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"
45 * Get the varying type stripped of the outermost array if we're processing
46 * a stage whose varyings are arrays indexed by a vertex number (such as
47 * geometry shader inputs).
49 static const glsl_type
*
50 get_varying_type(const ir_variable
*var
, gl_shader_stage stage
)
52 const glsl_type
*type
= var
->type
;
54 if (!var
->data
.patch
&&
55 ((var
->data
.mode
== ir_var_shader_out
&&
56 stage
== MESA_SHADER_TESS_CTRL
) ||
57 (var
->data
.mode
== ir_var_shader_in
&&
58 (stage
== MESA_SHADER_TESS_CTRL
|| stage
== MESA_SHADER_TESS_EVAL
||
59 stage
== MESA_SHADER_GEOMETRY
)))) {
60 assert(type
->is_array());
61 type
= type
->fields
.array
;
68 create_xfb_varying_names(void *mem_ctx
, const glsl_type
*t
, char **name
,
69 size_t name_length
, unsigned *count
,
70 const char *ifc_member_name
,
71 const glsl_type
*ifc_member_t
, char ***varying_names
)
73 if (t
->is_interface()) {
74 size_t new_length
= name_length
;
76 assert(ifc_member_name
&& ifc_member_t
);
77 ralloc_asprintf_rewrite_tail(name
, &new_length
, ".%s", ifc_member_name
);
79 create_xfb_varying_names(mem_ctx
, ifc_member_t
, name
, new_length
, count
,
80 NULL
, NULL
, varying_names
);
81 } else if (t
->is_record()) {
82 for (unsigned i
= 0; i
< t
->length
; i
++) {
83 const char *field
= t
->fields
.structure
[i
].name
;
84 size_t new_length
= name_length
;
86 ralloc_asprintf_rewrite_tail(name
, &new_length
, ".%s", field
);
88 create_xfb_varying_names(mem_ctx
, t
->fields
.structure
[i
].type
, name
,
89 new_length
, count
, NULL
, NULL
,
92 } else if (t
->without_array()->is_record() ||
93 t
->without_array()->is_interface() ||
94 (t
->is_array() && t
->fields
.array
->is_array())) {
95 for (unsigned i
= 0; i
< t
->length
; i
++) {
96 size_t new_length
= name_length
;
98 /* Append the subscript to the current variable name */
99 ralloc_asprintf_rewrite_tail(name
, &new_length
, "[%u]", i
);
101 create_xfb_varying_names(mem_ctx
, t
->fields
.array
, name
, new_length
,
102 count
, ifc_member_name
, ifc_member_t
,
106 (*varying_names
)[(*count
)++] = ralloc_strdup(mem_ctx
, *name
);
111 process_xfb_layout_qualifiers(void *mem_ctx
, const gl_linked_shader
*sh
,
112 struct gl_shader_program
*prog
,
113 unsigned *num_tfeedback_decls
,
114 char ***varying_names
)
116 bool has_xfb_qualifiers
= false;
118 /* We still need to enable transform feedback mode even if xfb_stride is
119 * only applied to a global out. Also we don't bother to propagate
120 * xfb_stride to interface block members so this will catch that case also.
122 for (unsigned j
= 0; j
< MAX_FEEDBACK_BUFFERS
; j
++) {
123 if (prog
->TransformFeedback
.BufferStride
[j
]) {
124 has_xfb_qualifiers
= true;
129 foreach_in_list(ir_instruction
, node
, sh
->ir
) {
130 ir_variable
*var
= node
->as_variable();
131 if (!var
|| var
->data
.mode
!= ir_var_shader_out
)
134 /* From the ARB_enhanced_layouts spec:
136 * "Any shader making any static use (after preprocessing) of any of
137 * these *xfb_* qualifiers will cause the shader to be in a
138 * transform feedback capturing mode and hence responsible for
139 * describing the transform feedback setup. This mode will capture
140 * any output selected by *xfb_offset*, directly or indirectly, to
141 * a transform feedback buffer."
143 if (var
->data
.explicit_xfb_buffer
|| var
->data
.explicit_xfb_stride
) {
144 has_xfb_qualifiers
= true;
147 if (var
->data
.explicit_xfb_offset
) {
148 *num_tfeedback_decls
+= var
->type
->varying_count();
149 has_xfb_qualifiers
= true;
153 if (*num_tfeedback_decls
== 0)
154 return has_xfb_qualifiers
;
157 *varying_names
= ralloc_array(mem_ctx
, char *, *num_tfeedback_decls
);
158 foreach_in_list(ir_instruction
, node
, sh
->ir
) {
159 ir_variable
*var
= node
->as_variable();
160 if (!var
|| var
->data
.mode
!= ir_var_shader_out
)
163 if (var
->data
.explicit_xfb_offset
) {
165 const glsl_type
*type
, *member_type
;
167 if (var
->data
.from_named_ifc_block
) {
168 type
= var
->get_interface_type();
170 /* Find the member type before it was altered by lowering */
171 const glsl_type
*type_wa
= type
->without_array();
173 type_wa
->fields
.structure
[type_wa
->field_index(var
->name
)].type
;
174 name
= ralloc_strdup(NULL
, type_wa
->name
);
178 name
= ralloc_strdup(NULL
, var
->name
);
180 create_xfb_varying_names(mem_ctx
, type
, &name
, strlen(name
), &i
,
181 var
->name
, member_type
, varying_names
);
186 assert(i
== *num_tfeedback_decls
);
187 return has_xfb_qualifiers
;
191 * Validate the types and qualifiers of an output from one stage against the
192 * matching input to another stage.
195 cross_validate_types_and_qualifiers(struct gl_context
*ctx
,
196 struct gl_shader_program
*prog
,
197 const ir_variable
*input
,
198 const ir_variable
*output
,
199 gl_shader_stage consumer_stage
,
200 gl_shader_stage producer_stage
)
202 /* Check that the types match between stages.
204 const glsl_type
*type_to_match
= input
->type
;
206 /* VS -> GS, VS -> TCS, VS -> TES, TES -> GS */
207 const bool extra_array_level
= (producer_stage
== MESA_SHADER_VERTEX
&&
208 consumer_stage
!= MESA_SHADER_FRAGMENT
) ||
209 consumer_stage
== MESA_SHADER_GEOMETRY
;
210 if (extra_array_level
) {
211 assert(type_to_match
->is_array());
212 type_to_match
= type_to_match
->fields
.array
;
215 if (type_to_match
!= output
->type
) {
216 /* There is a bit of a special case for gl_TexCoord. This
217 * built-in is unsized by default. Applications that variable
218 * access it must redeclare it with a size. There is some
219 * language in the GLSL spec that implies the fragment shader
220 * and vertex shader do not have to agree on this size. Other
221 * driver behave this way, and one or two applications seem to
224 * Neither declaration needs to be modified here because the array
225 * sizes are fixed later when update_array_sizes is called.
227 * From page 48 (page 54 of the PDF) of the GLSL 1.10 spec:
229 * "Unlike user-defined varying variables, the built-in
230 * varying variables don't have a strict one-to-one
231 * correspondence between the vertex language and the
232 * fragment language."
234 if (!output
->type
->is_array() || !is_gl_identifier(output
->name
)) {
236 "%s shader output `%s' declared as type `%s', "
237 "but %s shader input declared as type `%s'\n",
238 _mesa_shader_stage_to_string(producer_stage
),
241 _mesa_shader_stage_to_string(consumer_stage
),
247 /* Check that all of the qualifiers match between stages.
250 /* According to the OpenGL and OpenGLES GLSL specs, the centroid qualifier
251 * should match until OpenGL 4.3 and OpenGLES 3.1. The OpenGLES 3.0
252 * conformance test suite does not verify that the qualifiers must match.
253 * The deqp test suite expects the opposite (OpenGLES 3.1) behavior for
254 * OpenGLES 3.0 drivers, so we relax the checking in all cases.
256 if (false /* always skip the centroid check */ &&
257 prog
->data
->Version
< (prog
->IsES
? 310 : 430) &&
258 input
->data
.centroid
!= output
->data
.centroid
) {
260 "%s shader output `%s' %s centroid qualifier, "
261 "but %s shader input %s centroid qualifier\n",
262 _mesa_shader_stage_to_string(producer_stage
),
264 (output
->data
.centroid
) ? "has" : "lacks",
265 _mesa_shader_stage_to_string(consumer_stage
),
266 (input
->data
.centroid
) ? "has" : "lacks");
270 if (input
->data
.sample
!= output
->data
.sample
) {
272 "%s shader output `%s' %s sample qualifier, "
273 "but %s shader input %s sample qualifier\n",
274 _mesa_shader_stage_to_string(producer_stage
),
276 (output
->data
.sample
) ? "has" : "lacks",
277 _mesa_shader_stage_to_string(consumer_stage
),
278 (input
->data
.sample
) ? "has" : "lacks");
282 if (input
->data
.patch
!= output
->data
.patch
) {
284 "%s shader output `%s' %s patch qualifier, "
285 "but %s shader input %s patch qualifier\n",
286 _mesa_shader_stage_to_string(producer_stage
),
288 (output
->data
.patch
) ? "has" : "lacks",
289 _mesa_shader_stage_to_string(consumer_stage
),
290 (input
->data
.patch
) ? "has" : "lacks");
294 /* The GLSL 4.30 and GLSL ES 3.00 specifications say:
296 * "As only outputs need be declared with invariant, an output from
297 * one shader stage will still match an input of a subsequent stage
298 * without the input being declared as invariant."
300 * while GLSL 4.20 says:
302 * "For variables leaving one shader and coming into another shader,
303 * the invariant keyword has to be used in both shaders, or a link
304 * error will result."
306 * and GLSL ES 1.00 section 4.6.4 "Invariance and Linking" says:
308 * "The invariance of varyings that are declared in both the vertex
309 * and fragment shaders must match."
311 if (input
->data
.invariant
!= output
->data
.invariant
&&
312 prog
->data
->Version
< (prog
->IsES
? 300 : 430)) {
314 "%s shader output `%s' %s invariant qualifier, "
315 "but %s shader input %s invariant qualifier\n",
316 _mesa_shader_stage_to_string(producer_stage
),
318 (output
->data
.invariant
) ? "has" : "lacks",
319 _mesa_shader_stage_to_string(consumer_stage
),
320 (input
->data
.invariant
) ? "has" : "lacks");
324 /* GLSL >= 4.40 removes text requiring interpolation qualifiers
325 * to match cross stage, they must only match within the same stage.
327 * From page 84 (page 90 of the PDF) of the GLSL 4.40 spec:
329 * "It is a link-time error if, within the same stage, the interpolation
330 * qualifiers of variables of the same name do not match.
