2 * Copyright © 2012 Intel Corporation
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
21 * DEALINGS IN THE SOFTWARE.
25 * \file link_varyings.cpp
27 * Linker functions related specifically to linking varyings between shader
32 #include "main/mtypes.h"
33 #include "glsl_symbol_table.h"
34 #include "glsl_parser_extras.h"
35 #include "ir_optimization.h"
37 #include "link_varyings.h"
38 #include "main/macros.h"
39 #include "program/hash_table.h"
44 * Get the varying type stripped of the outermost array if we're processing
45 * a stage whose varyings are arrays indexed by a vertex number (such as
46 * geometry shader inputs).
48 static const glsl_type
*
49 get_varying_type(const ir_variable
*var
, gl_shader_stage stage
)
51 const glsl_type
*type
= var
->type
;
53 if (!var
->data
.patch
&&
54 ((var
->data
.mode
== ir_var_shader_out
&&
55 stage
== MESA_SHADER_TESS_CTRL
) ||
56 (var
->data
.mode
== ir_var_shader_in
&&
57 (stage
== MESA_SHADER_TESS_CTRL
|| stage
== MESA_SHADER_TESS_EVAL
||
58 stage
== MESA_SHADER_GEOMETRY
)))) {
59 assert(type
->is_array());
60 type
= type
->fields
.array
;
67 create_xfb_varying_names(void *mem_ctx
, const glsl_type
*t
, char **name
,
68 size_t name_length
, unsigned *count
,
69 const char *ifc_member_name
,
70 const glsl_type
*ifc_member_t
, char ***varying_names
)
72 if (t
->is_interface()) {
73 size_t new_length
= name_length
;
75 assert(ifc_member_name
&& ifc_member_t
);
76 ralloc_asprintf_rewrite_tail(name
, &new_length
, ".%s", ifc_member_name
);
78 create_xfb_varying_names(mem_ctx
, ifc_member_t
, name
, new_length
, count
,
79 NULL
, NULL
, varying_names
);
80 } else if (t
->is_record()) {
81 for (unsigned i
= 0; i
< t
->length
; i
++) {
82 const char *field
= t
->fields
.structure
[i
].name
;
83 size_t new_length
= name_length
;
85 ralloc_asprintf_rewrite_tail(name
, &new_length
, ".%s", field
);
87 create_xfb_varying_names(mem_ctx
, t
->fields
.structure
[i
].type
, name
,
88 new_length
, count
, NULL
, NULL
,
91 } else if (t
->without_array()->is_record() ||
92 t
->without_array()->is_interface() ||
93 (t
->is_array() && t
->fields
.array
->is_array())) {
94 for (unsigned i
= 0; i
< t
->length
; i
++) {
95 size_t new_length
= name_length
;
97 /* Append the subscript to the current variable name */
98 ralloc_asprintf_rewrite_tail(name
, &new_length
, "[%u]", i
);
100 create_xfb_varying_names(mem_ctx
, t
->fields
.array
, name
, new_length
,
101 count
, ifc_member_name
, ifc_member_t
,
105 (*varying_names
)[(*count
)++] = ralloc_strdup(mem_ctx
, *name
);
110 process_xfb_layout_qualifiers(void *mem_ctx
, const gl_shader
*sh
,
111 unsigned *num_tfeedback_decls
,
112 char ***varying_names
)
114 bool has_xfb_qualifiers
= false;
116 /* We still need to enable transform feedback mode even if xfb_stride is
117 * only applied to a global out. Also we don't bother to propagate
118 * xfb_stride to interface block members so this will catch that case also.
120 for (unsigned j
= 0; j
< MAX_FEEDBACK_BUFFERS
; j
++) {
121 if (sh
->TransformFeedback
.BufferStride
[j
]) {
122 has_xfb_qualifiers
= true;
126 foreach_in_list(ir_instruction
, node
, sh
->ir
) {
127 ir_variable
*var
= node
->as_variable();
128 if (!var
|| var
->data
.mode
!= ir_var_shader_out
)
131 /* From the ARB_enhanced_layouts spec:
133 * "Any shader making any static use (after preprocessing) of any of
134 * these *xfb_* qualifiers will cause the shader to be in a
135 * transform feedback capturing mode and hence responsible for
136 * describing the transform feedback setup. This mode will capture
137 * any output selected by *xfb_offset*, directly or indirectly, to
138 * a transform feedback buffer."
140 if (var
->data
.explicit_xfb_buffer
|| var
->data
.explicit_xfb_stride
) {
141 has_xfb_qualifiers
= true;
144 if (var
->data
.explicit_xfb_offset
) {
145 *num_tfeedback_decls
+= var
->type
->varying_count();
146 has_xfb_qualifiers
= true;
150 if (*num_tfeedback_decls
== 0)
151 return has_xfb_qualifiers
;
154 *varying_names
= ralloc_array(mem_ctx
, char *, *num_tfeedback_decls
);
155 foreach_in_list(ir_instruction
, node
, sh
->ir
) {
156 ir_variable
*var
= node
->as_variable();
157 if (!var
|| var
->data
.mode
!= ir_var_shader_out
)
160 if (var
->data
.explicit_xfb_offset
) {
162 const glsl_type
*type
, *member_type
;
164 if (var
->data
.from_named_ifc_block
) {
165 type
= var
->get_interface_type();
166 /* Find the member type before it was altered by lowering */
168 type
->fields
.structure
[type
->field_index(var
->name
)].type
;
169 name
= ralloc_strdup(NULL
, type
->without_array()->name
);
173 name
= ralloc_strdup(NULL
, var
->name
);
175 create_xfb_varying_names(mem_ctx
, type
, &name
, strlen(name
), &i
,
176 var
->name
, member_type
, varying_names
);
181 assert(i
== *num_tfeedback_decls
);
182 return has_xfb_qualifiers
;
186 * Validate the types and qualifiers of an output from one stage against the
187 * matching input to another stage.
190 cross_validate_types_and_qualifiers(struct gl_shader_program
*prog
,
191 const ir_variable
*input
,
192 const ir_variable
*output
,
193 gl_shader_stage consumer_stage
,
194 gl_shader_stage producer_stage
)
196 /* Check that the types match between stages.
198 const glsl_type
*type_to_match
= input
->type
;
200 /* VS -> GS, VS -> TCS, VS -> TES, TES -> GS */
201 const bool extra_array_level
= (producer_stage
== MESA_SHADER_VERTEX
&&
202 consumer_stage
!= MESA_SHADER_FRAGMENT
) ||
203 consumer_stage
== MESA_SHADER_GEOMETRY
;
204 if (extra_array_level
) {
205 assert(type_to_match
->is_array());
206 type_to_match
= type_to_match
->fields
.array
;
209 if (type_to_match
!= output
->type
) {
210 /* There is a bit of a special case for gl_TexCoord. This
211 * built-in is unsized by default. Applications that variable
212 * access it must redeclare it with a size. There is some
213 * language in the GLSL spec that implies the fragment shader
214 * and vertex shader do not have to agree on this size. Other
215 * driver behave this way, and one or two applications seem to
218 * Neither declaration needs to be modified here because the array
219 * sizes are fixed later when update_array_sizes is called.
221 * From page 48 (page 54 of the PDF) of the GLSL 1.10 spec:
223 * "Unlike user-defined varying variables, the built-in
224 * varying variables don't have a strict one-to-one
225 * correspondence between the vertex language and the
226 * fragment language."
228 if (!output
->type
->is_array() || !is_gl_identifier(output
->name
)) {
230 "%s shader output `%s' declared as type `%s', "
231 "but %s shader input declared as type `%s'\n",
232 _mesa_shader_stage_to_string(producer_stage
),
235 _mesa_shader_stage_to_string(consumer_stage
),
241 /* Check that all of the qualifiers match between stages.
244 /* According to the OpenGL and OpenGLES GLSL specs, the centroid qualifier
245 * should match until OpenGL 4.3 and OpenGLES 3.1. The OpenGLES 3.0
246 * conformance test suite does not verify that the qualifiers must match.
247 * The deqp test suite expects the opposite (OpenGLES 3.1) behavior for
248 * OpenGLES 3.0 drivers, so we relax the checking in all cases.
250 if (false /* always skip the centroid check */ &&
251 prog
->Version
< (prog
->IsES
? 310 : 430) &&
252 input
->data
.centroid
!= output
->data
.centroid
) {
254 "%s shader output `%s' %s centroid qualifier, "
255 "but %s shader input %s centroid qualifier\n",
256 _mesa_shader_stage_to_string(producer_stage
),
258 (output
->data
.centroid
) ? "has" : "lacks",
259 _mesa_shader_stage_to_string(consumer_stage
),
260 (input
->data
.centroid
) ? "has" : "lacks");
264 if (input
->data
.sample
!= output
->data
.sample
) {
266 "%s shader output `%s' %s sample qualifier, "
267 "but %s shader input %s sample qualifier\n",
268 _mesa_shader_stage_to_string(producer_stage
),
270 (output
->data
.sample
) ? "has" : "lacks",
271 _mesa_shader_stage_to_string(consumer_stage
),
272 (input
->data
.sample
) ? "has" : "lacks");
276 if (input
->data
.patch
!= output
->data
.patch
) {
278 "%s shader output `%s' %s patch qualifier, "
279 "but %s shader input %s patch qualifier\n",
280 _mesa_shader_stage_to_string(producer_stage
),
282 (output
->data
.patch
) ? "has" : "lacks",
283 _mesa_shader_stage_to_string(consumer_stage
),
284 (input
->data
.patch
) ? "has" : "lacks");
288 if (!prog
->IsES
&& input
->data
.invariant
!= output
->data
.invariant
) {
290 "%s shader output `%s' %s invariant qualifier, "
291 "but %s shader input %s invariant qualifier\n",
292 _mesa_shader_stage_to_string(producer_stage
),
294 (output
->data
.invariant
) ? "has" : "lacks",
295 _mesa_shader_stage_to_string(consumer_stage
),
296 (input
->data
.invariant
) ? "has" : "lacks");
300 /* GLSL >= 4.40 removes text requiring interpolation qualifiers
301 * to match cross stage, they must only match within the same stage.
