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 "util/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_linked_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
->info
.TransformFeedback
.BufferStride
[j
]) {
122 has_xfb_qualifiers
= true;
126 foreach_in_list(ir_instruction
, node
, sh
->ir
) {
127 ir_variable
*var
= node
->as_variable();
128 if (!var
|| var
->data
.mode
!= ir_var_shader_out
)
131 /* From the ARB_enhanced_layouts spec:
133 * "Any shader making any static use (after preprocessing) of any of
134 * these *xfb_* qualifiers will cause the shader to be in a
135 * transform feedback capturing mode and hence responsible for
136 * describing the transform feedback setup. This mode will capture
137 * any output selected by *xfb_offset*, directly or indirectly, to
138 * a transform feedback buffer."
140 if (var
->data
.explicit_xfb_buffer
|| var
->data
.explicit_xfb_stride
) {
141 has_xfb_qualifiers
= true;
144 if (var
->data
.explicit_xfb_offset
) {
145 *num_tfeedback_decls
+= var
->type
->varying_count();
146 has_xfb_qualifiers
= true;
150 if (*num_tfeedback_decls
== 0)
151 return has_xfb_qualifiers
;
154 *varying_names
= ralloc_array(mem_ctx
, char *, *num_tfeedback_decls
);
155 foreach_in_list(ir_instruction
, node
, sh
->ir
) {
156 ir_variable
*var
= node
->as_variable();
157 if (!var
|| var
->data
.mode
!= ir_var_shader_out
)
160 if (var
->data
.explicit_xfb_offset
) {
162 const glsl_type
*type
, *member_type
;
164 if (var
->data
.from_named_ifc_block
) {
165 type
= var
->get_interface_type();
166 /* Find the member type before it was altered by lowering */
168 type
->fields
.structure
[type
->field_index(var
->name
)].type
;
169 name
= ralloc_strdup(NULL
, type
->without_array()->name
);
173 name
= ralloc_strdup(NULL
, var
->name
);
175 create_xfb_varying_names(mem_ctx
, type
, &name
, strlen(name
), &i
,
176 var
->name
, member_type
, varying_names
);
181 assert(i
== *num_tfeedback_decls
);
182 return has_xfb_qualifiers
;
186 anonymous_struct_type_matches(const glsl_type
*output_type
,
187 const glsl_type
*to_match
)
189 while (output_type
->is_array() && to_match
->is_array()) {
190 /* if the lengths at each level don't match fail. */
191 if (output_type
->length
!= to_match
->length
)
193 output_type
= output_type
->fields
.array
;
194 to_match
= to_match
->fields
.array
;
197 if (output_type
->is_array() || to_match
->is_array())
199 return output_type
->is_anonymous() &&
200 to_match
->is_anonymous() &&
201 to_match
->record_compare(output_type
);
205 * Validate the types and qualifiers of an output from one stage against the
206 * matching input to another stage.
209 cross_validate_types_and_qualifiers(struct gl_shader_program
*prog
,
210 const ir_variable
*input
,
211 const ir_variable
*output
,
212 gl_shader_stage consumer_stage
,
213 gl_shader_stage producer_stage
)
215 /* Check that the types match between stages.
217 const glsl_type
*type_to_match
= input
->type
;
219 /* VS -> GS, VS -> TCS, VS -> TES, TES -> GS */
220 const bool extra_array_level
= (producer_stage
== MESA_SHADER_VERTEX
&&
221 consumer_stage
!= MESA_SHADER_FRAGMENT
) ||
222 consumer_stage
== MESA_SHADER_GEOMETRY
;
223 if (extra_array_level
) {
224 assert(type_to_match
->is_array());
225 type_to_match
= type_to_match
->fields
.array
;
228 if (type_to_match
!= output
->type
) {
229 /* There is a bit of a special case for gl_TexCoord. This
230 * built-in is unsized by default. Applications that variable
231 * access it must redeclare it with a size. There is some
232 * language in the GLSL spec that implies the fragment shader
233 * and vertex shader do not have to agree on this size. Other
234 * driver behave this way, and one or two applications seem to
237 * Neither declaration needs to be modified here because the array
238 * sizes are fixed later when update_array_sizes is called.
240 * From page 48 (page 54 of the PDF) of the GLSL 1.10 spec:
242 * "Unlike user-defined varying variables, the built-in
243 * varying variables don't have a strict one-to-one
244 * correspondence between the vertex language and the
245 * fragment language."
247 if (!output
->type
->is_array() || !is_gl_identifier(output
->name
)) {
248 bool anon_matches
= anonymous_struct_type_matches(output
->type
, type_to_match
);
252 "%s shader output `%s' declared as type `%s', "
253 "but %s shader input declared as type `%s'\n",
254 _mesa_shader_stage_to_string(producer_stage
),
257 _mesa_shader_stage_to_string(consumer_stage
),
264 /* Check that all of the qualifiers match between stages.
267 /* According to the OpenGL and OpenGLES GLSL specs, the centroid qualifier
268 * should match until OpenGL 4.3 and OpenGLES 3.1. The OpenGLES 3.0
269 * conformance test suite does not verify that the qualifiers must match.
270 * The deqp test suite expects the opposite (OpenGLES 3.1) behavior for
271 * OpenGLES 3.0 drivers, so we relax the checking in all cases.
273 if (false /* always skip the centroid check */ &&
274 prog
->Version
< (prog
->IsES
? 310 : 430) &&
275 input
->data
.centroid
!= output
->data
.centroid
) {
277 "%s shader output `%s' %s centroid qualifier, "
278 "but %s shader input %s centroid qualifier\n",
279 _mesa_shader_stage_to_string(producer_stage
),
281 (output
->data
.centroid
) ? "has" : "lacks",
282 _mesa_shader_stage_to_string(consumer_stage
),
283 (input
->data
.centroid
) ? "has" : "lacks");
287 if (input
->data
.sample
!= output
->data
.sample
) {
289 "%s shader output `%s' %s sample qualifier, "
290 "but %s shader input %s sample qualifier\n",
291 _mesa_shader_stage_to_string(producer_stage
),
293 (output
->data
.sample
) ? "has" : "lacks",
294 _mesa_shader_stage_to_string(consumer_stage
),
295 (input
->data
.sample
) ? "has" : "lacks");
299 if (input
->data
.patch
!= output
->data
.patch
) {
301 "%s shader output `%s' %s patch qualifier, "
302 "but %s shader input %s patch qualifier\n",
303 _mesa_shader_stage_to_string(producer_stage
),
305 (output
->data
.patch
) ? "has" : "lacks",
306 _mesa_shader_stage_to_string(consumer_stage
),
307 (input
->data
.patch
) ? "has" : "lacks");
311 /* The GLSL 4.30 and GLSL ES 3.00 specifications say:
313 * "As only outputs need be declared with invariant, an output from
314 * one shader stage will still match an input of a subsequent stage
315 * without the input being declared as invariant."
317 * while GLSL 4.20 says:
319 * "For variables leaving one shader and coming into another shader,
320 * the invariant keyword has to be used in both shaders, or a link
321 * error will result."
323 * and GLSL ES 1.00 section 4.6.4 "Invariance and Linking" says:
325 * "The invariance of varyings that are declared in both the vertex
326 * and fragment shaders must match."
328 if (input
->data
.invariant
!= output
->data
.invariant
&&
329 prog
->Version
< (prog
->IsES
? 300 : 430)) {
331 "%s shader output `%s' %s invariant qualifier, "
332 "but %s shader input %s invariant qualifier\n",
333 _mesa_shader_stage_to_string(producer_stage
),
335 (output
->data
.invariant
) ? "has" : "lacks",
336 _mesa_shader_stage_to_string(consumer_stage
),
337 (input
->data
.invariant
) ? "has" : "lacks");
341 /* GLSL >= 4.40 removes text requiring interpolation qualifiers
342 * to match cross stage, they must only match within the same stage.
344 * From page 84 (page 90 of the PDF) of the GLSL 4.40 spec:
346 * "It is a link-time error if, within the same stage, the interpolation
347 * qualifiers of variables of the same name do not match.
350 if (input
->data
.interpolation
!= output
->data
.interpolation
&&
351 prog
->Version
< 440) {
353 "%s shader output `%s' specifies %s "
354 "interpolation qualifier, "
355 "but %s shader input specifies %s "
356 "interpolation qualifier\n",
357 _mesa_shader_stage_to_string(producer_stage
),
359 interpolation_string(output
->data
.interpolation
),
360 _mesa_shader_stage_to_string(consumer_stage
),
361 interpolation_string(input
->data
.interpolation
));
367 * Validate front and back color outputs against single color input
370 cross_validate_front_and_back_color(struct gl_shader_program
*prog
,
371 const ir_variable
*input
,
372 const ir_variable
*front_color
,
373 const ir_variable
*back_color
,
374 gl_shader_stage consumer_stage
,
375 gl_shader_stage producer_stage
)
377 if (front_color
!= NULL
&& front_color
->data
.assigned
)
378 cross_validate_types_and_qualifiers(prog
, input
, front_color
,
379 consumer_stage
, producer_stage
);
381 if (back_color
!= NULL
&& back_color
->data
.assigned
)
382 cross_validate_types_and_qualifiers(prog
, input
, back_color
,
383 consumer_stage
, producer_stage
);
387 * Validate that outputs from one stage match inputs of another
390 cross_validate_outputs_to_inputs(struct gl_shader_program
*prog
,
391 gl_linked_shader
*producer
,
392 gl_linked_shader
*consumer
)
394 glsl_symbol_table parameters
;
395 ir_variable
*explicit_locations
[MAX_VARYINGS_INCL_PATCH
][4] =
398 /* Find all shader outputs in the "producer" stage.