332 * Section 4.3.9 (Interpolation) of the GLSL ES 3.00 spec says:
334 * "When no interpolation qualifier is present, smooth interpolation
337 * So we match variables where one is smooth and the other has no explicit
340 unsigned input_interpolation
= input
->data
.interpolation
;
341 unsigned output_interpolation
= output
->data
.interpolation
;
343 if (input_interpolation
== INTERP_MODE_NONE
)
344 input_interpolation
= INTERP_MODE_SMOOTH
;
345 if (output_interpolation
== INTERP_MODE_NONE
)
346 output_interpolation
= INTERP_MODE_SMOOTH
;
348 if (input_interpolation
!= output_interpolation
&&
349 prog
->data
->Version
< 440) {
350 if (!ctx
->Const
.AllowGLSLCrossStageInterpolationMismatch
) {
352 "%s shader output `%s' specifies %s "
353 "interpolation qualifier, "
354 "but %s shader input specifies %s "
355 "interpolation qualifier\n",
356 _mesa_shader_stage_to_string(producer_stage
),
358 interpolation_string(output
->data
.interpolation
),
359 _mesa_shader_stage_to_string(consumer_stage
),
360 interpolation_string(input
->data
.interpolation
));
364 "%s shader output `%s' specifies %s "
365 "interpolation qualifier, "
366 "but %s shader input specifies %s "
367 "interpolation qualifier\n",
368 _mesa_shader_stage_to_string(producer_stage
),
370 interpolation_string(output
->data
.interpolation
),
371 _mesa_shader_stage_to_string(consumer_stage
),
372 interpolation_string(input
->data
.interpolation
));
378 * Validate front and back color outputs against single color input
381 cross_validate_front_and_back_color(struct gl_context
*ctx
,
382 struct gl_shader_program
*prog
,
383 const ir_variable
*input
,
384 const ir_variable
*front_color
,
385 const ir_variable
*back_color
,
386 gl_shader_stage consumer_stage
,
387 gl_shader_stage producer_stage
)
389 if (front_color
!= NULL
&& front_color
->data
.assigned
)
390 cross_validate_types_and_qualifiers(ctx
, prog
, input
, front_color
,
391 consumer_stage
, producer_stage
);
393 if (back_color
!= NULL
&& back_color
->data
.assigned
)
394 cross_validate_types_and_qualifiers(ctx
, prog
, input
, back_color
,
395 consumer_stage
, producer_stage
);
399 compute_variable_location_slot(ir_variable
*var
, gl_shader_stage stage
)
401 unsigned location_start
= VARYING_SLOT_VAR0
;
404 case MESA_SHADER_VERTEX
:
405 if (var
->data
.mode
== ir_var_shader_in
)
406 location_start
= VERT_ATTRIB_GENERIC0
;
408 case MESA_SHADER_TESS_CTRL
:
409 case MESA_SHADER_TESS_EVAL
:
411 location_start
= VARYING_SLOT_PATCH0
;
413 case MESA_SHADER_FRAGMENT
:
414 if (var
->data
.mode
== ir_var_shader_out
)
415 location_start
= FRAG_RESULT_DATA0
;
421 return var
->data
.location
- location_start
;
424 struct explicit_location_info
{
426 unsigned numerical_type
;
427 unsigned interpolation
;
433 static inline unsigned
434 get_numerical_type(const glsl_type
*type
)
436 /* From the OpenGL 4.6 spec, section 4.4.1 Input Layout Qualifiers, Page 68,
437 * (Location aliasing):
439 * "Further, when location aliasing, the aliases sharing the location
440 * must have the same underlying numerical type (floating-point or
443 if (type
->is_float() || type
->is_double())
444 return GLSL_TYPE_FLOAT
;
445 return GLSL_TYPE_INT
;
449 check_location_aliasing(struct explicit_location_info explicit_locations
[][4],
453 unsigned location_limit
,
454 const glsl_type
*type
,
455 unsigned interpolation
,
459 gl_shader_program
*prog
,
460 gl_shader_stage stage
)
463 if (type
->without_array()->is_record()) {
464 /* The component qualifier can't be used on structs so just treat
465 * all component slots as used.
469 unsigned dmul
= type
->without_array()->is_64bit() ? 2 : 1;
470 last_comp
= component
+ type
->without_array()->vector_elements
* dmul
;
473 while (location
< location_limit
) {
476 struct explicit_location_info
*info
=
477 &explicit_locations
[location
][comp
];
480 /* Component aliasing is not alloed */
481 if (comp
>= component
&& comp
< last_comp
) {
483 "%s shader has multiple outputs explicitly "
484 "assigned to location %d and component %d\n",
485 _mesa_shader_stage_to_string(stage
),
489 /* For all other used components we need to have matching
490 * types, interpolation and auxiliary storage
492 if (info
->numerical_type
!=
493 get_numerical_type(type
->without_array())) {
495 "Varyings sharing the same location must "
496 "have the same underlying numerical type. "
497 "Location %u component %u\n",
502 if (info
->interpolation
!= interpolation
) {
504 "%s shader has multiple outputs at explicit "
505 "location %u with different interpolation "
507 _mesa_shader_stage_to_string(stage
), location
);
511 if (info
->centroid
!= centroid
||
512 info
->sample
!= sample
||
513 info
->patch
!= patch
) {
515 "%s shader has multiple outputs at explicit "
516 "location %u with different aux storage\n",
517 _mesa_shader_stage_to_string(stage
), location
);
521 } else if (comp
>= component
&& comp
< last_comp
) {
523 info
->numerical_type
= get_numerical_type(type
->without_array());
524 info
->interpolation
= interpolation
;
525 info
->centroid
= centroid
;
526 info
->sample
= sample
;
532 /* We need to do some special handling for doubles as dvec3 and
533 * dvec4 consume two consecutive locations. We don't need to
534 * worry about components beginning at anything other than 0 as
535 * the spec does not allow this for dvec3 and dvec4.
537 if (comp
== 4 && last_comp
> 4) {
538 last_comp
= last_comp
- 4;
539 /* Bump location index and reset the component index */
553 validate_explicit_variable_location(struct gl_context
*ctx
,
554 struct explicit_location_info explicit_locations
[][4],
556 gl_shader_program
*prog
,
557 gl_linked_shader
*sh
)
559 const glsl_type
*type
= get_varying_type(var
, sh
->Stage
);
560 unsigned num_elements
= type
->count_attribute_slots(false);
561 unsigned idx
= compute_variable_location_slot(var
, sh
->Stage
);
562 unsigned slot_limit
= idx
+ num_elements
;
564 /* Vertex shader inputs and fragment shader outputs are validated in
565 * assign_attribute_or_color_locations() so we should not attempt to
566 * validate them again here.
569 if (var
->data
.mode
== ir_var_shader_out
) {
570 assert(sh
->Stage
!= MESA_SHADER_FRAGMENT
);
572 ctx
->Const
.Program
[sh
->Stage
].MaxOutputComponents
/ 4;
574 assert(var
->data
.mode
== ir_var_shader_in
);
575 assert(sh
->Stage
!= MESA_SHADER_VERTEX
);
577 ctx
->Const
.Program
[sh
->Stage
].MaxInputComponents
/ 4;
580 if (slot_limit
> slot_max
) {
582 "Invalid location %u in %s shader\n",
583 idx
, _mesa_shader_stage_to_string(sh
->Stage
));
587 const glsl_type
*type_without_array
= type
->without_array();
588 if (type_without_array
->is_interface()) {
589 for (unsigned i
= 0; i
< type_without_array
->length
; i
++) {
590 glsl_struct_field
*field
= &type_without_array
->fields
.structure
[i
];
591 unsigned field_location
= field
->location
-
592 (field
->patch
? VARYING_SLOT_PATCH0
: VARYING_SLOT_VAR0
);
593 if (!check_location_aliasing(explicit_locations
, var
,
595 0, field_location
+ 1,
597 field
->interpolation
,
605 } else if (!check_location_aliasing(explicit_locations
, var
,
606 idx
, var
->data
.location_frac
,
608 var
->data
.interpolation
,
620 * Validate explicit locations for the inputs to the first stage and the
621 * outputs of the last stage in an SSO program (everything in between is
622 * validated in cross_validate_outputs_to_inputs).
625 validate_sso_explicit_locations(struct gl_context
*ctx
,
626 struct gl_shader_program
*prog
,
627 gl_shader_stage first_stage
,
628 gl_shader_stage last_stage
)
630 assert(prog
->SeparateShader
);
632 /* VS inputs and FS outputs are validated in
633 * assign_attribute_or_color_locations()
635 bool validate_first_stage
= first_stage
!= MESA_SHADER_VERTEX
;
636 bool validate_last_stage
= last_stage
!= MESA_SHADER_FRAGMENT
;
637 if (!validate_first_stage
&& !validate_last_stage
)
640 struct explicit_location_info explicit_locations
[MAX_VARYING
][4];
642 gl_shader_stage stages
[2] = { first_stage
, last_stage
};
643 bool validate_stage
[2] = { validate_first_stage
, validate_last_stage
};
644 ir_variable_mode var_direction
[2] = { ir_var_shader_in
, ir_var_shader_out
};
646 for (unsigned i
= 0; i
< 2; i
++) {
647 if (!validate_stage
[i
])
650 gl_shader_stage stage
= stages
[i
];
652 gl_linked_shader
*sh
= prog
->_LinkedShaders
[stage
];
655 memset(explicit_locations
, 0, sizeof(explicit_locations
));
657 foreach_in_list(ir_instruction
, node
, sh
->ir
) {
658 ir_variable
*const var
= node
->as_variable();
661 !var
->data
.explicit_location
||
662 var
->data
.location
< VARYING_SLOT_VAR0
||
663 var
->data
.mode
!= var_direction
[i
])
666 if (!validate_explicit_variable_location(
667 ctx
, explicit_locations
, var
, prog
, sh
)) {
675 * Validate that outputs from one stage match inputs of another
678 cross_validate_outputs_to_inputs(struct gl_context
*ctx
,
679 struct gl_shader_program
*prog
,
680 gl_linked_shader
*producer
,
681 gl_linked_shader
*consumer
)
683 glsl_symbol_table parameters
;
684 struct explicit_location_info explicit_locations
[MAX_VARYING
][4] = { 0 };
686 /* Find all shader outputs in the "producer" stage.
688 foreach_in_list(ir_instruction
, node
, producer
->ir
) {
689 ir_variable
*const var
= node
->as_variable();
691 if (var
== NULL
|| var
->data
.mode
!= ir_var_shader_out
)
694 if (!var
->data
.explicit_location
695 || var
->data
.location
< VARYING_SLOT_VAR0
)
696 parameters
.add_variable(var
);
698 /* User-defined varyings with explicit locations are handled
699 * differently because they do not need to have matching names.
701 if (!validate_explicit_variable_location(ctx
,
703 var
, prog
, producer
)) {
710 /* Find all shader inputs in the "consumer" stage. Any variables that have
711 * matching outputs already in the symbol table must have the same type and
714 * Exception: if the consumer is the geometry shader, then the inputs
715 * should be arrays and the type of the array element should match the type
716 * of the corresponding producer output.
718 foreach_in_list(ir_instruction
, node
, consumer
->ir
) {
719 ir_variable
*const input
= node
->as_variable();
721 if (input
== NULL
|| input
->data
.mode
!= ir_var_shader_in
)
724 if (strcmp(input
->name
, "gl_Color") == 0 && input
->data
.used
) {
725 const ir_variable
*const front_color
=
726 parameters
.get_variable("gl_FrontColor");
728 const ir_variable
*const back_color
=
729 parameters
.get_variable("gl_BackColor");
731 cross_validate_front_and_back_color(ctx
, prog
, input
,
732 front_color
, back_color
,
733 consumer
->Stage
, producer
->Stage
);
734 } else if (strcmp(input
->name
, "gl_SecondaryColor") == 0 && input
->data
.used
) {
735 const ir_variable
*const front_color
=
736 parameters
.get_variable("gl_FrontSecondaryColor");
738 const ir_variable
*const back_color
=
739 parameters
.get_variable("gl_BackSecondaryColor");
741 cross_validate_front_and_back_color(ctx
, prog
, input
,
742 front_color
, back_color
,
743 consumer
->Stage
, producer
->Stage
);
745 /* The rules for connecting inputs and outputs change in the presence
746 * of explicit locations. In this case, we no longer care about the
747 * names of the variables. Instead, we care only about the
748 * explicitly assigned location.
750 ir_variable
*output
= NULL
;
751 if (input
->data
.explicit_location
752 && input
->data
.location
>= VARYING_SLOT_VAR0
) {
754 const glsl_type
*type
= get_varying_type(input
, consumer
->Stage
);
755 unsigned num_elements
= type
->count_attribute_slots(false);
757 compute_variable_location_slot(input
, consumer
->Stage
);
758 unsigned slot_limit
= idx
+ num_elements
;
760 while (idx
< slot_limit
) {
761 if (idx
>= MAX_VARYING
) {
763 "Invalid location %u in %s shader\n", idx
,
764 _mesa_shader_stage_to_string(consumer
->Stage
));
768 output
= explicit_locations
[idx
][input
->data
.location_frac
].var
;
770 if (output
== NULL
||
771 input
->data
.location
!= output
->data
.location
) {
773 "%s shader input `%s' with explicit location "
774 "has no matching output\n",
775 _mesa_shader_stage_to_string(consumer
->Stage
),
782 output
= parameters
.get_variable(input
->name
);
785 if (output
!= NULL
) {
786 /* Interface blocks have their own validation elsewhere so don't
787 * try validating them here.
789 if (!(input
->get_interface_type() &&
790 output
->get_interface_type()))
791 cross_validate_types_and_qualifiers(ctx
, prog
, input
, output
,
795 /* Check for input vars with unmatched output vars in prev stage
796 * taking into account that interface blocks could have a matching
797 * output but with different name, so we ignore them.
799 assert(!input
->data
.assigned
);
800 if (input
->data
.used
&& !input
->get_interface_type() &&
801 !input
->data
.explicit_location
&& !prog
->SeparateShader
)
803 "%s shader input `%s' "
804 "has no matching output in the previous stage\n",
805 _mesa_shader_stage_to_string(consumer
->Stage
),
813 * Demote shader inputs and outputs that are not used in other stages, and
814 * remove them via dead code elimination.
