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 struct gl_shader_program
*prog
,
112 unsigned *num_tfeedback_decls
,
113 char ***varying_names
)
115 bool has_xfb_qualifiers
= false;
117 /* We still need to enable transform feedback mode even if xfb_stride is
118 * only applied to a global out. Also we don't bother to propagate
119 * xfb_stride to interface block members so this will catch that case also.
121 for (unsigned j
= 0; j
< MAX_FEEDBACK_BUFFERS
; j
++) {
122 if (prog
->TransformFeedback
.BufferStride
[j
]) {
123 has_xfb_qualifiers
= true;
128 foreach_in_list(ir_instruction
, node
, sh
->ir
) {
129 ir_variable
*var
= node
->as_variable();
130 if (!var
|| var
->data
.mode
!= ir_var_shader_out
)
133 /* From the ARB_enhanced_layouts spec:
135 * "Any shader making any static use (after preprocessing) of any of
136 * these *xfb_* qualifiers will cause the shader to be in a
137 * transform feedback capturing mode and hence responsible for
138 * describing the transform feedback setup. This mode will capture
139 * any output selected by *xfb_offset*, directly or indirectly, to
140 * a transform feedback buffer."
142 if (var
->data
.explicit_xfb_buffer
|| var
->data
.explicit_xfb_stride
) {
143 has_xfb_qualifiers
= true;
146 if (var
->data
.explicit_xfb_offset
) {
147 *num_tfeedback_decls
+= var
->type
->varying_count();
148 has_xfb_qualifiers
= true;
152 if (*num_tfeedback_decls
== 0)
153 return has_xfb_qualifiers
;
156 *varying_names
= ralloc_array(mem_ctx
, char *, *num_tfeedback_decls
);
157 foreach_in_list(ir_instruction
, node
, sh
->ir
) {
158 ir_variable
*var
= node
->as_variable();
159 if (!var
|| var
->data
.mode
!= ir_var_shader_out
)
162 if (var
->data
.explicit_xfb_offset
) {
164 const glsl_type
*type
, *member_type
;
166 if (var
->data
.from_named_ifc_block
) {
167 type
= var
->get_interface_type();
168 /* Find the member type before it was altered by lowering */
170 type
->fields
.structure
[type
->field_index(var
->name
)].type
;
171 name
= ralloc_strdup(NULL
, type
->without_array()->name
);
175 name
= ralloc_strdup(NULL
, var
->name
);
177 create_xfb_varying_names(mem_ctx
, type
, &name
, strlen(name
), &i
,
178 var
->name
, member_type
, varying_names
);
183 assert(i
== *num_tfeedback_decls
);
184 return has_xfb_qualifiers
;
188 anonymous_struct_type_matches(const glsl_type
*output_type
,
189 const glsl_type
*to_match
)
191 while (output_type
->is_array() && to_match
->is_array()) {
192 /* if the lengths at each level don't match fail. */
193 if (output_type
->length
!= to_match
->length
)
195 output_type
= output_type
->fields
.array
;
196 to_match
= to_match
->fields
.array
;
199 if (output_type
->is_array() || to_match
->is_array())
201 return output_type
->is_anonymous() &&
202 to_match
->is_anonymous() &&
203 to_match
->record_compare(output_type
);
207 * Validate the types and qualifiers of an output from one stage against the
208 * matching input to another stage.
211 cross_validate_types_and_qualifiers(struct gl_shader_program
*prog
,
212 const ir_variable
*input
,
213 const ir_variable
*output
,
214 gl_shader_stage consumer_stage
,
215 gl_shader_stage producer_stage
)
217 /* Check that the types match between stages.
219 const glsl_type
*type_to_match
= input
->type
;
221 /* VS -> GS, VS -> TCS, VS -> TES, TES -> GS */
222 const bool extra_array_level
= (producer_stage
== MESA_SHADER_VERTEX
&&
223 consumer_stage
!= MESA_SHADER_FRAGMENT
) ||
224 consumer_stage
== MESA_SHADER_GEOMETRY
;
225 if (extra_array_level
) {
226 assert(type_to_match
->is_array());
227 type_to_match
= type_to_match
->fields
.array
;
230 if (type_to_match
!= output
->type
) {
231 /* There is a bit of a special case for gl_TexCoord. This
232 * built-in is unsized by default. Applications that variable
233 * access it must redeclare it with a size. There is some
234 * language in the GLSL spec that implies the fragment shader
235 * and vertex shader do not have to agree on this size. Other
236 * driver behave this way, and one or two applications seem to
239 * Neither declaration needs to be modified here because the array
240 * sizes are fixed later when update_array_sizes is called.
242 * From page 48 (page 54 of the PDF) of the GLSL 1.10 spec:
244 * "Unlike user-defined varying variables, the built-in
245 * varying variables don't have a strict one-to-one
246 * correspondence between the vertex language and the
247 * fragment language."
