producer_stage
,
1491 gl_shader_stage consumer_stage
);
1493 void record(ir_variable
*producer_var
, ir_variable
*consumer_var
);
1494 unsigned assign_locations(struct gl_shader_program
*prog
,
1495 uint8_t components
[],
1496 uint64_t reserved_slots
);
1497 void store_locations() const;
1500 bool is_varying_packing_safe(const glsl_type
*type
,
1501 const ir_variable
*var
) const;
1504 * If true, this driver disables varying packing, so all varyings need to
1505 * be aligned on slot boundaries, and take up a number of slots equal to
1506 * their number of matrix columns times their array size.
1508 * Packing may also be disabled because our current packing method is not
1509 * safe in SSO or versions of OpenGL where interpolation qualifiers are not
1510 * guaranteed to match across stages.
1512 const bool disable_varying_packing
;
1515 * If true, this driver has transform feedback enabled. The transform
1516 * feedback code requires at least some packing be done even when varying
1517 * packing is disabled, fortunately where transform feedback requires
1518 * packing it's safe to override the disabled setting. See
1519 * is_varying_packing_safe().
1521 const bool xfb_enabled
;
1523 const bool enhanced_layouts_enabled
;
1526 * Enum representing the order in which varyings are packed within a
1529 * Currently we pack vec4's first, then vec2's, then scalar values, then
1530 * vec3's. This order ensures that the only vectors that are at risk of
1531 * having to be "double parked" (split between two adjacent varying slots)
1534 enum packing_order_enum
{
1537 PACKING_ORDER_SCALAR
,
1541 static unsigned compute_packing_class(const ir_variable
*var
);
1542 static packing_order_enum
compute_packing_order(const ir_variable
*var
);
1543 static int match_comparator(const void *x_generic
, const void *y_generic
);
1544 static int xfb_comparator(const void *x_generic
, const void *y_generic
);
1547 * Structure recording the relationship between a single producer output
1548 * and a single consumer input.
1552 * Packing class for this varying, computed by compute_packing_class().
1554 unsigned packing_class
;
1557 * Packing order for this varying, computed by compute_packing_order().
1559 packing_order_enum packing_order
;
1560 unsigned num_components
;
1563 * The output variable in the producer stage.
1565 ir_variable
*producer_var
;
1568 * The input variable in the consumer stage.
1570 ir_variable
*consumer_var
;
1573 * The location which has been assigned for this varying. This is
1574 * expressed in multiples of a float, with the first generic varying
1575 * (i.e. the one referred to by VARYING_SLOT_VAR0) represented by the
1578 unsigned generic_location
;
1582 * The number of elements in the \c matches array that are currently in
1585 unsigned num_matches
;
1588 * The number of elements that were set aside for the \c matches array when
1591 unsigned matches_capacity
;
1593 gl_shader_stage producer_stage
;
1594 gl_shader_stage consumer_stage
;
1597 } /* anonymous namespace */
1599 varying_matches::varying_matches(bool disable_varying_packing
,
1601 bool enhanced_layouts_enabled
,
1602 gl_shader_stage producer_stage
,
1603 gl_shader_stage consumer_stage
)
1604 : disable_varying_packing(disable_varying_packing
),
1605 xfb_enabled(xfb_enabled
),
1606 enhanced_layouts_enabled(enhanced_layouts_enabled
),
1607 producer_stage(producer_stage
),
1608 consumer_stage(consumer_stage
)
1610 /* Note: this initial capacity is rather arbitrarily chosen to be large
1611 * enough for many cases without wasting an unreasonable amount of space.
1612 * varying_matches::record() will resize the array if there are more than
1613 * this number of varyings.
1615 this->matches_capacity
= 8;
1616 this->matches
= (match
*)
1617 malloc(sizeof(*this->matches
) * this->matches_capacity
);
1618 this->num_matches
= 0;
1622 varying_matches::~varying_matches()
1624 free(this->matches
);
1629 * Packing is always safe on individual arrays, structures, and matrices. It
1630 * is also safe if the varying is only used for transform feedback.
1633 varying_matches::is_varying_packing_safe(const glsl_type
*type
,
1634 const ir_variable
*var
) const
1636 if (consumer_stage
== MESA_SHADER_TESS_EVAL
||
1637 consumer_stage
== MESA_SHADER_TESS_CTRL
||
1638 producer_stage
== MESA_SHADER_TESS_CTRL
)
1641 return xfb_enabled
&& (type
->is_array() || type
->is_struct() ||
1642 type
->is_matrix() || var
->data
.is_xfb_only
);
1647 * Record the given producer/consumer variable pair in the list of variables
1648 * that should later be assigned locations.
1650 * It is permissible for \c consumer_var to be NULL (this happens if a
1651 * variable is output by the producer and consumed by transform feedback, but
1652 * not consumed by the consumer).
1654 * If \c producer_var has already been paired up with a consumer_var, or
1655 * producer_var is part of fixed pipeline functionality (and hence already has
1656 * a location assigned), this function has no effect.
1658 * Note: as a side effect this function may change the interpolation type of
1659 * \c producer_var, but only when the change couldn't possibly affect
1663 varying_matches::record(ir_variable
*producer_var
, ir_variable
*consumer_var
)
1665 assert(producer_var
!= NULL
|| consumer_var
!= NULL
);
1667 if ((producer_var
&& (!producer_var
->data
.is_unmatched_generic_inout
||
1668 producer_var
->data
.explicit_location
)) ||
1669 (consumer_var
&& (!consumer_var
->data
.is_unmatched_generic_inout
||
1670 consumer_var
->data
.explicit_location
))) {
1671 /* Either a location already exists for this variable (since it is part
1672 * of fixed functionality), or it has already been recorded as part of a
1678 bool needs_flat_qualifier
= consumer_var
== NULL
&&
1679 (producer_var
->type
->contains_integer() ||
1680 producer_var
->type
->contains_double());
1682 if (!disable_varying_packing
&&
1683 (needs_flat_qualifier
||
1684 (consumer_stage
!= MESA_SHADER_NONE
&& consumer_stage
!= MESA_SHADER_FRAGMENT
))) {
1685 /* Since this varying is not being consumed by the fragment shader, its
1686 * interpolation type varying cannot possibly affect rendering.
1687 * Also, this variable is non-flat and is (or contains) an integer
1689 * If the consumer stage is unknown, don't modify the interpolation
1690 * type as it could affect rendering later with separate shaders.
1692 * lower_packed_varyings requires all integer varyings to flat,
1693 * regardless of where they appear. We can trivially satisfy that
1694 * requirement by changing the interpolation type to flat here.