817 remove_unused_shader_inputs_and_outputs(bool is_separate_shader_object
,
818 gl_linked_shader
*sh
,
819 enum ir_variable_mode mode
)
821 if (is_separate_shader_object
)
824 foreach_in_list(ir_instruction
, node
, sh
->ir
) {
825 ir_variable
*const var
= node
->as_variable();
827 if (var
== NULL
|| var
->data
.mode
!= int(mode
))
830 /* A shader 'in' or 'out' variable is only really an input or output if
831 * its value is used by other shader stages. This will cause the
832 * variable to have a location assigned.
834 if (var
->data
.is_unmatched_generic_inout
&& !var
->data
.is_xfb_only
) {
835 assert(var
->data
.mode
!= ir_var_temporary
);
837 /* Assign zeros to demoted inputs to allow more optimizations. */
838 if (var
->data
.mode
== ir_var_shader_in
&& !var
->constant_value
)
839 var
->constant_value
= ir_constant::zero(var
, var
->type
);
841 var
->data
.mode
= ir_var_auto
;
845 /* Eliminate code that is now dead due to unused inputs/outputs being
848 while (do_dead_code(sh
->ir
, false))
854 * Initialize this object based on a string that was passed to
855 * glTransformFeedbackVaryings.
857 * If the input is mal-formed, this call still succeeds, but it sets
858 * this->var_name to a mal-formed input, so tfeedback_decl::find_output_var()
859 * will fail to find any matching variable.
862 tfeedback_decl::init(struct gl_context
*ctx
, const void *mem_ctx
,
865 /* We don't have to be pedantic about what is a valid GLSL variable name,
866 * because any variable with an invalid name can't exist in the IR anyway.
870 this->orig_name
= input
;
871 this->lowered_builtin_array_variable
= none
;
872 this->skip_components
= 0;
873 this->next_buffer_separator
= false;
874 this->matched_candidate
= NULL
;
879 if (ctx
->Extensions
.ARB_transform_feedback3
) {
880 /* Parse gl_NextBuffer. */
881 if (strcmp(input
, "gl_NextBuffer") == 0) {
882 this->next_buffer_separator
= true;
886 /* Parse gl_SkipComponents. */
887 if (strcmp(input
, "gl_SkipComponents1") == 0)
888 this->skip_components
= 1;
889 else if (strcmp(input
, "gl_SkipComponents2") == 0)
890 this->skip_components
= 2;
891 else if (strcmp(input
, "gl_SkipComponents3") == 0)
892 this->skip_components
= 3;
893 else if (strcmp(input
, "gl_SkipComponents4") == 0)
894 this->skip_components
= 4;
896 if (this->skip_components
)
900 /* Parse a declaration. */
901 const char *base_name_end
;
902 long subscript
= parse_program_resource_name(input
, &base_name_end
);
903 this->var_name
= ralloc_strndup(mem_ctx
, input
, base_name_end
- input
);
904 if (this->var_name
== NULL
) {
905 _mesa_error_no_memory(__func__
);
909 if (subscript
>= 0) {
910 this->array_subscript
= subscript
;
911 this->is_subscripted
= true;
913 this->is_subscripted
= false;
916 /* For drivers that lower gl_ClipDistance to gl_ClipDistanceMESA, this
917 * class must behave specially to account for the fact that gl_ClipDistance
918 * is converted from a float[8] to a vec4[2].
920 if (ctx
->Const
.ShaderCompilerOptions
[MESA_SHADER_VERTEX
].LowerCombinedClipCullDistance
&&
921 strcmp(this->var_name
, "gl_ClipDistance") == 0) {
922 this->lowered_builtin_array_variable
= clip_distance
;
924 if (ctx
->Const
.ShaderCompilerOptions
[MESA_SHADER_VERTEX
].LowerCombinedClipCullDistance
&&
925 strcmp(this->var_name
, "gl_CullDistance") == 0) {
926 this->lowered_builtin_array_variable
= cull_distance
;
929 if (ctx
->Const
.LowerTessLevel
&&
930 (strcmp(this->var_name
, "gl_TessLevelOuter") == 0))
931 this->lowered_builtin_array_variable
= tess_level_outer
;
932 if (ctx
->Const
.LowerTessLevel
&&
933 (strcmp(this->var_name
, "gl_TessLevelInner") == 0))
934 this->lowered_builtin_array_variable
= tess_level_inner
;
939 * Determine whether two tfeedback_decl objects refer to the same variable and
940 * array index (if applicable).
943 tfeedback_decl::is_same(const tfeedback_decl
&x
, const tfeedback_decl
&y
)
945 assert(x
.is_varying() && y
.is_varying());
947 if (strcmp(x
.var_name
, y
.var_name
) != 0)
949 if (x
.is_subscripted
!= y
.is_subscripted
)
951 if (x
.is_subscripted
&& x
.array_subscript
!= y
.array_subscript
)
958 * Assign a location and stream ID for this tfeedback_decl object based on the
959 * transform feedback candidate found by find_candidate.
961 * If an error occurs, the error is reported through linker_error() and false
965 tfeedback_decl::assign_location(struct gl_context
*ctx
,
966 struct gl_shader_program
*prog
)
968 assert(this->is_varying());
970 unsigned fine_location
971 = this->matched_candidate
->toplevel_var
->data
.location
* 4
972 + this->matched_candidate
->toplevel_var
->data
.location_frac
973 + this->matched_candidate
->offset
;
974 const unsigned dmul
=
975 this->matched_candidate
->type
->without_array()->is_64bit() ? 2 : 1;
977 if (this->matched_candidate
->type
->is_array()) {
979 const unsigned matrix_cols
=
980 this->matched_candidate
->type
->fields
.array
->matrix_columns
;
981 const unsigned vector_elements
=
982 this->matched_candidate
->type
->fields
.array
->vector_elements
;
983 unsigned actual_array_size
;
984 switch (this->lowered_builtin_array_variable
) {
986 actual_array_size
= prog
->last_vert_prog
?
987 prog
->last_vert_prog
->info
.clip_distance_array_size
: 0;
990 actual_array_size
= prog
->last_vert_prog
?
991 prog
->last_vert_prog
->info
.cull_distance_array_size
: 0;
993 case tess_level_outer
:
994 actual_array_size
= 4;
996 case tess_level_inner
:
997 actual_array_size
= 2;
1001 actual_array_size
= this->matched_candidate
->type
->array_size();
1005 if (this->is_subscripted
) {
1006 /* Check array bounds. */
1007 if (this->array_subscript
>= actual_array_size
) {
1008 linker_error(prog
, "Transform feedback varying %s has index "
1009 "%i, but the array size is %u.",
1010 this->orig_name
, this->array_subscript
,
1014 unsigned array_elem_size
= this->lowered_builtin_array_variable
?
1015 1 : vector_elements
* matrix_cols
* dmul
;
1016 fine_location
+= array_elem_size
* this->array_subscript
;
1019 this->size
= actual_array_size
;
1021 this->vector_elements
= vector_elements
;
1022 this->matrix_columns
= matrix_cols
;
1023 if (this->lowered_builtin_array_variable
)
1024 this->type
= GL_FLOAT
;
1026 this->type
= this->matched_candidate
->type
->fields
.array
->gl_type
;
1028 /* Regular variable (scalar, vector, or matrix) */
1029 if (this->is_subscripted
) {
1030 linker_error(prog
, "Transform feedback varying %s requested, "
1031 "but %s is not an array.",
1032 this->orig_name
, this->var_name
);
1036 this->vector_elements
= this->matched_candidate
->type
->vector_elements
;
1037 this->matrix_columns
= this->matched_candidate
->type
->matrix_columns
;
1038 this->type
= this->matched_candidate
->type
->gl_type
;
1040 this->location
= fine_location
/ 4;
1041 this->location_frac
= fine_location
% 4;
1043 /* From GL_EXT_transform_feedback:
1044 * A program will fail to link if:
1046 * * the total number of components to capture in any varying
1047 * variable in <varyings> is greater than the constant
1048 * MAX_TRANSFORM_FEEDBACK_SEPARATE_COMPONENTS_EXT and the
1049 * buffer mode is SEPARATE_ATTRIBS_EXT;
1051 if (prog
->TransformFeedback
.BufferMode
== GL_SEPARATE_ATTRIBS
&&
1052 this->num_components() >
1053 ctx
->Const
.MaxTransformFeedbackSeparateComponents
) {
1054 linker_error(prog
, "Transform feedback varying %s exceeds "
1055 "MAX_TRANSFORM_FEEDBACK_SEPARATE_COMPONENTS.",
1060 /* Only transform feedback varyings can be assigned to non-zero streams,
1061 * so assign the stream id here.
1063 this->stream_id
= this->matched_candidate
->toplevel_var
->data
.stream
;
1065 unsigned array_offset
= this->array_subscript
* 4 * dmul
;
1066 unsigned struct_offset
= this->matched_candidate
->offset
* 4 * dmul
;
1067 this->buffer
= this->matched_candidate
->toplevel_var
->data
.xfb_buffer
;
1068 this->offset
= this->matched_candidate
->toplevel_var
->data
.offset
+
1069 array_offset
+ struct_offset
;
1076 tfeedback_decl::get_num_outputs() const
1078 if (!this->is_varying()) {
1081 return (this->num_components() + this->location_frac
+ 3)/4;
1086 * Update gl_transform_feedback_info to reflect this tfeedback_decl.
1088 * If an error occurs, the error is reported through linker_error() and false
1092 tfeedback_decl::store(struct gl_context
*ctx
, struct gl_shader_program
*prog
,
1093 struct gl_transform_feedback_info
*info
,
1094 unsigned buffer
, unsigned buffer_index
,
1095 const unsigned max_outputs
, bool *explicit_stride
,
1096 bool has_xfb_qualifiers
) const
1098 unsigned xfb_offset
= 0;
1099 unsigned size
= this->size
;
1100 /* Handle gl_SkipComponents. */
1101 if (this->skip_components
) {
1102 info
->Buffers
[buffer
].Stride
+= this->skip_components
;
1103 size
= this->skip_components
;
1107 if (this->next_buffer_separator
) {
1112 if (has_xfb_qualifiers
) {
1113 xfb_offset
= this->offset
/ 4;
1115 xfb_offset
= info
->Buffers
[buffer
].Stride
;
1117 info
->Varyings
[info
->NumVarying
].Offset
= xfb_offset
* 4;
1120 unsigned location
= this->location
;
1121 unsigned location_frac
= this->location_frac
;
1122 unsigned num_components
= this->num_components();
1123 while (num_components
> 0) {
1124 unsigned output_size
= MIN2(num_components
, 4 - location_frac
);
1125 assert((info
->NumOutputs
== 0 && max_outputs
== 0) ||
1126 info
->NumOutputs
< max_outputs
);
1128 /* From the ARB_enhanced_layouts spec:
1130 * "If such a block member or variable is not written during a shader
1131 * invocation, the buffer contents at the assigned offset will be
1132 * undefined. Even if there are no static writes to a variable or
1133 * member that is assigned a transform feedback offset, the space is
1134 * still allocated in the buffer and still affects the stride."
1136 if (this->is_varying_written()) {
1137 info
->Outputs
[info
->NumOutputs
].ComponentOffset
= location_frac
;
1138 info
->Outputs
[info
->NumOutputs
].OutputRegister
= location
;
1139 info
->Outputs
[info
->NumOutputs
].NumComponents
= output_size
;
1140 info
->Outputs
[info
->NumOutputs
].StreamId
= stream_id
;
1141 info
->Outputs
[info
->NumOutputs
].OutputBuffer
= buffer
;
1142 info
->Outputs
[info
->NumOutputs
].DstOffset
= xfb_offset
;
1145 info
->Buffers
[buffer
].Stream
= this->stream_id
;
1146 xfb_offset
+= output_size
;
1148 num_components
-= output_size
;
1154 if (explicit_stride
&& explicit_stride
[buffer
]) {
1155 if (this->is_64bit() && info
->Buffers
[buffer
].Stride
% 2) {
1156 linker_error(prog
, "invalid qualifier xfb_stride=%d must be a "
1157 "multiple of 8 as its applied to a type that is or "
1158 "contains a double.",
1159 info
->Buffers
[buffer
].Stride
* 4);
1163 if ((this->offset
/ 4) / info
->Buffers
[buffer
].Stride
!=
1164 (xfb_offset
- 1) / info
->Buffers
[buffer
].Stride
) {
1165 linker_error(prog
, "xfb_offset (%d) overflows xfb_stride (%d) for "
1166 "buffer (%d)", xfb_offset
* 4,
1167 info
->Buffers
[buffer
].Stride
* 4, buffer
);
1171 info
->Buffers
[buffer
].Stride
= xfb_offset
;
1174 /* From GL_EXT_transform_feedback:
1175 * A program will fail to link if:
1177 * * the total number of components to capture is greater than
1178 * the constant MAX_TRANSFORM_FEEDBACK_INTERLEAVED_COMPONENTS_EXT
1179 * and the buffer mode is INTERLEAVED_ATTRIBS_EXT.