400 foreach_in_list(ir_instruction
, node
, producer
->ir
) {
401 ir_variable
*const var
= node
->as_variable();
403 if (var
== NULL
|| var
->data
.mode
!= ir_var_shader_out
)
406 if (!var
->data
.explicit_location
407 || var
->data
.location
< VARYING_SLOT_VAR0
)
408 parameters
.add_variable(var
);
410 /* User-defined varyings with explicit locations are handled
411 * differently because they do not need to have matching names.
413 const glsl_type
*type
= get_varying_type(var
, producer
->Stage
);
414 unsigned num_elements
= type
->count_attribute_slots(false);
415 unsigned idx
= var
->data
.location
- VARYING_SLOT_VAR0
;
416 unsigned slot_limit
= idx
+ num_elements
;
419 if (type
->without_array()->is_record()) {
420 /* The component qualifier can't be used on structs so just treat
421 * all component slots as used.
425 unsigned dmul
= type
->without_array()->is_64bit() ? 2 : 1;
426 last_comp
= var
->data
.location_frac
+
427 type
->without_array()->vector_elements
* dmul
;
430 while (idx
< slot_limit
) {
431 unsigned i
= var
->data
.location_frac
;
432 while (i
< last_comp
) {
433 if (explicit_locations
[idx
][i
] != NULL
) {
435 "%s shader has multiple outputs explicitly "
436 "assigned to location %d and component %d\n",
437 _mesa_shader_stage_to_string(producer
->Stage
),
438 idx
, var
->data
.location_frac
);
442 /* Make sure all component at this location have the same type.
444 for (unsigned j
= 0; j
< 4; j
++) {
445 if (explicit_locations
[idx
][j
] &&
446 (explicit_locations
[idx
][j
]->type
->without_array()
447 ->base_type
!= type
->without_array()->base_type
)) {
449 "Varyings sharing the same location must "
450 "have the same underlying numerical type. "
451 "Location %u component %u\n", idx
,
452 var
->data
.location_frac
);
457 explicit_locations
[idx
][i
] = var
;
460 /* We need to do some special handling for doubles as dvec3 and
461 * dvec4 consume two consecutive locations. We don't need to
462 * worry about components beginning at anything other than 0 as
463 * the spec does not allow this for dvec3 and dvec4.
465 if (i
== 4 && last_comp
> 4) {
466 last_comp
= last_comp
- 4;
467 /* Bump location index and reset the component index */
478 /* Find all shader inputs in the "consumer" stage. Any variables that have
479 * matching outputs already in the symbol table must have the same type and
482 * Exception: if the consumer is the geometry shader, then the inputs
483 * should be arrays and the type of the array element should match the type
484 * of the corresponding producer output.
486 foreach_in_list(ir_instruction
, node
, consumer
->ir
) {
487 ir_variable
*const input
= node
->as_variable();
489 if (input
== NULL
|| input
->data
.mode
!= ir_var_shader_in
)
492 if (strcmp(input
->name
, "gl_Color") == 0 && input
->data
.used
) {
493 const ir_variable
*const front_color
=
494 parameters
.get_variable("gl_FrontColor");
496 const ir_variable
*const back_color
=
497 parameters
.get_variable("gl_BackColor");
499 cross_validate_front_and_back_color(prog
, input
,
500 front_color
, back_color
,
501 consumer
->Stage
, producer
->Stage
);
502 } else if (strcmp(input
->name
, "gl_SecondaryColor") == 0 && input
->data
.used
) {
503 const ir_variable
*const front_color
=
504 parameters
.get_variable("gl_FrontSecondaryColor");
506 const ir_variable
*const back_color
=
507 parameters
.get_variable("gl_BackSecondaryColor");
509 cross_validate_front_and_back_color(prog
, input
,
510 front_color
, back_color
,
511 consumer
->Stage
, producer
->Stage
);
513 /* The rules for connecting inputs and outputs change in the presence
514 * of explicit locations. In this case, we no longer care about the
515 * names of the variables. Instead, we care only about the
516 * explicitly assigned location.
518 ir_variable
*output
= NULL
;
519 if (input
->data
.explicit_location
520 && input
->data
.location
>= VARYING_SLOT_VAR0
) {
522 const glsl_type
*type
= get_varying_type(input
, consumer
->Stage
);
523 unsigned num_elements
= type
->count_attribute_slots(false);
524 unsigned idx
= input
->data
.location
- VARYING_SLOT_VAR0
;
525 unsigned slot_limit
= idx
+ num_elements
;
527 while (idx
< slot_limit
) {
528 output
= explicit_locations
[idx
][input
->data
.location_frac
];
530 if (output
== NULL
||
531 input
->data
.location
!= output
->data
.location
) {
533 "%s shader input `%s' with explicit location "
534 "has no matching output\n",
535 _mesa_shader_stage_to_string(consumer
->Stage
),
542 output
= parameters
.get_variable(input
->name
);
545 if (output
!= NULL
) {
546 /* Interface blocks have their own validation elsewhere so don't
547 * try validating them here.
549 if (!(input
->get_interface_type() &&
550 output
->get_interface_type()))
551 cross_validate_types_and_qualifiers(prog
, input
, output
,
555 /* Check for input vars with unmatched output vars in prev stage
556 * taking into account that interface blocks could have a matching
557 * output but with different name, so we ignore them.
559 assert(!input
->data
.assigned
);
560 if (input
->data
.used
&& !input
->get_interface_type() &&
561 !input
->data
.explicit_location
&& !prog
->SeparateShader
)
563 "%s shader input `%s' "
564 "has no matching output in the previous stage\n",
565 _mesa_shader_stage_to_string(consumer
->Stage
),
573 * Demote shader inputs and outputs that are not used in other stages, and
574 * remove them via dead code elimination.
577 remove_unused_shader_inputs_and_outputs(bool is_separate_shader_object
,
578 gl_linked_shader
*sh
,
579 enum ir_variable_mode mode
)
581 if (is_separate_shader_object
)
584 foreach_in_list(ir_instruction
, node
, sh
->ir
) {
585 ir_variable
*const var
= node
->as_variable();
587 if (var
== NULL
|| var
->data
.mode
!= int(mode
))
590 /* A shader 'in' or 'out' variable is only really an input or output if
591 * its value is used by other shader stages. This will cause the
592 * variable to have a location assigned.
594 if (var
->data
.is_unmatched_generic_inout
&& !var
->data
.is_xfb_only
) {
595 assert(var
->data
.mode
!= ir_var_temporary
);
597 /* Assign zeros to demoted inputs to allow more optimizations. */
598 if (var
->data
.mode
== ir_var_shader_in
&& !var
->constant_value
)
599 var
->constant_value
= ir_constant::zero(var
, var
->type
);
601 var
->data
.mode
= ir_var_auto
;
605 /* Eliminate code that is now dead due to unused inputs/outputs being
608 while (do_dead_code(sh
->ir
, false))
614 * Initialize this object based on a string that was passed to
615 * glTransformFeedbackVaryings.
617 * If the input is mal-formed, this call still succeeds, but it sets
618 * this->var_name to a mal-formed input, so tfeedback_decl::find_output_var()
619 * will fail to find any matching variable.
622 tfeedback_decl::init(struct gl_context
*ctx
, const void *mem_ctx
,
625 /* We don't have to be pedantic about what is a valid GLSL variable name,
626 * because any variable with an invalid name can't exist in the IR anyway.
630 this->orig_name
= input
;
631 this->lowered_builtin_array_variable
= none
;
632 this->skip_components
= 0;
633 this->next_buffer_separator
= false;
634 this->matched_candidate
= NULL
;
639 if (ctx
->Extensions
.ARB_transform_feedback3
) {
640 /* Parse gl_NextBuffer. */
641 if (strcmp(input
, "gl_NextBuffer") == 0) {
642 this->next_buffer_separator
= true;
646 /* Parse gl_SkipComponents. */
647 if (strcmp(input
, "gl_SkipComponents1") == 0)
648 this->skip_components
= 1;
649 else if (strcmp(input
, "gl_SkipComponents2") == 0)
650 this->skip_components
= 2;
651 else if (strcmp(input
, "gl_SkipComponents3") == 0)
652 this->skip_components
= 3;
653 else if (strcmp(input
, "gl_SkipComponents4") == 0)
654 this->skip_components
= 4;
656 if (this->skip_components
)
660 /* Parse a declaration. */
661 const char *base_name_end
;
662 long subscript
= parse_program_resource_name(input
, &base_name_end
);
663 this->var_name
= ralloc_strndup(mem_ctx
, input
, base_name_end
- input
);
664 if (this->var_name
== NULL
) {
665 _mesa_error_no_memory(__func__
);
669 if (subscript
>= 0) {
670 this->array_subscript
= subscript
;
671 this->is_subscripted
= true;
673 this->is_subscripted
= false;
676 /* For drivers that lower gl_ClipDistance to gl_ClipDistanceMESA, this
677 * class must behave specially to account for the fact that gl_ClipDistance
678 * is converted from a float[8] to a vec4[2].
680 if (ctx
->Const
.ShaderCompilerOptions
[MESA_SHADER_VERTEX
].LowerCombinedClipCullDistance
&&
681 strcmp(this->var_name
, "gl_ClipDistance") == 0) {
682 this->lowered_builtin_array_variable
= clip_distance
;
684 if (ctx
->Const
.ShaderCompilerOptions
[MESA_SHADER_VERTEX
].LowerCombinedClipCullDistance
&&
685 strcmp(this->var_name
, "gl_CullDistance") == 0) {
686 this->lowered_builtin_array_variable
= cull_distance
;
689 if (ctx
->Const
.LowerTessLevel
&&
690 (strcmp(this->var_name
, "gl_TessLevelOuter") == 0))
691 this->lowered_builtin_array_variable
= tess_level_outer
;
692 if (ctx
->Const
.LowerTessLevel
&&
693 (strcmp(this->var_name
, "gl_TessLevelInner") == 0))
694 this->lowered_builtin_array_variable
= tess_level_inner
;
699 * Determine whether two tfeedback_decl objects refer to the same variable and
700 * array index (if applicable).