1697 producer_var
->data
.centroid
= false;
1698 producer_var
->data
.sample
= false;
1699 producer_var
->data
.interpolation
= INTERP_MODE_FLAT
;
1703 consumer_var
->data
.centroid
= false;
1704 consumer_var
->data
.sample
= false;
1705 consumer_var
->data
.interpolation
= INTERP_MODE_FLAT
;
1709 if (this->num_matches
== this->matches_capacity
) {
1710 this->matches_capacity
*= 2;
1711 this->matches
= (match
*)
1712 realloc(this->matches
,
1713 sizeof(*this->matches
) * this->matches_capacity
);
1716 /* We must use the consumer to compute the packing class because in GL4.4+
1717 * there is no guarantee interpolation qualifiers will match across stages.
1719 * From Section 4.5 (Interpolation Qualifiers) of the GLSL 4.30 spec:
1721 * "The type and presence of interpolation qualifiers of variables with
1722 * the same name declared in all linked shaders for the same cross-stage
1723 * interface must match, otherwise the link command will fail.
1725 * When comparing an output from one stage to an input of a subsequent
1726 * stage, the input and output don't match if their interpolation
1727 * qualifiers (or lack thereof) are not the same."
1729 * This text was also in at least revison 7 of the 4.40 spec but is no
1730 * longer in revision 9 and not in the 4.50 spec.
1732 const ir_variable
*const var
= (consumer_var
!= NULL
)
1733 ? consumer_var
: producer_var
;
1734 const gl_shader_stage stage
= (consumer_var
!= NULL
)
1735 ? consumer_stage
: producer_stage
;
1736 const glsl_type
*type
= get_varying_type(var
, stage
);
1738 if (producer_var
&& consumer_var
&&
1739 consumer_var
->data
.must_be_shader_input
) {
1740 producer_var
->data
.must_be_shader_input
= 1;
1743 this->matches
[this->num_matches
].packing_class
1744 = this->compute_packing_class(var
);
1745 this->matches
[this->num_matches
].packing_order
1746 = this->compute_packing_order(var
);
1747 if ((this->disable_varying_packing
&& !is_varying_packing_safe(type
, var
)) ||
1748 var
->data
.must_be_shader_input
) {
1749 unsigned slots
= type
->count_attribute_slots(false);
1750 this->matches
[this->num_matches
].num_components
= slots
* 4;
1752 this->matches
[this->num_matches
].num_components
1753 = type
->component_slots();
1756 this->matches
[this->num_matches
].producer_var
= producer_var
;
1757 this->matches
[this->num_matches
].consumer_var
= consumer_var
;
1758 this->num_matches
++;
1760 producer_var
->data
.is_unmatched_generic_inout
= 0;
1762 consumer_var
->data
.is_unmatched_generic_inout
= 0;
1767 * Choose locations for all of the variable matches that were previously
1768 * passed to varying_matches::record().
1769 * \param components returns array[slot] of number of components used
1770 * per slot (1, 2, 3 or 4)
1771 * \param reserved_slots bitmask indicating which varying slots are already
1773 * \return number of slots (4-element vectors) allocated
1776 varying_matches::assign_locations(struct gl_shader_program
*prog
,
1777 uint8_t components
[],
1778 uint64_t reserved_slots
)
1780 /* If packing has been disabled then we cannot safely sort the varyings by
1781 * class as it may mean we are using a version of OpenGL where
1782 * interpolation qualifiers are not guaranteed to be matching across
1783 * shaders, sorting in this case could result in mismatching shader
1785 * When packing is disabled the sort orders varyings used by transform
1786 * feedback first, but also depends on *undefined behaviour* of qsort to
1787 * reverse the order of the varyings. See: xfb_comparator().
1789 if (!this->disable_varying_packing
) {
1790 /* Sort varying matches into an order that makes them easy to pack. */
1791 qsort(this->matches
, this->num_matches
, sizeof(*this->matches
),
1792 &varying_matches::match_comparator
);
1794 /* Only sort varyings that are only used by transform feedback. */
1795 qsort(this->matches
, this->num_matches
, sizeof(*this->matches
),
1796 &varying_matches::xfb_comparator
);
1799 unsigned generic_location
= 0;
1800 unsigned generic_patch_location
= MAX_VARYING
*4;
1801 bool previous_var_xfb_only
= false;
1802 unsigned previous_packing_class
= ~0u;
1804 /* For tranform feedback separate mode, we know the number of attributes
1805 * is <= the number of buffers. So packing isn't critical. In fact,
1806 * packing vec3 attributes can cause trouble because splitting a vec3
1807 * effectively creates an additional transform feedback output. The
1808 * extra TFB output may exceed device driver limits.
1810 const bool dont_pack_vec3
=
1811 (prog
->TransformFeedback
.BufferMode
== GL_SEPARATE_ATTRIBS
&&
1812 prog
->TransformFeedback
.NumVarying
> 0);
1814 for (unsigned i
= 0; i
< this->num_matches
; i
++) {
1815 unsigned *location
= &generic_location
;
1816 const ir_variable
*var
;
1817 const glsl_type
*type
;
1818 bool is_vertex_input
= false;
1820 if (matches
[i
].consumer_var
) {
1821 var
= matches
[i
].consumer_var
;
1822 type
= get_varying_type(var
, consumer_stage
);
1823 if (consumer_stage
== MESA_SHADER_VERTEX
)
1824 is_vertex_input
= true;
1826 var
= matches
[i
].producer_var
;
1827 type
= get_varying_type(var
, producer_stage
);
1830 if (var
->data
.patch
)
1831 location
= &generic_patch_location
;
1833 /* Advance to the next slot if this varying has a different packing
1834 * class than the previous one, and we're not already on a slot
1837 * Also advance to the next slot if packing is disabled. This makes sure
1838 * we don't assign varyings the same locations which is possible
1839 * because we still pack individual arrays, records and matrices even
1840 * when packing is disabled. Note we don't advance to the next slot if
1841 * we can pack varyings together that are only used for transform
1844 if (var
->data
.must_be_shader_input
||
1845 (this->disable_varying_packing
&&
1846 !(previous_var_xfb_only
&& var
->data
.is_xfb_only
)) ||
1847 (previous_packing_class
!= this->matches
[i
].packing_class
) ||
1848 (this->matches
[i
].packing_order
== PACKING_ORDER_VEC3
&&
1850 *location
= ALIGN(*location
, 4);
1853 previous_var_xfb_only
= var
->data
.is_xfb_only
;
1854 previous_packing_class
= this->matches
[i
].packing_class
;
1856 /* The number of components taken up by this variable. For vertex shader
1857 * inputs, we use the number of slots * 4, as they have different
1860 unsigned num_components
= is_vertex_input
?