1181 * From GL_ARB_enhanced_layouts:
1183 * "The resulting stride (implicit or explicit) must be less than or
1184 * equal to the implementation-dependent constant
1185 * gl_MaxTransformFeedbackInterleavedComponents."
1187 if ((prog
->TransformFeedback
.BufferMode
== GL_INTERLEAVED_ATTRIBS
||
1188 has_xfb_qualifiers
) &&
1189 info
->Buffers
[buffer
].Stride
>
1190 ctx
->Const
.MaxTransformFeedbackInterleavedComponents
) {
1191 linker_error(prog
, "The MAX_TRANSFORM_FEEDBACK_INTERLEAVED_COMPONENTS "
1192 "limit has been exceeded.");
1197 info
->Varyings
[info
->NumVarying
].Name
= ralloc_strdup(prog
,
1199 info
->Varyings
[info
->NumVarying
].Type
= this->type
;
1200 info
->Varyings
[info
->NumVarying
].Size
= size
;
1201 info
->Varyings
[info
->NumVarying
].BufferIndex
= buffer_index
;
1203 info
->Buffers
[buffer
].NumVaryings
++;
1209 const tfeedback_candidate
*
1210 tfeedback_decl::find_candidate(gl_shader_program
*prog
,
1211 hash_table
*tfeedback_candidates
)
1213 const char *name
= this->var_name
;
1214 switch (this->lowered_builtin_array_variable
) {
1216 name
= this->var_name
;
1219 name
= "gl_ClipDistanceMESA";
1222 name
= "gl_CullDistanceMESA";
1224 case tess_level_outer
:
1225 name
= "gl_TessLevelOuterMESA";
1227 case tess_level_inner
:
1228 name
= "gl_TessLevelInnerMESA";
1231 hash_entry
*entry
= _mesa_hash_table_search(tfeedback_candidates
, name
);
1233 this->matched_candidate
= entry
?
1234 (const tfeedback_candidate
*) entry
->data
: NULL
;
1236 if (!this->matched_candidate
) {
1237 /* From GL_EXT_transform_feedback:
1238 * A program will fail to link if:
1240 * * any variable name specified in the <varyings> array is not
1241 * declared as an output in the geometry shader (if present) or
1242 * the vertex shader (if no geometry shader is present);
1244 linker_error(prog
, "Transform feedback varying %s undeclared.",
1248 return this->matched_candidate
;
1253 * Parse all the transform feedback declarations that were passed to
1254 * glTransformFeedbackVaryings() and store them in tfeedback_decl objects.
1256 * If an error occurs, the error is reported through linker_error() and false
1260 parse_tfeedback_decls(struct gl_context
*ctx
, struct gl_shader_program
*prog
,
1261 const void *mem_ctx
, unsigned num_names
,
1262 char **varying_names
, tfeedback_decl
*decls
)
1264 for (unsigned i
= 0; i
< num_names
; ++i
) {
1265 decls
[i
].init(ctx
, mem_ctx
, varying_names
[i
]);
1267 if (!decls
[i
].is_varying())
1270 /* From GL_EXT_transform_feedback:
1271 * A program will fail to link if:
1273 * * any two entries in the <varyings> array specify the same varying
1276 * We interpret this to mean "any two entries in the <varyings> array
1277 * specify the same varying variable and array index", since transform
1278 * feedback of arrays would be useless otherwise.
1280 for (unsigned j
= 0; j
< i
; ++j
) {
1281 if (decls
[j
].is_varying()) {
1282 if (tfeedback_decl::is_same(decls
[i
], decls
[j
])) {
1283 linker_error(prog
, "Transform feedback varying %s specified "
1284 "more than once.", varying_names
[i
]);
1295 cmp_xfb_offset(const void * x_generic
, const void * y_generic
)
1297 tfeedback_decl
*x
= (tfeedback_decl
*) x_generic
;
1298 tfeedback_decl
*y
= (tfeedback_decl
*) y_generic
;
1300 if (x
->get_buffer() != y
->get_buffer())
1301 return x
->get_buffer() - y
->get_buffer();
1302 return x
->get_offset() - y
->get_offset();
1306 * Store transform feedback location assignments into
1307 * prog->sh.LinkedTransformFeedback based on the data stored in
1310 * If an error occurs, the error is reported through linker_error() and false
1314 store_tfeedback_info(struct gl_context
*ctx
, struct gl_shader_program
*prog
,
1315 unsigned num_tfeedback_decls
,
1316 tfeedback_decl
*tfeedback_decls
, bool has_xfb_qualifiers
)
1318 if (!prog
->last_vert_prog
)
1321 /* Make sure MaxTransformFeedbackBuffers is less than 32 so the bitmask for
1322 * tracking the number of buffers doesn't overflow.
1324 assert(ctx
->Const
.MaxTransformFeedbackBuffers
< 32);
1326 bool separate_attribs_mode
=
1327 prog
->TransformFeedback
.BufferMode
== GL_SEPARATE_ATTRIBS
;
1329 struct gl_program
*xfb_prog
= prog
->last_vert_prog
;
1330 xfb_prog
->sh
.LinkedTransformFeedback
=
1331 rzalloc(xfb_prog
, struct gl_transform_feedback_info
);
1333 /* The xfb_offset qualifier does not have to be used in increasing order
1334 * however some drivers expect to receive the list of transform feedback
1335 * declarations in order so sort it now for convenience.
1337 if (has_xfb_qualifiers
) {
1338 qsort(tfeedback_decls
, num_tfeedback_decls
, sizeof(*tfeedback_decls
),
1342 xfb_prog
->sh
.LinkedTransformFeedback
->Varyings
=
1343 rzalloc_array(xfb_prog
, struct gl_transform_feedback_varying_info
,
1344 num_tfeedback_decls
);
1346 unsigned num_outputs
= 0;
1347 for (unsigned i
= 0; i
< num_tfeedback_decls
; ++i
) {
1348 if (tfeedback_decls
[i
].is_varying_written())
1349 num_outputs
+= tfeedback_decls
[i
].get_num_outputs();
1352 xfb_prog
->sh
.LinkedTransformFeedback
->Outputs
=
1353 rzalloc_array(xfb_prog
, struct gl_transform_feedback_output
,
1356 unsigned num_buffers
= 0;
1357 unsigned buffers
= 0;
1359 if (!has_xfb_qualifiers
&& separate_attribs_mode
) {
1360 /* GL_SEPARATE_ATTRIBS */
1361 for (unsigned i
= 0; i
< num_tfeedback_decls
; ++i
) {
1362 if (!tfeedback_decls
[i
].store(ctx
, prog
,
1363 xfb_prog
->sh
.LinkedTransformFeedback
,
1364 num_buffers
, num_buffers
, num_outputs
,
1365 NULL
, has_xfb_qualifiers
))
1368 buffers
|= 1 << num_buffers
;
1373 /* GL_INVERLEAVED_ATTRIBS */
1374 int buffer_stream_id
= -1;
1376 num_tfeedback_decls
? tfeedback_decls
[0].get_buffer() : 0;
1377 bool explicit_stride
[MAX_FEEDBACK_BUFFERS
] = { false };
1379 /* Apply any xfb_stride global qualifiers */
1380 if (has_xfb_qualifiers
) {
1381 for (unsigned j
= 0; j
< MAX_FEEDBACK_BUFFERS
; j
++) {
1382 if (prog
->TransformFeedback
.BufferStride
[j
]) {
1383 explicit_stride
[j
] = true;
1384 xfb_prog
->sh
.LinkedTransformFeedback
->Buffers
[j
].Stride
=
1385 prog
->TransformFeedback
.BufferStride
[j
] / 4;
1390 for (unsigned i
= 0; i
< num_tfeedback_decls
; ++i
) {
1391 if (has_xfb_qualifiers
&&
1392 buffer
!= tfeedback_decls
[i
].get_buffer()) {
1393 /* we have moved to the next buffer so reset stream id */
1394 buffer_stream_id
= -1;
1398 if (tfeedback_decls
[i
].is_next_buffer_separator()) {
1399 if (!tfeedback_decls
[i
].store(ctx
, prog
,
1400 xfb_prog
->sh
.LinkedTransformFeedback
,
1401 buffer
, num_buffers
, num_outputs
,
1402 explicit_stride
, has_xfb_qualifiers
))
1405 buffer_stream_id
= -1;
1409 if (has_xfb_qualifiers
) {
1410 buffer
= tfeedback_decls
[i
].get_buffer();
1412 buffer
= num_buffers
;
1415 if (tfeedback_decls
[i
].is_varying()) {
1416 if (buffer_stream_id
== -1) {
1417 /* First varying writing to this buffer: remember its stream */
1418 buffer_stream_id
= (int) tfeedback_decls
[i
].get_stream_id();
1420 /* Only mark a buffer as active when there is a varying
1421 * attached to it. This behaviour is based on a revised version
1422 * of section 13.2.2 of the GL 4.6 spec.
1424 buffers
|= 1 << buffer
;
1425 } else if (buffer_stream_id
!=
1426 (int) tfeedback_decls
[i
].get_stream_id()) {
1427 /* Varying writes to the same buffer from a different stream */
1429 "Transform feedback can't capture varyings belonging "
1430 "to different vertex streams in a single buffer. "
1431 "Varying %s writes to buffer from stream %u, other "
1432 "varyings in the same buffer write from stream %u.",
1433 tfeedback_decls
[i
].name(),
1434 tfeedback_decls
[i
].get_stream_id(),
1440 if (!tfeedback_decls
[i
].store(ctx
, prog
,
1441 xfb_prog
->sh
.LinkedTransformFeedback
,
1442 buffer
, num_buffers
, num_outputs
,
1443 explicit_stride
, has_xfb_qualifiers
))
1448 assert(xfb_prog
->sh
.LinkedTransformFeedback
->NumOutputs
== num_outputs
);
1450 xfb_prog
->sh
.LinkedTransformFeedback
->ActiveBuffers
= buffers
;
1457 * Data structure recording the relationship between outputs of one shader
1458 * stage (the "producer") and inputs of another (the "consumer").
1460 class varying_matches
1463 varying_matches(bool disable_varying_packing
, bool xfb_enabled
,
1464 bool enhanced_layouts_enabled
,
1465 gl_shader_stage producer_stage
,
1466 gl_shader_stage consumer_stage
);
1468 void record(ir_variable
*producer_var
, ir_variable
*consumer_var
);
1469 unsigned assign_locations(struct gl_shader_program
*prog
,
1470 uint8_t components
[],
1471 uint64_t reserved_slots
);
1472 void store_locations() const;
1475 bool is_varying_packing_safe(const glsl_type
*type
,
1476 const ir_variable
*var
) const;
1479 * If true, this driver disables varying packing, so all varyings need to
1480 * be aligned on slot boundaries, and take up a number of slots equal to
1481 * their number of matrix columns times their array size.
1483 * Packing may also be disabled because our current packing method is not
1484 * safe in SSO or versions of OpenGL where interpolation qualifiers are not
1485 * guaranteed to match across stages.
1487 const bool disable_varying_packing
;
1490 * If true, this driver has transform feedback enabled. The transform
1491 * feedback code requires at least some packing be done even when varying
1492 * packing is disabled, fortunately where transform feedback requires
1493 * packing it's safe to override the disabled setting. See
1494 * is_varying_packing_safe().