703 tfeedback_decl::is_same(const tfeedback_decl
&x
, const tfeedback_decl
&y
)
705 assert(x
.is_varying() && y
.is_varying());
707 if (strcmp(x
.var_name
, y
.var_name
) != 0)
709 if (x
.is_subscripted
!= y
.is_subscripted
)
711 if (x
.is_subscripted
&& x
.array_subscript
!= y
.array_subscript
)
718 * Assign a location and stream ID for this tfeedback_decl object based on the
719 * transform feedback candidate found by find_candidate.
721 * If an error occurs, the error is reported through linker_error() and false
725 tfeedback_decl::assign_location(struct gl_context
*ctx
,
726 struct gl_shader_program
*prog
)
728 assert(this->is_varying());
730 unsigned fine_location
731 = this->matched_candidate
->toplevel_var
->data
.location
* 4
732 + this->matched_candidate
->toplevel_var
->data
.location_frac
733 + this->matched_candidate
->offset
;
734 const unsigned dmul
=
735 this->matched_candidate
->type
->without_array()->is_64bit() ? 2 : 1;
737 if (this->matched_candidate
->type
->is_array()) {
739 const unsigned matrix_cols
=
740 this->matched_candidate
->type
->fields
.array
->matrix_columns
;
741 const unsigned vector_elements
=
742 this->matched_candidate
->type
->fields
.array
->vector_elements
;
743 unsigned actual_array_size
;
744 switch (this->lowered_builtin_array_variable
) {
746 actual_array_size
= prog
->LastClipDistanceArraySize
;
749 actual_array_size
= prog
->LastCullDistanceArraySize
;
751 case tess_level_outer
:
752 actual_array_size
= 4;
754 case tess_level_inner
:
755 actual_array_size
= 2;
759 actual_array_size
= this->matched_candidate
->type
->array_size();
763 if (this->is_subscripted
) {
764 /* Check array bounds. */
765 if (this->array_subscript
>= actual_array_size
) {
766 linker_error(prog
, "Transform feedback varying %s has index "
767 "%i, but the array size is %u.",
768 this->orig_name
, this->array_subscript
,
772 unsigned array_elem_size
= this->lowered_builtin_array_variable
?
773 1 : vector_elements
* matrix_cols
* dmul
;
774 fine_location
+= array_elem_size
* this->array_subscript
;
777 this->size
= actual_array_size
;
779 this->vector_elements
= vector_elements
;
780 this->matrix_columns
= matrix_cols
;
781 if (this->lowered_builtin_array_variable
)
782 this->type
= GL_FLOAT
;
784 this->type
= this->matched_candidate
->type
->fields
.array
->gl_type
;
786 /* Regular variable (scalar, vector, or matrix) */
787 if (this->is_subscripted
) {
788 linker_error(prog
, "Transform feedback varying %s requested, "
789 "but %s is not an array.",
790 this->orig_name
, this->var_name
);
794 this->vector_elements
= this->matched_candidate
->type
->vector_elements
;
795 this->matrix_columns
= this->matched_candidate
->type
->matrix_columns
;
796 this->type
= this->matched_candidate
->type
->gl_type
;
798 this->location
= fine_location
/ 4;
799 this->location_frac
= fine_location
% 4;
801 /* From GL_EXT_transform_feedback:
802 * A program will fail to link if:
804 * * the total number of components to capture in any varying
805 * variable in <varyings> is greater than the constant
806 * MAX_TRANSFORM_FEEDBACK_SEPARATE_COMPONENTS_EXT and the
807 * buffer mode is SEPARATE_ATTRIBS_EXT;
809 if (prog
->TransformFeedback
.BufferMode
== GL_SEPARATE_ATTRIBS
&&
810 this->num_components() >
811 ctx
->Const
.MaxTransformFeedbackSeparateComponents
) {
812 linker_error(prog
, "Transform feedback varying %s exceeds "
813 "MAX_TRANSFORM_FEEDBACK_SEPARATE_COMPONENTS.",
818 /* Only transform feedback varyings can be assigned to non-zero streams,
819 * so assign the stream id here.
821 this->stream_id
= this->matched_candidate
->toplevel_var
->data
.stream
;
823 unsigned array_offset
= this->array_subscript
* 4 * dmul
;
824 unsigned struct_offset
= this->matched_candidate
->offset
* 4 * dmul
;
825 this->buffer
= this->matched_candidate
->toplevel_var
->data
.xfb_buffer
;
826 this->offset
= this->matched_candidate
->toplevel_var
->data
.offset
+
827 array_offset
+ struct_offset
;
834 tfeedback_decl::get_num_outputs() const
836 if (!this->is_varying()) {
839 return (this->num_components() + this->location_frac
+ 3)/4;
844 * Update gl_transform_feedback_info to reflect this tfeedback_decl.
846 * If an error occurs, the error is reported through linker_error() and false
850 tfeedback_decl::store(struct gl_context
*ctx
, struct gl_shader_program
*prog
,
851 struct gl_transform_feedback_info
*info
,
852 unsigned buffer
, unsigned buffer_index
,
853 const unsigned max_outputs
, bool *explicit_stride
,
854 bool has_xfb_qualifiers
) const
856 unsigned xfb_offset
= 0;
857 unsigned size
= this->size
;
858 /* Handle gl_SkipComponents. */
859 if (this->skip_components
) {
860 info
->Buffers
[buffer
].Stride
+= this->skip_components
;
861 size
= this->skip_components
;
865 if (this->next_buffer_separator
) {
870 if (has_xfb_qualifiers
) {
871 xfb_offset
= this->offset
/ 4;
873 xfb_offset
= info
->Buffers
[buffer
].Stride
;
875 info
->Varyings
[info
->NumVarying
].Offset
= xfb_offset
* 4;
878 unsigned location
= this->location
;
879 unsigned location_frac
= this->location_frac
;
880 unsigned num_components
= this->num_components();
881 while (num_components
> 0) {
882 unsigned output_size
= MIN2(num_components
, 4 - location_frac
);
883 assert((info
->NumOutputs
== 0 && max_outputs
== 0) ||
884 info
->NumOutputs
< max_outputs
);
886 /* From the ARB_enhanced_layouts spec:
888 * "If such a block member or variable is not written during a shader
889 * invocation, the buffer contents at the assigned offset will be
890 * undefined. Even if there are no static writes to a variable or
891 * member that is assigned a transform feedback offset, the space is
892 * still allocated in the buffer and still affects the stride."
894 if (this->is_varying_written()) {
895 info
->Outputs
[info
->NumOutputs
].ComponentOffset
= location_frac
;
896 info
->Outputs
[info
->NumOutputs
].OutputRegister
= location
;
897 info
->Outputs
[info
->NumOutputs
].NumComponents
= output_size
;
898 info
->Outputs
[info
->NumOutputs
].StreamId
= stream_id
;
899 info
->Outputs
[info
->NumOutputs
].OutputBuffer
= buffer
;
900 info
->Outputs
[info
->NumOutputs
].DstOffset
= xfb_offset
;
903 info
->Buffers
[buffer
].Stream
= this->stream_id
;
904 xfb_offset
+= output_size
;
906 num_components
-= output_size
;
912 if (explicit_stride
&& explicit_stride
[buffer
]) {
913 if (this->is_64bit() && info
->Buffers
[buffer
].Stride
% 2) {
914 linker_error(prog
, "invalid qualifier xfb_stride=%d must be a "
915 "multiple of 8 as its applied to a type that is or "
916 "contains a double.",
917 info
->Buffers
[buffer
].Stride
* 4);
921 if ((this->offset
/ 4) / info
->Buffers
[buffer
].Stride
!=
922 (xfb_offset
- 1) / info
->Buffers
[buffer
].Stride
) {
923 linker_error(prog
, "xfb_offset (%d) overflows xfb_stride (%d) for "
924 "buffer (%d)", xfb_offset
* 4,
925 info
->Buffers
[buffer
].Stride
* 4, buffer
);
929 info
->Buffers
[buffer
].Stride
= xfb_offset
;
932 /* From GL_EXT_transform_feedback:
933 * A program will fail to link if:
935 * * the total number of components to capture is greater than
936 * the constant MAX_TRANSFORM_FEEDBACK_INTERLEAVED_COMPONENTS_EXT
937 * and the buffer mode is INTERLEAVED_ATTRIBS_EXT.
939 * From GL_ARB_enhanced_layouts:
941 * "The resulting stride (implicit or explicit) must be less than or
942 * equal to the implementation-dependent constant
943 * gl_MaxTransformFeedbackInterleavedComponents."
945 if ((prog
->TransformFeedback
.BufferMode
== GL_INTERLEAVED_ATTRIBS
||
946 has_xfb_qualifiers
) &&
947 info
->Buffers
[buffer
].Stride
>
948 ctx
->Const
.MaxTransformFeedbackInterleavedComponents
) {
949 linker_error(prog
, "The MAX_TRANSFORM_FEEDBACK_INTERLEAVED_COMPONENTS "
950 "limit has been exceeded.");
955 info
->Varyings
[info
->NumVarying
].Name
= ralloc_strdup(prog
,
957 info
->Varyings
[info
->NumVarying
].Type
= this->type
;
958 info
->Varyings
[info
->NumVarying
].Size
= size
;
959 info
->Varyings
[info
->NumVarying
].BufferIndex
= buffer_index
;
961 info
->Buffers
[buffer
].NumVaryings
++;
967 const tfeedback_candidate
*
968 tfeedback_decl::find_candidate(gl_shader_program
*prog
,
969 hash_table
*tfeedback_candidates
)
971 const char *name
= this->var_name
;
972 switch (this->lowered_builtin_array_variable
) {
974 name
= this->var_name
;
977 name
= "gl_ClipDistanceMESA";
980 name
= "gl_CullDistanceMESA";
982 case tess_level_outer
:
983 name
= "gl_TessLevelOuterMESA";
985 case tess_level_inner
:
986 name
= "gl_TessLevelInnerMESA";
989 hash_entry
*entry
= _mesa_hash_table_search(tfeedback_candidates
, name
);
991 this->matched_candidate
= entry
?