249 if (!output
->type
->is_array() || !is_gl_identifier(output
->name
)) {
250 bool anon_matches
= anonymous_struct_type_matches(output
->type
, type_to_match
);
254 "%s shader output `%s' declared as type `%s', "
255 "but %s shader input declared as type `%s'\n",
256 _mesa_shader_stage_to_string(producer_stage
),
259 _mesa_shader_stage_to_string(consumer_stage
),
266 /* Check that all of the qualifiers match between stages.
269 /* According to the OpenGL and OpenGLES GLSL specs, the centroid qualifier
270 * should match until OpenGL 4.3 and OpenGLES 3.1. The OpenGLES 3.0
271 * conformance test suite does not verify that the qualifiers must match.
272 * The deqp test suite expects the opposite (OpenGLES 3.1) behavior for
273 * OpenGLES 3.0 drivers, so we relax the checking in all cases.
275 if (false /* always skip the centroid check */ &&
276 prog
->data
->Version
< (prog
->IsES
? 310 : 430) &&
277 input
->data
.centroid
!= output
->data
.centroid
) {
279 "%s shader output `%s' %s centroid qualifier, "
280 "but %s shader input %s centroid qualifier\n",
281 _mesa_shader_stage_to_string(producer_stage
),
283 (output
->data
.centroid
) ? "has" : "lacks",
284 _mesa_shader_stage_to_string(consumer_stage
),
285 (input
->data
.centroid
) ? "has" : "lacks");
289 if (input
->data
.sample
!= output
->data
.sample
) {
291 "%s shader output `%s' %s sample qualifier, "
292 "but %s shader input %s sample qualifier\n",
293 _mesa_shader_stage_to_string(producer_stage
),
295 (output
->data
.sample
) ? "has" : "lacks",
296 _mesa_shader_stage_to_string(consumer_stage
),
297 (input
->data
.sample
) ? "has" : "lacks");
301 if (input
->data
.patch
!= output
->data
.patch
) {
303 "%s shader output `%s' %s patch qualifier, "
304 "but %s shader input %s patch qualifier\n",
305 _mesa_shader_stage_to_string(producer_stage
),
307 (output
->data
.patch
) ? "has" : "lacks",
308 _mesa_shader_stage_to_string(consumer_stage
),
309 (input
->data
.patch
) ? "has" : "lacks");
313 /* The GLSL 4.30 and GLSL ES 3.00 specifications say:
315 * "As only outputs need be declared with invariant, an output from
316 * one shader stage will still match an input of a subsequent stage
317 * without the input being declared as invariant."
319 * while GLSL 4.20 says:
321 * "For variables leaving one shader and coming into another shader,
322 * the invariant keyword has to be used in both shaders, or a link
323 * error will result."
325 * and GLSL ES 1.00 section 4.6.4 "Invariance and Linking" says:
327 * "The invariance of varyings that are declared in both the vertex
328 * and fragment shaders must match."
330 if (input
->data
.invariant
!= output
->data
.invariant
&&
331 prog
->data
->Version
< (prog
->IsES
? 300 : 430)) {
333 "%s shader output `%s' %s invariant qualifier, "
334 "but %s shader input %s invariant qualifier\n",
335 _mesa_shader_stage_to_string(producer_stage
),
337 (output
->data
.invariant
) ? "has" : "lacks",
338 _mesa_shader_stage_to_string(consumer_stage
),
339 (input
->data
.invariant
) ? "has" : "lacks");
343 /* GLSL >= 4.40 removes text requiring interpolation qualifiers
344 * to match cross stage, they must only match within the same stage.
346 * From page 84 (page 90 of the PDF) of the GLSL 4.40 spec:
348 * "It is a link-time error if, within the same stage, the interpolation
349 * qualifiers of variables of the same name do not match.
352 if (input
->data
.interpolation
!= output
->data
.interpolation
&&
353 prog
->data
->Version
< 440) {
355 "%s shader output `%s' specifies %s "
356 "interpolation qualifier, "
357 "but %s shader input specifies %s "
358 "interpolation qualifier\n",
359 _mesa_shader_stage_to_string(producer_stage
),
361 interpolation_string(output
->data
.interpolation
),
362 _mesa_shader_stage_to_string(consumer_stage
),
363 interpolation_string(input
->data
.interpolation
));
369 * Validate front and back color outputs against single color input
372 cross_validate_front_and_back_color(struct gl_shader_program
*prog
,
373 const ir_variable
*input
,
374 const ir_variable
*front_color
,
375 const ir_variable
*back_color
,
376 gl_shader_stage consumer_stage
,
377 gl_shader_stage producer_stage
)
379 if (front_color
!= NULL
&& front_color
->data
.assigned
)
380 cross_validate_types_and_qualifiers(prog
, input
, front_color
,
381 consumer_stage
, producer_stage
);
383 if (back_color
!= NULL
&& back_color
->data
.assigned
)
384 cross_validate_types_and_qualifiers(prog
, input
, back_color
,
385 consumer_stage
, producer_stage
);
389 * Validate that outputs from one stage match inputs of another
392 cross_validate_outputs_to_inputs(struct gl_shader_program
*prog
,
393 gl_linked_shader
*producer
,
394 gl_linked_shader
*consumer
)
396 glsl_symbol_table parameters
;
397 ir_variable
*explicit_locations
[MAX_VARYINGS_INCL_PATCH
][4] =
400 /* Find all shader outputs in the "producer" stage.