1496 const bool xfb_enabled
;
1498 const bool enhanced_layouts_enabled
;
1501 * Enum representing the order in which varyings are packed within a
1504 * Currently we pack vec4's first, then vec2's, then scalar values, then
1505 * vec3's. This order ensures that the only vectors that are at risk of
1506 * having to be "double parked" (split between two adjacent varying slots)
1509 enum packing_order_enum
{
1512 PACKING_ORDER_SCALAR
,
1516 static unsigned compute_packing_class(const ir_variable
*var
);
1517 static packing_order_enum
compute_packing_order(const ir_variable
*var
);
1518 static int match_comparator(const void *x_generic
, const void *y_generic
);
1519 static int xfb_comparator(const void *x_generic
, const void *y_generic
);
1522 * Structure recording the relationship between a single producer output
1523 * and a single consumer input.
1527 * Packing class for this varying, computed by compute_packing_class().
1529 unsigned packing_class
;
1532 * Packing order for this varying, computed by compute_packing_order().
1534 packing_order_enum packing_order
;
1535 unsigned num_components
;
1538 * The output variable in the producer stage.
1540 ir_variable
*producer_var
;
1543 * The input variable in the consumer stage.
1545 ir_variable
*consumer_var
;
1548 * The location which has been assigned for this varying. This is
1549 * expressed in multiples of a float, with the first generic varying
1550 * (i.e. the one referred to by VARYING_SLOT_VAR0) represented by the
1553 unsigned generic_location
;
1557 * The number of elements in the \c matches array that are currently in
1560 unsigned num_matches
;
1563 * The number of elements that were set aside for the \c matches array when
1566 unsigned matches_capacity
;
1568 gl_shader_stage producer_stage
;
1569 gl_shader_stage consumer_stage
;
1572 } /* anonymous namespace */
1574 varying_matches::varying_matches(bool disable_varying_packing
,
1576 bool enhanced_layouts_enabled
,
1577 gl_shader_stage producer_stage
,
1578 gl_shader_stage consumer_stage
)
1579 : disable_varying_packing(disable_varying_packing
),
1580 xfb_enabled(xfb_enabled
),
1581 enhanced_layouts_enabled(enhanced_layouts_enabled
),
1582 producer_stage(producer_stage
),
1583 consumer_stage(consumer_stage
)
1585 /* Note: this initial capacity is rather arbitrarily chosen to be large
1586 * enough for many cases without wasting an unreasonable amount of space.
1587 * varying_matches::record() will resize the array if there are more than
1588 * this number of varyings.
1590 this->matches_capacity
= 8;
1591 this->matches
= (match
*)
1592 malloc(sizeof(*this->matches
) * this->matches_capacity
);
1593 this->num_matches
= 0;
1597 varying_matches::~varying_matches()
1599 free(this->matches
);
1604 * Packing is always safe on individual arrays, structures, and matrices. It
1605 * is also safe if the varying is only used for transform feedback.
1608 varying_matches::is_varying_packing_safe(const glsl_type
*type
,
1609 const ir_variable
*var
) const
1611 if (consumer_stage
== MESA_SHADER_TESS_EVAL
||
1612 consumer_stage
== MESA_SHADER_TESS_CTRL
||
1613 producer_stage
== MESA_SHADER_TESS_CTRL
)
1616 return xfb_enabled
&& (type
->is_array() || type
->is_record() ||
1617 type
->is_matrix() || var
->data
.is_xfb_only
);
1622 * Record the given producer/consumer variable pair in the list of variables
1623 * that should later be assigned locations.
1625 * It is permissible for \c consumer_var to be NULL (this happens if a
1626 * variable is output by the producer and consumed by transform feedback, but
1627 * not consumed by the consumer).
1629 * If \c producer_var has already been paired up with a consumer_var, or
1630 * producer_var is part of fixed pipeline functionality (and hence already has
1631 * a location assigned), this function has no effect.
1633 * Note: as a side effect this function may change the interpolation type of
1634 * \c producer_var, but only when the change couldn't possibly affect
1638 varying_matches::record(ir_variable
*producer_var
, ir_variable
*consumer_var
)
1640 assert(producer_var
!= NULL
|| consumer_var
!= NULL
);
1642 if ((producer_var
&& (!producer_var
->data
.is_unmatched_generic_inout
||
1643 producer_var
->data
.explicit_location
)) ||
1644 (consumer_var
&& (!consumer_var
->data
.is_unmatched_generic_inout
||
1645 consumer_var
->data
.explicit_location
))) {
1646 /* Either a location already exists for this variable (since it is part
1647 * of fixed functionality), or it has already been recorded as part of a
1653 bool needs_flat_qualifier
= consumer_var
== NULL
&&
1654 (producer_var
->type
->contains_integer() ||
1655 producer_var
->type
->contains_double());
1657 if (!disable_varying_packing
&&
1658 (needs_flat_qualifier
||
1659 (consumer_stage
!= MESA_SHADER_NONE
&& consumer_stage
!= MESA_SHADER_FRAGMENT
))) {
1660 /* Since this varying is not being consumed by the fragment shader, its
1661 * interpolation type varying cannot possibly affect rendering.
1662 * Also, this variable is non-flat and is (or contains) an integer
1664 * If the consumer stage is unknown, don't modify the interpolation
1665 * type as it could affect rendering later with separate shaders.
1667 * lower_packed_varyings requires all integer varyings to flat,
1668 * regardless of where they appear. We can trivially satisfy that
1669 * requirement by changing the interpolation type to flat here.
1672 producer_var
->data
.centroid
= false;
1673 producer_var
->data
.sample
= false;
1674 producer_var
->data
.interpolation
= INTERP_MODE_FLAT
;
1678 consumer_var
->data
.centroid
= false;
1679 consumer_var
->data
.sample
= false;
1680 consumer_var
->data
.interpolation
= INTERP_MODE_FLAT
;
1684 if (this->num_matches
== this->matches_capacity
) {
1685 this->matches_capacity
*= 2;
1686 this->matches
= (match
*)
1687 realloc(this->matches
,
1688 sizeof(*this->matches
) * this->matches_capacity
);
1691 /* We must use the consumer to compute the packing class because in GL4.4+
1692 * there is no guarantee interpolation qualifiers will match across stages.
1694 * From Section 4.5 (Interpolation Qualifiers) of the GLSL 4.30 spec:
1696 * "The type and presence of interpolation qualifiers of variables with
1697 * the same name declared in all linked shaders for the same cross-stage
1698 * interface must match, otherwise the link command will fail.
1700 * When comparing an output from one stage to an input of a subsequent
1701 * stage, the input and output don't match if their interpolation
1702 * qualifiers (or lack thereof) are not the same."
1704 * This text was also in at least revison 7 of the 4.40 spec but is no
1705 * longer in revision 9 and not in the 4.50 spec.
1707 const ir_variable
*const var
= (consumer_var
!= NULL
)
1708 ? consumer_var
: producer_var
;
1709 const gl_shader_stage stage
= (consumer_var
!= NULL
)
1710 ? consumer_stage
: producer_stage
;
1711 const glsl_type
*type
= get_varying_type(var
, stage
);
1713 if (producer_var
&& consumer_var
&&
1714 consumer_var
->data
.must_be_shader_input
) {
1715 producer_var
->data
.must_be_shader_input
= 1;
1718 this->matches
[this->num_matches
].packing_class
1719 = this->compute_packing_class(var
);
1720 this->matches
[this->num_matches
].packing_order
1721 = this->compute_packing_order(var
);
1722 if ((this->disable_varying_packing
&& !is_varying_packing_safe(type
, var
)) ||
1723 var
->data
.must_be_shader_input
) {
1724 unsigned slots
= type
->count_attribute_slots(false);
1725 this->matches
[this->num_matches
].num_components
= slots
* 4;
1727 this->matches
[this->num_matches
].num_components
1728 = type
->component_slots();
1731 this->matches
[this->num_matches
].producer_var
= producer_var
;
1732 this->matches
[this->num_matches
].consumer_var
= consumer_var
;
1733 this->num_matches
++;
1735 producer_var
->data
.is_unmatched_generic_inout
= 0;
1737 consumer_var
->data
.is_unmatched_generic_inout
= 0;
1742 * Choose locations for all of the variable matches that were previously
1743 * passed to varying_matches::record().
1744 * \param components returns array[slot] of number of components used
1745 * per slot (1, 2, 3 or 4)
1746 * \param reserved_slots bitmask indicating which varying slots are already
1748 * \return number of slots (4-element vectors) allocated
1751 varying_matches::assign_locations(struct gl_shader_program
*prog
,
1752 uint8_t components
[],
1753 uint64_t reserved_slots
)
1755 /* If packing has been disabled then we cannot safely sort the varyings by
1756 * class as it may mean we are using a version of OpenGL where
1757 * interpolation qualifiers are not guaranteed to be matching across
1758 * shaders, sorting in this case could result in mismatching shader
1760 * When packing is disabled the sort orders varyings used by transform
1761 * feedback first, but also depends on *undefined behaviour* of qsort to
1762 * reverse the order of the varyings. See: xfb_comparator().
1764 if (!this->disable_varying_packing
) {
1765 /* Sort varying matches into an order that makes them easy to pack. */
1766 qsort(this->matches
, this->num_matches
, sizeof(*this->matches
),
1767 &varying_matches::match_comparator
);
1769 /* Only sort varyings that are only used by transform feedback. */
1770 qsort(this->matches
, this->num_matches
, sizeof(*this->matches
),
1771 &varying_matches::xfb_comparator
);
1774 unsigned generic_location
= 0;
1775 unsigned generic_patch_location
= MAX_VARYING
*4;
1776 bool previous_var_xfb_only
= false;
1777 unsigned previous_packing_class
= ~0u;
1779 /* For tranform feedback separate mode, we know the number of attributes
1780 * is <= the number of buffers. So packing isn't critical. In fact,
1781 * packing vec3 attributes can cause trouble because splitting a vec3
1782 * effectively creates an additional transform feedback output. The
1783 * extra TFB output may exceed device driver limits.
1785 const bool dont_pack_vec3
=
1786 (prog
->TransformFeedback
.BufferMode
== GL_SEPARATE_ATTRIBS
&&
1787 prog
->TransformFeedback
.NumVarying
> 0);
1789 for (unsigned i
= 0; i
< this->num_matches
; i
++) {
1790 unsigned *location
= &generic_location
;
1791 const ir_variable
*var
;
1792 const glsl_type
*type
;
1793 bool is_vertex_input
= false;
1795 if (matches
[i
].consumer_var
) {
1796 var
= matches
[i
].consumer_var
;
1797 type
= get_varying_type(var
, consumer_stage
);
1798 if (consumer_stage
== MESA_SHADER_VERTEX
)
1799 is_vertex_input
= true;
1801 var
= matches
[i
].producer_var
;
1802 type
= get_varying_type(var
, producer_stage
);
1805 if (var
->data
.patch
)
1806 location
= &generic_patch_location
;
1808 /* Advance to the next slot if this varying has a different packing
1809 * class than the previous one, and we're not already on a slot
1812 * Also advance to the next slot if packing is disabled. This makes sure
1813 * we don't assign varyings the same locations which is possible
1814 * because we still pack individual arrays, records and matrices even
1815 * when packing is disabled. Note we don't advance to the next slot if
1816 * we can pack varyings together that are only used for transform
1819 if (var
->data
.must_be_shader_input
||
1820 (this->disable_varying_packing
&&
1821 !(previous_var_xfb_only
&& var
->data
.is_xfb_only
)) ||
1822 (previous_packing_class
!= this->matches
[i
].packing_class
) ||
1823 (this->matches
[i
].packing_order
== PACKING_ORDER_VEC3
&&
1825 *location
= ALIGN(*location
, 4);
1828 previous_var_xfb_only
= var
->data
.is_xfb_only
;
1829 previous_packing_class
= this->matches
[i
].packing_class
;
1831 /* The number of components taken up by this variable. For vertex shader
1832 * inputs, we use the number of slots * 4, as they have different
1835 unsigned num_components
= is_vertex_input
?
1836 type
->count_attribute_slots(is_vertex_input
) * 4 :
1837 this->matches
[i
].num_components
;
1839 /* The last slot for this variable, inclusive. */
1840 unsigned slot_end
= *location
+ num_components
- 1;
1842 /* FIXME: We could be smarter in the below code and loop back over
1843 * trying to fill any locations that we skipped because we couldn't pack
1844 * the varying between an explicit location. For now just let the user
1845 * hit the linking error if we run out of room and suggest they use
1846 * explicit locations.