992 (const tfeedback_candidate
*) entry
->data
: NULL
;
994 if (!this->matched_candidate
) {
995 /* From GL_EXT_transform_feedback:
996 * A program will fail to link if:
998 * * any variable name specified in the <varyings> array is not
999 * declared as an output in the geometry shader (if present) or
1000 * the vertex shader (if no geometry shader is present);
1002 linker_error(prog
, "Transform feedback varying %s undeclared.",
1006 return this->matched_candidate
;
1011 * Parse all the transform feedback declarations that were passed to
1012 * glTransformFeedbackVaryings() and store them in tfeedback_decl objects.
1014 * If an error occurs, the error is reported through linker_error() and false
1018 parse_tfeedback_decls(struct gl_context
*ctx
, struct gl_shader_program
*prog
,
1019 const void *mem_ctx
, unsigned num_names
,
1020 char **varying_names
, tfeedback_decl
*decls
)
1022 for (unsigned i
= 0; i
< num_names
; ++i
) {
1023 decls
[i
].init(ctx
, mem_ctx
, varying_names
[i
]);
1025 if (!decls
[i
].is_varying())
1028 /* From GL_EXT_transform_feedback:
1029 * A program will fail to link if:
1031 * * any two entries in the <varyings> array specify the same varying
1034 * We interpret this to mean "any two entries in the <varyings> array
1035 * specify the same varying variable and array index", since transform
1036 * feedback of arrays would be useless otherwise.
1038 for (unsigned j
= 0; j
< i
; ++j
) {
1039 if (!decls
[j
].is_varying())
1042 if (tfeedback_decl::is_same(decls
[i
], decls
[j
])) {
1043 linker_error(prog
, "Transform feedback varying %s specified "
1044 "more than once.", varying_names
[i
]);
1054 cmp_xfb_offset(const void * x_generic
, const void * y_generic
)
1056 tfeedback_decl
*x
= (tfeedback_decl
*) x_generic
;
1057 tfeedback_decl
*y
= (tfeedback_decl
*) y_generic
;
1059 if (x
->get_buffer() != y
->get_buffer())
1060 return x
->get_buffer() - y
->get_buffer();
1061 return x
->get_offset() - y
->get_offset();
1065 * Store transform feedback location assignments into
1066 * prog->LinkedTransformFeedback based on the data stored in tfeedback_decls.
1068 * If an error occurs, the error is reported through linker_error() and false
1072 store_tfeedback_info(struct gl_context
*ctx
, struct gl_shader_program
*prog
,
1073 unsigned num_tfeedback_decls
,
1074 tfeedback_decl
*tfeedback_decls
, bool has_xfb_qualifiers
)
1076 /* Make sure MaxTransformFeedbackBuffers is less than 32 so the bitmask for
1077 * tracking the number of buffers doesn't overflow.
1079 assert(ctx
->Const
.MaxTransformFeedbackBuffers
< 32);
1081 bool separate_attribs_mode
=
1082 prog
->TransformFeedback
.BufferMode
== GL_SEPARATE_ATTRIBS
;
1084 ralloc_free(prog
->LinkedTransformFeedback
.Varyings
);
1085 ralloc_free(prog
->LinkedTransformFeedback
.Outputs
);
1087 memset(&prog
->LinkedTransformFeedback
, 0,
1088 sizeof(prog
->LinkedTransformFeedback
));
1090 /* The xfb_offset qualifier does not have to be used in increasing order
1091 * however some drivers expect to receive the list of transform feedback
1092 * declarations in order so sort it now for convenience.
1094 if (has_xfb_qualifiers
)
1095 qsort(tfeedback_decls
, num_tfeedback_decls
, sizeof(*tfeedback_decls
),
1098 prog
->LinkedTransformFeedback
.Varyings
=
1100 struct gl_transform_feedback_varying_info
,
1101 num_tfeedback_decls
);
1103 unsigned num_outputs
= 0;
1104 for (unsigned i
= 0; i
< num_tfeedback_decls
; ++i
) {
1105 if (tfeedback_decls
[i
].is_varying_written())
1106 num_outputs
+= tfeedback_decls
[i
].get_num_outputs();
1109 prog
->LinkedTransformFeedback
.Outputs
=
1111 struct gl_transform_feedback_output
,
1114 unsigned num_buffers
= 0;
1115 unsigned buffers
= 0;
1117 if (!has_xfb_qualifiers
&& separate_attribs_mode
) {
1118 /* GL_SEPARATE_ATTRIBS */
1119 for (unsigned i
= 0; i
< num_tfeedback_decls
; ++i
) {
1120 if (!tfeedback_decls
[i
].store(ctx
, prog
, &prog
->LinkedTransformFeedback
,
1121 num_buffers
, num_buffers
, num_outputs
,
1122 NULL
, has_xfb_qualifiers
))
1125 buffers
|= 1 << num_buffers
;
1130 /* GL_INVERLEAVED_ATTRIBS */
1131 int buffer_stream_id
= -1;
1133 num_tfeedback_decls
? tfeedback_decls
[0].get_buffer() : 0;
1134 bool explicit_stride
[MAX_FEEDBACK_BUFFERS
] = { false };
1136 /* Apply any xfb_stride global qualifiers */
1137 if (has_xfb_qualifiers
) {
1138 for (unsigned j
= 0; j
< MAX_FEEDBACK_BUFFERS
; j
++) {
1139 if (prog
->LinkedTransformFeedback
.BufferStride
[j
]) {
1141 explicit_stride
[j
] = true;
1142 prog
->LinkedTransformFeedback
.Buffers
[j
].Stride
=
1143 prog
->LinkedTransformFeedback
.BufferStride
[j
] / 4;
1148 for (unsigned i
= 0; i
< num_tfeedback_decls
; ++i
) {
1149 if (has_xfb_qualifiers
&&
1150 buffer
!= tfeedback_decls
[i
].get_buffer()) {
1151 /* we have moved to the next buffer so reset stream id */
1152 buffer_stream_id
= -1;
1156 if (tfeedback_decls
[i
].is_next_buffer_separator()) {
1157 if (!tfeedback_decls
[i
].store(ctx
, prog
,
1158 &prog
->LinkedTransformFeedback
,
1159 buffer
, num_buffers
, num_outputs
,
1160 explicit_stride
, has_xfb_qualifiers
))
1163 buffer_stream_id
= -1;
1165 } else if (tfeedback_decls
[i
].is_varying()) {
1166 if (buffer_stream_id
== -1) {
1167 /* First varying writing to this buffer: remember its stream */
1168 buffer_stream_id
= (int) tfeedback_decls
[i
].get_stream_id();
1169 } else if (buffer_stream_id
!=
1170 (int) tfeedback_decls
[i
].get_stream_id()) {
1171 /* Varying writes to the same buffer from a different stream */
1173 "Transform feedback can't capture varyings belonging "
1174 "to different vertex streams in a single buffer. "
1175 "Varying %s writes to buffer from stream %u, other "
1176 "varyings in the same buffer write from stream %u.",
1177 tfeedback_decls
[i
].name(),
1178 tfeedback_decls
[i
].get_stream_id(),
1184 if (has_xfb_qualifiers
) {
1185 buffer
= tfeedback_decls
[i
].get_buffer();
1187 buffer
= num_buffers
;
1189 buffers
|= 1 << buffer
;
1191 if (!tfeedback_decls
[i
].store(ctx
, prog
,
1192 &prog
->LinkedTransformFeedback
,
1193 buffer
, num_buffers
, num_outputs
,
1194 explicit_stride
, has_xfb_qualifiers
))
1199 assert(prog
->LinkedTransformFeedback
.NumOutputs
== num_outputs
);
1201 prog
->LinkedTransformFeedback
.ActiveBuffers
= buffers
;
1208 * Data structure recording the relationship between outputs of one shader
1209 * stage (the "producer") and inputs of another (the "consumer").
1211 class varying_matches
1214 varying_matches(bool disable_varying_packing
, bool xfb_enabled
,
1215 gl_shader_stage producer_stage
,
1216 gl_shader_stage consumer_stage
);
1218 void record(ir_variable
*producer_var
, ir_variable
*consumer_var
);
1219 unsigned assign_locations(struct gl_shader_program
*prog
,
1220 uint64_t reserved_slots
);
1221 void store_locations() const;
1224 bool is_varying_packing_safe(const glsl_type
*type
,
1225 const ir_variable
*var
);
1228 * If true, this driver disables varying packing, so all varyings need to
1229 * be aligned on slot boundaries, and take up a number of slots equal to
1230 * their number of matrix columns times their array size.
1232 * Packing may also be disabled because our current packing method is not
1233 * safe in SSO or versions of OpenGL where interpolation qualifiers are not
1234 * guaranteed to match across stages.
1236 const bool disable_varying_packing
;
1239 * If true, this driver has transform feedback enabled. The transform
1240 * feedback code requires at least some packing be done even when varying
1241 * packing is disabled, fortunately where transform feedback requires
1242 * packing it's safe to override the disabled setting. See
1243 * is_varying_packing_safe().