402 foreach_in_list(ir_instruction
, node
, producer
->ir
) {
403 ir_variable
*const var
= node
->as_variable();
405 if (var
== NULL
|| var
->data
.mode
!= ir_var_shader_out
)
408 if (!var
->data
.explicit_location
409 || var
->data
.location
< VARYING_SLOT_VAR0
)
410 parameters
.add_variable(var
);
412 /* User-defined varyings with explicit locations are handled
413 * differently because they do not need to have matching names.
415 const glsl_type
*type
= get_varying_type(var
, producer
->Stage
);
416 unsigned num_elements
= type
->count_attribute_slots(false);
417 unsigned idx
= var
->data
.location
- VARYING_SLOT_VAR0
;
418 unsigned slot_limit
= idx
+ num_elements
;
421 if (type
->without_array()->is_record()) {
422 /* The component qualifier can't be used on structs so just treat
423 * all component slots as used.
427 unsigned dmul
= type
->without_array()->is_64bit() ? 2 : 1;
428 last_comp
= var
->data
.location_frac
+
429 type
->without_array()->vector_elements
* dmul
;
432 while (idx
< slot_limit
) {
433 unsigned i
= var
->data
.location_frac
;
434 while (i
< last_comp
) {
435 if (explicit_locations
[idx
][i
] != NULL
) {
437 "%s shader has multiple outputs explicitly "
438 "assigned to location %d and component %d\n",
439 _mesa_shader_stage_to_string(producer
->Stage
),
440 idx
, var
->data
.location_frac
);
444 /* Make sure all component at this location have the same type.
446 for (unsigned j
= 0; j
< 4; j
++) {
447 if (explicit_locations
[idx
][j
] &&
448 (explicit_locations
[idx
][j
]->type
->without_array()
449 ->base_type
!= type
->without_array()->base_type
)) {
451 "Varyings sharing the same location must "
452 "have the same underlying numerical type. "
453 "Location %u component %u\n", idx
,
454 var
->data
.location_frac
);
459 explicit_locations
[idx
][i
] = var
;
462 /* We need to do some special handling for doubles as dvec3 and
463 * dvec4 consume two consecutive locations. We don't need to
464 * worry about components beginning at anything other than 0 as
465 * the spec does not allow this for dvec3 and dvec4.
467 if (i
== 4 && last_comp
> 4) {
468 last_comp
= last_comp
- 4;
469 /* Bump location index and reset the component index */
480 /* Find all shader inputs in the "consumer" stage. Any variables that have
481 * matching outputs already in the symbol table must have the same type and
484 * Exception: if the consumer is the geometry shader, then the inputs
485 * should be arrays and the type of the array element should match the type
486 * of the corresponding producer output.
488 foreach_in_list(ir_instruction
, node
, consumer
->ir
) {
489 ir_variable
*const input
= node
->as_variable();
491 if (input
== NULL
|| input
->data
.mode
!= ir_var_shader_in
)
494 if (strcmp(input
->name
, "gl_Color") == 0 && input
->data
.used
) {
495 const ir_variable
*const front_color
=
496 parameters
.get_variable("gl_FrontColor");
498 const ir_variable
*const back_color
=
499 parameters
.get_variable("gl_BackColor");
501 cross_validate_front_and_back_color(prog
, input
,
502 front_color
, back_color
,
503 consumer
->Stage
, producer
->Stage
);
504 } else if (strcmp(input
->name
, "gl_SecondaryColor") == 0 && input
->data
.used
) {
505 const ir_variable
*const front_color
=
506 parameters
.get_variable("gl_FrontSecondaryColor");
508 const ir_variable
*const back_color
=
509 parameters
.get_variable("gl_BackSecondaryColor");
511 cross_validate_front_and_back_color(prog
, input
,
512 front_color
, back_color
,
513 consumer
->Stage
, producer
->Stage
);
515 /* The rules for connecting inputs and outputs change in the presence
516 * of explicit locations. In this case, we no longer care about the
517 * names of the variables. Instead, we care only about the
518 * explicitly assigned location.
520 ir_variable
*output
= NULL
;
521 if (input
->data
.explicit_location
522 && input
->data
.location
>= VARYING_SLOT_VAR0
) {
524 const glsl_type
*type
= get_varying_type(input
, consumer
->Stage
);
525 unsigned num_elements
= type
->count_attribute_slots(false);
526 unsigned idx
= input
->data
.location
- VARYING_SLOT_VAR0
;
527 unsigned slot_limit
= idx
+ num_elements
;
529 while (idx
< slot_limit
) {
530 output
= explicit_locations
[idx
][input
->data
.location_frac
];
532 if (output
== NULL
||
533 input
->data
.location
!= output
->data
.location
) {
535 "%s shader input `%s' with explicit location "
536 "has no matching output\n",
537 _mesa_shader_stage_to_string(consumer
->Stage
),
544 output
= parameters
.get_variable(input
->name
);
547 if (output
!= NULL
) {
548 /* Interface blocks have their own validation elsewhere so don't
549 * try validating them here.