1861 type
->count_attribute_slots(is_vertex_input
) * 4 :
1862 this->matches
[i
].num_components
;
1864 /* The last slot for this variable, inclusive. */
1865 unsigned slot_end
= *location
+ num_components
- 1;
1867 /* FIXME: We could be smarter in the below code and loop back over
1868 * trying to fill any locations that we skipped because we couldn't pack
1869 * the varying between an explicit location. For now just let the user
1870 * hit the linking error if we run out of room and suggest they use
1871 * explicit locations.
1873 while (slot_end
< MAX_VARYING
* 4u) {
1874 const unsigned slots
= (slot_end
/ 4u) - (*location
/ 4u) + 1;
1875 const uint64_t slot_mask
= ((1ull << slots
) - 1) << (*location
/ 4u);
1879 if ((reserved_slots
& slot_mask
) == 0) {
1883 *location
= ALIGN(*location
+ 1, 4);
1884 slot_end
= *location
+ num_components
- 1;
1887 if (!var
->data
.patch
&& slot_end
>= MAX_VARYING
* 4u) {
1888 linker_error(prog
, "insufficient contiguous locations available for "
1889 "%s it is possible an array or struct could not be "
1890 "packed between varyings with explicit locations. Try "
1891 "using an explicit location for arrays and structs.",
1895 if (slot_end
< MAX_VARYINGS_INCL_PATCH
* 4u) {
1896 for (unsigned j
= *location
/ 4u; j
< slot_end
/ 4u; j
++)
1898 components
[slot_end
/ 4u] = (slot_end
& 3) + 1;
1901 this->matches
[i
].generic_location
= *location
;
1903 *location
= slot_end
+ 1;
1906 return (generic_location
+ 3) / 4;
1911 * Update the producer and consumer shaders to reflect the locations
1912 * assignments that were made by varying_matches::assign_locations().
1915 varying_matches::store_locations() const
1917 /* Check is location needs to be packed with lower_packed_varyings() or if
1918 * we can just use ARB_enhanced_layouts packing.
1920 bool pack_loc
[MAX_VARYINGS_INCL_PATCH
] = { 0 };
1921 const glsl_type
*loc_type
[MAX_VARYINGS_INCL_PATCH
][4] = { {NULL
, NULL
} };
1923 for (unsigned i
= 0; i
< this->num_matches
; i
++) {
1924 ir_variable
*producer_var
= this->matches
[i
].producer_var
;
1925 ir_variable
*consumer_var
= this->matches
[i
].consumer_var
;
1926 unsigned generic_location
= this->matches
[i
].generic_location
;
1927 unsigned slot
= generic_location
/ 4;
1928 unsigned offset
= generic_location
% 4;
1931 producer_var
->data
.location
= VARYING_SLOT_VAR0
+ slot
;
1932 producer_var
->data
.location_frac
= offset
;
1936 assert(consumer_var
->data
.location
== -1);
1937 consumer_var
->data
.location
= VARYING_SLOT_VAR0
+ slot
;
1938 consumer_var
->data
.location_frac
= offset
;
1941 /* Find locations suitable for native packing via
1942 * ARB_enhanced_layouts.
1944 if (producer_var
&& consumer_var
) {
1945 if (enhanced_layouts_enabled
) {
1946 const glsl_type
*type
=
1947 get_varying_type(producer_var
, producer_stage
);
1948 if (type
->is_array() || type
->is_matrix() || type
->is_struct() ||
1949 type
->is_double()) {
1950 unsigned comp_slots
= type
->component_slots() + offset
;
1951 unsigned slots
= comp_slots
/ 4;
1955 for (unsigned j
= 0; j
< slots
; j
++) {
1956 pack_loc
[slot
+ j
] = true;
1958 } else if (offset
+ type
->vector_elements
> 4) {
1959 pack_loc
[slot
] = true;
1960 pack_loc
[slot
+ 1] = true;
1962 loc_type
[slot
][offset
] = type
;
1968 /* Attempt to use ARB_enhanced_layouts for more efficient packing if
1971 if (enhanced_layouts_enabled
) {
1972 for (unsigned i
= 0; i
< this->num_matches
; i
++) {
1973 ir_variable
*producer_var
= this->matches
[i
].producer_var
;
1974 ir_variable
*consumer_var
= this->matches
[i
].consumer_var
;
1975 unsigned generic_location
= this->matches
[i
].generic_location
;
1976 unsigned slot
= generic_location
/ 4;
1978 if (pack_loc
[slot
] || !producer_var
|| !consumer_var
)
1981 const glsl_type
*type
=
1982 get_varying_type(producer_var
, producer_stage
);
1983 bool type_match
= true;
1984 for (unsigned j
= 0; j
< 4; j
++) {
1985 if (loc_type
[slot
][j
]) {
1986 if (type
->base_type
!= loc_type
[slot
][j
]->base_type
)
1992 producer_var
->data
.explicit_location
= 1;
1993 consumer_var
->data
.explicit_location
= 1;
1994 producer_var
->data
.explicit_component
= 1;
1995 consumer_var
->data
.explicit_component
= 1;
2003 * Compute the "packing class" of the given varying. This is an unsigned
2004 * integer with the property that two variables in the same packing class can
2005 * be safely backed into the same vec4.
2008 varying_matches::compute_packing_class(const ir_variable
*var
)
2010 /* Without help from the back-end, there is no way to pack together
2011 * variables with different interpolation types, because
2012 * lower_packed_varyings must choose exactly one interpolation type for
2013 * each packed varying it creates.
2015 * However, we can safely pack together floats, ints, and uints, because:
2017 * - varyings of base type "int" and "uint" must use the "flat"
2018 * interpolation type, which can only occur in GLSL 1.30 and above.
2020 * - On platforms that support GLSL 1.30 and above, lower_packed_varyings
2021 * can store flat floats as ints without losing any information (using
2022 * the ir_unop_bitcast_* opcodes).
2024 * Therefore, the packing class depends only on the interpolation type.
2026 const unsigned interp
= var
->is_interpolation_flat()
2027 ? unsigned(INTERP_MODE_FLAT
) : var
->data
.interpolation
;
2029 assert(interp
< (1 << 3));
2031 const unsigned packing_class
= (interp
<< 0) |
2032 (var
->data
.centroid
<< 3) |
2033 (var
->data
.sample
<< 4) |
2034 (var
->data
.patch
<< 5) |
2035 (var
->data
.must_be_shader_input
<< 6);
2037 return packing_class
;
2042 * Compute the "packing order" of the given varying. This is a sort key we
2043 * use to determine when to attempt to pack the given varying relative to
2044 * other varyings in the same packing class.