1245 const bool xfb_enabled
;
1248 * Enum representing the order in which varyings are packed within a
1251 * Currently we pack vec4's first, then vec2's, then scalar values, then
1252 * vec3's. This order ensures that the only vectors that are at risk of
1253 * having to be "double parked" (split between two adjacent varying slots)
1256 enum packing_order_enum
{
1259 PACKING_ORDER_SCALAR
,
1263 static unsigned compute_packing_class(const ir_variable
*var
);
1264 static packing_order_enum
compute_packing_order(const ir_variable
*var
);
1265 static int match_comparator(const void *x_generic
, const void *y_generic
);
1266 static int xfb_comparator(const void *x_generic
, const void *y_generic
);
1269 * Structure recording the relationship between a single producer output
1270 * and a single consumer input.
1274 * Packing class for this varying, computed by compute_packing_class().
1276 unsigned packing_class
;
1279 * Packing order for this varying, computed by compute_packing_order().
1281 packing_order_enum packing_order
;
1282 unsigned num_components
;
1285 * The output variable in the producer stage.
1287 ir_variable
*producer_var
;
1290 * The input variable in the consumer stage.
1292 ir_variable
*consumer_var
;
1295 * The location which has been assigned for this varying. This is
1296 * expressed in multiples of a float, with the first generic varying
1297 * (i.e. the one referred to by VARYING_SLOT_VAR0) represented by the
1300 unsigned generic_location
;
1304 * The number of elements in the \c matches array that are currently in
1307 unsigned num_matches
;
1310 * The number of elements that were set aside for the \c matches array when
1313 unsigned matches_capacity
;
1315 gl_shader_stage producer_stage
;
1316 gl_shader_stage consumer_stage
;
1319 } /* anonymous namespace */
1321 varying_matches::varying_matches(bool disable_varying_packing
,
1323 gl_shader_stage producer_stage
,
1324 gl_shader_stage consumer_stage
)
1325 : disable_varying_packing(disable_varying_packing
),
1326 xfb_enabled(xfb_enabled
),
1327 producer_stage(producer_stage
),
1328 consumer_stage(consumer_stage
)
1330 /* Note: this initial capacity is rather arbitrarily chosen to be large
1331 * enough for many cases without wasting an unreasonable amount of space.
1332 * varying_matches::record() will resize the array if there are more than
1333 * this number of varyings.
1335 this->matches_capacity
= 8;
1336 this->matches
= (match
*)
1337 malloc(sizeof(*this->matches
) * this->matches_capacity
);
1338 this->num_matches
= 0;
1342 varying_matches::~varying_matches()
1344 free(this->matches
);
1349 * Packing is always safe on individual arrays, structures, and matrices. It
1350 * is also safe if the varying is only used for transform feedback.
1353 varying_matches::is_varying_packing_safe(const glsl_type
*type
,
1354 const ir_variable
*var
)
1356 if (consumer_stage
== MESA_SHADER_TESS_EVAL
||
1357 consumer_stage
== MESA_SHADER_TESS_CTRL
||
1358 producer_stage
== MESA_SHADER_TESS_CTRL
)
1361 return xfb_enabled
&& (type
->is_array() || type
->is_record() ||
1362 type
->is_matrix() || var
->data
.is_xfb_only
);
1367 * Record the given producer/consumer variable pair in the list of variables
1368 * that should later be assigned locations.
1370 * It is permissible for \c consumer_var to be NULL (this happens if a
1371 * variable is output by the producer and consumed by transform feedback, but
1372 * not consumed by the consumer).
1374 * If \c producer_var has already been paired up with a consumer_var, or
1375 * producer_var is part of fixed pipeline functionality (and hence already has
1376 * a location assigned), this function has no effect.
1378 * Note: as a side effect this function may change the interpolation type of
1379 * \c producer_var, but only when the change couldn't possibly affect
1383 varying_matches::record(ir_variable
*producer_var
, ir_variable
*consumer_var
)
1385 assert(producer_var
!= NULL
|| consumer_var
!= NULL
);
1387 if ((producer_var
&& (!producer_var
->data
.is_unmatched_generic_inout
||
1388 producer_var
->data
.explicit_location
)) ||
1389 (consumer_var
&& (!consumer_var
->data
.is_unmatched_generic_inout
||
1390 consumer_var
->data
.explicit_location
))) {
1391 /* Either a location already exists for this variable (since it is part
1392 * of fixed functionality), or it has already been recorded as part of a
1398 bool needs_flat_qualifier
= consumer_var
== NULL
&&
1399 (producer_var
->type
->contains_integer() ||
1400 producer_var
->type
->contains_double());
1402 if (!disable_varying_packing
&&
1403 (needs_flat_qualifier
||
1404 (consumer_stage
!= -1 && consumer_stage
!= MESA_SHADER_FRAGMENT
))) {
1405 /* Since this varying is not being consumed by the fragment shader, its
1406 * interpolation type varying cannot possibly affect rendering.
1407 * Also, this variable is non-flat and is (or contains) an integer
1409 * If the consumer stage is unknown, don't modify the interpolation
1410 * type as it could affect rendering later with separate shaders.
1412 * lower_packed_varyings requires all integer varyings to flat,
1413 * regardless of where they appear. We can trivially satisfy that
1414 * requirement by changing the interpolation type to flat here.
1417 producer_var
->data
.centroid
= false;
1418 producer_var
->data
.sample
= false;
1419 producer_var
->data
.interpolation
= INTERP_MODE_FLAT
;
1423 consumer_var
->data
.centroid
= false;
1424 consumer_var
->data
.sample
= false;
1425 consumer_var
->data
.interpolation
= INTERP_MODE_FLAT
;
1429 if (this->num_matches
== this->matches_capacity
) {
1430 this->matches_capacity
*= 2;
1431 this->matches
= (match
*)
1432 realloc(this->matches
,
1433 sizeof(*this->matches
) * this->matches_capacity
);
1436 /* We must use the consumer to compute the packing class because in GL4.4+
1437 * there is no guarantee interpolation qualifiers will match across stages.
1439 * From Section 4.5 (Interpolation Qualifiers) of the GLSL 4.30 spec:
1441 * "The type and presence of interpolation qualifiers of variables with
1442 * the same name declared in all linked shaders for the same cross-stage
1443 * interface must match, otherwise the link command will fail.
1445 * When comparing an output from one stage to an input of a subsequent
1446 * stage, the input and output don't match if their interpolation
1447 * qualifiers (or lack thereof) are not the same."
1449 * This text was also in at least revison 7 of the 4.40 spec but is no
1450 * longer in revision 9 and not in the 4.50 spec.
1452 const ir_variable
*const var
= (consumer_var
!= NULL
)
1453 ? consumer_var
: producer_var
;
1454 const gl_shader_stage stage
= (consumer_var
!= NULL
)
1455 ? consumer_stage
: producer_stage
;
1456 const glsl_type
*type
= get_varying_type(var
, stage
);
1458 this->matches
[this->num_matches
].packing_class
1459 = this->compute_packing_class(var
);
1460 this->matches
[this->num_matches
].packing_order
1461 = this->compute_packing_order(var
);
1462 if (this->disable_varying_packing
&& !is_varying_packing_safe(type
, var
)) {
1463 unsigned slots
= type
->count_attribute_slots(false);
1464 this->matches
[this->num_matches
].num_components
= slots
* 4;
1466 this->matches
[this->num_matches
].num_components
1467 = type
->component_slots();
1469 this->matches
[this->num_matches
].producer_var
= producer_var
;
1470 this->matches
[this->num_matches
].consumer_var
= consumer_var
;
1471 this->num_matches
++;
1473 producer_var
->data
.is_unmatched_generic_inout
= 0;
1475 consumer_var
->data
.is_unmatched_generic_inout
= 0;
1480 * Choose locations for all of the variable matches that were previously
1481 * passed to varying_matches::record().
1484 varying_matches::assign_locations(struct gl_shader_program
*prog
,
1485 uint64_t reserved_slots
)
1487 /* If packing has been disabled then we cannot safely sort the varyings by
1488 * class as it may mean we are using a version of OpenGL where
1489 * interpolation qualifiers are not guaranteed to be matching across
1490 * shaders, sorting in this case could result in mismatching shader
1492 * When packing is disabled the sort orders varyings used by transform
1493 * feedback first, but also depends on *undefined behaviour* of qsort to
1494 * reverse the order of the varyings. See: xfb_comparator().
1496 if (!this->disable_varying_packing
) {
1497 /* Sort varying matches into an order that makes them easy to pack. */
1498 qsort(this->matches
, this->num_matches
, sizeof(*this->matches
),
1499 &varying_matches::match_comparator
);
1501 /* Only sort varyings that are only used by transform feedback. */
1502 qsort(this->matches
, this->num_matches
, sizeof(*this->matches
),
1503 &varying_matches::xfb_comparator
);
1506 unsigned generic_location
= 0;
1507 unsigned generic_patch_location
= MAX_VARYING
*4;
1508 bool previous_var_xfb_only
= false;
1510 for (unsigned i
= 0; i
< this->num_matches
; i
++) {
1511 unsigned *location
= &generic_location
;
1513 const ir_variable
*var
;
1514 const glsl_type
*type
;
1515 bool is_vertex_input
= false;
1516 if (matches
[i
].consumer_var
) {
1517 var
= matches
[i
].consumer_var
;
1518 type
= get_varying_type(var
, consumer_stage
);
1519 if (consumer_stage
== MESA_SHADER_VERTEX
)
1520 is_vertex_input
= true;
1522 var
= matches
[i
].producer_var
;
1523 type
= get_varying_type(var
, producer_stage
);
1526 if (var
->data
.patch
)
1527 location
= &generic_patch_location
;
1529 /* Advance to the next slot if this varying has a different packing
1530 * class than the previous one, and we're not already on a slot
1533 * Also advance to the next slot if packing is disabled. This makes sure
1534 * we don't assign varyings the same locations which is possible
1535 * because we still pack individual arrays, records and matrices even
1536 * when packing is disabled. Note we don't advance to the next slot if
1537 * we can pack varyings together that are only used for transform
1540 if ((this->disable_varying_packing
&&
1541 !(previous_var_xfb_only
&& var
->data
.is_xfb_only
)) ||
1542 (i
> 0 && this->matches
[i
- 1].packing_class
1543 != this->matches
[i
].packing_class
)) {
1544 *location
= ALIGN(*location
, 4);
1547 previous_var_xfb_only
= var
->data
.is_xfb_only
;
1549 unsigned num_elements
= type
->count_attribute_slots(is_vertex_input
);
1551 if (this->disable_varying_packing
&&
1552 !is_varying_packing_safe(type
, var
))
1555 slot_end
= type
->without_array()->vector_elements
;
1556 slot_end
+= *location
- 1;
1558 /* FIXME: We could be smarter in the below code and loop back over
1559 * trying to fill any locations that we skipped because we couldn't pack
1560 * the varying between an explicit location. For now just let the user
1561 * hit the linking error if we run out of room and suggest they use
1562 * explicit locations.