551 if (!(input
->get_interface_type() &&
552 output
->get_interface_type()))
553 cross_validate_types_and_qualifiers(prog
, input
, output
,
557 /* Check for input vars with unmatched output vars in prev stage
558 * taking into account that interface blocks could have a matching
559 * output but with different name, so we ignore them.
561 assert(!input
->data
.assigned
);
562 if (input
->data
.used
&& !input
->get_interface_type() &&
563 !input
->data
.explicit_location
&& !prog
->SeparateShader
)
565 "%s shader input `%s' "
566 "has no matching output in the previous stage\n",
567 _mesa_shader_stage_to_string(consumer
->Stage
),
575 * Demote shader inputs and outputs that are not used in other stages, and
576 * remove them via dead code elimination.
579 remove_unused_shader_inputs_and_outputs(bool is_separate_shader_object
,
580 gl_linked_shader
*sh
,
581 enum ir_variable_mode mode
)
583 if (is_separate_shader_object
)
586 foreach_in_list(ir_instruction
, node
, sh
->ir
) {
587 ir_variable
*const var
= node
->as_variable();
589 if (var
== NULL
|| var
->data
.mode
!= int(mode
))
592 /* A shader 'in' or 'out' variable is only really an input or output if
593 * its value is used by other shader stages. This will cause the
594 * variable to have a location assigned.
596 if (var
->data
.is_unmatched_generic_inout
&& !var
->data
.is_xfb_only
) {
597 assert(var
->data
.mode
!= ir_var_temporary
);
599 /* Assign zeros to demoted inputs to allow more optimizations. */
600 if (var
->data
.mode
== ir_var_shader_in
&& !var
->constant_value
)
601 var
->constant_value
= ir_constant::zero(var
, var
->type
);
603 var
->data
.mode
= ir_var_auto
;
607 /* Eliminate code that is now dead due to unused inputs/outputs being
610 while (do_dead_code(sh
->ir
, false))
616 * Initialize this object based on a string that was passed to
617 * glTransformFeedbackVaryings.
619 * If the input is mal-formed, this call still succeeds, but it sets
620 * this->var_name to a mal-formed input, so tfeedback_decl::find_output_var()
621 * will fail to find any matching variable.
624 tfeedback_decl::init(struct gl_context
*ctx
, const void *mem_ctx
,
627 /* We don't have to be pedantic about what is a valid GLSL variable name,
628 * because any variable with an invalid name can't exist in the IR anyway.
632 this->orig_name
= input
;
633 this->lowered_builtin_array_variable
= none
;
634 this->skip_components
= 0;
635 this->next_buffer_separator
= false;
636 this->matched_candidate
= NULL
;
641 if (ctx
->Extensions
.ARB_transform_feedback3
) {
642 /* Parse gl_NextBuffer. */
643 if (strcmp(input
, "gl_NextBuffer") == 0) {
644 this->next_buffer_separator
= true;
648 /* Parse gl_SkipComponents. */
649 if (strcmp(input
, "gl_SkipComponents1") == 0)
650 this->skip_components
= 1;
651 else if (strcmp(input
, "gl_SkipComponents2") == 0)
652 this->skip_components
= 2;
653 else if (strcmp(input
, "gl_SkipComponents3") == 0)
654 this->skip_components
= 3;
655 else if (strcmp(input
, "gl_SkipComponents4") == 0)
656 this->skip_components
= 4;
658 if (this->skip_components
)
662 /* Parse a declaration. */
663 const char *base_name_end
;
664 long subscript
= parse_program_resource_name(input
, &base_name_end
);
665 this->var_name
= ralloc_strndup(mem_ctx
, input
, base_name_end
- input
);
666 if (this->var_name
== NULL
) {
667 _mesa_error_no_memory(__func__
);
671 if (subscript
>= 0) {
672 this->array_subscript
= subscript
;
673 this->is_subscripted
= true;
675 this->is_subscripted
= false;
678 /* For drivers that lower gl_ClipDistance to gl_ClipDistanceMESA, this
679 * class must behave specially to account for the fact that gl_ClipDistance
680 * is converted from a float[8] to a vec4[2].