1848 while (slot_end
< MAX_VARYING
* 4u) {
1849 const unsigned slots
= (slot_end
/ 4u) - (*location
/ 4u) + 1;
1850 const uint64_t slot_mask
= ((1ull << slots
) - 1) << (*location
/ 4u);
1854 if ((reserved_slots
& slot_mask
) == 0) {
1858 *location
= ALIGN(*location
+ 1, 4);
1859 slot_end
= *location
+ num_components
- 1;
1862 if (!var
->data
.patch
&& slot_end
>= MAX_VARYING
* 4u) {
1863 linker_error(prog
, "insufficient contiguous locations available for "
1864 "%s it is possible an array or struct could not be "
1865 "packed between varyings with explicit locations. Try "
1866 "using an explicit location for arrays and structs.",
1870 if (slot_end
< MAX_VARYINGS_INCL_PATCH
* 4u) {
1871 for (unsigned j
= *location
/ 4u; j
< slot_end
/ 4u; j
++)
1873 components
[slot_end
/ 4u] = (slot_end
& 3) + 1;
1876 this->matches
[i
].generic_location
= *location
;
1878 *location
= slot_end
+ 1;
1881 return (generic_location
+ 3) / 4;
1886 * Update the producer and consumer shaders to reflect the locations
1887 * assignments that were made by varying_matches::assign_locations().
1890 varying_matches::store_locations() const
1892 /* Check is location needs to be packed with lower_packed_varyings() or if
1893 * we can just use ARB_enhanced_layouts packing.
1895 bool pack_loc
[MAX_VARYINGS_INCL_PATCH
] = { 0 };
1896 const glsl_type
*loc_type
[MAX_VARYINGS_INCL_PATCH
][4] = { {NULL
, NULL
} };
1898 for (unsigned i
= 0; i
< this->num_matches
; i
++) {
1899 ir_variable
*producer_var
= this->matches
[i
].producer_var
;
1900 ir_variable
*consumer_var
= this->matches
[i
].consumer_var
;
1901 unsigned generic_location
= this->matches
[i
].generic_location
;
1902 unsigned slot
= generic_location
/ 4;
1903 unsigned offset
= generic_location
% 4;
1906 producer_var
->data
.location
= VARYING_SLOT_VAR0
+ slot
;
1907 producer_var
->data
.location_frac
= offset
;
1911 assert(consumer_var
->data
.location
== -1);
1912 consumer_var
->data
.location
= VARYING_SLOT_VAR0
+ slot
;
1913 consumer_var
->data
.location_frac
= offset
;
1916 /* Find locations suitable for native packing via
1917 * ARB_enhanced_layouts.
1919 if (producer_var
&& consumer_var
) {
1920 if (enhanced_layouts_enabled
) {
1921 const glsl_type
*type
=
1922 get_varying_type(producer_var
, producer_stage
);
1923 if (type
->is_array() || type
->is_matrix() || type
->is_record() ||
1924 type
->is_double()) {
1925 unsigned comp_slots
= type
->component_slots() + offset
;
1926 unsigned slots
= comp_slots
/ 4;
1930 for (unsigned j
= 0; j
< slots
; j
++) {
1931 pack_loc
[slot
+ j
] = true;
1933 } else if (offset
+ type
->vector_elements
> 4) {
1934 pack_loc
[slot
] = true;
1935 pack_loc
[slot
+ 1] = true;
1937 loc_type
[slot
][offset
] = type
;
1943 /* Attempt to use ARB_enhanced_layouts for more efficient packing if
1946 if (enhanced_layouts_enabled
) {
1947 for (unsigned i
= 0; i
< this->num_matches
; i
++) {
1948 ir_variable
*producer_var
= this->matches
[i
].producer_var
;
1949 ir_variable
*consumer_var
= this->matches
[i
].consumer_var
;
1950 unsigned generic_location
= this->matches
[i
].generic_location
;
1951 unsigned slot
= generic_location
/ 4;
1953 if (pack_loc
[slot
] || !producer_var
|| !consumer_var
)
1956 const glsl_type
*type
=
1957 get_varying_type(producer_var
, producer_stage
);
1958 bool type_match
= true;
1959 for (unsigned j
= 0; j
< 4; j
++) {
1960 if (loc_type
[slot
][j
]) {
1961 if (type
->base_type
!= loc_type
[slot
][j
]->base_type
)
1967 producer_var
->data
.explicit_location
= 1;
1968 consumer_var
->data
.explicit_location
= 1;
1969 producer_var
->data
.explicit_component
= 1;
1970 consumer_var
->data
.explicit_component
= 1;
1978 * Compute the "packing class" of the given varying. This is an unsigned
1979 * integer with the property that two variables in the same packing class can
1980 * be safely backed into the same vec4.
1983 varying_matches::compute_packing_class(const ir_variable
*var
)
1985 /* Without help from the back-end, there is no way to pack together
1986 * variables with different interpolation types, because
1987 * lower_packed_varyings must choose exactly one interpolation type for
1988 * each packed varying it creates.
1990 * However, we can safely pack together floats, ints, and uints, because:
1992 * - varyings of base type "int" and "uint" must use the "flat"
1993 * interpolation type, which can only occur in GLSL 1.30 and above.
1995 * - On platforms that support GLSL 1.30 and above, lower_packed_varyings
1996 * can store flat floats as ints without losing any information (using
1997 * the ir_unop_bitcast_* opcodes).
1999 * Therefore, the packing class depends only on the interpolation type.
2001 const unsigned interp
= var
->is_interpolation_flat()
2002 ? unsigned(INTERP_MODE_FLAT
) : var
->data
.interpolation
;
2004 assert(interp
< (1 << 3));
2006 const unsigned packing_class
= (interp
<< 0) |
2007 (var
->data
.centroid
<< 3) |
2008 (var
->data
.sample
<< 4) |
2009 (var
->data
.patch
<< 5) |
2010 (var
->data
.must_be_shader_input
<< 6);
2012 return packing_class
;
2017 * Compute the "packing order" of the given varying. This is a sort key we
2018 * use to determine when to attempt to pack the given varying relative to
2019 * other varyings in the same packing class.
2021 varying_matches::packing_order_enum
2022 varying_matches::compute_packing_order(const ir_variable
*var
)
2024 const glsl_type
*element_type
= var
->type
;
2026 while (element_type
->is_array()) {
2027 element_type
= element_type
->fields
.array
;
2030 switch (element_type
->component_slots() % 4) {
2031 case 1: return PACKING_ORDER_SCALAR
;
2032 case 2: return PACKING_ORDER_VEC2
;
2033 case 3: return PACKING_ORDER_VEC3
;
2034 case 0: return PACKING_ORDER_VEC4
;
2036 assert(!"Unexpected value of vector_elements");
2037 return PACKING_ORDER_VEC4
;
2043 * Comparison function passed to qsort() to sort varyings by packing_class and
2044 * then by packing_order.
2047 varying_matches::match_comparator(const void *x_generic
, const void *y_generic
)
2049 const match
*x
= (const match
*) x_generic
;
2050 const match
*y
= (const match
*) y_generic
;
2052 if (x
->packing_class
!= y
->packing_class
)
2053 return x
->packing_class
- y
->packing_class
;
2054 return x
->packing_order
- y
->packing_order
;
2059 * Comparison function passed to qsort() to sort varyings used only by
2060 * transform feedback when packing of other varyings is disabled.
2063 varying_matches::xfb_comparator(const void *x_generic
, const void *y_generic
)
2065 const match
*x
= (const match
*) x_generic
;
2067 if (x
->producer_var
!= NULL
&& x
->producer_var
->data
.is_xfb_only
)
2068 return match_comparator(x_generic
, y_generic
);
2070 /* FIXME: When the comparator returns 0 it means the elements being
2071 * compared are equivalent. However the qsort documentation says:
2073 * "The order of equivalent elements is undefined."
2075 * In practice the sort ends up reversing the order of the varyings which
2076 * means locations are also assigned in this reversed order and happens to
2077 * be what we want. This is also whats happening in
2078 * varying_matches::match_comparator().
2085 * Is the given variable a varying variable to be counted against the
2086 * limit in ctx->Const.MaxVarying?
2087 * This includes variables such as texcoords, colors and generic
2088 * varyings, but excludes variables such as gl_FrontFacing and gl_FragCoord.
2091 var_counts_against_varying_limit(gl_shader_stage stage
, const ir_variable
*var
)
2093 /* Only fragment shaders will take a varying variable as an input */
2094 if (stage
== MESA_SHADER_FRAGMENT
&&
2095 var
->data
.mode
== ir_var_shader_in
) {
2096 switch (var
->data
.location
) {
2097 case VARYING_SLOT_POS
:
2098 case VARYING_SLOT_FACE
:
2099 case VARYING_SLOT_PNTC
:
2110 * Visitor class that generates tfeedback_candidate structs describing all
2111 * possible targets of transform feedback.
2113 * tfeedback_candidate structs are stored in the hash table
2114 * tfeedback_candidates, which is passed to the constructor. This hash table
2115 * maps varying names to instances of the tfeedback_candidate struct.
2117 class tfeedback_candidate_generator
: public program_resource_visitor
2120 tfeedback_candidate_generator(void *mem_ctx
,
2121 hash_table
*tfeedback_candidates
)
2123 tfeedback_candidates(tfeedback_candidates
),
2129 void process(ir_variable
*var
)
2131 /* All named varying interface blocks should be flattened by now */
2132 assert(!var
->is_interface_instance());
2134 this->toplevel_var
= var
;
2135 this->varying_floats
= 0;
2136 program_resource_visitor::process(var
, false);
2140 virtual void visit_field(const glsl_type
*type
, const char *name
,
2141 bool /* row_major */,
2142 const glsl_type
* /* record_type */,
2143 const enum glsl_interface_packing
,
2144 bool /* last_field */)
2146 assert(!type
->without_array()->is_record());
2147 assert(!type
->without_array()->is_interface());
2149 tfeedback_candidate
*candidate
2150 = rzalloc(this->mem_ctx
, tfeedback_candidate
);
2151 candidate
->toplevel_var
= this->toplevel_var
;
2152 candidate
->type
= type
;
2153 candidate
->offset
= this->varying_floats
;
2154 _mesa_hash_table_insert(this->tfeedback_candidates
,
2155 ralloc_strdup(this->mem_ctx
, name
),
2157 this->varying_floats
+= type
->component_slots();
2161 * Memory context used to allocate hash table keys and values.
2163 void * const mem_ctx
;
2166 * Hash table in which tfeedback_candidate objects should be stored.
2168 hash_table
* const tfeedback_candidates
;
2171 * Pointer to the toplevel variable that is being traversed.
2173 ir_variable
*toplevel_var
;
2176 * Total number of varying floats that have been visited so far. This is
2177 * used to determine the offset to each varying within the toplevel
2180 unsigned varying_floats
;
2187 populate_consumer_input_sets(void *mem_ctx
, exec_list
*ir
,
2188 hash_table
*consumer_inputs
,
2189 hash_table
*consumer_interface_inputs
,
2190 ir_variable
*consumer_inputs_with_locations
[VARYING_SLOT_TESS_MAX
])
2192 memset(consumer_inputs_with_locations
,
2194 sizeof(consumer_inputs_with_locations
[0]) * VARYING_SLOT_TESS_MAX
);
2196 foreach_in_list(ir_instruction
, node
, ir
) {
2197 ir_variable
*const input_var
= node
->as_variable();
2199 if (input_var
!= NULL
&& input_var
->data
.mode
== ir_var_shader_in
) {
2200 /* All interface blocks should have been lowered by this point */
2201 assert(!input_var
->type
->is_interface());
2203 if (input_var
->data
.explicit_location
) {
2204 /* assign_varying_locations only cares about finding the
2205 * ir_variable at the start of a contiguous location block.
2207 * - For !producer, consumer_inputs_with_locations isn't used.
2209 * - For !consumer, consumer_inputs_with_locations is empty.
2211 * For consumer && producer, if you were trying to set some
2212 * ir_variable to the middle of a location block on the other side
2213 * of producer/consumer, cross_validate_outputs_to_inputs() should
2214 * be link-erroring due to either type mismatch or location
2215 * overlaps. If the variables do match up, then they've got a
2216 * matching data.location and you only looked at
2217 * consumer_inputs_with_locations[var->data.location], not any
2218 * following entries for the array/structure.