303 * From page 84 (page 90 of the PDF) of the GLSL 4.40 spec:
305 * "It is a link-time error if, within the same stage, the interpolation
306 * qualifiers of variables of the same name do not match.
309 if (input
->data
.interpolation
!= output
->data
.interpolation
&&
310 prog
->Version
< 440) {
312 "%s shader output `%s' specifies %s "
313 "interpolation qualifier, "
314 "but %s shader input specifies %s "
315 "interpolation qualifier\n",
316 _mesa_shader_stage_to_string(producer_stage
),
318 interpolation_string(output
->data
.interpolation
),
319 _mesa_shader_stage_to_string(consumer_stage
),
320 interpolation_string(input
->data
.interpolation
));
326 * Validate front and back color outputs against single color input
329 cross_validate_front_and_back_color(struct gl_shader_program
*prog
,
330 const ir_variable
*input
,
331 const ir_variable
*front_color
,
332 const ir_variable
*back_color
,
333 gl_shader_stage consumer_stage
,
334 gl_shader_stage producer_stage
)
336 if (front_color
!= NULL
&& front_color
->data
.assigned
)
337 cross_validate_types_and_qualifiers(prog
, input
, front_color
,
338 consumer_stage
, producer_stage
);
340 if (back_color
!= NULL
&& back_color
->data
.assigned
)
341 cross_validate_types_and_qualifiers(prog
, input
, back_color
,
342 consumer_stage
, producer_stage
);
346 * Validate that outputs from one stage match inputs of another
349 cross_validate_outputs_to_inputs(struct gl_shader_program
*prog
,
350 gl_shader
*producer
, gl_shader
*consumer
)
352 glsl_symbol_table parameters
;
353 ir_variable
*explicit_locations
[MAX_VARYING
][4] = { {NULL
, NULL
} };
355 /* Find all shader outputs in the "producer" stage.
357 foreach_in_list(ir_instruction
, node
, producer
->ir
) {
358 ir_variable
*const var
= node
->as_variable();
360 if ((var
== NULL
) || (var
->data
.mode
!= ir_var_shader_out
))
363 if (!var
->data
.explicit_location
364 || var
->data
.location
< VARYING_SLOT_VAR0
)
365 parameters
.add_variable(var
);
367 /* User-defined varyings with explicit locations are handled
368 * differently because they do not need to have matching names.
370 const glsl_type
*type
= get_varying_type(var
, producer
->Stage
);
371 unsigned num_elements
= type
->count_attribute_slots(false);
372 unsigned idx
= var
->data
.location
- VARYING_SLOT_VAR0
;
373 unsigned slot_limit
= idx
+ num_elements
;
376 if (var
->type
->without_array()->is_record()) {
377 /* The component qualifier can't be used on structs so just treat
378 * all component slots as used.
382 unsigned dmul
= var
->type
->is_double() ? 2 : 1;
383 last_comp
= var
->data
.location_frac
+
384 var
->type
->without_array()->vector_elements
* dmul
;
387 while (idx
< slot_limit
) {
388 for (unsigned i
= var
->data
.location_frac
; i
< last_comp
; i
++) {
389 if (explicit_locations
[idx
][i
] != NULL
) {
391 "%s shader has multiple outputs explicitly "
392 "assigned to location %d and component %d\n",
393 _mesa_shader_stage_to_string(producer
->Stage
),
394 idx
, var
->data
.location_frac
);
398 /* Make sure all component at this location have the same type.
400 for (unsigned j
= 0; j
< 4; j
++) {
401 if (explicit_locations
[idx
][j
] &&
402 (explicit_locations
[idx
][j
]->type
->without_array()
403 ->base_type
!= var
->type
->without_array()->base_type
)) {
405 "Varyings sharing the same location must "
406 "have the same underlying numerical type. "
407 "Location %u component %u\n", idx
,
408 var
->data
.location_frac
);
413 explicit_locations
[idx
][i
] = var
;
415 /* We need to do some special handling for doubles as dvec3 and
416 * dvec4 consume two consecutive locations. We don't need to
417 * worry about components beginning at anything other than 0 as
418 * the spec does not allow this for dvec3 and dvec4.
420 if (i
== 3 && last_comp
> 4) {
421 last_comp
= last_comp
- 4;
422 /* Bump location index and reset the component index */
433 /* Find all shader inputs in the "consumer" stage. Any variables that have
434 * matching outputs already in the symbol table must have the same type and
437 * Exception: if the consumer is the geometry shader, then the inputs
438 * should be arrays and the type of the array element should match the type
439 * of the corresponding producer output.
441 foreach_in_list(ir_instruction
, node
, consumer
->ir
) {
442 ir_variable
*const input
= node
->as_variable();
444 if ((input
== NULL
) || (input
->data
.mode
!= ir_var_shader_in
))
447 if (strcmp(input
->name
, "gl_Color") == 0 && input
->data
.used
) {
448 const ir_variable
*const front_color
=
449 parameters
.get_variable("gl_FrontColor");
451 const ir_variable
*const back_color
=
452 parameters
.get_variable("gl_BackColor");
454 cross_validate_front_and_back_color(prog
, input
,
455 front_color
, back_color
,
456 consumer
->Stage
, producer
->Stage
);
457 } else if (strcmp(input
->name
, "gl_SecondaryColor") == 0 && input
->data
.used
) {
458 const ir_variable
*const front_color
=
459 parameters
.get_variable("gl_FrontSecondaryColor");
461 const ir_variable
*const back_color
=
462 parameters
.get_variable("gl_BackSecondaryColor");
464 cross_validate_front_and_back_color(prog
, input
,
465 front_color
, back_color
,
466 consumer
->Stage
, producer
->Stage
);
468 /* The rules for connecting inputs and outputs change in the presence
469 * of explicit locations. In this case, we no longer care about the
470 * names of the variables. Instead, we care only about the
471 * explicitly assigned location.
473 ir_variable
*output
= NULL
;
474 if (input
->data
.explicit_location
475 && input
->data
.location
>= VARYING_SLOT_VAR0
) {
477 const glsl_type
*type
= get_varying_type(input
, consumer
->Stage
);
478 unsigned num_elements
= type
->count_attribute_slots(false);
479 unsigned idx
= input
->data
.location
- VARYING_SLOT_VAR0
;
480 unsigned slot_limit
= idx
+ num_elements
;
482 while (idx
< slot_limit
) {
483 output
= explicit_locations
[idx
][input
->data
.location_frac
];
485 if (output
== NULL
||
486 input
->data
.location
!= output
->data
.location
) {
488 "%s shader input `%s' with explicit location "
489 "has no matching output\n",
490 _mesa_shader_stage_to_string(consumer
->Stage
),
497 output
= parameters
.get_variable(input
->name
);
500 if (output
!= NULL
) {
501 /* Interface blocks have their own validation elsewhere so don't
502 * try validating them here.
504 if (!(input
->get_interface_type() &&
505 output
->get_interface_type()))
506 cross_validate_types_and_qualifiers(prog
, input
, output
,
510 /* Check for input vars with unmatched output vars in prev stage
511 * taking into account that interface blocks could have a matching
512 * output but with different name, so we ignore them.
514 assert(!input
->data
.assigned
);
515 if (input
->data
.used
&& !input
->get_interface_type() &&
516 !input
->data
.explicit_location
&& !prog
->SeparateShader
)
518 "%s shader input `%s' "
519 "has no matching output in the previous stage\n",
520 _mesa_shader_stage_to_string(consumer
->Stage
),
528 * Demote shader inputs and outputs that are not used in other stages, and
529 * remove them via dead code elimination.
532 remove_unused_shader_inputs_and_outputs(bool is_separate_shader_object
,
534 enum ir_variable_mode mode
)
536 if (is_separate_shader_object
)
539 foreach_in_list(ir_instruction
, node
, sh
->ir
) {
540 ir_variable
*const var
= node
->as_variable();
542 if ((var
== NULL
) || (var
->data
.mode
!= int(mode
)))
545 /* A shader 'in' or 'out' variable is only really an input or output if
546 * its value is used by other shader stages. This will cause the
547 * variable to have a location assigned.
549 if (var
->data
.is_unmatched_generic_inout
&& !var
->data
.is_xfb_only
) {
550 assert(var
->data
.mode
!= ir_var_temporary
);
551 var
->data
.mode
= ir_var_auto
;
555 /* Eliminate code that is now dead due to unused inputs/outputs being
558 while (do_dead_code(sh
->ir
, false))
564 * Initialize this object based on a string that was passed to
565 * glTransformFeedbackVaryings.
567 * If the input is mal-formed, this call still succeeds, but it sets
568 * this->var_name to a mal-formed input, so tfeedback_decl::find_output_var()
569 * will fail to find any matching variable.
572 tfeedback_decl::init(struct gl_context
*ctx
, const void *mem_ctx
,
575 /* We don't have to be pedantic about what is a valid GLSL variable name,
576 * because any variable with an invalid name can't exist in the IR anyway.