682 if (ctx
->Const
.ShaderCompilerOptions
[MESA_SHADER_VERTEX
].LowerCombinedClipCullDistance
&&
683 strcmp(this->var_name
, "gl_ClipDistance") == 0) {
684 this->lowered_builtin_array_variable
= clip_distance
;
686 if (ctx
->Const
.ShaderCompilerOptions
[MESA_SHADER_VERTEX
].LowerCombinedClipCullDistance
&&
687 strcmp(this->var_name
, "gl_CullDistance") == 0) {
688 this->lowered_builtin_array_variable
= cull_distance
;
691 if (ctx
->Const
.LowerTessLevel
&&
692 (strcmp(this->var_name
, "gl_TessLevelOuter") == 0))
693 this->lowered_builtin_array_variable
= tess_level_outer
;
694 if (ctx
->Const
.LowerTessLevel
&&
695 (strcmp(this->var_name
, "gl_TessLevelInner") == 0))
696 this->lowered_builtin_array_variable
= tess_level_inner
;
701 * Determine whether two tfeedback_decl objects refer to the same variable and
702 * array index (if applicable).
705 tfeedback_decl::is_same(const tfeedback_decl
&x
, const tfeedback_decl
&y
)
707 assert(x
.is_varying() && y
.is_varying());
709 if (strcmp(x
.var_name
, y
.var_name
) != 0)
711 if (x
.is_subscripted
!= y
.is_subscripted
)
713 if (x
.is_subscripted
&& x
.array_subscript
!= y
.array_subscript
)
720 * Assign a location and stream ID for this tfeedback_decl object based on the
721 * transform feedback candidate found by find_candidate.
723 * If an error occurs, the error is reported through linker_error() and false
727 tfeedback_decl::assign_location(struct gl_context
*ctx
,
728 struct gl_shader_program
*prog
)
730 assert(this->is_varying());
732 unsigned fine_location
733 = this->matched_candidate
->toplevel_var
->data
.location
* 4
734 + this->matched_candidate
->toplevel_var
->data
.location_frac
735 + this->matched_candidate
->offset
;
736 const unsigned dmul
=
737 this->matched_candidate
->type
->without_array()->is_64bit() ? 2 : 1;
739 if (this->matched_candidate
->type
->is_array()) {
741 const unsigned matrix_cols
=
742 this->matched_candidate
->type
->fields
.array
->matrix_columns
;
743 const unsigned vector_elements
=
744 this->matched_candidate
->type
->fields
.array
->vector_elements
;
745 unsigned actual_array_size
;
746 switch (this->lowered_builtin_array_variable
) {
748 actual_array_size
= prog
->last_vert_prog
?
749 prog
->last_vert_prog
->info
.clip_distance_array_size
: 0;
752 actual_array_size
= prog
->last_vert_prog
?
753 prog
->last_vert_prog
->info
.cull_distance_array_size
: 0;
755 case tess_level_outer
:
756 actual_array_size
= 4;
758 case tess_level_inner
:
759 actual_array_size
= 2;
763 actual_array_size
= this->matched_candidate
->type
->array_size();
767 if (this->is_subscripted
) {
768 /* Check array bounds. */
769 if (this->array_subscript
>= actual_array_size
) {
770 linker_error(prog
, "Transform feedback varying %s has index "
771 "%i, but the array size is %u.",
772 this->orig_name
, this->array_subscript
,
776 unsigned array_elem_size
= this->lowered_builtin_array_variable
?
777 1 : vector_elements
* matrix_cols
* dmul
;
778 fine_location
+= array_elem_size
* this->array_subscript
;
781 this->size
= actual_array_size
;
783 this->vector_elements
= vector_elements
;
784 this->matrix_columns
= matrix_cols
;
785 if (this->lowered_builtin_array_variable
)
786 this->type
= GL_FLOAT
;
788 this->type
= this->matched_candidate
->type
->fields
.array
->gl_type
;
790 /* Regular variable (scalar, vector, or matrix) */
791 if (this->is_subscripted
) {
792 linker_error(prog
, "Transform feedback varying %s requested, "
793 "but %s is not an array.",
794 this->orig_name
, this->var_name
);
798 this->vector_elements
= this->matched_candidate
->type
->vector_elements
;
799 this->matrix_columns
= this->matched_candidate
->type
->matrix_columns
;
800 this->type
= this->matched_candidate
->type
->gl_type
;
802 this->location
= fine_location
/ 4;
803 this->location_frac
= fine_location
% 4;
805 /* From GL_EXT_transform_feedback:
806 * A program will fail to link if:
808 * * the total number of components to capture in any varying
809 * variable in <varyings> is greater than the constant
810 * MAX_TRANSFORM_FEEDBACK_SEPARATE_COMPONENTS_EXT and the
811 * buffer mode is SEPARATE_ATTRIBS_EXT;
813 if (prog
->TransformFeedback
.BufferMode
== GL_SEPARATE_ATTRIBS
&&
814 this->num_components() >
815 ctx
->Const
.MaxTransformFeedbackSeparateComponents
) {
816 linker_error(prog
, "Transform feedback varying %s exceeds "
817 "MAX_TRANSFORM_FEEDBACK_SEPARATE_COMPONENTS.",
822 /* Only transform feedback varyings can be assigned to non-zero streams,
823 * so assign the stream id here.