1564 for (unsigned j
= 0; j
< num_elements
; j
++) {
1565 while ((slot_end
< MAX_VARYING
* 4u) &&
1566 ((reserved_slots
& (UINT64_C(1) << *location
/ 4u) ||
1567 (reserved_slots
& (UINT64_C(1) << slot_end
/ 4u))))) {
1569 *location
= ALIGN(*location
+ 1, 4);
1570 slot_end
= *location
;
1572 /* reset the counter and try again */
1576 /* Increase the slot to make sure there is enough room for next
1579 if (this->disable_varying_packing
&&
1580 !is_varying_packing_safe(type
, var
))
1583 slot_end
+= type
->without_array()->vector_elements
;
1586 if (!var
->data
.patch
&& *location
>= MAX_VARYING
* 4u) {
1587 linker_error(prog
, "insufficient contiguous locations available for "
1588 "%s it is possible an array or struct could not be "
1589 "packed between varyings with explicit locations. Try "
1590 "using an explicit location for arrays and structs.",
1594 this->matches
[i
].generic_location
= *location
;
1596 *location
+= this->matches
[i
].num_components
;
1599 return (generic_location
+ 3) / 4;
1604 * Update the producer and consumer shaders to reflect the locations
1605 * assignments that were made by varying_matches::assign_locations().
1608 varying_matches::store_locations() const
1610 for (unsigned i
= 0; i
< this->num_matches
; i
++) {
1611 ir_variable
*producer_var
= this->matches
[i
].producer_var
;
1612 ir_variable
*consumer_var
= this->matches
[i
].consumer_var
;
1613 unsigned generic_location
= this->matches
[i
].generic_location
;
1614 unsigned slot
= generic_location
/ 4;
1615 unsigned offset
= generic_location
% 4;
1618 producer_var
->data
.location
= VARYING_SLOT_VAR0
+ slot
;
1619 producer_var
->data
.location_frac
= offset
;
1623 assert(consumer_var
->data
.location
== -1);
1624 consumer_var
->data
.location
= VARYING_SLOT_VAR0
+ slot
;
1625 consumer_var
->data
.location_frac
= offset
;
1632 * Compute the "packing class" of the given varying. This is an unsigned
1633 * integer with the property that two variables in the same packing class can
1634 * be safely backed into the same vec4.
1637 varying_matches::compute_packing_class(const ir_variable
*var
)
1639 /* Without help from the back-end, there is no way to pack together
1640 * variables with different interpolation types, because
1641 * lower_packed_varyings must choose exactly one interpolation type for
1642 * each packed varying it creates.
1644 * However, we can safely pack together floats, ints, and uints, because:
1646 * - varyings of base type "int" and "uint" must use the "flat"
1647 * interpolation type, which can only occur in GLSL 1.30 and above.
1649 * - On platforms that support GLSL 1.30 and above, lower_packed_varyings
1650 * can store flat floats as ints without losing any information (using
1651 * the ir_unop_bitcast_* opcodes).
1653 * Therefore, the packing class depends only on the interpolation type.
1655 unsigned packing_class
= var
->data
.centroid
| (var
->data
.sample
<< 1) |
1656 (var
->data
.patch
<< 2);
1658 packing_class
+= var
->is_interpolation_flat()
1659 ? unsigned(INTERP_MODE_FLAT
) : var
->data
.interpolation
;
1660 return packing_class
;
1665 * Compute the "packing order" of the given varying. This is a sort key we
1666 * use to determine when to attempt to pack the given varying relative to
1667 * other varyings in the same packing class.
1669 varying_matches::packing_order_enum
1670 varying_matches::compute_packing_order(const ir_variable
*var
)
1672 const glsl_type
*element_type
= var
->type
;
1674 while (element_type
->base_type
== GLSL_TYPE_ARRAY
) {
1675 element_type
= element_type
->fields
.array
;
1678 switch (element_type
->component_slots() % 4) {
1679 case 1: return PACKING_ORDER_SCALAR
;
1680 case 2: return PACKING_ORDER_VEC2
;
1681 case 3: return PACKING_ORDER_VEC3
;
1682 case 0: return PACKING_ORDER_VEC4
;
1684 assert(!"Unexpected value of vector_elements");
1685 return PACKING_ORDER_VEC4
;
1691 * Comparison function passed to qsort() to sort varyings by packing_class and
1692 * then by packing_order.
1695 varying_matches::match_comparator(const void *x_generic
, const void *y_generic
)
1697 const match
*x
= (const match
*) x_generic
;
1698 const match
*y
= (const match
*) y_generic
;
1700 if (x
->packing_class
!= y
->packing_class
)
1701 return x
->packing_class
- y
->packing_class
;
1702 return x
->packing_order
- y
->packing_order
;
1707 * Comparison function passed to qsort() to sort varyings used only by
1708 * transform feedback when packing of other varyings is disabled.
1711 varying_matches::xfb_comparator(const void *x_generic
, const void *y_generic
)
1713 const match
*x
= (const match
*) x_generic
;
1715 if (x
->producer_var
!= NULL
&& x
->producer_var
->data
.is_xfb_only
)
1716 return match_comparator(x_generic
, y_generic
);
1718 /* FIXME: When the comparator returns 0 it means the elements being
1719 * compared are equivalent. However the qsort documentation says:
1721 * "The order of equivalent elements is undefined."
1723 * In practice the sort ends up reversing the order of the varyings which
1724 * means locations are also assigned in this reversed order and happens to
1725 * be what we want. This is also whats happening in
1726 * varying_matches::match_comparator().
1733 * Is the given variable a varying variable to be counted against the
1734 * limit in ctx->Const.MaxVarying?
1735 * This includes variables such as texcoords, colors and generic
1736 * varyings, but excludes variables such as gl_FrontFacing and gl_FragCoord.
1739 var_counts_against_varying_limit(gl_shader_stage stage
, const ir_variable
*var
)
1741 /* Only fragment shaders will take a varying variable as an input */
1742 if (stage
== MESA_SHADER_FRAGMENT
&&
1743 var
->data
.mode
== ir_var_shader_in
) {
1744 switch (var
->data
.location
) {
1745 case VARYING_SLOT_POS
:
1746 case VARYING_SLOT_FACE
:
1747 case VARYING_SLOT_PNTC
:
1758 * Visitor class that generates tfeedback_candidate structs describing all
1759 * possible targets of transform feedback.
1761 * tfeedback_candidate structs are stored in the hash table
1762 * tfeedback_candidates, which is passed to the constructor. This hash table
1763 * maps varying names to instances of the tfeedback_candidate struct.
1765 class tfeedback_candidate_generator
: public program_resource_visitor
1768 tfeedback_candidate_generator(void *mem_ctx
,
1769 hash_table
*tfeedback_candidates
)
1771 tfeedback_candidates(tfeedback_candidates
),
1777 void process(ir_variable
*var
)
1779 /* All named varying interface blocks should be flattened by now */
1780 assert(!var
->is_interface_instance());
1782 this->toplevel_var
= var
;
1783 this->varying_floats
= 0;
1784 program_resource_visitor::process(var
);
1788 virtual void visit_field(const glsl_type
*type
, const char *name
,
1791 assert(!type
->without_array()->is_record());
1792 assert(!type
->without_array()->is_interface());
1796 tfeedback_candidate
*candidate
1797 = rzalloc(this->mem_ctx
, tfeedback_candidate
);
1798 candidate
->toplevel_var
= this->toplevel_var
;
1799 candidate
->type
= type
;
1800 candidate
->offset
= this->varying_floats
;
1801 _mesa_hash_table_insert(this->tfeedback_candidates
,
1802 ralloc_strdup(this->mem_ctx
, name
),
1804 this->varying_floats
+= type
->component_slots();
1808 * Memory context used to allocate hash table keys and values.
1810 void * const mem_ctx
;
1813 * Hash table in which tfeedback_candidate objects should be stored.
1815 hash_table
* const tfeedback_candidates
;
1818 * Pointer to the toplevel variable that is being traversed.