580 this->orig_name
= input
;
581 this->lowered_builtin_array_variable
= none
;
582 this->skip_components
= 0;
583 this->next_buffer_separator
= false;
584 this->matched_candidate
= NULL
;
589 if (ctx
->Extensions
.ARB_transform_feedback3
) {
590 /* Parse gl_NextBuffer. */
591 if (strcmp(input
, "gl_NextBuffer") == 0) {
592 this->next_buffer_separator
= true;
596 /* Parse gl_SkipComponents. */
597 if (strcmp(input
, "gl_SkipComponents1") == 0)
598 this->skip_components
= 1;
599 else if (strcmp(input
, "gl_SkipComponents2") == 0)
600 this->skip_components
= 2;
601 else if (strcmp(input
, "gl_SkipComponents3") == 0)
602 this->skip_components
= 3;
603 else if (strcmp(input
, "gl_SkipComponents4") == 0)
604 this->skip_components
= 4;
606 if (this->skip_components
)
610 /* Parse a declaration. */
611 const char *base_name_end
;
612 long subscript
= parse_program_resource_name(input
, &base_name_end
);
613 this->var_name
= ralloc_strndup(mem_ctx
, input
, base_name_end
- input
);
614 if (this->var_name
== NULL
) {
615 _mesa_error_no_memory(__func__
);
619 if (subscript
>= 0) {
620 this->array_subscript
= subscript
;
621 this->is_subscripted
= true;
623 this->is_subscripted
= false;
626 /* For drivers that lower gl_ClipDistance to gl_ClipDistanceMESA, this
627 * class must behave specially to account for the fact that gl_ClipDistance
628 * is converted from a float[8] to a vec4[2].
630 if (ctx
->Const
.ShaderCompilerOptions
[MESA_SHADER_VERTEX
].LowerCombinedClipCullDistance
&&
631 strcmp(this->var_name
, "gl_ClipDistance") == 0) {
632 this->lowered_builtin_array_variable
= clip_distance
;
635 if (ctx
->Const
.LowerTessLevel
&&
636 (strcmp(this->var_name
, "gl_TessLevelOuter") == 0))
637 this->lowered_builtin_array_variable
= tess_level_outer
;
638 if (ctx
->Const
.LowerTessLevel
&&
639 (strcmp(this->var_name
, "gl_TessLevelInner") == 0))
640 this->lowered_builtin_array_variable
= tess_level_inner
;
645 * Determine whether two tfeedback_decl objects refer to the same variable and
646 * array index (if applicable).
649 tfeedback_decl::is_same(const tfeedback_decl
&x
, const tfeedback_decl
&y
)
651 assert(x
.is_varying() && y
.is_varying());
653 if (strcmp(x
.var_name
, y
.var_name
) != 0)
655 if (x
.is_subscripted
!= y
.is_subscripted
)
657 if (x
.is_subscripted
&& x
.array_subscript
!= y
.array_subscript
)
664 * Assign a location and stream ID for this tfeedback_decl object based on the
665 * transform feedback candidate found by find_candidate.
667 * If an error occurs, the error is reported through linker_error() and false
671 tfeedback_decl::assign_location(struct gl_context
*ctx
,
672 struct gl_shader_program
*prog
)
674 assert(this->is_varying());
676 unsigned fine_location
677 = this->matched_candidate
->toplevel_var
->data
.location
* 4
678 + this->matched_candidate
->toplevel_var
->data
.location_frac
679 + this->matched_candidate
->offset
;
680 const unsigned dmul
=
681 this->matched_candidate
->type
->without_array()->is_double() ? 2 : 1;
683 if (this->matched_candidate
->type
->is_array()) {
685 const unsigned matrix_cols
=
686 this->matched_candidate
->type
->fields
.array
->matrix_columns
;
687 const unsigned vector_elements
=
688 this->matched_candidate
->type
->fields
.array
->vector_elements
;
689 unsigned actual_array_size
;
690 switch (this->lowered_builtin_array_variable
) {
692 actual_array_size
= prog
->LastClipDistanceArraySize
;
694 case tess_level_outer
:
695 actual_array_size
= 4;
697 case tess_level_inner
:
698 actual_array_size
= 2;
702 actual_array_size
= this->matched_candidate
->type
->array_size();
706 if (this->is_subscripted
) {
707 /* Check array bounds. */
708 if (this->array_subscript
>= actual_array_size
) {
709 linker_error(prog
, "Transform feedback varying %s has index "
710 "%i, but the array size is %u.",
711 this->orig_name
, this->array_subscript
,
715 unsigned array_elem_size
= this->lowered_builtin_array_variable
?
716 1 : vector_elements
* matrix_cols
* dmul
;
717 fine_location
+= array_elem_size
* this->array_subscript
;
720 this->size
= actual_array_size
;
722 this->vector_elements
= vector_elements
;
723 this->matrix_columns
= matrix_cols
;
724 if (this->lowered_builtin_array_variable
)
725 this->type
= GL_FLOAT
;
727 this->type
= this->matched_candidate
->type
->fields
.array
->gl_type
;
729 /* Regular variable (scalar, vector, or matrix) */
730 if (this->is_subscripted
) {
731 linker_error(prog
, "Transform feedback varying %s requested, "
732 "but %s is not an array.",
733 this->orig_name
, this->var_name
);
737 this->vector_elements
= this->matched_candidate
->type
->vector_elements
;
738 this->matrix_columns
= this->matched_candidate
->type
->matrix_columns
;
739 this->type
= this->matched_candidate
->type
->gl_type
;
741 this->location
= fine_location
/ 4;
742 this->location_frac
= fine_location
% 4;
744 /* From GL_EXT_transform_feedback:
745 * A program will fail to link if:
747 * * the total number of components to capture in any varying
748 * variable in <varyings> is greater than the constant
749 * MAX_TRANSFORM_FEEDBACK_SEPARATE_COMPONENTS_EXT and the
750 * buffer mode is SEPARATE_ATTRIBS_EXT;
752 if (prog
->TransformFeedback
.BufferMode
== GL_SEPARATE_ATTRIBS
&&
753 this->num_components() >
754 ctx
->Const
.MaxTransformFeedbackSeparateComponents
) {
755 linker_error(prog
, "Transform feedback varying %s exceeds "
756 "MAX_TRANSFORM_FEEDBACK_SEPARATE_COMPONENTS.",
761 /* Only transform feedback varyings can be assigned to non-zero streams,
762 * so assign the stream id here.
764 this->stream_id
= this->matched_candidate
->toplevel_var
->data
.stream
;
766 unsigned array_offset
= this->array_subscript
* 4 * dmul
;
767 unsigned struct_offset
= this->matched_candidate
->offset
* 4 * dmul
;
768 this->buffer
= this->matched_candidate
->toplevel_var
->data
.xfb_buffer
;
769 this->offset
= this->matched_candidate
->toplevel_var
->data
.offset
+
770 array_offset
+ struct_offset
;
777 tfeedback_decl::get_num_outputs() const
779 if (!this->is_varying()) {
782 return (this->num_components() + this->location_frac
+ 3)/4;
787 * Update gl_transform_feedback_info to reflect this tfeedback_decl.
789 * If an error occurs, the error is reported through linker_error() and false
793 tfeedback_decl::store(struct gl_context
*ctx
, struct gl_shader_program
*prog
,
794 struct gl_transform_feedback_info
*info
,
795 unsigned buffer
, unsigned buffer_index
,
796 const unsigned max_outputs
, bool *explicit_stride
,
797 bool has_xfb_qualifiers
) const
799 assert(!this->next_buffer_separator
);
801 /* Handle gl_SkipComponents. */
802 if (this->skip_components
) {
803 info
->Buffers
[buffer
].Stride
+= this->skip_components
;
807 unsigned xfb_offset
= 0;
808 if (has_xfb_qualifiers
) {
809 xfb_offset
= this->offset
/ 4;
811 xfb_offset
= info
->Buffers
[buffer
].Stride
;
813 info
->Varyings
[info
->NumVarying
].Offset
= xfb_offset
* 4;
815 unsigned location
= this->location
;
816 unsigned location_frac
= this->location_frac
;
817 unsigned num_components
= this->num_components();
818 while (num_components
> 0) {
819 unsigned output_size
= MIN2(num_components
, 4 - location_frac
);
820 assert((info
->NumOutputs
== 0 && max_outputs
== 0) ||
821 info
->NumOutputs
< max_outputs
);
823 /* From the ARB_enhanced_layouts spec:
825 * "If such a block member or variable is not written during a shader
826 * invocation, the buffer contents at the assigned offset will be
827 * undefined. Even if there are no static writes to a variable or
828 * member that is assigned a transform feedback offset, the space is
829 * still allocated in the buffer and still affects the stride."
831 if (this->is_varying_written()) {
832 info
->Outputs
[info
->NumOutputs
].ComponentOffset
= location_frac
;
833 info
->Outputs
[info
->NumOutputs
].OutputRegister
= location
;
834 info
->Outputs
[info
->NumOutputs
].NumComponents
= output_size
;
835 info
->Outputs
[info
->NumOutputs
].StreamId
= stream_id
;
836 info
->Outputs
[info
->NumOutputs
].OutputBuffer
= buffer
;
837 info
->Outputs
[info
->NumOutputs
].DstOffset
= xfb_offset
;
840 info
->Buffers
[buffer
].Stream
= this->stream_id
;
841 xfb_offset
+= output_size
;
843 num_components
-= output_size
;
848 if (explicit_stride
&& explicit_stride
[buffer
]) {
849 if (this->is_double() && info
->Buffers
[buffer
].Stride
% 2) {
850 linker_error(prog
, "invalid qualifier xfb_stride=%d must be a "
851 "multiple of 8 as its applied to a type that is or "
852 "contains a double.",
853 info
->Buffers
[buffer
].Stride
* 4);
857 if ((this->offset
/ 4) / info
->Buffers
[buffer
].Stride
!=
858 (xfb_offset
- 1) / info
->Buffers
[buffer
].Stride
) {
859 linker_error(prog
, "xfb_offset (%d) overflows xfb_stride (%d) for "
860 "buffer (%d)", xfb_offset
* 4,
861 info
->Buffers
[buffer
].Stride
* 4, buffer
);
865 info
->Buffers
[buffer
].Stride
= xfb_offset
;
868 /* From GL_EXT_transform_feedback:
869 * A program will fail to link if:
871 * * the total number of components to capture is greater than
872 * the constant MAX_TRANSFORM_FEEDBACK_INTERLEAVED_COMPONENTS_EXT
873 * and the buffer mode is INTERLEAVED_ATTRIBS_EXT.