825 this->stream_id
= this->matched_candidate
->toplevel_var
->data
.stream
;
827 unsigned array_offset
= this->array_subscript
* 4 * dmul
;
828 unsigned struct_offset
= this->matched_candidate
->offset
* 4 * dmul
;
829 this->buffer
= this->matched_candidate
->toplevel_var
->data
.xfb_buffer
;
830 this->offset
= this->matched_candidate
->toplevel_var
->data
.offset
+
831 array_offset
+ struct_offset
;
838 tfeedback_decl::get_num_outputs() const
840 if (!this->is_varying()) {
843 return (this->num_components() + this->location_frac
+ 3)/4;
848 * Update gl_transform_feedback_info to reflect this tfeedback_decl.
850 * If an error occurs, the error is reported through linker_error() and false
854 tfeedback_decl::store(struct gl_context
*ctx
, struct gl_shader_program
*prog
,
855 struct gl_transform_feedback_info
*info
,
856 unsigned buffer
, unsigned buffer_index
,
857 const unsigned max_outputs
, bool *explicit_stride
,
858 bool has_xfb_qualifiers
) const
860 unsigned xfb_offset
= 0;
861 unsigned size
= this->size
;
862 /* Handle gl_SkipComponents. */
863 if (this->skip_components
) {
864 info
->Buffers
[buffer
].Stride
+= this->skip_components
;
865 size
= this->skip_components
;
869 if (this->next_buffer_separator
) {
874 if (has_xfb_qualifiers
) {
875 xfb_offset
= this->offset
/ 4;
877 xfb_offset
= info
->Buffers
[buffer
].Stride
;
879 info
->Varyings
[info
->NumVarying
].Offset
= xfb_offset
* 4;
882 unsigned location
= this->location
;
883 unsigned location_frac
= this->location_frac
;
884 unsigned num_components
= this->num_components();
885 while (num_components
> 0) {
886 unsigned output_size
= MIN2(num_components
, 4 - location_frac
);
887 assert((info
->NumOutputs
== 0 && max_outputs
== 0) ||
888 info
->NumOutputs
< max_outputs
);
890 /* From the ARB_enhanced_layouts spec:
892 * "If such a block member or variable is not written during a shader
893 * invocation, the buffer contents at the assigned offset will be
894 * undefined. Even if there are no static writes to a variable or
895 * member that is assigned a transform feedback offset, the space is
896 * still allocated in the buffer and still affects the stride."
898 if (this->is_varying_written()) {
899 info
->Outputs
[info
->NumOutputs
].ComponentOffset
= location_frac
;
900 info
->Outputs
[info
->NumOutputs
].OutputRegister
= location
;
901 info
->Outputs
[info
->NumOutputs
].NumComponents
= output_size
;
902 info
->Outputs
[info
->NumOutputs
].StreamId
= stream_id
;
903 info
->Outputs
[info
->NumOutputs
].OutputBuffer
= buffer
;
904 info
->Outputs
[info
->NumOutputs
].DstOffset
= xfb_offset
;
907 info
->Buffers
[buffer
].Stream
= this->stream_id
;
908 xfb_offset
+= output_size
;
910 num_components
-= output_size
;
916 if (explicit_stride
&& explicit_stride
[buffer
]) {
917 if (this->is_64bit() && info
->Buffers
[buffer
].Stride
% 2) {
918 linker_error(prog
, "invalid qualifier xfb_stride=%d must be a "
919 "multiple of 8 as its applied to a type that is or "
920 "contains a double.",
921 info
->Buffers
[buffer
].Stride
* 4);
925 if ((this->offset
/ 4) / info
->Buffers
[buffer
].Stride
!=
926 (xfb_offset
- 1) / info
->Buffers
[buffer
].Stride
) {
927 linker_error(prog
, "xfb_offset (%d) overflows xfb_stride (%d) for "
928 "buffer (%d)", xfb_offset
* 4,
929 info
->Buffers
[buffer
].Stride
* 4, buffer
);
933 info
->Buffers
[buffer
].Stride
= xfb_offset
;
936 /* From GL_EXT_transform_feedback:
937 * A program will fail to link if:
939 * * the total number of components to capture is greater than
940 * the constant MAX_TRANSFORM_FEEDBACK_INTERLEAVED_COMPONENTS_EXT
941 * and the buffer mode is INTERLEAVED_ATTRIBS_EXT.
943 * From GL_ARB_enhanced_layouts:
945 * "The resulting stride (implicit or explicit) must be less than or
946 * equal to the implementation-dependent constant
947 * gl_MaxTransformFeedbackInterleavedComponents."