1820 ir_variable
*toplevel_var
;
1823 * Total number of varying floats that have been visited so far. This is
1824 * used to determine the offset to each varying within the toplevel
1827 unsigned varying_floats
;
1834 populate_consumer_input_sets(void *mem_ctx
, exec_list
*ir
,
1835 hash_table
*consumer_inputs
,
1836 hash_table
*consumer_interface_inputs
,
1837 ir_variable
*consumer_inputs_with_locations
[VARYING_SLOT_TESS_MAX
])
1839 memset(consumer_inputs_with_locations
,
1841 sizeof(consumer_inputs_with_locations
[0]) * VARYING_SLOT_TESS_MAX
);
1843 foreach_in_list(ir_instruction
, node
, ir
) {
1844 ir_variable
*const input_var
= node
->as_variable();
1846 if (input_var
!= NULL
&& input_var
->data
.mode
== ir_var_shader_in
) {
1847 /* All interface blocks should have been lowered by this point */
1848 assert(!input_var
->type
->is_interface());
1850 if (input_var
->data
.explicit_location
) {
1851 /* assign_varying_locations only cares about finding the
1852 * ir_variable at the start of a contiguous location block.
1854 * - For !producer, consumer_inputs_with_locations isn't used.
1856 * - For !consumer, consumer_inputs_with_locations is empty.
1858 * For consumer && producer, if you were trying to set some
1859 * ir_variable to the middle of a location block on the other side
1860 * of producer/consumer, cross_validate_outputs_to_inputs() should
1861 * be link-erroring due to either type mismatch or location
1862 * overlaps. If the variables do match up, then they've got a
1863 * matching data.location and you only looked at
1864 * consumer_inputs_with_locations[var->data.location], not any
1865 * following entries for the array/structure.
1867 consumer_inputs_with_locations
[input_var
->data
.location
] =
1869 } else if (input_var
->get_interface_type() != NULL
) {
1870 char *const iface_field_name
=
1871 ralloc_asprintf(mem_ctx
, "%s.%s",
1872 input_var
->get_interface_type()->without_array()->name
,
1874 _mesa_hash_table_insert(consumer_interface_inputs
,
1875 iface_field_name
, input_var
);
1877 _mesa_hash_table_insert(consumer_inputs
,
1878 ralloc_strdup(mem_ctx
, input_var
->name
),
1886 * Find a variable from the consumer that "matches" the specified variable
1888 * This function only finds inputs with names that match. There is no
1889 * validation (here) that the types, etc. are compatible.
1892 get_matching_input(void *mem_ctx
,
1893 const ir_variable
*output_var
,
1894 hash_table
*consumer_inputs
,
1895 hash_table
*consumer_interface_inputs
,
1896 ir_variable
*consumer_inputs_with_locations
[VARYING_SLOT_TESS_MAX
])
1898 ir_variable
*input_var
;
1900 if (output_var
->data
.explicit_location
) {
1901 input_var
= consumer_inputs_with_locations
[output_var
->data
.location
];
1902 } else if (output_var
->get_interface_type() != NULL
) {
1903 char *const iface_field_name
=
1904 ralloc_asprintf(mem_ctx
, "%s.%s",
1905 output_var
->get_interface_type()->without_array()->name
,
1907 hash_entry
*entry
= _mesa_hash_table_search(consumer_interface_inputs
, iface_field_name
);
1908 input_var
= entry
? (ir_variable
*) entry
->data
: NULL
;
1910 hash_entry
*entry
= _mesa_hash_table_search(consumer_inputs
, output_var
->name
);
1911 input_var
= entry
? (ir_variable
*) entry
->data
: NULL
;
1914 return (input_var
== NULL
|| input_var
->data
.mode
!= ir_var_shader_in
)
1921 io_variable_cmp(const void *_a
, const void *_b
)
1923 const ir_variable
*const a
= *(const ir_variable
**) _a
;
1924 const ir_variable
*const b
= *(const ir_variable
**) _b
;
1926 if (a
->data
.explicit_location
&& b
->data
.explicit_location
)
1927 return b
->data
.location
- a
->data
.location
;
1929 if (a
->data
.explicit_location
&& !b
->data
.explicit_location
)
1932 if (!a
->data
.explicit_location
&& b
->data
.explicit_location
)
1935 return -strcmp(a
->name
, b
->name
);
1939 * Sort the shader IO variables into canonical order
1942 canonicalize_shader_io(exec_list
*ir
, enum ir_variable_mode io_mode
)
1944 ir_variable
*var_table
[MAX_PROGRAM_OUTPUTS
* 4];
1945 unsigned num_variables
= 0;
1947 foreach_in_list(ir_instruction
, node
, ir
) {
1948 ir_variable
*const var
= node
->as_variable();
1950 if (var
== NULL
|| var
->data
.mode
!= io_mode
)
1953 /* If we have already encountered more I/O variables that could
1954 * successfully link, bail.
1956 if (num_variables
== ARRAY_SIZE(var_table
))
1959 var_table
[num_variables
++] = var
;
1962 if (num_variables
== 0)
1965 /* Sort the list in reverse order (io_variable_cmp handles this). Later
1966 * we're going to push the variables on to the IR list as a stack, so we
1967 * want the last variable (in canonical order) to be first in the list.
1969 qsort(var_table
, num_variables
, sizeof(var_table
[0]), io_variable_cmp
);
1971 /* Remove the variable from it's current location in the IR, and put it at
1974 for (unsigned i
= 0; i
< num_variables
; i
++) {
1975 var_table
[i
]->remove();
1976 ir
->push_head(var_table
[i
]);
1981 * Generate a bitfield map of the explicit locations for shader varyings.
1983 * Note: For Tessellation shaders we are sitting right on the limits of the
1984 * 64 bit map. Per-vertex and per-patch both have separate location domains
1985 * with a max of MAX_VARYING.
1988 reserved_varying_slot(struct gl_linked_shader
*stage
,
1989 ir_variable_mode io_mode
)
1991 assert(io_mode
== ir_var_shader_in
|| io_mode
== ir_var_shader_out
);
1992 /* Avoid an overflow of the returned value */
1993 assert(MAX_VARYINGS_INCL_PATCH
<= 64);
2001 foreach_in_list(ir_instruction
, node
, stage
->ir
) {
2002 ir_variable
*const var
= node
->as_variable();
2004 if (var
== NULL
|| var
->data
.mode
!= io_mode
||
2005 !var
->data
.explicit_location
||
2006 var
->data
.location
< VARYING_SLOT_VAR0
)
2009 var_slot
= var
->data
.location
- VARYING_SLOT_VAR0
;
2011 unsigned num_elements
= get_varying_type(var
, stage
->Stage
)
2012 ->count_attribute_slots(stage
->Stage
== MESA_SHADER_VERTEX
);
2013 for (unsigned i
= 0; i
< num_elements
; i
++) {
2014 if (var_slot
>= 0 && var_slot
< MAX_VARYINGS_INCL_PATCH
)
2015 slots
|= UINT64_C(1) << var_slot
;
2025 * Assign locations for all variables that are produced in one pipeline stage
2026 * (the "producer") and consumed in the next stage (the "consumer").
2028 * Variables produced by the producer may also be consumed by transform
2031 * \param num_tfeedback_decls is the number of declarations indicating
2032 * variables that may be consumed by transform feedback.
2034 * \param tfeedback_decls is a pointer to an array of tfeedback_decl objects
2035 * representing the result of parsing the strings passed to
2036 * glTransformFeedbackVaryings(). assign_location() will be called for
2037 * each of these objects that matches one of the outputs of the
2040 * When num_tfeedback_decls is nonzero, it is permissible for the consumer to
2041 * be NULL. In this case, varying locations are assigned solely based on the
2042 * requirements of transform feedback.
2045 assign_varying_locations(struct gl_context
*ctx
,
2047 struct gl_shader_program
*prog
,
2048 gl_linked_shader
*producer
,
2049 gl_linked_shader
*consumer
,
2050 unsigned num_tfeedback_decls
,
2051 tfeedback_decl
*tfeedback_decls
,
2052 const uint64_t reserved_slots
)
2054 /* Tessellation shaders treat inputs and outputs as shared memory and can
2055 * access inputs and outputs of other invocations.
2056 * Therefore, they can't be lowered to temps easily (and definitely not
2059 bool unpackable_tess
=
2060 (consumer
&& consumer
->Stage
== MESA_SHADER_TESS_EVAL
) ||
2061 (consumer
&& consumer
->Stage
== MESA_SHADER_TESS_CTRL
) ||
2062 (producer
&& producer
->Stage
== MESA_SHADER_TESS_CTRL
);
2064 /* Transform feedback code assumes varying arrays are packed, so if the
2065 * driver has disabled varying packing, make sure to at least enable
2066 * packing required by transform feedback.
2069 ctx
->Extensions
.EXT_transform_feedback
&& !unpackable_tess
;
2071 /* Disable packing on outward facing interfaces for SSO because in ES we
2072 * need to retain the unpacked varying information for draw time
2075 * Packing is still enabled on individual arrays, structs, and matrices as
2076 * these are required by the transform feedback code and it is still safe
2077 * to do so. We also enable packing when a varying is only used for
2078 * transform feedback and its not a SSO.
2080 bool disable_varying_packing
=
2081 ctx
->Const
.DisableVaryingPacking
|| unpackable_tess
;
2082 if (prog
->SeparateShader
&& (producer
== NULL
|| consumer
== NULL
))
2083 disable_varying_packing
= true;
2085 varying_matches
matches(disable_varying_packing
, xfb_enabled
,
2086 producer
? producer
->Stage
: (gl_shader_stage
)-1,
2087 consumer
? consumer
->Stage
: (gl_shader_stage
)-1);
2088 hash_table
*tfeedback_candidates
=
2089 _mesa_hash_table_create(NULL
, _mesa_key_hash_string
,
2090 _mesa_key_string_equal
);
2091 hash_table
*consumer_inputs
=
2092 _mesa_hash_table_create(NULL
, _mesa_key_hash_string
,
2093 _mesa_key_string_equal
);
2094 hash_table
*consumer_interface_inputs
=
2095 _mesa_hash_table_create(NULL
, _mesa_key_hash_string
,
2096 _mesa_key_string_equal
);
2097 ir_variable
*consumer_inputs_with_locations
[VARYING_SLOT_TESS_MAX
] = {
2101 unsigned consumer_vertices
= 0;
2102 if (consumer
&& consumer
->Stage
== MESA_SHADER_GEOMETRY
)
2103 consumer_vertices
= prog
->Geom
.VerticesIn
;
2105 /* Operate in a total of four passes.