875 * From GL_ARB_enhanced_layouts:
877 * "The resulting stride (implicit or explicit) must be less than or
878 * equal to the implementation-dependent constant
879 * gl_MaxTransformFeedbackInterleavedComponents."
881 if ((prog
->TransformFeedback
.BufferMode
== GL_INTERLEAVED_ATTRIBS
||
882 has_xfb_qualifiers
) &&
883 info
->Buffers
[buffer
].Stride
>
884 ctx
->Const
.MaxTransformFeedbackInterleavedComponents
) {
885 linker_error(prog
, "The MAX_TRANSFORM_FEEDBACK_INTERLEAVED_COMPONENTS "
886 "limit has been exceeded.");
890 info
->Varyings
[info
->NumVarying
].Name
= ralloc_strdup(prog
,
892 info
->Varyings
[info
->NumVarying
].Type
= this->type
;
893 info
->Varyings
[info
->NumVarying
].Size
= this->size
;
894 info
->Varyings
[info
->NumVarying
].BufferIndex
= buffer_index
;
896 info
->Buffers
[buffer
].NumVaryings
++;
902 const tfeedback_candidate
*
903 tfeedback_decl::find_candidate(gl_shader_program
*prog
,
904 hash_table
*tfeedback_candidates
)
906 const char *name
= this->var_name
;
907 switch (this->lowered_builtin_array_variable
) {
909 name
= this->var_name
;
912 name
= "gl_ClipDistanceMESA";
914 case tess_level_outer
:
915 name
= "gl_TessLevelOuterMESA";
917 case tess_level_inner
:
918 name
= "gl_TessLevelInnerMESA";
921 this->matched_candidate
= (const tfeedback_candidate
*)
922 hash_table_find(tfeedback_candidates
, name
);
923 if (!this->matched_candidate
) {
924 /* From GL_EXT_transform_feedback:
925 * A program will fail to link if:
927 * * any variable name specified in the <varyings> array is not
928 * declared as an output in the geometry shader (if present) or
929 * the vertex shader (if no geometry shader is present);
931 linker_error(prog
, "Transform feedback varying %s undeclared.",
934 return this->matched_candidate
;
939 * Parse all the transform feedback declarations that were passed to
940 * glTransformFeedbackVaryings() and store them in tfeedback_decl objects.
942 * If an error occurs, the error is reported through linker_error() and false
946 parse_tfeedback_decls(struct gl_context
*ctx
, struct gl_shader_program
*prog
,
947 const void *mem_ctx
, unsigned num_names
,
948 char **varying_names
, tfeedback_decl
*decls
)
950 for (unsigned i
= 0; i
< num_names
; ++i
) {
951 decls
[i
].init(ctx
, mem_ctx
, varying_names
[i
]);
953 if (!decls
[i
].is_varying())
956 /* From GL_EXT_transform_feedback:
957 * A program will fail to link if:
959 * * any two entries in the <varyings> array specify the same varying
962 * We interpret this to mean "any two entries in the <varyings> array
963 * specify the same varying variable and array index", since transform
964 * feedback of arrays would be useless otherwise.
966 for (unsigned j
= 0; j
< i
; ++j
) {
967 if (!decls
[j
].is_varying())
970 if (tfeedback_decl::is_same(decls
[i
], decls
[j
])) {
971 linker_error(prog
, "Transform feedback varying %s specified "
972 "more than once.", varying_names
[i
]);
982 cmp_xfb_offset(const void * x_generic
, const void * y_generic
)
984 tfeedback_decl
*x
= (tfeedback_decl
*) x_generic
;
985 tfeedback_decl
*y
= (tfeedback_decl
*) y_generic
;
987 if (x
->get_buffer() != y
->get_buffer())
988 return x
->get_buffer() - y
->get_buffer();
989 return x
->get_offset() - y
->get_offset();
993 * Store transform feedback location assignments into
994 * prog->LinkedTransformFeedback based on the data stored in tfeedback_decls.
996 * If an error occurs, the error is reported through linker_error() and false
1000 store_tfeedback_info(struct gl_context
*ctx
, struct gl_shader_program
*prog
,
1001 unsigned num_tfeedback_decls
,
1002 tfeedback_decl
*tfeedback_decls
, bool has_xfb_qualifiers
)
1004 /* Make sure MaxTransformFeedbackBuffers is less than 32 so the bitmask for
1005 * tracking the number of buffers doesn't overflow.
1007 assert(ctx
->Const
.MaxTransformFeedbackBuffers
< 32);
1009 bool separate_attribs_mode
=
1010 prog
->TransformFeedback
.BufferMode
== GL_SEPARATE_ATTRIBS
;
1012 ralloc_free(prog
->LinkedTransformFeedback
.Varyings
);
1013 ralloc_free(prog
->LinkedTransformFeedback
.Outputs
);
1015 memset(&prog
->LinkedTransformFeedback
, 0,
1016 sizeof(prog
->LinkedTransformFeedback
));
1018 /* The xfb_offset qualifier does not have to be used in increasing order
1019 * however some drivers expect to receive the list of transform feedback
1020 * declarations in order so sort it now for convenience.
1022 if (has_xfb_qualifiers
)
1023 qsort(tfeedback_decls
, num_tfeedback_decls
, sizeof(*tfeedback_decls
),
1026 prog
->LinkedTransformFeedback
.Varyings
=
1028 struct gl_transform_feedback_varying_info
,
1029 num_tfeedback_decls
);
1031 unsigned num_outputs
= 0;
1032 for (unsigned i
= 0; i
< num_tfeedback_decls
; ++i
) {
1033 if (tfeedback_decls
[i
].is_varying_written())
1034 num_outputs
+= tfeedback_decls
[i
].get_num_outputs();
1037 prog
->LinkedTransformFeedback
.Outputs
=
1039 struct gl_transform_feedback_output
,
1042 unsigned num_buffers
= 0;
1043 unsigned buffers
= 0;
1045 if (!has_xfb_qualifiers
&& separate_attribs_mode
) {
1046 /* GL_SEPARATE_ATTRIBS */
1047 for (unsigned i
= 0; i
< num_tfeedback_decls
; ++i
) {
1048 if (!tfeedback_decls
[i
].store(ctx
, prog
, &prog
->LinkedTransformFeedback
,
1049 num_buffers
, num_buffers
, num_outputs
,
1050 NULL
, has_xfb_qualifiers
))
1053 buffers
|= 1 << num_buffers
;
1058 /* GL_INVERLEAVED_ATTRIBS */
1059 int buffer_stream_id
= -1;
1061 num_tfeedback_decls
? tfeedback_decls
[0].get_buffer() : 0;
1062 bool explicit_stride
[MAX_FEEDBACK_BUFFERS
] = { false };
1064 /* Apply any xfb_stride global qualifiers */
1065 if (has_xfb_qualifiers
) {
1066 for (unsigned j
= 0; j
< MAX_FEEDBACK_BUFFERS
; j
++) {
1067 if (prog
->TransformFeedback
.BufferStride
[j
]) {
1069 explicit_stride
[j
] = true;
1070 prog
->LinkedTransformFeedback
.Buffers
[j
].Stride
=
1071 prog
->TransformFeedback
.BufferStride
[j
] / 4;
1076 for (unsigned i
= 0; i
< num_tfeedback_decls
; ++i
) {
1077 if (has_xfb_qualifiers
&&
1078 buffer
!= tfeedback_decls
[i
].get_buffer()) {
1079 /* we have moved to the next buffer so reset stream id */
1080 buffer_stream_id
= -1;
1084 if (tfeedback_decls
[i
].is_next_buffer_separator()) {
1086 buffer_stream_id
= -1;
1088 } else if (buffer_stream_id
== -1) {
1089 /* First varying writing to this buffer: remember its stream */
1090 buffer_stream_id
= (int) tfeedback_decls
[i
].get_stream_id();
1091 } else if (buffer_stream_id
!=
1092 (int) tfeedback_decls
[i
].get_stream_id()) {
1093 /* Varying writes to the same buffer from a different stream */
1095 "Transform feedback can't capture varyings belonging "
1096 "to different vertex streams in a single buffer. "
1097 "Varying %s writes to buffer from stream %u, other "
1098 "varyings in the same buffer write from stream %u.",
1099 tfeedback_decls
[i
].name(),
1100 tfeedback_decls
[i
].get_stream_id(),
1105 if (has_xfb_qualifiers
) {
1106 buffer
= tfeedback_decls
[i
].get_buffer();
1108 buffer
= num_buffers
;
1110 buffers
|= 1 << buffer
;
1112 if (!tfeedback_decls
[i
].store(ctx
, prog
,
1113 &prog
->LinkedTransformFeedback
,
1114 buffer
, num_buffers
, num_outputs
,
1115 explicit_stride
, has_xfb_qualifiers
))
1120 assert(prog
->LinkedTransformFeedback
.NumOutputs
== num_outputs
);
1122 prog
->LinkedTransformFeedback
.ActiveBuffers
= buffers
;
1129 * Data structure recording the relationship between outputs of one shader
1130 * stage (the "producer") and inputs of another (the "consumer").
1132 class varying_matches
1135 varying_matches(bool disable_varying_packing
, bool xfb_enabled
,
1136 gl_shader_stage producer_stage
,
1137 gl_shader_stage consumer_stage
);
1139 void record(ir_variable
*producer_var
, ir_variable
*consumer_var
);
1140 unsigned assign_locations(struct gl_shader_program
*prog
,
1141 uint64_t reserved_slots
, bool separate_shader
);
1142 void store_locations() const;
1145 bool is_varying_packing_safe(const glsl_type
*type
,
1146 const ir_variable
*var
);
1149 * If true, this driver disables varying packing, so all varyings need to
1150 * be aligned on slot boundaries, and take up a number of slots equal to
1151 * their number of matrix columns times their array size.