2220 consumer_inputs_with_locations
[input_var
->data
.location
] =
2222 } else if (input_var
->get_interface_type() != NULL
) {
2223 char *const iface_field_name
=
2224 ralloc_asprintf(mem_ctx
, "%s.%s",
2225 input_var
->get_interface_type()->without_array()->name
,
2227 _mesa_hash_table_insert(consumer_interface_inputs
,
2228 iface_field_name
, input_var
);
2230 _mesa_hash_table_insert(consumer_inputs
,
2231 ralloc_strdup(mem_ctx
, input_var
->name
),
2239 * Find a variable from the consumer that "matches" the specified variable
2241 * This function only finds inputs with names that match. There is no
2242 * validation (here) that the types, etc. are compatible.
2245 get_matching_input(void *mem_ctx
,
2246 const ir_variable
*output_var
,
2247 hash_table
*consumer_inputs
,
2248 hash_table
*consumer_interface_inputs
,
2249 ir_variable
*consumer_inputs_with_locations
[VARYING_SLOT_TESS_MAX
])
2251 ir_variable
*input_var
;
2253 if (output_var
->data
.explicit_location
) {
2254 input_var
= consumer_inputs_with_locations
[output_var
->data
.location
];
2255 } else if (output_var
->get_interface_type() != NULL
) {
2256 char *const iface_field_name
=
2257 ralloc_asprintf(mem_ctx
, "%s.%s",
2258 output_var
->get_interface_type()->without_array()->name
,
2260 hash_entry
*entry
= _mesa_hash_table_search(consumer_interface_inputs
, iface_field_name
);
2261 input_var
= entry
? (ir_variable
*) entry
->data
: NULL
;
2263 hash_entry
*entry
= _mesa_hash_table_search(consumer_inputs
, output_var
->name
);
2264 input_var
= entry
? (ir_variable
*) entry
->data
: NULL
;
2267 return (input_var
== NULL
|| input_var
->data
.mode
!= ir_var_shader_in
)
2274 io_variable_cmp(const void *_a
, const void *_b
)
2276 const ir_variable
*const a
= *(const ir_variable
**) _a
;
2277 const ir_variable
*const b
= *(const ir_variable
**) _b
;
2279 if (a
->data
.explicit_location
&& b
->data
.explicit_location
)
2280 return b
->data
.location
- a
->data
.location
;
2282 if (a
->data
.explicit_location
&& !b
->data
.explicit_location
)
2285 if (!a
->data
.explicit_location
&& b
->data
.explicit_location
)
2288 return -strcmp(a
->name
, b
->name
);
2292 * Sort the shader IO variables into canonical order
2295 canonicalize_shader_io(exec_list
*ir
, enum ir_variable_mode io_mode
)
2297 ir_variable
*var_table
[MAX_PROGRAM_OUTPUTS
* 4];
2298 unsigned num_variables
= 0;
2300 foreach_in_list(ir_instruction
, node
, ir
) {
2301 ir_variable
*const var
= node
->as_variable();
2303 if (var
== NULL
|| var
->data
.mode
!= io_mode
)
2306 /* If we have already encountered more I/O variables that could
2307 * successfully link, bail.
2309 if (num_variables
== ARRAY_SIZE(var_table
))
2312 var_table
[num_variables
++] = var
;
2315 if (num_variables
== 0)
2318 /* Sort the list in reverse order (io_variable_cmp handles this). Later
2319 * we're going to push the variables on to the IR list as a stack, so we
2320 * want the last variable (in canonical order) to be first in the list.
2322 qsort(var_table
, num_variables
, sizeof(var_table
[0]), io_variable_cmp
);
2324 /* Remove the variable from it's current location in the IR, and put it at
2327 for (unsigned i
= 0; i
< num_variables
; i
++) {
2328 var_table
[i
]->remove();
2329 ir
->push_head(var_table
[i
]);
2334 * Generate a bitfield map of the explicit locations for shader varyings.
2336 * Note: For Tessellation shaders we are sitting right on the limits of the
2337 * 64 bit map. Per-vertex and per-patch both have separate location domains
2338 * with a max of MAX_VARYING.
2341 reserved_varying_slot(struct gl_linked_shader
*stage
,
2342 ir_variable_mode io_mode
)
2344 assert(io_mode
== ir_var_shader_in
|| io_mode
== ir_var_shader_out
);
2345 /* Avoid an overflow of the returned value */
2346 assert(MAX_VARYINGS_INCL_PATCH
<= 64);
2354 foreach_in_list(ir_instruction
, node
, stage
->ir
) {
2355 ir_variable
*const var
= node
->as_variable();
2357 if (var
== NULL
|| var
->data
.mode
!= io_mode
||
2358 !var
->data
.explicit_location
||
2359 var
->data
.location
< VARYING_SLOT_VAR0
)
2362 var_slot
= var
->data
.location
- VARYING_SLOT_VAR0
;
2364 unsigned num_elements
= get_varying_type(var
, stage
->Stage
)
2365 ->count_attribute_slots(io_mode
== ir_var_shader_in
&&
2366 stage
->Stage
== MESA_SHADER_VERTEX
);
2367 for (unsigned i
= 0; i
< num_elements
; i
++) {
2368 if (var_slot
>= 0 && var_slot
< MAX_VARYINGS_INCL_PATCH
)
2369 slots
|= UINT64_C(1) << var_slot
;
2379 * Assign locations for all variables that are produced in one pipeline stage
2380 * (the "producer") and consumed in the next stage (the "consumer").
2382 * Variables produced by the producer may also be consumed by transform
2385 * \param num_tfeedback_decls is the number of declarations indicating
2386 * variables that may be consumed by transform feedback.
2388 * \param tfeedback_decls is a pointer to an array of tfeedback_decl objects
2389 * representing the result of parsing the strings passed to
2390 * glTransformFeedbackVaryings(). assign_location() will be called for
2391 * each of these objects that matches one of the outputs of the
2394 * When num_tfeedback_decls is nonzero, it is permissible for the consumer to
2395 * be NULL. In this case, varying locations are assigned solely based on the
2396 * requirements of transform feedback.
2399 assign_varying_locations(struct gl_context
*ctx
,
2401 struct gl_shader_program
*prog
,
2402 gl_linked_shader
*producer
,
2403 gl_linked_shader
*consumer
,
2404 unsigned num_tfeedback_decls
,
2405 tfeedback_decl
*tfeedback_decls
,
2406 const uint64_t reserved_slots
)
2408 /* Tessellation shaders treat inputs and outputs as shared memory and can
2409 * access inputs and outputs of other invocations.
2410 * Therefore, they can't be lowered to temps easily (and definitely not
2413 bool unpackable_tess
=
2414 (consumer
&& consumer
->Stage
== MESA_SHADER_TESS_EVAL
) ||
2415 (consumer
&& consumer
->Stage
== MESA_SHADER_TESS_CTRL
) ||
2416 (producer
&& producer
->Stage
== MESA_SHADER_TESS_CTRL
);
2418 /* Transform feedback code assumes varying arrays are packed, so if the
2419 * driver has disabled varying packing, make sure to at least enable
2420 * packing required by transform feedback.
2423 ctx
->Extensions
.EXT_transform_feedback
&& !unpackable_tess
;
2425 /* Disable packing on outward facing interfaces for SSO because in ES we
2426 * need to retain the unpacked varying information for draw time
2429 * Packing is still enabled on individual arrays, structs, and matrices as
2430 * these are required by the transform feedback code and it is still safe
2431 * to do so. We also enable packing when a varying is only used for
2432 * transform feedback and its not a SSO.
2434 bool disable_varying_packing
=
2435 ctx
->Const
.DisableVaryingPacking
|| unpackable_tess
;
2436 if (prog
->SeparateShader
&& (producer
== NULL
|| consumer
== NULL
))
2437 disable_varying_packing
= true;
2439 varying_matches
matches(disable_varying_packing
, xfb_enabled
,
2440 ctx
->Extensions
.ARB_enhanced_layouts
,
2441 producer
? producer
->Stage
: MESA_SHADER_NONE
,
2442 consumer
? consumer
->Stage
: MESA_SHADER_NONE
);
2443 hash_table
*tfeedback_candidates
=
2444 _mesa_hash_table_create(NULL
, _mesa_key_hash_string
,
2445 _mesa_key_string_equal
);
2446 hash_table
*consumer_inputs
=
2447 _mesa_hash_table_create(NULL
, _mesa_key_hash_string
,
2448 _mesa_key_string_equal
);
2449 hash_table
*consumer_interface_inputs
=
2450 _mesa_hash_table_create(NULL
, _mesa_key_hash_string
,
2451 _mesa_key_string_equal
);
2452 ir_variable
*consumer_inputs_with_locations
[VARYING_SLOT_TESS_MAX
] = {
2456 unsigned consumer_vertices
= 0;
2457 if (consumer
&& consumer
->Stage
== MESA_SHADER_GEOMETRY
)
2458 consumer_vertices
= prog
->Geom
.VerticesIn
;
2460 /* Operate in a total of four passes.
2462 * 1. Sort inputs / outputs into a canonical order. This is necessary so
2463 * that inputs / outputs of separable shaders will be assigned
2464 * predictable locations regardless of the order in which declarations
2465 * appeared in the shader source.
2467 * 2. Assign locations for any matching inputs and outputs.
2469 * 3. Mark output variables in the producer that do not have locations as
2470 * not being outputs. This lets the optimizer eliminate them.
2472 * 4. Mark input variables in the consumer that do not have locations as
2473 * not being inputs. This lets the optimizer eliminate them.
2476 canonicalize_shader_io(consumer
->ir
, ir_var_shader_in
);
2479 canonicalize_shader_io(producer
->ir
, ir_var_shader_out
);
2482 linker::populate_consumer_input_sets(mem_ctx
, consumer
->ir
,
2484 consumer_interface_inputs
,
2485 consumer_inputs_with_locations
);
2488 foreach_in_list(ir_instruction
, node
, producer
->ir
) {
2489 ir_variable
*const output_var
= node
->as_variable();
2491 if (output_var
== NULL
|| output_var
->data
.mode
!= ir_var_shader_out
)
2494 /* Only geometry shaders can use non-zero streams */
2495 assert(output_var
->data
.stream
== 0 ||
2496 (output_var
->data
.stream
< MAX_VERTEX_STREAMS
&&
2497 producer
->Stage
== MESA_SHADER_GEOMETRY
));
2499 if (num_tfeedback_decls
> 0) {
2500 tfeedback_candidate_generator
g(mem_ctx
, tfeedback_candidates
);
2501 g
.process(output_var
);
2504 ir_variable
*const input_var
=
2505 linker::get_matching_input(mem_ctx
, output_var
, consumer_inputs
,
2506 consumer_interface_inputs
,
2507 consumer_inputs_with_locations
);
2509 /* If a matching input variable was found, add this output (and the
2510 * input) to the set. If this is a separable program and there is no
2511 * consumer stage, add the output.
2513 * Always add TCS outputs. They are shared by all invocations
2514 * within a patch and can be used as shared memory.
2516 if (input_var
|| (prog
->SeparateShader
&& consumer
== NULL
) ||
2517 producer
->Stage
== MESA_SHADER_TESS_CTRL
) {
2518 matches
.record(output_var
, input_var
);
2521 /* Only stream 0 outputs can be consumed in the next stage */
2522 if (input_var
&& output_var
->data
.stream
!= 0) {
2523 linker_error(prog
, "output %s is assigned to stream=%d but "
2524 "is linked to an input, which requires stream=0",
2525 output_var
->name
, output_var
->data
.stream
);
2530 /* If there's no producer stage, then this must be a separable program.
2531 * For example, we may have a program that has just a fragment shader.
2532 * Later this program will be used with some arbitrary vertex (or
2533 * geometry) shader program. This means that locations must be assigned
2534 * for all the inputs.
2536 foreach_in_list(ir_instruction
, node
, consumer
->ir
) {
2537 ir_variable
*const input_var
= node
->as_variable();
2538 if (input_var
&& input_var
->data
.mode
== ir_var_shader_in
) {
2539 matches
.record(NULL
, input_var
);
2544 for (unsigned i
= 0; i
< num_tfeedback_decls
; ++i
) {
2545 if (!tfeedback_decls
[i
].is_varying())
2548 const tfeedback_candidate
*matched_candidate
2549 = tfeedback_decls
[i
].find_candidate(prog
, tfeedback_candidates
);
2551 if (matched_candidate
== NULL
) {
2552 _mesa_hash_table_destroy(tfeedback_candidates
, NULL
);
2556 /* Mark xfb varyings as always active */
2557 matched_candidate
->toplevel_var
->data
.always_active_io
= 1;
2559 /* Mark any corresponding inputs as always active also. We must do this
2560 * because we have a NIR pass that lowers vectors to scalars and another
2561 * that removes unused varyings.