949 if ((prog
->TransformFeedback
.BufferMode
== GL_INTERLEAVED_ATTRIBS
||
950 has_xfb_qualifiers
) &&
951 info
->Buffers
[buffer
].Stride
>
952 ctx
->Const
.MaxTransformFeedbackInterleavedComponents
) {
953 linker_error(prog
, "The MAX_TRANSFORM_FEEDBACK_INTERLEAVED_COMPONENTS "
954 "limit has been exceeded.");
959 info
->Varyings
[info
->NumVarying
].Name
= ralloc_strdup(prog
,
961 info
->Varyings
[info
->NumVarying
].Type
= this->type
;
962 info
->Varyings
[info
->NumVarying
].Size
= size
;
963 info
->Varyings
[info
->NumVarying
].BufferIndex
= buffer_index
;
965 info
->Buffers
[buffer
].NumVaryings
++;
971 const tfeedback_candidate
*
972 tfeedback_decl::find_candidate(gl_shader_program
*prog
,
973 hash_table
*tfeedback_candidates
)
975 const char *name
= this->var_name
;
976 switch (this->lowered_builtin_array_variable
) {
978 name
= this->var_name
;
981 name
= "gl_ClipDistanceMESA";
984 name
= "gl_CullDistanceMESA";
986 case tess_level_outer
:
987 name
= "gl_TessLevelOuterMESA";
989 case tess_level_inner
:
990 name
= "gl_TessLevelInnerMESA";
993 hash_entry
*entry
= _mesa_hash_table_search(tfeedback_candidates
, name
);
995 this->matched_candidate
= entry
?
996 (const tfeedback_candidate
*) entry
->data
: NULL
;
998 if (!this->matched_candidate
) {
999 /* From GL_EXT_transform_feedback:
1000 * A program will fail to link if:
1002 * * any variable name specified in the <varyings> array is not
1003 * declared as an output in the geometry shader (if present) or
1004 * the vertex shader (if no geometry shader is present);
1006 linker_error(prog
, "Transform feedback varying %s undeclared.",
1010 return this->matched_candidate
;
1015 * Parse all the transform feedback declarations that were passed to
1016 * glTransformFeedbackVaryings() and store them in tfeedback_decl objects.
1018 * If an error occurs, the error is reported through linker_error() and false
1022 parse_tfeedback_decls(struct gl_context
*ctx
, struct gl_shader_program
*prog
,
1023 const void *mem_ctx
, unsigned num_names
,
1024 char **varying_names
, tfeedback_decl
*decls
)
1026 for (unsigned i
= 0; i
< num_names
; ++i
) {
1027 decls
[i
].init(ctx
, mem_ctx
, varying_names
[i
]);
1029 if (!decls
[i
].is_varying())
1032 /* From GL_EXT_transform_feedback:
1033 * A program will fail to link if:
1035 * * any two entries in the <varyings> array specify the same varying
1038 * We interpret this to mean "any two entries in the <varyings> array
1039 * specify the same varying variable and array index", since transform
1040 * feedback of arrays would be useless otherwise.
1042 for (unsigned j
= 0; j
< i
; ++j
) {
1043 if (!decls
[j
].is_varying())
1046 if (tfeedback_decl::is_same(decls
[i
], decls
[j
])) {
1047 linker_error(prog
, "Transform feedback varying %s specified "
1048 "more than once.", varying_names
[i
]);
1058 cmp_xfb_offset(const void * x_generic
, const void * y_generic
)
1060 tfeedback_decl
*x
= (tfeedback_decl
*) x_generic
;
1061 tfeedback_decl
*y
= (tfeedback_decl
*) y_generic
;
1063 if (x
->get_buffer() != y
->get_buffer())
1064 return x
->get_buffer() - y
->get_buffer();
1065 return x
->get_offset() - y
->get_offset();
1069 * Store transform feedback location assignments into
1070 * prog->sh.LinkedTransformFeedback based on the data stored in
1073 * If an error occurs, the error is reported through linker_error() and false
1077 store_tfeedback_info(struct gl_context
*ctx
, struct gl_shader_program
*prog
,
1078 unsigned num_tfeedback_decls
,
1079 tfeedback_decl
*tfeedback_decls
, bool has_xfb_qualifiers
)
1081 if (!prog
->last_vert_prog
)
1084 /* Make sure MaxTransformFeedbackBuffers is less than 32 so the bitmask for
1085 * tracking the number of buffers doesn't overflow.
1087 assert(ctx
->Const
.MaxTransformFeedbackBuffers
< 32);
1089 bool separate_attribs_mode
=
1090 prog
->TransformFeedback
.BufferMode
== GL_SEPARATE_ATTRIBS
;
1092 struct gl_program
*xfb_prog
= prog
->last_vert_prog
;
1093 xfb_prog
->sh
.LinkedTransformFeedback
=
1094 rzalloc(xfb_prog
, struct gl_transform_feedback_info
);
1096 /* The xfb_offset qualifier does not have to be used in increasing order
1097 * however some drivers expect to receive the list of transform feedback
1098 * declarations in order so sort it now for convenience.