2046 varying_matches::packing_order_enum
2047 varying_matches::compute_packing_order(const ir_variable
*var
)
2049 const glsl_type
*element_type
= var
->type
;
2051 while (element_type
->is_array()) {
2052 element_type
= element_type
->fields
.array
;
2055 switch (element_type
->component_slots() % 4) {
2056 case 1: return PACKING_ORDER_SCALAR
;
2057 case 2: return PACKING_ORDER_VEC2
;
2058 case 3: return PACKING_ORDER_VEC3
;
2059 case 0: return PACKING_ORDER_VEC4
;
2061 assert(!"Unexpected value of vector_elements");
2062 return PACKING_ORDER_VEC4
;
2068 * Comparison function passed to qsort() to sort varyings by packing_class and
2069 * then by packing_order.
2072 varying_matches::match_comparator(const void *x_generic
, const void *y_generic
)
2074 const match
*x
= (const match
*) x_generic
;
2075 const match
*y
= (const match
*) y_generic
;
2077 if (x
->packing_class
!= y
->packing_class
)
2078 return x
->packing_class
- y
->packing_class
;
2079 return x
->packing_order
- y
->packing_order
;
2084 * Comparison function passed to qsort() to sort varyings used only by
2085 * transform feedback when packing of other varyings is disabled.
2088 varying_matches::xfb_comparator(const void *x_generic
, const void *y_generic
)
2090 const match
*x
= (const match
*) x_generic
;
2092 if (x
->producer_var
!= NULL
&& x
->producer_var
->data
.is_xfb_only
)
2093 return match_comparator(x_generic
, y_generic
);
2095 /* FIXME: When the comparator returns 0 it means the elements being
2096 * compared are equivalent. However the qsort documentation says:
2098 * "The order of equivalent elements is undefined."
2100 * In practice the sort ends up reversing the order of the varyings which
2101 * means locations are also assigned in this reversed order and happens to
2102 * be what we want. This is also whats happening in
2103 * varying_matches::match_comparator().
2110 * Is the given variable a varying variable to be counted against the
2111 * limit in ctx->Const.MaxVarying?
2112 * This includes variables such as texcoords, colors and generic
2113 * varyings, but excludes variables such as gl_FrontFacing and gl_FragCoord.
2116 var_counts_against_varying_limit(gl_shader_stage stage
, const ir_variable
*var
)
2118 /* Only fragment shaders will take a varying variable as an input */
2119 if (stage
== MESA_SHADER_FRAGMENT
&&
2120 var
->data
.mode
== ir_var_shader_in
) {
2121 switch (var
->data
.location
) {
2122 case VARYING_SLOT_POS
:
2123 case VARYING_SLOT_FACE
:
2124 case VARYING_SLOT_PNTC
:
2135 * Visitor class that generates tfeedback_candidate structs describing all
2136 * possible targets of transform feedback.
2138 * tfeedback_candidate structs are stored in the hash table
2139 * tfeedback_candidates, which is passed to the constructor. This hash table
2140 * maps varying names to instances of the tfeedback_candidate struct.
2142 class tfeedback_candidate_generator
: public program_resource_visitor
2145 tfeedback_candidate_generator(void *mem_ctx
,
2146 hash_table
*tfeedback_candidates
,
2147 gl_shader_stage stage
)
2149 tfeedback_candidates(tfeedback_candidates
),
2156 void process(ir_variable
*var
)
2158 /* All named varying interface blocks should be flattened by now */
2159 assert(!var
->is_interface_instance());
2160 assert(var
->data
.mode
== ir_var_shader_out
);
2162 this->toplevel_var
= var
;
2163 this->varying_floats
= 0;
2164 const glsl_type
*t
=
2165 var
->data
.from_named_ifc_block
? var
->get_interface_type() : var
->type
;
2166 if (!var
->data
.patch
&& stage
== MESA_SHADER_TESS_CTRL
) {
2167 assert(t
->is_array());
2168 t
= t
->fields
.array
;
2170 program_resource_visitor::process(var
, t
, false);
2174 virtual void visit_field(const glsl_type
*type
, const char *name
,
2175 bool /* row_major */,
2176 const glsl_type
* /* record_type */,
2177 const enum glsl_interface_packing
,
2178 bool /* last_field */)
2180 assert(!type
->without_array()->is_struct());
2181 assert(!type
->without_array()->is_interface());
2183 tfeedback_candidate
*candidate
2184 = rzalloc(this->mem_ctx
, tfeedback_candidate
);
2185 candidate
->toplevel_var
= this->toplevel_var
;
2186 candidate
->type
= type
;
2187 candidate
->offset
= this->varying_floats
;
2188 _mesa_hash_table_insert(this->tfeedback_candidates
,
2189 ralloc_strdup(this->mem_ctx
, name
),
2191 this->varying_floats
+= type
->component_slots();
2195 * Memory context used to allocate hash table keys and values.
2197 void * const mem_ctx
;
2200 * Hash table in which tfeedback_candidate objects should be stored.
2202 hash_table
* const tfeedback_candidates
;
2204 gl_shader_stage stage
;
2207 * Pointer to the toplevel variable that is being traversed.
2209 ir_variable
*toplevel_var
;
2212 * Total number of varying floats that have been visited so far. This is
2213 * used to determine the offset to each varying within the toplevel
2216 unsigned varying_floats
;
2223 populate_consumer_input_sets(void *mem_ctx
, exec_list
*ir
,
2224 hash_table
*consumer_inputs
,
2225 hash_table
*consumer_interface_inputs
,
2226 ir_variable
*consumer_inputs_with_locations
[VARYING_SLOT_TESS_MAX
])
2228 memset(consumer_inputs_with_locations
,
2230 sizeof(consumer_inputs_with_locations
[0]) * VARYING_SLOT_TESS_MAX
);
2232 foreach_in_list(ir_instruction
, node
, ir
) {
2233 ir_variable
*const input_var
= node
->as_variable();
2235 if (input_var
!= NULL
&& input_var
->data
.mode
== ir_var_shader_in
) {
2236 /* All interface blocks should have been lowered by this point */
2237 assert(!input_var
->type
->is_interface());
2239 if (input_var
->data
.explicit_location
) {
2240 /* assign_varying_locations only cares about finding the
2241 * ir_variable at the start of a contiguous location block.
2243 * - For !producer, consumer_inputs_with_locations isn't used.
2245 * - For !consumer, consumer_inputs_with_locations is empty.