2562 * We don't split varyings marked as always active because there is no
2563 * point in doing so. This means we need to mark both sides of the
2564 * interface as always active otherwise we will have a mismatch and
2565 * start removing things we shouldn't.
2567 ir_variable
*const input_var
=
2568 linker::get_matching_input(mem_ctx
, matched_candidate
->toplevel_var
,
2570 consumer_interface_inputs
,
2571 consumer_inputs_with_locations
);
2573 input_var
->data
.always_active_io
= 1;
2575 if (matched_candidate
->toplevel_var
->data
.is_unmatched_generic_inout
) {
2576 matched_candidate
->toplevel_var
->data
.is_xfb_only
= 1;
2577 matches
.record(matched_candidate
->toplevel_var
, NULL
);
2581 _mesa_hash_table_destroy(consumer_inputs
, NULL
);
2582 _mesa_hash_table_destroy(consumer_interface_inputs
, NULL
);
2584 uint8_t components
[MAX_VARYINGS_INCL_PATCH
] = {0};
2585 const unsigned slots_used
= matches
.assign_locations(
2586 prog
, components
, reserved_slots
);
2587 matches
.store_locations();
2589 for (unsigned i
= 0; i
< num_tfeedback_decls
; ++i
) {
2590 if (tfeedback_decls
[i
].is_varying()) {
2591 if (!tfeedback_decls
[i
].assign_location(ctx
, prog
)) {
2592 _mesa_hash_table_destroy(tfeedback_candidates
, NULL
);
2597 _mesa_hash_table_destroy(tfeedback_candidates
, NULL
);
2599 if (consumer
&& producer
) {
2600 foreach_in_list(ir_instruction
, node
, consumer
->ir
) {
2601 ir_variable
*const var
= node
->as_variable();
2603 if (var
&& var
->data
.mode
== ir_var_shader_in
&&
2604 var
->data
.is_unmatched_generic_inout
) {
2605 if (!prog
->IsES
&& prog
->data
->Version
<= 120) {
2606 /* On page 25 (page 31 of the PDF) of the GLSL 1.20 spec:
2608 * Only those varying variables used (i.e. read) in
2609 * the fragment shader executable must be written to
2610 * by the vertex shader executable; declaring
2611 * superfluous varying variables in a vertex shader is
2614 * We interpret this text as meaning that the VS must
2615 * write the variable for the FS to read it. See
2616 * "glsl1-varying read but not written" in piglit.
2618 linker_error(prog
, "%s shader varying %s not written "
2620 _mesa_shader_stage_to_string(consumer
->Stage
),
2622 _mesa_shader_stage_to_string(producer
->Stage
));
2624 linker_warning(prog
, "%s shader varying %s not written "
2626 _mesa_shader_stage_to_string(consumer
->Stage
),
2628 _mesa_shader_stage_to_string(producer
->Stage
));
2633 /* Now that validation is done its safe to remove unused varyings. As
2634 * we have both a producer and consumer its safe to remove unused
2635 * varyings even if the program is a SSO because the stages are being
2636 * linked together i.e. we have a multi-stage SSO.
2638 remove_unused_shader_inputs_and_outputs(false, producer
,
2640 remove_unused_shader_inputs_and_outputs(false, consumer
,
2645 lower_packed_varyings(mem_ctx
, slots_used
, components
, ir_var_shader_out
,
2646 0, producer
, disable_varying_packing
,
2651 lower_packed_varyings(mem_ctx
, slots_used
, components
, ir_var_shader_in
,
2652 consumer_vertices
, consumer
,
2653 disable_varying_packing
, xfb_enabled
);
2660 check_against_output_limit(struct gl_context
*ctx
,
2661 struct gl_shader_program
*prog
,
2662 gl_linked_shader
*producer
,
2663 unsigned num_explicit_locations
)
2665 unsigned output_vectors
= num_explicit_locations
;
2667 foreach_in_list(ir_instruction
, node
, producer
->ir
) {
2668 ir_variable
*const var
= node
->as_variable();
2670 if (var
&& !var
->data
.explicit_location
&&
2671 var
->data
.mode
== ir_var_shader_out
&&
2672 var_counts_against_varying_limit(producer
->Stage
, var
)) {
2673 /* outputs for fragment shader can't be doubles */
2674 output_vectors
+= var
->type
->count_attribute_slots(false);
2678 assert(producer
->Stage
!= MESA_SHADER_FRAGMENT
);
2679 unsigned max_output_components
=
2680 ctx
->Const
.Program
[producer
->Stage
].MaxOutputComponents
;
2682 const unsigned output_components
= output_vectors
* 4;
2683 if (output_components
> max_output_components
) {
2684 if (ctx
->API
== API_OPENGLES2
|| prog
->IsES
)
2685 linker_error(prog
, "%s shader uses too many output vectors "
2687 _mesa_shader_stage_to_string(producer
->Stage
),
2689 max_output_components
/ 4);
2691 linker_error(prog
, "%s shader uses too many output components "
2693 _mesa_shader_stage_to_string(producer
->Stage
),
2695 max_output_components
);
2704 check_against_input_limit(struct gl_context
*ctx
,
2705 struct gl_shader_program
*prog
,
2706 gl_linked_shader
*consumer
,
2707 unsigned num_explicit_locations
)
2709 unsigned input_vectors
= num_explicit_locations
;
2711 foreach_in_list(ir_instruction
, node
, consumer
->ir
) {
2712 ir_variable
*const var
= node
->as_variable();
2714 if (var
&& !var
->data
.explicit_location
&&
2715 var
->data
.mode
== ir_var_shader_in
&&
2716 var_counts_against_varying_limit(consumer
->Stage
, var
)) {
2717 /* vertex inputs aren't varying counted */
2718 input_vectors
+= var
->type
->count_attribute_slots(false);
2722 assert(consumer
->Stage
!= MESA_SHADER_VERTEX
);
2723 unsigned max_input_components
=
2724 ctx
->Const
.Program
[consumer
->Stage
].MaxInputComponents
;
2726 const unsigned input_components
= input_vectors
* 4;
2727 if (input_components
> max_input_components
) {
2728 if (ctx
->API
== API_OPENGLES2
|| prog
->IsES
)
2729 linker_error(prog
, "%s shader uses too many input vectors "
2731 _mesa_shader_stage_to_string(consumer
->Stage
),
2733 max_input_components
/ 4);
2735 linker_error(prog
, "%s shader uses too many input components "
2737 _mesa_shader_stage_to_string(consumer
->Stage
),
2739 max_input_components
);
2748 link_varyings(struct gl_shader_program
*prog
, unsigned first
, unsigned last
,
2749 struct gl_context
*ctx
, void *mem_ctx
)
2751 bool has_xfb_qualifiers
= false;
2752 unsigned num_tfeedback_decls
= 0;
2753 char **varying_names
= NULL
;
2754 tfeedback_decl
*tfeedback_decls
= NULL
;
2756 /* From the ARB_enhanced_layouts spec:
2758 * "If the shader used to record output variables for transform feedback
2759 * varyings uses the "xfb_buffer", "xfb_offset", or "xfb_stride" layout
2760 * qualifiers, the values specified by TransformFeedbackVaryings are
2761 * ignored, and the set of variables captured for transform feedback is
2762 * instead derived from the specified layout qualifiers."
2764 for (int i
= MESA_SHADER_FRAGMENT
- 1; i
>= 0; i
--) {
2765 /* Find last stage before fragment shader */
2766 if (prog
->_LinkedShaders
[i
]) {
2767 has_xfb_qualifiers
=
2768 process_xfb_layout_qualifiers(mem_ctx
, prog
->_LinkedShaders
[i
],
2769 prog
, &num_tfeedback_decls
,
2775 if (!has_xfb_qualifiers
) {
2776 num_tfeedback_decls
= prog
->TransformFeedback
.NumVarying
;
2777 varying_names
= prog
->TransformFeedback
.VaryingNames
;
2780 if (num_tfeedback_decls
!= 0) {
2781 /* From GL_EXT_transform_feedback:
2782 * A program will fail to link if:
2784 * * the <count> specified by TransformFeedbackVaryingsEXT is
2785 * non-zero, but the program object has no vertex or geometry
2788 if (first
>= MESA_SHADER_FRAGMENT
) {
2789 linker_error(prog
, "Transform feedback varyings specified, but "
2790 "no vertex, tessellation, or geometry shader is "
2795 tfeedback_decls
= rzalloc_array(mem_ctx
, tfeedback_decl
,
2796 num_tfeedback_decls
);
2797 if (!parse_tfeedback_decls(ctx
, prog
, mem_ctx
, num_tfeedback_decls
,
2798 varying_names
, tfeedback_decls
))
2802 /* If there is no fragment shader we need to set transform feedback.
2804 * For SSO we also need to assign output locations. We assign them here
2805 * because we need to do it for both single stage programs and multi stage
2808 if (last
< MESA_SHADER_FRAGMENT
&&
2809 (num_tfeedback_decls
!= 0 || prog
->SeparateShader
)) {
2810 const uint64_t reserved_out_slots
=
2811 reserved_varying_slot(prog
->_LinkedShaders
[last
], ir_var_shader_out
);
2812 if (!assign_varying_locations(ctx
, mem_ctx
, prog
,
2813 prog
->_LinkedShaders
[last
], NULL
,
2814 num_tfeedback_decls
, tfeedback_decls
,
2815 reserved_out_slots
))
2819 if (last
<= MESA_SHADER_FRAGMENT
) {
2820 /* Remove unused varyings from the first/last stage unless SSO */
2821 remove_unused_shader_inputs_and_outputs(prog
->SeparateShader
,
2822 prog
->_LinkedShaders
[first
],
2824 remove_unused_shader_inputs_and_outputs(prog
->SeparateShader
,
2825 prog
->_LinkedShaders
[last
],
2828 /* If the program is made up of only a single stage */
2829 if (first
== last
) {
2830 gl_linked_shader
*const sh
= prog
->_LinkedShaders
[last
];
2832 do_dead_builtin_varyings(ctx
, NULL
, sh
, 0, NULL
);
2833 do_dead_builtin_varyings(ctx
, sh
, NULL
, num_tfeedback_decls
,
2836 if (prog
->SeparateShader
) {
2837 const uint64_t reserved_slots
=
2838 reserved_varying_slot(sh
, ir_var_shader_in
);
2840 /* Assign input locations for SSO, output locations are already
2843 if (!assign_varying_locations(ctx
, mem_ctx
, prog
,
2844 NULL
/* producer */,
2846 0 /* num_tfeedback_decls */,
2847 NULL
/* tfeedback_decls */,
2852 /* Linking the stages in the opposite order (from fragment to vertex)
2853 * ensures that inter-shader outputs written to in an earlier stage
2854 * are eliminated if they are (transitively) not used in a later
2858 for (int i
= next
- 1; i
>= 0; i
--) {
2859 if (prog
->_LinkedShaders
[i
] == NULL
&& i
!= 0)
2862 gl_linked_shader
*const sh_i
= prog
->_LinkedShaders
[i
];
2863 gl_linked_shader
*const sh_next
= prog
->_LinkedShaders
[next
];
2865 const uint64_t reserved_out_slots
=
2866 reserved_varying_slot(sh_i
, ir_var_shader_out
);
2867 const uint64_t reserved_in_slots
=
2868 reserved_varying_slot(sh_next
, ir_var_shader_in
);
2870 do_dead_builtin_varyings(ctx
, sh_i
, sh_next
,
2871 next
== MESA_SHADER_FRAGMENT
? num_tfeedback_decls
: 0,
2874 if (!assign_varying_locations(ctx
, mem_ctx
, prog
, sh_i
, sh_next
,
2875 next
== MESA_SHADER_FRAGMENT
? num_tfeedback_decls
: 0,
2877 reserved_out_slots
| reserved_in_slots
))
2880 /* This must be done after all dead varyings are eliminated. */
2882 unsigned slots_used
= _mesa_bitcount_64(reserved_out_slots
);
2883 if (!check_against_output_limit(ctx
, prog
, sh_i
, slots_used
)) {
2888 unsigned slots_used
= _mesa_bitcount_64(reserved_in_slots
);
2889 if (!check_against_input_limit(ctx
, prog
, sh_next
, slots_used
))
2897 if (!store_tfeedback_info(ctx
, prog
, num_tfeedback_decls
, tfeedback_decls
,
2898 has_xfb_qualifiers
))