1153 * Packing may also be disabled because our current packing method is not
1154 * safe in SSO or versions of OpenGL where interpolation qualifiers are not
1155 * guaranteed to match across stages.
1157 const bool disable_varying_packing
;
1160 * If true, this driver has transform feedback enabled. The transform
1161 * feedback code requires at least some packing be done even when varying
1162 * packing is disabled, fortunately where transform feedback requires
1163 * packing it's safe to override the disabled setting. See
1164 * is_varying_packing_safe().
1166 const bool xfb_enabled
;
1169 * Enum representing the order in which varyings are packed within a
1172 * Currently we pack vec4's first, then vec2's, then scalar values, then
1173 * vec3's. This order ensures that the only vectors that are at risk of
1174 * having to be "double parked" (split between two adjacent varying slots)
1177 enum packing_order_enum
{
1180 PACKING_ORDER_SCALAR
,
1184 static unsigned compute_packing_class(const ir_variable
*var
);
1185 static packing_order_enum
compute_packing_order(const ir_variable
*var
);
1186 static int match_comparator(const void *x_generic
, const void *y_generic
);
1187 static int xfb_comparator(const void *x_generic
, const void *y_generic
);
1190 * Structure recording the relationship between a single producer output
1191 * and a single consumer input.
1195 * Packing class for this varying, computed by compute_packing_class().
1197 unsigned packing_class
;
1200 * Packing order for this varying, computed by compute_packing_order().
1202 packing_order_enum packing_order
;
1203 unsigned num_components
;
1206 * The output variable in the producer stage.
1208 ir_variable
*producer_var
;
1211 * The input variable in the consumer stage.
1213 ir_variable
*consumer_var
;
1216 * The location which has been assigned for this varying. This is
1217 * expressed in multiples of a float, with the first generic varying
1218 * (i.e. the one referred to by VARYING_SLOT_VAR0) represented by the
1221 unsigned generic_location
;
1225 * The number of elements in the \c matches array that are currently in
1228 unsigned num_matches
;
1231 * The number of elements that were set aside for the \c matches array when
1234 unsigned matches_capacity
;
1236 gl_shader_stage producer_stage
;
1237 gl_shader_stage consumer_stage
;
1240 } /* anonymous namespace */
1242 varying_matches::varying_matches(bool disable_varying_packing
,
1244 gl_shader_stage producer_stage
,
1245 gl_shader_stage consumer_stage
)
1246 : disable_varying_packing(disable_varying_packing
),
1247 xfb_enabled(xfb_enabled
),
1248 producer_stage(producer_stage
),
1249 consumer_stage(consumer_stage
)
1251 /* Note: this initial capacity is rather arbitrarily chosen to be large
1252 * enough for many cases without wasting an unreasonable amount of space.
1253 * varying_matches::record() will resize the array if there are more than
1254 * this number of varyings.
1256 this->matches_capacity
= 8;
1257 this->matches
= (match
*)
1258 malloc(sizeof(*this->matches
) * this->matches_capacity
);
1259 this->num_matches
= 0;
1263 varying_matches::~varying_matches()
1265 free(this->matches
);
1270 * Packing is always safe on individual arrays, structure and matices. It is
1271 * also safe if the varying is only used for transform feedback.
1274 varying_matches::is_varying_packing_safe(const glsl_type
*type
,
1275 const ir_variable
*var
)
1277 if (consumer_stage
== MESA_SHADER_TESS_EVAL
||
1278 consumer_stage
== MESA_SHADER_TESS_CTRL
||
1279 producer_stage
== MESA_SHADER_TESS_CTRL
)
1282 return xfb_enabled
&& (type
->is_array() || type
->is_record() ||
1283 type
->is_matrix() || var
->data
.is_xfb_only
);
1288 * Record the given producer/consumer variable pair in the list of variables
1289 * that should later be assigned locations.
1291 * It is permissible for \c consumer_var to be NULL (this happens if a
1292 * variable is output by the producer and consumed by transform feedback, but
1293 * not consumed by the consumer).
1295 * If \c producer_var has already been paired up with a consumer_var, or
1296 * producer_var is part of fixed pipeline functionality (and hence already has
1297 * a location assigned), this function has no effect.
1299 * Note: as a side effect this function may change the interpolation type of
1300 * \c producer_var, but only when the change couldn't possibly affect
1304 varying_matches::record(ir_variable
*producer_var
, ir_variable
*consumer_var
)
1306 assert(producer_var
!= NULL
|| consumer_var
!= NULL
);
1308 if ((producer_var
&& (!producer_var
->data
.is_unmatched_generic_inout
||
1309 producer_var
->data
.explicit_location
)) ||
1310 (consumer_var
&& (!consumer_var
->data
.is_unmatched_generic_inout
||
1311 consumer_var
->data
.explicit_location
))) {
1312 /* Either a location already exists for this variable (since it is part
1313 * of fixed functionality), or it has already been recorded as part of a
1319 bool needs_flat_qualifier
= consumer_var
== NULL
&&
1320 (producer_var
->type
->contains_integer() ||
1321 producer_var
->type
->contains_double());
1323 if (needs_flat_qualifier
||
1324 (consumer_stage
!= -1 && consumer_stage
!= MESA_SHADER_FRAGMENT
)) {
1325 /* Since this varying is not being consumed by the fragment shader, its
1326 * interpolation type varying cannot possibly affect rendering.
1327 * Also, this variable is non-flat and is (or contains) an integer
1329 * If the consumer stage is unknown, don't modify the interpolation
1330 * type as it could affect rendering later with separate shaders.
1332 * lower_packed_varyings requires all integer varyings to flat,
1333 * regardless of where they appear. We can trivially satisfy that
1334 * requirement by changing the interpolation type to flat here.
1337 producer_var
->data
.centroid
= false;
1338 producer_var
->data
.sample
= false;
1339 producer_var
->data
.interpolation
= INTERP_QUALIFIER_FLAT
;
1343 consumer_var
->data
.centroid
= false;
1344 consumer_var
->data
.sample
= false;
1345 consumer_var
->data
.interpolation
= INTERP_QUALIFIER_FLAT
;
1349 if (this->num_matches
== this->matches_capacity
) {
1350 this->matches_capacity
*= 2;
1351 this->matches
= (match
*)
1352 realloc(this->matches
,
1353 sizeof(*this->matches
) * this->matches_capacity
);
1356 const ir_variable
*const var
= (producer_var
!= NULL
)
1357 ? producer_var
: consumer_var
;
1358 const gl_shader_stage stage
= (producer_var
!= NULL
)
1359 ? producer_stage
: consumer_stage
;
1360 const glsl_type
*type
= get_varying_type(var
, stage
);
1362 this->matches
[this->num_matches
].packing_class
1363 = this->compute_packing_class(var
);
1364 this->matches
[this->num_matches
].packing_order
1365 = this->compute_packing_order(var
);
1366 if (this->disable_varying_packing
&& !is_varying_packing_safe(type
, var
)) {
1367 unsigned slots
= type
->count_attribute_slots(false);
1368 this->matches
[this->num_matches
].num_components
= slots
* 4;
1370 this->matches
[this->num_matches
].num_components
1371 = type
->component_slots();
1373 this->matches
[this->num_matches
].producer_var
= producer_var
;
1374 this->matches
[this->num_matches
].consumer_var
= consumer_var
;
1375 this->num_matches
++;
1377 producer_var
->data
.is_unmatched_generic_inout
= 0;
1379 consumer_var
->data
.is_unmatched_generic_inout
= 0;
1384 * Choose locations for all of the variable matches that were previously
1385 * passed to varying_matches::record().
1388 varying_matches::assign_locations(struct gl_shader_program
*prog
,
1389 uint64_t reserved_slots
,
1390 bool separate_shader
)
1392 /* If packing has been disabled then we cannot safely sort the varyings by
1393 * class as it may mean we are using a version of OpenGL where
1394 * interpolation qualifiers are not guaranteed to be matching across
1395 * shaders, sorting in this case could result in mismatching shader
1397 * When packing is disabled the sort orders varyings used by transform
1398 * feedback first, but also depends on *undefined behaviour* of qsort to
1399 * reverse the order of the varyings. See: xfb_comparator().
1401 if (!this->disable_varying_packing
) {
1402 /* Sort varying matches into an order that makes them easy to pack. */
1403 qsort(this->matches
, this->num_matches
, sizeof(*this->matches
),
1404 &varying_matches::match_comparator
);
1406 /* Only sort varyings that are only used by transform feedback. */
1407 qsort(this->matches
, this->num_matches
, sizeof(*this->matches
),
1408 &varying_matches::xfb_comparator
);
1411 unsigned generic_location
= 0;
1412 unsigned generic_patch_location
= MAX_VARYING
*4;
1413 bool previous_var_xfb_only
= false;
1415 for (unsigned i
= 0; i
< this->num_matches
; i
++) {
1416 unsigned *location
= &generic_location
;
1418 const ir_variable
*var
;
1419 const glsl_type
*type
;
1420 bool is_vertex_input
= false;
1421 if (matches
[i
].consumer_var
) {
1422 var
= matches
[i
].consumer_var
;
1423 type
= get_varying_type(var
, consumer_stage
);
1424 if (consumer_stage
== MESA_SHADER_VERTEX
)
1425 is_vertex_input
= true;
1427 var
= matches
[i
].producer_var
;
1428 type
= get_varying_type(var
, producer_stage
);
1431 if (var
->data
.patch
)
1432 location
= &generic_patch_location
;
1434 /* Advance to the next slot if this varying has a different packing
1435 * class than the previous one, and we're not already on a slot
1438 * Also advance to the next slot if packing is disabled. This makes sure
1439 * we don't assign varyings the same locations which is possible
1440 * because we still pack individual arrays, records and matrices even
1441 * when packing is disabled. Note we don't advance to the next slot if
1442 * we can pack varyings together that are only used for transform
1445 if ((this->disable_varying_packing
&&
1446 !(previous_var_xfb_only
&& var
->data
.is_xfb_only
)) ||
1447 (i
> 0 && this->matches
[i
- 1].packing_class
1448 != this->matches
[i
].packing_class
)) {
1449 *location
= ALIGN(*location
, 4);
1452 previous_var_xfb_only
= var
->data
.is_xfb_only
;
1454 unsigned num_elements
= type
->count_attribute_slots(is_vertex_input
);
1456 if (this->disable_varying_packing
&&
1457 !is_varying_packing_safe(type
, var
))
1460 slot_end
= type
->without_array()->vector_elements
;
1461 slot_end
+= *location
- 1;
1463 /* FIXME: We could be smarter in the below code and loop back over
1464 * trying to fill any locations that we skipped because we couldn't pack
1465 * the varying between an explicit location. For now just let the user
1466 * hit the linking error if we run out of room and suggest they use
1467 * explicit locations.