2107 * 1. Sort inputs / outputs into a canonical order. This is necessary so
2108 * that inputs / outputs of separable shaders will be assigned
2109 * predictable locations regardless of the order in which declarations
2110 * appeared in the shader source.
2112 * 2. Assign locations for any matching inputs and outputs.
2114 * 3. Mark output variables in the producer that do not have locations as
2115 * not being outputs. This lets the optimizer eliminate them.
2117 * 4. Mark input variables in the consumer that do not have locations as
2118 * not being inputs. This lets the optimizer eliminate them.
2121 canonicalize_shader_io(consumer
->ir
, ir_var_shader_in
);
2124 canonicalize_shader_io(producer
->ir
, ir_var_shader_out
);
2127 linker::populate_consumer_input_sets(mem_ctx
, consumer
->ir
,
2129 consumer_interface_inputs
,
2130 consumer_inputs_with_locations
);
2133 foreach_in_list(ir_instruction
, node
, producer
->ir
) {
2134 ir_variable
*const output_var
= node
->as_variable();
2136 if (output_var
== NULL
|| output_var
->data
.mode
!= ir_var_shader_out
)
2139 /* Only geometry shaders can use non-zero streams */
2140 assert(output_var
->data
.stream
== 0 ||
2141 (output_var
->data
.stream
< MAX_VERTEX_STREAMS
&&
2142 producer
->Stage
== MESA_SHADER_GEOMETRY
));
2144 if (num_tfeedback_decls
> 0) {
2145 tfeedback_candidate_generator
g(mem_ctx
, tfeedback_candidates
);
2146 g
.process(output_var
);
2149 ir_variable
*const input_var
=
2150 linker::get_matching_input(mem_ctx
, output_var
, consumer_inputs
,
2151 consumer_interface_inputs
,
2152 consumer_inputs_with_locations
);
2154 /* If a matching input variable was found, add this output (and the
2155 * input) to the set. If this is a separable program and there is no
2156 * consumer stage, add the output.
2158 * Always add TCS outputs. They are shared by all invocations
2159 * within a patch and can be used as shared memory.
2161 if (input_var
|| (prog
->SeparateShader
&& consumer
== NULL
) ||
2162 producer
->Stage
== MESA_SHADER_TESS_CTRL
) {
2163 matches
.record(output_var
, input_var
);
2166 /* Only stream 0 outputs can be consumed in the next stage */
2167 if (input_var
&& output_var
->data
.stream
!= 0) {
2168 linker_error(prog
, "output %s is assigned to stream=%d but "
2169 "is linked to an input, which requires stream=0",
2170 output_var
->name
, output_var
->data
.stream
);
2175 /* If there's no producer stage, then this must be a separable program.
2176 * For example, we may have a program that has just a fragment shader.
2177 * Later this program will be used with some arbitrary vertex (or
2178 * geometry) shader program. This means that locations must be assigned
2179 * for all the inputs.
2181 foreach_in_list(ir_instruction
, node
, consumer
->ir
) {
2182 ir_variable
*const input_var
= node
->as_variable();
2184 if (input_var
== NULL
|| input_var
->data
.mode
!= ir_var_shader_in
)
2187 matches
.record(NULL
, input_var
);
2191 _mesa_hash_table_destroy(consumer_inputs
, NULL
);
2192 _mesa_hash_table_destroy(consumer_interface_inputs
, NULL
);
2194 for (unsigned i
= 0; i
< num_tfeedback_decls
; ++i
) {
2195 if (!tfeedback_decls
[i
].is_varying())
2198 const tfeedback_candidate
*matched_candidate
2199 = tfeedback_decls
[i
].find_candidate(prog
, tfeedback_candidates
);
2201 if (matched_candidate
== NULL
) {
2202 _mesa_hash_table_destroy(tfeedback_candidates
, NULL
);
2206 if (matched_candidate
->toplevel_var
->data
.is_unmatched_generic_inout
) {
2207 matched_candidate
->toplevel_var
->data
.is_xfb_only
= 1;
2208 matches
.record(matched_candidate
->toplevel_var
, NULL
);
2212 const unsigned slots_used
= matches
.assign_locations(prog
, reserved_slots
);
2213 matches
.store_locations();
2215 for (unsigned i
= 0; i
< num_tfeedback_decls
; ++i
) {
2216 if (!tfeedback_decls
[i
].is_varying())
2219 if (!tfeedback_decls
[i
].assign_location(ctx
, prog
)) {
2220 _mesa_hash_table_destroy(tfeedback_candidates
, NULL
);
2224 _mesa_hash_table_destroy(tfeedback_candidates
, NULL
);
2226 if (consumer
&& producer
) {
2227 foreach_in_list(ir_instruction
, node
, consumer
->ir
) {
2228 ir_variable
*const var
= node
->as_variable();
2230 if (var
&& var
->data
.mode
== ir_var_shader_in
&&
2231 var
->data
.is_unmatched_generic_inout
) {
2232 if (!prog
->IsES
&& prog
->Version
<= 120) {
2233 /* On page 25 (page 31 of the PDF) of the GLSL 1.20 spec:
2235 * Only those varying variables used (i.e. read) in
2236 * the fragment shader executable must be written to
2237 * by the vertex shader executable; declaring
2238 * superfluous varying variables in a vertex shader is
2241 * We interpret this text as meaning that the VS must
2242 * write the variable for the FS to read it. See
2243 * "glsl1-varying read but not written" in piglit.
2245 linker_error(prog
, "%s shader varying %s not written "
2247 _mesa_shader_stage_to_string(consumer
->Stage
),
2249 _mesa_shader_stage_to_string(producer
->Stage
));
2251 linker_warning(prog
, "%s shader varying %s not written "
2253 _mesa_shader_stage_to_string(consumer
->Stage
),
2255 _mesa_shader_stage_to_string(producer
->Stage
));
2260 /* Now that validation is done its safe to remove unused varyings. As
2261 * we have both a producer and consumer its safe to remove unused
2262 * varyings even if the program is a SSO because the stages are being
2263 * linked together i.e. we have a multi-stage SSO.
2265 remove_unused_shader_inputs_and_outputs(false, producer
,
2267 remove_unused_shader_inputs_and_outputs(false, consumer
,
2272 lower_packed_varyings(mem_ctx
, slots_used
, ir_var_shader_out
,
2273 0, producer
, disable_varying_packing
,
2278 lower_packed_varyings(mem_ctx
, slots_used
, ir_var_shader_in
,
2279 consumer_vertices
, consumer
,
2280 disable_varying_packing
, xfb_enabled
);
2287 check_against_output_limit(struct gl_context
*ctx
,
2288 struct gl_shader_program
*prog
,
2289 gl_linked_shader
*producer
,
2290 unsigned num_explicit_locations
)
2292 unsigned output_vectors
= num_explicit_locations
;
2294 foreach_in_list(ir_instruction
, node
, producer
->ir
) {
2295 ir_variable
*const var
= node
->as_variable();
2297 if (var
&& !var
->data
.explicit_location
&&
2298 var
->data
.mode
== ir_var_shader_out
&&
2299 var_counts_against_varying_limit(producer
->Stage
, var
)) {
2300 /* outputs for fragment shader can't be doubles */
2301 output_vectors
+= var
->type
->count_attribute_slots(false);
2305 assert(producer
->Stage
!= MESA_SHADER_FRAGMENT
);
2306 unsigned max_output_components
=
2307 ctx
->Const
.Program
[producer
->Stage
].MaxOutputComponents
;
2309 const unsigned output_components
= output_vectors
* 4;
2310 if (output_components
> max_output_components
) {
2311 if (ctx
->API
== API_OPENGLES2
|| prog
->IsES
)
2312 linker_error(prog
, "%s shader uses too many output vectors "
2314 _mesa_shader_stage_to_string(producer
->Stage
),
2316 max_output_components
/ 4);
2318 linker_error(prog
, "%s shader uses too many output components "
2320 _mesa_shader_stage_to_string(producer
->Stage
),
2322 max_output_components
);
2331 check_against_input_limit(struct gl_context
*ctx
,
2332 struct gl_shader_program
*prog
,
2333 gl_linked_shader
*consumer
,
2334 unsigned num_explicit_locations
)
2336 unsigned input_vectors
= num_explicit_locations
;
2338 foreach_in_list(ir_instruction
, node
, consumer
->ir
) {
2339 ir_variable
*const var
= node
->as_variable();
2341 if (var
&& !var
->data
.explicit_location
&&
2342 var
->data
.mode
== ir_var_shader_in
&&
2343 var_counts_against_varying_limit(consumer
->Stage
, var
)) {
2344 /* vertex inputs aren't varying counted */
2345 input_vectors
+= var
->type
->count_attribute_slots(false);
2349 assert(consumer
->Stage
!= MESA_SHADER_VERTEX
);
2350 unsigned max_input_components
=
2351 ctx
->Const
.Program
[consumer
->Stage
].MaxInputComponents
;
2353 const unsigned input_components
= input_vectors
* 4;
2354 if (input_components
> max_input_components
) {
2355 if (ctx
->API
== API_OPENGLES2
|| prog
->IsES
)
2356 linker_error(prog
, "%s shader uses too many input vectors "
2358 _mesa_shader_stage_to_string(consumer
->Stage
),
2360 max_input_components
/ 4);
2362 linker_error(prog
, "%s shader uses too many input components "
2364 _mesa_shader_stage_to_string(consumer
->Stage
),
2366 max_input_components
);