2247 * For consumer && producer, if you were trying to set some
2248 * ir_variable to the middle of a location block on the other side
2249 * of producer/consumer, cross_validate_outputs_to_inputs() should
2250 * be link-erroring due to either type mismatch or location
2251 * overlaps. If the variables do match up, then they've got a
2252 * matching data.location and you only looked at
2253 * consumer_inputs_with_locations[var->data.location], not any
2254 * following entries for the array/structure.
2256 consumer_inputs_with_locations
[input_var
->data
.location
] =
2258 } else if (input_var
->get_interface_type() != NULL
) {
2259 char *const iface_field_name
=
2260 ralloc_asprintf(mem_ctx
, "%s.%s",
2261 input_var
->get_interface_type()->without_array()->name
,
2263 _mesa_hash_table_insert(consumer_interface_inputs
,
2264 iface_field_name
, input_var
);
2266 _mesa_hash_table_insert(consumer_inputs
,
2267 ralloc_strdup(mem_ctx
, input_var
->name
),
2275 * Find a variable from the consumer that "matches" the specified variable
2277 * This function only finds inputs with names that match. There is no
2278 * validation (here) that the types, etc. are compatible.
2281 get_matching_input(void *mem_ctx
,
2282 const ir_variable
*output_var
,
2283 hash_table
*consumer_inputs
,
2284 hash_table
*consumer_interface_inputs
,
2285 ir_variable
*consumer_inputs_with_locations
[VARYING_SLOT_TESS_MAX
])
2287 ir_variable
*input_var
;
2289 if (output_var
->data
.explicit_location
) {
2290 input_var
= consumer_inputs_with_locations
[output_var
->data
.location
];
2291 } else if (output_var
->get_interface_type() != NULL
) {
2292 char *const iface_field_name
=
2293 ralloc_asprintf(mem_ctx
, "%s.%s",
2294 output_var
->get_interface_type()->without_array()->name
,
2296 hash_entry
*entry
= _mesa_hash_table_search(consumer_interface_inputs
, iface_field_name
);
2297 input_var
= entry
? (ir_variable
*) entry
->data
: NULL
;
2299 hash_entry
*entry
= _mesa_hash_table_search(consumer_inputs
, output_var
->name
);
2300 input_var
= entry
? (ir_variable
*) entry
->data
: NULL
;
2303 return (input_var
== NULL
|| input_var
->data
.mode
!= ir_var_shader_in
)
2310 io_variable_cmp(const void *_a
, const void *_b
)
2312 const ir_variable
*const a
= *(const ir_variable
**) _a
;
2313 const ir_variable
*const b
= *(const ir_variable
**) _b
;
2315 if (a
->data
.explicit_location
&& b
->data
.explicit_location
)
2316 return b
->data
.location
- a
->data
.location
;
2318 if (a
->data
.explicit_location
&& !b
->data
.explicit_location
)
2321 if (!a
->data
.explicit_location
&& b
->data
.explicit_location
)
2324 return -strcmp(a
->name
, b
->name
);
2328 * Sort the shader IO variables into canonical order
2331 canonicalize_shader_io(exec_list
*ir
, enum ir_variable_mode io_mode
)
2333 ir_variable
*var_table
[MAX_PROGRAM_OUTPUTS
* 4];
2334 unsigned num_variables
= 0;
2336 foreach_in_list(ir_instruction
, node
, ir
) {
2337 ir_variable
*const var
= node
->as_variable();
2339 if (var
== NULL
|| var
->data
.mode
!= io_mode
)
2342 /* If we have already encountered more I/O variables that could
2343 * successfully link, bail.
2345 if (num_variables
== ARRAY_SIZE(var_table
))
2348 var_table
[num_variables
++] = var
;
2351 if (num_variables
== 0)
2354 /* Sort the list in reverse order (io_variable_cmp handles this). Later
2355 * we're going to push the variables on to the IR list as a stack, so we
2356 * want the last variable (in canonical order) to be first in the list.
2358 qsort(var_table
, num_variables
, sizeof(var_table
[0]), io_variable_cmp
);
2360 /* Remove the variable from it's current location in the IR, and put it at
2363 for (unsigned i
= 0; i
< num_variables
; i
++) {
2364 var_table
[i
]->remove();
2365 ir
->push_head(var_table
[i
]);
2370 * Generate a bitfield map of the explicit locations for shader varyings.
2372 * Note: For Tessellation shaders we are sitting right on the limits of the
2373 * 64 bit map. Per-vertex and per-patch both have separate location domains
2374 * with a max of MAX_VARYING.
2377 reserved_varying_slot(struct gl_linked_shader
*stage
,
2378 ir_variable_mode io_mode
)
2380 assert(io_mode
== ir_var_shader_in
|| io_mode
== ir_var_shader_out
);
2381 /* Avoid an overflow of the returned value */
2382 assert(MAX_VARYINGS_INCL_PATCH
<= 64);
2390 foreach_in_list(ir_instruction
, node
, stage
->ir
) {
2391 ir_variable
*const var
= node
->as_variable();
2393 if (var
== NULL
|| var
->data
.mode
!= io_mode
||
2394 !var
->data
.explicit_location
||
2395 var
->data
.location
< VARYING_SLOT_VAR0
)
2398 var_slot
= var
->data
.location
- VARYING_SLOT_VAR0
;
2400 unsigned num_elements
= get_varying_type(var
, stage
->Stage
)
2401 ->count_attribute_slots(io_mode
== ir_var_shader_in
&&
2402 stage
->Stage
== MESA_SHADER_VERTEX
);
2403 for (unsigned i
= 0; i
< num_elements
; i
++) {
2404 if (var_slot
>= 0 && var_slot
< MAX_VARYINGS_INCL_PATCH
)
2405 slots
|= UINT64_C(1) << var_slot
;
2415 * Assign locations for all variables that are produced in one pipeline stage
2416 * (the "producer") and consumed in the next stage (the "consumer").
2418 * Variables produced by the producer may also be consumed by transform
2421 * \param num_tfeedback_decls is the number of declarations indicating
2422 * variables that may be consumed by transform feedback.
2424 * \param tfeedback_decls is a pointer to an array of tfeedback_decl objects
2425 * representing the result of parsing the strings passed to
2426 * glTransformFeedbackVaryings(). assign_location() will be called for
2427 * each of these objects that matches one of the outputs of the
2430 * When num_tfeedback_decls is nonzero, it is permissible for the consumer to
2431 * be NULL. In this case, varying locations are assigned solely based on the
2432 * requirements of transform feedback.
2435 assign_varying_locations(struct gl_context
*ctx
,
2437 struct gl_shader_program
*prog
,
2438 gl_linked_shader
*producer
,
2439 gl_linked_shader
*consumer
,
2440 unsigned num_tfeedback_decls
,
2441 tfeedback_decl
*tfeedback_decls
,
2442 const uint64_t reserved_slots
)
2444 /* Tessellation shaders treat inputs and outputs as shared memory and can
2445 * access inputs and outputs of other invocations.