1469 for (unsigned j
= 0; j
< num_elements
; j
++) {
1470 while ((slot_end
< MAX_VARYING
* 4u) &&
1471 ((reserved_slots
& (UINT64_C(1) << *location
/ 4u) ||
1472 (reserved_slots
& (UINT64_C(1) << slot_end
/ 4u))))) {
1474 *location
= ALIGN(*location
+ 1, 4);
1475 slot_end
= *location
;
1477 /* reset the counter and try again */
1481 /* Increase the slot to make sure there is enough room for next
1484 if (this->disable_varying_packing
&&
1485 !is_varying_packing_safe(type
, var
))
1488 slot_end
+= type
->without_array()->vector_elements
;
1491 if (!var
->data
.patch
&& *location
>= MAX_VARYING
* 4u) {
1492 linker_error(prog
, "insufficient contiguous locations available for "
1493 "%s it is possible an array or struct could not be "
1494 "packed between varyings with explicit locations. Try "
1495 "using an explicit location for arrays and structs.",
1499 this->matches
[i
].generic_location
= *location
;
1501 *location
+= this->matches
[i
].num_components
;
1504 return (generic_location
+ 3) / 4;
1509 * Update the producer and consumer shaders to reflect the locations
1510 * assignments that were made by varying_matches::assign_locations().
1513 varying_matches::store_locations() const
1515 for (unsigned i
= 0; i
< this->num_matches
; i
++) {
1516 ir_variable
*producer_var
= this->matches
[i
].producer_var
;
1517 ir_variable
*consumer_var
= this->matches
[i
].consumer_var
;
1518 unsigned generic_location
= this->matches
[i
].generic_location
;
1519 unsigned slot
= generic_location
/ 4;
1520 unsigned offset
= generic_location
% 4;
1523 producer_var
->data
.location
= VARYING_SLOT_VAR0
+ slot
;
1524 producer_var
->data
.location_frac
= offset
;
1528 assert(consumer_var
->data
.location
== -1);
1529 consumer_var
->data
.location
= VARYING_SLOT_VAR0
+ slot
;
1530 consumer_var
->data
.location_frac
= offset
;
1537 * Compute the "packing class" of the given varying. This is an unsigned
1538 * integer with the property that two variables in the same packing class can
1539 * be safely backed into the same vec4.
1542 varying_matches::compute_packing_class(const ir_variable
*var
)
1544 /* Without help from the back-end, there is no way to pack together
1545 * variables with different interpolation types, because
1546 * lower_packed_varyings must choose exactly one interpolation type for
1547 * each packed varying it creates.
1549 * However, we can safely pack together floats, ints, and uints, because:
1551 * - varyings of base type "int" and "uint" must use the "flat"
1552 * interpolation type, which can only occur in GLSL 1.30 and above.
1554 * - On platforms that support GLSL 1.30 and above, lower_packed_varyings
1555 * can store flat floats as ints without losing any information (using
1556 * the ir_unop_bitcast_* opcodes).
1558 * Therefore, the packing class depends only on the interpolation type.
1560 unsigned packing_class
= var
->data
.centroid
| (var
->data
.sample
<< 1) |
1561 (var
->data
.patch
<< 2);
1563 packing_class
+= var
->data
.interpolation
;
1564 return packing_class
;
1569 * Compute the "packing order" of the given varying. This is a sort key we
1570 * use to determine when to attempt to pack the given varying relative to
1571 * other varyings in the same packing class.
1573 varying_matches::packing_order_enum
1574 varying_matches::compute_packing_order(const ir_variable
*var
)
1576 const glsl_type
*element_type
= var
->type
;
1578 while (element_type
->base_type
== GLSL_TYPE_ARRAY
) {
1579 element_type
= element_type
->fields
.array
;
1582 switch (element_type
->component_slots() % 4) {
1583 case 1: return PACKING_ORDER_SCALAR
;
1584 case 2: return PACKING_ORDER_VEC2
;
1585 case 3: return PACKING_ORDER_VEC3
;
1586 case 0: return PACKING_ORDER_VEC4
;
1588 assert(!"Unexpected value of vector_elements");
1589 return PACKING_ORDER_VEC4
;
1595 * Comparison function passed to qsort() to sort varyings by packing_class and
1596 * then by packing_order.
1599 varying_matches::match_comparator(const void *x_generic
, const void *y_generic
)
1601 const match
*x
= (const match
*) x_generic
;
1602 const match
*y
= (const match
*) y_generic
;
1604 if (x
->packing_class
!= y
->packing_class
)
1605 return x
->packing_class
- y
->packing_class
;
1606 return x
->packing_order
- y
->packing_order
;
1611 * Comparison function passed to qsort() to sort varyings used only by
1612 * transform feedback when packing of other varyings is disabled.
1615 varying_matches::xfb_comparator(const void *x_generic
, const void *y_generic
)
1617 const match
*x
= (const match
*) x_generic
;
1619 if (x
->producer_var
!= NULL
&& x
->producer_var
->data
.is_xfb_only
)
1620 return match_comparator(x_generic
, y_generic
);
1622 /* FIXME: When the comparator returns 0 it means the elements being
1623 * compared are equivalent. However the qsort documentation says:
1625 * "The order of equivalent elements is undefined."
1627 * In practice the sort ends up reversing the order of the varyings which
1628 * means locations are also assigned in this reversed order and happens to
1629 * be what we want. This is also whats happening in
1630 * varying_matches::match_comparator().
1637 * Is the given variable a varying variable to be counted against the
1638 * limit in ctx->Const.MaxVarying?
1639 * This includes variables such as texcoords, colors and generic
1640 * varyings, but excludes variables such as gl_FrontFacing and gl_FragCoord.
1643 var_counts_against_varying_limit(gl_shader_stage stage
, const ir_variable
*var
)
1645 /* Only fragment shaders will take a varying variable as an input */
1646 if (stage
== MESA_SHADER_FRAGMENT
&&
1647 var
->data
.mode
== ir_var_shader_in
) {
1648 switch (var
->data
.location
) {
1649 case VARYING_SLOT_POS
:
1650 case VARYING_SLOT_FACE
:
1651 case VARYING_SLOT_PNTC
:
1662 * Visitor class that generates tfeedback_candidate structs describing all
1663 * possible targets of transform feedback.
1665 * tfeedback_candidate structs are stored in the hash table
1666 * tfeedback_candidates, which is passed to the constructor. This hash table
1667 * maps varying names to instances of the tfeedback_candidate struct.
1669 class tfeedback_candidate_generator
: public program_resource_visitor
1672 tfeedback_candidate_generator(void *mem_ctx
,
1673 hash_table
*tfeedback_candidates
)
1675 tfeedback_candidates(tfeedback_candidates
),
1681 void process(ir_variable
*var
)
1683 /* All named varying interface blocks should be flattened by now */
1684 assert(!var
->is_interface_instance());
1686 this->toplevel_var
= var
;
1687 this->varying_floats
= 0;
1688 program_resource_visitor::process(var
);
1692 virtual void visit_field(const glsl_type
*type
, const char *name
,
1695 assert(!type
->without_array()->is_record());
1696 assert(!type
->without_array()->is_interface());
1700 tfeedback_candidate
*candidate
1701 = rzalloc(this->mem_ctx
, tfeedback_candidate
);
1702 candidate
->toplevel_var
= this->toplevel_var
;
1703 candidate
->type
= type
;
1704 candidate
->offset
= this->varying_floats
;
1705 hash_table_insert(this->tfeedback_candidates
, candidate
,
1706 ralloc_strdup(this->mem_ctx
, name
));
1707 this->varying_floats
+= type
->component_slots();
1711 * Memory context used to allocate hash table keys and values.
1713 void * const mem_ctx
;
1716 * Hash table in which tfeedback_candidate objects should be stored.
1718 hash_table
* const tfeedback_candidates
;
1721 * Pointer to the toplevel variable that is being traversed.
1723 ir_variable
*toplevel_var
;
1726 * Total number of varying floats that have been visited so far. This is
1727 * used to determine the offset to each varying within the toplevel
1730 unsigned varying_floats
;
1737 populate_consumer_input_sets(void *mem_ctx
, exec_list
*ir
,
1738 hash_table
*consumer_inputs
,
1739 hash_table
*consumer_interface_inputs
,
1740 ir_variable
*consumer_inputs_with_locations
[VARYING_SLOT_TESS_MAX
])
1742 memset(consumer_inputs_with_locations
,
1744 sizeof(consumer_inputs_with_locations
[0]) * VARYING_SLOT_TESS_MAX
);
1746 foreach_in_list(ir_instruction
, node
, ir
) {
1747 ir_variable
*const input_var
= node
->as_variable();
1749 if ((input_var
!= NULL
) && (input_var
->data
.mode
== ir_var_shader_in
)) {
1750 /* All interface blocks should have been lowered by this point */
1751 assert(!input_var
->type
->is_interface());
1753 if (input_var
->data
.explicit_location
) {
1754 /* assign_varying_locations only cares about finding the
1755 * ir_variable at the start of a contiguous location block.