2446 * Therefore, they can't be lowered to temps easily (and definitely not
2449 bool unpackable_tess
=
2450 (consumer
&& consumer
->Stage
== MESA_SHADER_TESS_EVAL
) ||
2451 (consumer
&& consumer
->Stage
== MESA_SHADER_TESS_CTRL
) ||
2452 (producer
&& producer
->Stage
== MESA_SHADER_TESS_CTRL
);
2454 /* Transform feedback code assumes varying arrays are packed, so if the
2455 * driver has disabled varying packing, make sure to at least enable
2456 * packing required by transform feedback.
2459 ctx
->Extensions
.EXT_transform_feedback
&& !unpackable_tess
;
2461 /* Disable packing on outward facing interfaces for SSO because in ES we
2462 * need to retain the unpacked varying information for draw time
2465 * Packing is still enabled on individual arrays, structs, and matrices as
2466 * these are required by the transform feedback code and it is still safe
2467 * to do so. We also enable packing when a varying is only used for
2468 * transform feedback and its not a SSO.
2470 bool disable_varying_packing
=
2471 ctx
->Const
.DisableVaryingPacking
|| unpackable_tess
;
2472 if (prog
->SeparateShader
&& (producer
== NULL
|| consumer
== NULL
))
2473 disable_varying_packing
= true;
2475 varying_matches
matches(disable_varying_packing
, xfb_enabled
,
2476 ctx
->Extensions
.ARB_enhanced_layouts
,
2477 producer
? producer
->Stage
: MESA_SHADER_NONE
,
2478 consumer
? consumer
->Stage
: MESA_SHADER_NONE
);
2479 hash_table
*tfeedback_candidates
=
2480 _mesa_hash_table_create(NULL
, _mesa_key_hash_string
,
2481 _mesa_key_string_equal
);
2482 hash_table
*consumer_inputs
=
2483 _mesa_hash_table_create(NULL
, _mesa_key_hash_string
,
2484 _mesa_key_string_equal
);
2485 hash_table
*consumer_interface_inputs
=
2486 _mesa_hash_table_create(NULL
, _mesa_key_hash_string
,
2487 _mesa_key_string_equal
);
2488 ir_variable
*consumer_inputs_with_locations
[VARYING_SLOT_TESS_MAX
] = {
2492 unsigned consumer_vertices
= 0;
2493 if (consumer
&& consumer
->Stage
== MESA_SHADER_GEOMETRY
)
2494 consumer_vertices
= prog
->Geom
.VerticesIn
;
2496 /* Operate in a total of four passes.
2498 * 1. Sort inputs / outputs into a canonical order. This is necessary so
2499 * that inputs / outputs of separable shaders will be assigned
2500 * predictable locations regardless of the order in which declarations
2501 * appeared in the shader source.
2503 * 2. Assign locations for any matching inputs and outputs.
2505 * 3. Mark output variables in the producer that do not have locations as
2506 * not being outputs. This lets the optimizer eliminate them.
2508 * 4. Mark input variables in the consumer that do not have locations as
2509 * not being inputs. This lets the optimizer eliminate them.
2512 canonicalize_shader_io(consumer
->ir
, ir_var_shader_in
);
2515 canonicalize_shader_io(producer
->ir
, ir_var_shader_out
);
2518 linker::populate_consumer_input_sets(mem_ctx
, consumer
->ir
,
2520 consumer_interface_inputs
,
2521 consumer_inputs_with_locations
);
2524 foreach_in_list(ir_instruction
, node
, producer
->ir
) {
2525 ir_variable
*const output_var
= node
->as_variable();
2527 if (output_var
== NULL
|| output_var
->data
.mode
!= ir_var_shader_out
)
2530 /* Only geometry shaders can use non-zero streams */
2531 assert(output_var
->data
.stream
== 0 ||
2532 (output_var
->data
.stream
< MAX_VERTEX_STREAMS
&&
2533 producer
->Stage
== MESA_SHADER_GEOMETRY
));
2535 if (num_tfeedback_decls
> 0) {
2536 tfeedback_candidate_generator
g(mem_ctx
, tfeedback_candidates
, producer
->Stage
);
2537 /* From OpenGL 4.6 (Core Profile) spec, section 11.1.2.1
2538 * ("Vertex Shader Variables / Output Variables")
2540 * "Each program object can specify a set of output variables from
2541 * one shader to be recorded in transform feedback mode (see
2542 * section 13.3). The variables that can be recorded are those
2543 * emitted by the first active shader, in order, from the
2547 * * tessellation evaluation shader
2548 * * tessellation control shader
2551 * But on OpenGL ES 3.2, section 11.1.2.1 ("Vertex Shader
2552 * Variables / Output Variables") tessellation control shader is
2553 * not included in the stages list.
2555 if (!prog
->IsES
|| producer
->Stage
!= MESA_SHADER_TESS_CTRL
) {
2556 g
.process(output_var
);
2560 ir_variable
*const input_var
=
2561 linker::get_matching_input(mem_ctx
, output_var
, consumer_inputs
,
2562 consumer_interface_inputs
,
2563 consumer_inputs_with_locations
);
2565 /* If a matching input variable was found, add this output (and the
2566 * input) to the set. If this is a separable program and there is no
2567 * consumer stage, add the output.
2569 * Always add TCS outputs. They are shared by all invocations
2570 * within a patch and can be used as shared memory.
2572 if (input_var
|| (prog
->SeparateShader
&& consumer
== NULL
) ||
2573 producer
->Stage
== MESA_SHADER_TESS_CTRL
) {
2574 matches
.record(output_var
, input_var
);
2577 /* Only stream 0 outputs can be consumed in the next stage */
2578 if (input_var
&& output_var
->data
.stream
!= 0) {
2579 linker_error(prog
, "output %s is assigned to stream=%d but "
2580 "is linked to an input, which requires stream=0",
2581 output_var
->name
, output_var
->data
.stream
);
2586 /* If there's no producer stage, then this must be a separable program.
2587 * For example, we may have a program that has just a fragment shader.
2588 * Later this program will be used with some arbitrary vertex (or
2589 * geometry) shader program. This means that locations must be assigned
2590 * for all the inputs.
2592 foreach_in_list(ir_instruction
, node
, consumer
->ir
) {
2593 ir_variable
*const input_var
= node
->as_variable();
2594 if (input_var
&& input_var
->data
.mode
== ir_var_shader_in
) {
2595 matches
.record(NULL
, input_var
);
2600 for (unsigned i
= 0; i
< num_tfeedback_decls
; ++i
) {
2601 if (!tfeedback_decls
[i
].is_varying())
2604 const tfeedback_candidate
*matched_candidate
2605 = tfeedback_decls
[i
].find_candidate(prog
, tfeedback_candidates
);
2607 if (matched_candidate
== NULL
) {
2608 _mesa_hash_table_destroy(tfeedback_candidates
, NULL
);
2612 /* Mark xfb varyings as always active */
2613 matched_candidate
->toplevel_var
->data
.always_active_io
= 1;
2615 /* Mark any corresponding inputs as always active also. We must do this
2616 * because we have a NIR pass that lowers vectors to scalars and another
2617 * that removes unused varyings.
2618 * We don't split varyings marked as always active because there is no
2619 * point in doing so. This means we need to mark both sides of the
2620 * interface as always active otherwise we will have a mismatch and
2621 * start removing things we shouldn't.
2623 ir_variable
*const input_var
=
2624 linker::get_matching_input(mem_ctx
, matched_candidate
->toplevel_var
,
2626 consumer_interface_inputs
,
2627 consumer_inputs_with_locations
);
2629 input_var
->data
.always_active_io
= 1;
2631 if (matched_candidate
->toplevel_var
->data
.is_unmatched_generic_inout
) {
2632 matched_candidate
->toplevel_var
->data
.is_xfb_only
= 1;
2633 matches
.record(matched_candidate
->toplevel_var
, NULL
);
2637 _mesa_hash_table_destroy(consumer_inputs
, NULL
);
2638 _mesa_hash_table_destroy(consumer_interface_inputs
, NULL
);
2640 uint8_t components
[MAX_VARYINGS_INCL_PATCH
] = {0};
2641 const unsigned slots_used
= matches
.assign_locations(
2642 prog
, components
, reserved_slots
);
2643 matches
.store_locations();
2645 for (unsigned i
= 0; i
< num_tfeedback_decls
; ++i
) {
2646 if (tfeedback_decls
[i
].is_varying()) {
2647 if (!tfeedback_decls
[i
].assign_location(ctx
, prog
)) {
2648 _mesa_hash_table_destroy(tfeedback_candidates
, NULL
);
2653 _mesa_hash_table_destroy(tfeedback_candidates
, NULL
);
2655 if (consumer
&& producer
) {
2656 foreach_in_list(ir_instruction
, node
, consumer
->ir
) {
2657 ir_variable
*const var
= node
->as_variable();
2659 if (var
&& var
->data
.mode
== ir_var_shader_in
&&
2660 var
->data
.is_unmatched_generic_inout
) {
2661 if (!prog
->IsES
&& prog
->data
->Version
<= 120) {
2662 /* On page 25 (page 31 of the PDF) of the GLSL 1.20 spec:
2664 * Only those varying variables used (i.e. read) in
2665 * the fragment shader executable must be written to
2666 * by the vertex shader executable; declaring
2667 * superfluous varying variables in a vertex shader is
2670 * We interpret this text as meaning that the VS must
2671 * write the variable for the FS to read it. See
2672 * "glsl1-varying read but not written" in piglit.
2674 linker_error(prog
, "%s shader varying %s not written "
2676 _mesa_shader_stage_to_string(consumer
->Stage
),
2678 _mesa_shader_stage_to_string(producer
->Stage
));
2680 linker_warning(prog
, "%s shader varying %s not written "
2682 _mesa_shader_stage_to_string(consumer
->Stage
),
2684 _mesa_shader_stage_to_string(producer
->Stage
));
2689 /* Now that validation is done its safe to remove unused varyings. As
2690 * we have both a producer and consumer its safe to remove unused
2691 * varyings even if the program is a SSO because the stages are being
2692 * linked together i.e. we have a multi-stage SSO.
2694 remove_unused_shader_inputs_and_outputs(false, producer
,
2696 remove_unused_shader_inputs_and_outputs(false, consumer
,
2701 lower_packed_varyings(mem_ctx
, slots_used
, components
, ir_var_shader_out
,
2702 0, producer
, disable_varying_packing
,
2707 lower_packed_varyings(mem_ctx
, slots_used
, components
, ir_var_shader_in
,
2708 consumer_vertices
, consumer
,
2709 disable_varying_packing
, xfb_enabled
);
2716 check_against_output_limit(struct gl_context
*ctx
,
2717 struct gl_shader_program
*prog
,
2718 gl_linked_shader
*producer
,
2719 unsigned num_explicit_locations
)
2721 unsigned output_vectors
= num_explicit_locations
;
2723 foreach_in_list(ir_instruction
, node
, producer
->ir
) {
2724 ir_variable
*const var
= node
->as_variable();
2726 if (var
&& !var
->data
.explicit_location
&&
2727 var
->data
.mode
== ir_var_shader_out
&&
2728 var_counts_against_varying_limit(producer
->Stage
, var
)) {
2729 /* outputs for fragment shader can't be doubles */
2730 output_vectors
+= var
->type
->count_attribute_slots(false);
2734 assert(producer
->Stage
!= MESA_SHADER_FRAGMENT
);
2735 unsigned max_output_components
=
2736 ctx
->Const
.Program
[producer
->Stage
].MaxOutputComponents
;
2738 const unsigned output_components
= output_vectors
* 4;
2739 if (output_components
> max_output_components
) {
2740 if (ctx
->API
== API_OPENGLES2
|| prog
->IsES
)
2741 linker_error(prog
, "%s shader uses too many output vectors "
2743 _mesa_shader_stage_to_string(producer
->Stage
),
2745 max_output_components
/ 4);
2747 linker_error(prog
, "%s shader uses too many output components "
2749 _mesa_shader_stage_to_string(producer
->Stage
),
2751 max_output_components
);
2760 check_against_input_limit(struct gl_context
*ctx
,
2761 struct gl_shader_program
*prog
,
2762 gl_linked_shader
*consumer
,
2763 unsigned num_explicit_locations
)
2765 unsigned input_vectors
= num_explicit_locations
;
2767 foreach_in_list(ir_instruction
, node
, consumer
->ir
) {
2768 ir_variable
*const var
= node
->as_variable();
2770 if (var
&& !var
->data
.explicit_location
&&
2771 var
->data
.mode
== ir_var_shader_in
&&
2772 var_counts_against_varying_limit(consumer
->Stage
, var
)) {
2773 /* vertex inputs aren't varying counted */
2774 input_vectors
+= var
->type
->count_attribute_slots(false);
2778 assert(consumer
->Stage
!= MESA_SHADER_VERTEX
);
2779 unsigned max_input_components
=
2780 ctx
->Const
.Program
[consumer
->Stage
].MaxInputComponents
;
2782 const unsigned input_components
= input_vectors
* 4;
2783 if (input_components
> max_input_components
) {
2784 if (ctx
->API
== API_OPENGLES2
|| prog
->IsES
)
2785 linker_error(prog
, "%s shader uses too many input vectors "
2787 _mesa_shader_stage_to_string(consumer
->Stage
),
2789 max_input_components
/ 4);
2791 linker_error(prog
, "%s shader uses too many input components "
2793 _mesa_shader_stage_to_string(consumer
->Stage
),
2795 max_input_components
);
2804 link_varyings(struct gl_shader_program
*prog
, unsigned first
, unsigned last
,
2805 struct gl_context
*ctx
, void *mem_ctx
)
2807 bool has_xfb_qualifiers
= false;
2808 unsigned num_tfeedback_decls
= 0;
2809 char **varying_names
= NULL
;
2810 tfeedback_decl
*tfeedback_decls
= NULL
;
2812 /* From the ARB_enhanced_layouts spec:
2814 * "If the shader used to record output variables for transform feedback
2815 * varyings uses the "xfb_buffer", "xfb_offset", or "xfb_stride" layout
2816 * qualifiers, the values specified by TransformFeedbackVaryings are
2817 * ignored, and the set of variables captured for transform feedback is
2818 * instead derived from the specified layout qualifiers."