1757 * - For !producer, consumer_inputs_with_locations isn't used.
1759 * - For !consumer, consumer_inputs_with_locations is empty.
1761 * For consumer && producer, if you were trying to set some
1762 * ir_variable to the middle of a location block on the other side
1763 * of producer/consumer, cross_validate_outputs_to_inputs() should
1764 * be link-erroring due to either type mismatch or location
1765 * overlaps. If the variables do match up, then they've got a
1766 * matching data.location and you only looked at
1767 * consumer_inputs_with_locations[var->data.location], not any
1768 * following entries for the array/structure.
1770 consumer_inputs_with_locations
[input_var
->data
.location
] =
1772 } else if (input_var
->get_interface_type() != NULL
) {
1773 char *const iface_field_name
=
1774 ralloc_asprintf(mem_ctx
, "%s.%s",
1775 input_var
->get_interface_type()->without_array()->name
,
1777 hash_table_insert(consumer_interface_inputs
, input_var
,
1780 hash_table_insert(consumer_inputs
, input_var
,
1781 ralloc_strdup(mem_ctx
, input_var
->name
));
1788 * Find a variable from the consumer that "matches" the specified variable
1790 * This function only finds inputs with names that match. There is no
1791 * validation (here) that the types, etc. are compatible.
1794 get_matching_input(void *mem_ctx
,
1795 const ir_variable
*output_var
,
1796 hash_table
*consumer_inputs
,
1797 hash_table
*consumer_interface_inputs
,
1798 ir_variable
*consumer_inputs_with_locations
[VARYING_SLOT_TESS_MAX
])
1800 ir_variable
*input_var
;
1802 if (output_var
->data
.explicit_location
) {
1803 input_var
= consumer_inputs_with_locations
[output_var
->data
.location
];
1804 } else if (output_var
->get_interface_type() != NULL
) {
1805 char *const iface_field_name
=
1806 ralloc_asprintf(mem_ctx
, "%s.%s",
1807 output_var
->get_interface_type()->without_array()->name
,
1810 (ir_variable
*) hash_table_find(consumer_interface_inputs
,
1814 (ir_variable
*) hash_table_find(consumer_inputs
, output_var
->name
);
1817 return (input_var
== NULL
|| input_var
->data
.mode
!= ir_var_shader_in
)
1824 io_variable_cmp(const void *_a
, const void *_b
)
1826 const ir_variable
*const a
= *(const ir_variable
**) _a
;
1827 const ir_variable
*const b
= *(const ir_variable
**) _b
;
1829 if (a
->data
.explicit_location
&& b
->data
.explicit_location
)
1830 return b
->data
.location
- a
->data
.location
;
1832 if (a
->data
.explicit_location
&& !b
->data
.explicit_location
)
1835 if (!a
->data
.explicit_location
&& b
->data
.explicit_location
)
1838 return -strcmp(a
->name
, b
->name
);
1842 * Sort the shader IO variables into canonical order
1845 canonicalize_shader_io(exec_list
*ir
, enum ir_variable_mode io_mode
)
1847 ir_variable
*var_table
[MAX_PROGRAM_OUTPUTS
* 4];
1848 unsigned num_variables
= 0;
1850 foreach_in_list(ir_instruction
, node
, ir
) {
1851 ir_variable
*const var
= node
->as_variable();
1853 if (var
== NULL
|| var
->data
.mode
!= io_mode
)
1856 /* If we have already encountered more I/O variables that could
1857 * successfully link, bail.
1859 if (num_variables
== ARRAY_SIZE(var_table
))
1862 var_table
[num_variables
++] = var
;
1865 if (num_variables
== 0)
1868 /* Sort the list in reverse order (io_variable_cmp handles this). Later
1869 * we're going to push the variables on to the IR list as a stack, so we
1870 * want the last variable (in canonical order) to be first in the list.
1872 qsort(var_table
, num_variables
, sizeof(var_table
[0]), io_variable_cmp
);
1874 /* Remove the variable from it's current location in the IR, and put it at
1877 for (unsigned i
= 0; i
< num_variables
; i
++) {
1878 var_table
[i
]->remove();
1879 ir
->push_head(var_table
[i
]);
1884 * Generate a bitfield map of the explicit locations for shader varyings.
1886 * In theory a 32 bits value will be enough but a 64 bits value is future proof.
1889 reserved_varying_slot(struct gl_shader
*stage
, ir_variable_mode io_mode
)
1891 assert(io_mode
== ir_var_shader_in
|| io_mode
== ir_var_shader_out
);
1892 assert(MAX_VARYING
<= 64); /* avoid an overflow of the returned value */
1900 foreach_in_list(ir_instruction
, node
, stage
->ir
) {
1901 ir_variable
*const var
= node
->as_variable();
1903 if (var
== NULL
|| var
->data
.mode
!= io_mode
||
1904 !var
->data
.explicit_location
||
1905 var
->data
.location
< VARYING_SLOT_VAR0
)
1908 var_slot
= var
->data
.location
- VARYING_SLOT_VAR0
;
1910 unsigned num_elements
= get_varying_type(var
, stage
->Stage
)
1911 ->count_attribute_slots(stage
->Stage
== MESA_SHADER_VERTEX
);
1912 for (unsigned i
= 0; i
< num_elements
; i
++) {
1913 if (var_slot
>= 0 && var_slot
< MAX_VARYING
)
1914 slots
|= UINT64_C(1) << var_slot
;
1924 * Assign locations for all variables that are produced in one pipeline stage
1925 * (the "producer") and consumed in the next stage (the "consumer").
1927 * Variables produced by the producer may also be consumed by transform
1930 * \param num_tfeedback_decls is the number of declarations indicating
1931 * variables that may be consumed by transform feedback.
1933 * \param tfeedback_decls is a pointer to an array of tfeedback_decl objects
1934 * representing the result of parsing the strings passed to
1935 * glTransformFeedbackVaryings(). assign_location() will be called for
1936 * each of these objects that matches one of the outputs of the
1939 * When num_tfeedback_decls is nonzero, it is permissible for the consumer to
1940 * be NULL. In this case, varying locations are assigned solely based on the
1941 * requirements of transform feedback.
1944 assign_varying_locations(struct gl_context
*ctx
,
1946 struct gl_shader_program
*prog
,
1947 gl_shader
*producer
, gl_shader
*consumer
,
1948 unsigned num_tfeedback_decls
,
1949 tfeedback_decl
*tfeedback_decls
)
1951 /* Tessellation shaders treat inputs and outputs as shared memory and can
1952 * access inputs and outputs of other invocations.
1953 * Therefore, they can't be lowered to temps easily (and definitely not
1956 bool unpackable_tess
=
1957 (consumer
&& consumer
->Stage
== MESA_SHADER_TESS_EVAL
) ||
1958 (consumer
&& consumer
->Stage
== MESA_SHADER_TESS_CTRL
) ||
1959 (producer
&& producer
->Stage
== MESA_SHADER_TESS_CTRL
);
1961 /* Transform feedback code assumes varying arrays are packed, so if the
1962 * driver has disabled varying packing, make sure to at least enable
1963 * packing required by transform feedback.
1966 ctx
->Extensions
.EXT_transform_feedback
&& !unpackable_tess
;
1968 /* Disable varying packing for GL 4.4+ as there is no guarantee
1969 * that interpolation qualifiers will match between shaders in these
1970 * versions. We also disable packing on outerward facing interfaces for
1971 * SSO because in ES we need to retain the unpacked varying information
1972 * for draw time validation. For desktop GL we could allow packing for
1973 * versions < 4.4 but its just safer not to do packing.
1975 * Packing is still enabled on individual arrays, structs, and matrices as
1976 * these are required by the transform feedback code and it is still safe
1977 * to do so. We also enable packing when a varying is only used for
1978 * transform feedback and its not a SSO.
1980 * Varying packing currently only packs together varyings with matching
1981 * interpolation qualifiers as the backends assume all packed components
1982 * are to be processed in the same way. Therefore we cannot do packing in
1983 * these versions of GL without the risk of mismatching interfaces.
1985 * From Section 4.5 (Interpolation Qualifiers) of the GLSL 4.30 spec:
1987 * "The type and presence of interpolation qualifiers of variables with
1988 * the same name declared in all linked shaders for the same cross-stage
1989 * interface must match, otherwise the link command will fail.
1991 * When comparing an output from one stage to an input of a subsequent
1992 * stage, the input and output don't match if their interpolation
1993 * qualifiers (or lack thereof) are not the same."
1995 * This text was also in at least revison 7 of the 4.40 spec but is no
1996 * longer in revision 9 and not in the 4.50 spec.
1998 bool disable_varying_packing
=
1999 ctx
->Const
.DisableVaryingPacking
|| unpackable_tess
;
2000 if ((ctx
->API
== API_OPENGL_CORE
&& ctx
->Version
>= 44) ||
2001 (prog
->SeparateShader
&& (producer
== NULL
|| consumer
== NULL
)))
2002 disable_varying_packing
= true;
2004 varying_matches
matches(disable_varying_packing
, xfb_enabled
,
2005 producer
? producer
->Stage
: (gl_shader_stage
)-1,
2006 consumer
? consumer
->Stage
: (gl_shader_stage
)-1);
2007 hash_table
*tfeedback_candidates
2008 = hash_table_ctor(0, hash_table_string_hash
, hash_table_string_compare
);
2009 hash_table
*consumer_inputs
2010 = hash_table_ctor(0, hash_table_string_hash
, hash_table_string_compare
);
2011 hash_table
*consumer_interface_inputs
2012 = hash_table_ctor(0, hash_table_string_hash
, hash_table_string_compare
);
2013 ir_variable
*consumer_inputs_with_locations
[VARYING_SLOT_TESS_MAX
] = {
2017 unsigned consumer_vertices
= 0;
2018 if (consumer
&& consumer
->Stage
== MESA_SHADER_GEOMETRY
)
2019 consumer_vertices
= prog
->Geom
.VerticesIn
;
2021 /* Operate in a total of four passes.