2820 for (int i
= MESA_SHADER_FRAGMENT
- 1; i
>= 0; i
--) {
2821 /* Find last stage before fragment shader */
2822 if (prog
->_LinkedShaders
[i
]) {
2823 has_xfb_qualifiers
=
2824 process_xfb_layout_qualifiers(mem_ctx
, prog
->_LinkedShaders
[i
],
2825 prog
, &num_tfeedback_decls
,
2831 if (!has_xfb_qualifiers
) {
2832 num_tfeedback_decls
= prog
->TransformFeedback
.NumVarying
;
2833 varying_names
= prog
->TransformFeedback
.VaryingNames
;
2836 if (num_tfeedback_decls
!= 0) {
2837 /* From GL_EXT_transform_feedback:
2838 * A program will fail to link if:
2840 * * the <count> specified by TransformFeedbackVaryingsEXT is
2841 * non-zero, but the program object has no vertex or geometry
2844 if (first
>= MESA_SHADER_FRAGMENT
) {
2845 linker_error(prog
, "Transform feedback varyings specified, but "
2846 "no vertex, tessellation, or geometry shader is "
2851 tfeedback_decls
= rzalloc_array(mem_ctx
, tfeedback_decl
,
2852 num_tfeedback_decls
);
2853 if (!parse_tfeedback_decls(ctx
, prog
, mem_ctx
, num_tfeedback_decls
,
2854 varying_names
, tfeedback_decls
))
2858 /* If there is no fragment shader we need to set transform feedback.
2860 * For SSO we also need to assign output locations. We assign them here
2861 * because we need to do it for both single stage programs and multi stage
2864 if (last
< MESA_SHADER_FRAGMENT
&&
2865 (num_tfeedback_decls
!= 0 || prog
->SeparateShader
)) {
2866 const uint64_t reserved_out_slots
=
2867 reserved_varying_slot(prog
->_LinkedShaders
[last
], ir_var_shader_out
);
2868 if (!assign_varying_locations(ctx
, mem_ctx
, prog
,
2869 prog
->_LinkedShaders
[last
], NULL
,
2870 num_tfeedback_decls
, tfeedback_decls
,
2871 reserved_out_slots
))
2875 if (last
<= MESA_SHADER_FRAGMENT
) {
2876 /* Remove unused varyings from the first/last stage unless SSO */
2877 remove_unused_shader_inputs_and_outputs(prog
->SeparateShader
,
2878 prog
->_LinkedShaders
[first
],
2880 remove_unused_shader_inputs_and_outputs(prog
->SeparateShader
,
2881 prog
->_LinkedShaders
[last
],
2884 /* If the program is made up of only a single stage */
2885 if (first
== last
) {
2886 gl_linked_shader
*const sh
= prog
->_LinkedShaders
[last
];
2888 do_dead_builtin_varyings(ctx
, NULL
, sh
, 0, NULL
);
2889 do_dead_builtin_varyings(ctx
, sh
, NULL
, num_tfeedback_decls
,
2892 if (prog
->SeparateShader
) {
2893 const uint64_t reserved_slots
=
2894 reserved_varying_slot(sh
, ir_var_shader_in
);
2896 /* Assign input locations for SSO, output locations are already
2899 if (!assign_varying_locations(ctx
, mem_ctx
, prog
,
2900 NULL
/* producer */,
2902 0 /* num_tfeedback_decls */,
2903 NULL
/* tfeedback_decls */,
2908 /* Linking the stages in the opposite order (from fragment to vertex)
2909 * ensures that inter-shader outputs written to in an earlier stage
2910 * are eliminated if they are (transitively) not used in a later
2914 for (int i
= next
- 1; i
>= 0; i
--) {
2915 if (prog
->_LinkedShaders
[i
] == NULL
&& i
!= 0)
2918 gl_linked_shader
*const sh_i
= prog
->_LinkedShaders
[i
];
2919 gl_linked_shader
*const sh_next
= prog
->_LinkedShaders
[next
];
2921 const uint64_t reserved_out_slots
=
2922 reserved_varying_slot(sh_i
, ir_var_shader_out
);
2923 const uint64_t reserved_in_slots
=
2924 reserved_varying_slot(sh_next
, ir_var_shader_in
);
2926 do_dead_builtin_varyings(ctx
, sh_i
, sh_next
,
2927 next
== MESA_SHADER_FRAGMENT
? num_tfeedback_decls
: 0,
2930 if (!assign_varying_locations(ctx
, mem_ctx
, prog
, sh_i
, sh_next
,
2931 next
== MESA_SHADER_FRAGMENT
? num_tfeedback_decls
: 0,
2933 reserved_out_slots
| reserved_in_slots
))
2936 /* This must be done after all dead varyings are eliminated. */
2938 unsigned slots_used
= util_bitcount64(reserved_out_slots
);
2939 if (!check_against_output_limit(ctx
, prog
, sh_i
, slots_used
)) {
2944 unsigned slots_used
= util_bitcount64(reserved_in_slots
);
2945 if (!check_against_input_limit(ctx
, prog
, sh_next
, slots_used
))
2953 if (!store_tfeedback_info(ctx
, prog
, num_tfeedback_decls
, tfeedback_decls
,
2954 has_xfb_qualifiers
))