2023 * 1. Sort inputs / outputs into a canonical order. This is necessary so
2024 * that inputs / outputs of separable shaders will be assigned
2025 * predictable locations regardless of the order in which declarations
2026 * appeared in the shader source.
2028 * 2. Assign locations for any matching inputs and outputs.
2030 * 3. Mark output variables in the producer that do not have locations as
2031 * not being outputs. This lets the optimizer eliminate them.
2033 * 4. Mark input variables in the consumer that do not have locations as
2034 * not being inputs. This lets the optimizer eliminate them.
2037 canonicalize_shader_io(consumer
->ir
, ir_var_shader_in
);
2040 canonicalize_shader_io(producer
->ir
, ir_var_shader_out
);
2043 linker::populate_consumer_input_sets(mem_ctx
, consumer
->ir
,
2045 consumer_interface_inputs
,
2046 consumer_inputs_with_locations
);
2049 foreach_in_list(ir_instruction
, node
, producer
->ir
) {
2050 ir_variable
*const output_var
= node
->as_variable();
2052 if ((output_var
== NULL
) ||
2053 (output_var
->data
.mode
!= ir_var_shader_out
))
2056 /* Only geometry shaders can use non-zero streams */
2057 assert(output_var
->data
.stream
== 0 ||
2058 (output_var
->data
.stream
< MAX_VERTEX_STREAMS
&&
2059 producer
->Stage
== MESA_SHADER_GEOMETRY
));
2061 if (num_tfeedback_decls
> 0) {
2062 tfeedback_candidate_generator
g(mem_ctx
, tfeedback_candidates
);
2063 g
.process(output_var
);
2066 ir_variable
*const input_var
=
2067 linker::get_matching_input(mem_ctx
, output_var
, consumer_inputs
,
2068 consumer_interface_inputs
,
2069 consumer_inputs_with_locations
);
2071 /* If a matching input variable was found, add this ouptut (and the
2072 * input) to the set. If this is a separable program and there is no
2073 * consumer stage, add the output.
2075 * Always add TCS outputs. They are shared by all invocations
2076 * within a patch and can be used as shared memory.
2078 if (input_var
|| (prog
->SeparateShader
&& consumer
== NULL
) ||
2079 producer
->Type
== GL_TESS_CONTROL_SHADER
) {
2080 matches
.record(output_var
, input_var
);
2083 /* Only stream 0 outputs can be consumed in the next stage */
2084 if (input_var
&& output_var
->data
.stream
!= 0) {
2085 linker_error(prog
, "output %s is assigned to stream=%d but "
2086 "is linked to an input, which requires stream=0",
2087 output_var
->name
, output_var
->data
.stream
);
2092 /* If there's no producer stage, then this must be a separable program.
2093 * For example, we may have a program that has just a fragment shader.
2094 * Later this program will be used with some arbitrary vertex (or
2095 * geometry) shader program. This means that locations must be assigned
2096 * for all the inputs.
2098 foreach_in_list(ir_instruction
, node
, consumer
->ir
) {
2099 ir_variable
*const input_var
= node
->as_variable();
2101 if ((input_var
== NULL
) ||
2102 (input_var
->data
.mode
!= ir_var_shader_in
))
2105 matches
.record(NULL
, input_var
);
2109 for (unsigned i
= 0; i
< num_tfeedback_decls
; ++i
) {
2110 if (!tfeedback_decls
[i
].is_varying())
2113 const tfeedback_candidate
*matched_candidate
2114 = tfeedback_decls
[i
].find_candidate(prog
, tfeedback_candidates
);
2116 if (matched_candidate
== NULL
) {
2117 hash_table_dtor(tfeedback_candidates
);
2118 hash_table_dtor(consumer_inputs
);
2119 hash_table_dtor(consumer_interface_inputs
);
2123 if (matched_candidate
->toplevel_var
->data
.is_unmatched_generic_inout
) {
2124 matched_candidate
->toplevel_var
->data
.is_xfb_only
= 1;
2125 matches
.record(matched_candidate
->toplevel_var
, NULL
);
2129 const uint64_t reserved_slots
=
2130 reserved_varying_slot(producer
, ir_var_shader_out
) |
2131 reserved_varying_slot(consumer
, ir_var_shader_in
);
2133 const unsigned slots_used
= matches
.assign_locations(prog
, reserved_slots
,
2134 prog
->SeparateShader
);
2135 matches
.store_locations();
2137 for (unsigned i
= 0; i
< num_tfeedback_decls
; ++i
) {
2138 if (!tfeedback_decls
[i
].is_varying())
2141 if (!tfeedback_decls
[i
].assign_location(ctx
, prog
)) {
2142 hash_table_dtor(tfeedback_candidates
);
2143 hash_table_dtor(consumer_inputs
);
2144 hash_table_dtor(consumer_interface_inputs
);
2149 hash_table_dtor(tfeedback_candidates
);
2150 hash_table_dtor(consumer_inputs
);
2151 hash_table_dtor(consumer_interface_inputs
);
2153 if (consumer
&& producer
) {
2154 foreach_in_list(ir_instruction
, node
, consumer
->ir
) {
2155 ir_variable
*const var
= node
->as_variable();
2157 if (var
&& var
->data
.mode
== ir_var_shader_in
&&
2158 var
->data
.is_unmatched_generic_inout
) {
2159 if (!prog
->IsES
&& prog
->Version
<= 120) {
2160 /* On page 25 (page 31 of the PDF) of the GLSL 1.20 spec:
2162 * Only those varying variables used (i.e. read) in
2163 * the fragment shader executable must be written to
2164 * by the vertex shader executable; declaring
2165 * superfluous varying variables in a vertex shader is
2168 * We interpret this text as meaning that the VS must
2169 * write the variable for the FS to read it. See
2170 * "glsl1-varying read but not written" in piglit.
2172 linker_error(prog
, "%s shader varying %s not written "
2174 _mesa_shader_stage_to_string(consumer
->Stage
),
2176 _mesa_shader_stage_to_string(producer
->Stage
));
2178 linker_warning(prog
, "%s shader varying %s not written "
2180 _mesa_shader_stage_to_string(consumer
->Stage
),
2182 _mesa_shader_stage_to_string(producer
->Stage
));
2187 /* Now that validation is done its safe to remove unused varyings. As
2188 * we have both a producer and consumer its safe to remove unused
2189 * varyings even if the program is a SSO because the stages are being
2190 * linked together i.e. we have a multi-stage SSO.
2192 remove_unused_shader_inputs_and_outputs(false, producer
,
2194 remove_unused_shader_inputs_and_outputs(false, consumer
,
2199 lower_packed_varyings(mem_ctx
, slots_used
, ir_var_shader_out
,
2200 0, producer
, disable_varying_packing
,
2205 lower_packed_varyings(mem_ctx
, slots_used
, ir_var_shader_in
,
2206 consumer_vertices
, consumer
,
2207 disable_varying_packing
, xfb_enabled
);
2214 check_against_output_limit(struct gl_context
*ctx
,
2215 struct gl_shader_program
*prog
,
2216 gl_shader
*producer
)
2218 unsigned output_vectors
= 0;
2220 foreach_in_list(ir_instruction
, node
, producer
->ir
) {
2221 ir_variable
*const var
= node
->as_variable();
2223 if (var
&& var
->data
.mode
== ir_var_shader_out
&&
2224 var_counts_against_varying_limit(producer
->Stage
, var
)) {
2225 /* outputs for fragment shader can't be doubles */
2226 output_vectors
+= var
->type
->count_attribute_slots(false);
2230 assert(producer
->Stage
!= MESA_SHADER_FRAGMENT
);
2231 unsigned max_output_components
=
2232 ctx
->Const
.Program
[producer
->Stage
].MaxOutputComponents
;
2234 const unsigned output_components
= output_vectors
* 4;
2235 if (output_components
> max_output_components
) {
2236 if (ctx
->API
== API_OPENGLES2
|| prog
->IsES
)
2237 linker_error(prog
, "%s shader uses too many output vectors "
2239 _mesa_shader_stage_to_string(producer
->Stage
),
2241 max_output_components
/ 4);
2243 linker_error(prog
, "%s shader uses too many output components "
2245 _mesa_shader_stage_to_string(producer
->Stage
),
2247 max_output_components
);
2256 check_against_input_limit(struct gl_context
*ctx
,
2257 struct gl_shader_program
*prog
,
2258 gl_shader
*consumer
)
2260 unsigned input_vectors
= 0;
2262 foreach_in_list(ir_instruction
, node
, consumer
->ir
) {
2263 ir_variable
*const var
= node
->as_variable();
2265 if (var
&& var
->data
.mode
== ir_var_shader_in
&&
2266 var_counts_against_varying_limit(consumer
->Stage
, var
)) {
2267 /* vertex inputs aren't varying counted */
2268 input_vectors
+= var
->type
->count_attribute_slots(false);
2272 assert(consumer
->Stage
!= MESA_SHADER_VERTEX
);
2273 unsigned max_input_components
=
2274 ctx
->Const
.Program
[consumer
->Stage
].MaxInputComponents
;
2276 const unsigned input_components
= input_vectors
* 4;
2277 if (input_components
> max_input_components
) {
2278 if (ctx
->API
== API_OPENGLES2
|| prog
->IsES
)
2279 linker_error(prog
, "%s shader uses too many input vectors "
2281 _mesa_shader_stage_to_string(consumer
->Stage
),
2283 max_input_components
/ 4);
2285 linker_error(prog
, "%s shader uses too many input components "
2287 _mesa_shader_stage_to_string(consumer
->Stage
),
2289 max_input_components
);