1100 if (has_xfb_qualifiers
)
1101 qsort(tfeedback_decls
, num_tfeedback_decls
, sizeof(*tfeedback_decls
),
1104 xfb_prog
->sh
.LinkedTransformFeedback
->Varyings
=
1105 rzalloc_array(xfb_prog
, struct gl_transform_feedback_varying_info
,
1106 num_tfeedback_decls
);
1108 unsigned num_outputs
= 0;
1109 for (unsigned i
= 0; i
< num_tfeedback_decls
; ++i
) {
1110 if (tfeedback_decls
[i
].is_varying_written())
1111 num_outputs
+= tfeedback_decls
[i
].get_num_outputs();
1114 xfb_prog
->sh
.LinkedTransformFeedback
->Outputs
=
1115 rzalloc_array(xfb_prog
, struct gl_transform_feedback_output
,
1118 unsigned num_buffers
= 0;
1119 unsigned buffers
= 0;
1121 if (!has_xfb_qualifiers
&& separate_attribs_mode
) {
1122 /* GL_SEPARATE_ATTRIBS */
1123 for (unsigned i
= 0; i
< num_tfeedback_decls
; ++i
) {
1124 if (!tfeedback_decls
[i
].store(ctx
, prog
,
1125 xfb_prog
->sh
.LinkedTransformFeedback
,
1126 num_buffers
, num_buffers
, num_outputs
,
1127 NULL
, has_xfb_qualifiers
))
1130 buffers
|= 1 << num_buffers
;
1135 /* GL_INVERLEAVED_ATTRIBS */
1136 int buffer_stream_id
= -1;
1138 num_tfeedback_decls
? tfeedback_decls
[0].get_buffer() : 0;
1139 bool explicit_stride
[MAX_FEEDBACK_BUFFERS
] = { false };
1141 /* Apply any xfb_stride global qualifiers */
1142 if (has_xfb_qualifiers
) {
1143 for (unsigned j
= 0; j
< MAX_FEEDBACK_BUFFERS
; j
++) {
1144 if (prog
->TransformFeedback
.BufferStride
[j
]) {
1146 explicit_stride
[j
] = true;
1147 xfb_prog
->sh
.LinkedTransformFeedback
->Buffers
[j
].Stride
=
1148 prog
->TransformFeedback
.BufferStride
[j
] / 4;
1153 for (unsigned i
= 0; i
< num_tfeedback_decls
; ++i
) {
1154 if (has_xfb_qualifiers
&&
1155 buffer
!= tfeedback_decls
[i
].get_buffer()) {
1156 /* we have moved to the next buffer so reset stream id */
1157 buffer_stream_id
= -1;
1161 if (tfeedback_decls
[i
].is_next_buffer_separator()) {
1162 if (!tfeedback_decls
[i
].store(ctx
, prog
,
1163 xfb_prog
->sh
.LinkedTransformFeedback
,
1164 buffer
, num_buffers
, num_outputs
,
1165 explicit_stride
, has_xfb_qualifiers
))
1168 buffer_stream_id
= -1;
1170 } else if (tfeedback_decls
[i
].is_varying()) {
1171 if (buffer_stream_id
== -1) {
1172 /* First varying writing to this buffer: remember its stream */
1173 buffer_stream_id
= (int) tfeedback_decls
[i
].get_stream_id();
1174 } else if (buffer_stream_id
!=
1175 (int) tfeedback_decls
[i
].get_stream_id()) {
1176 /* Varying writes to the same buffer from a different stream */
1178 "Transform feedback can't capture varyings belonging "
1179 "to different vertex streams in a single buffer. "
1180 "Varying %s writes to buffer from stream %u, other "
1181 "varyings in the same buffer write from stream %u.",
1182 tfeedback_decls
[i
].name(),
1183 tfeedback_decls
[i
].get_stream_id(),
1189 if (has_xfb_qualifiers
) {
1190 buffer
= tfeedback_decls
[i
].get_buffer();
1192 buffer
= num_buffers
;
1194 buffers
|= 1 << buffer
;
1196 if (!tfeedback_decls
[i
].store(ctx
, prog
,
1197 xfb_prog
->sh
.LinkedTransformFeedback
,
1198 buffer
, num_buffers
, num_outputs
,
1199 explicit_stride
, has_xfb_qualifiers
))
1204 assert(xfb_prog
->sh
.LinkedTransformFeedback
->NumOutputs
== num_outputs
);
1206 xfb_prog
->sh
.LinkedTransformFeedback
->ActiveBuffers
= buffers
;
1213 * Data structure recording the relationship between outputs of one shader
1214 * stage (the "producer") and inputs of another (the "consumer").
1216 class varying_matches
1219 varying_matches(bool disable_varying_packing
, bool xfb_enabled
,
1220 gl_shader_stage producer_stage
,
1221 gl_shader_stage consumer_stage
);
1223 void record(ir_variable
*producer_var
, ir_variable
*consumer_var
);
1224 unsigned assign_locations(struct gl_shader_program
*prog
,
1225 uint8_t *components
,
1226 uint64_t reserved_slots
);
1227 void store_locations() const;
1230 bool is_varying_packing_safe(const glsl_type
*type
,
1231 const ir_variable
*var
);
1234 * If true, this driver disables varying packing, so all varyings need to
1235 * be aligned on slot boundaries, and take up a number of slots equal to
1236 * their number of matrix columns times their array size.
1238 * Packing may also be disabled because our current packing method is not
1239 * safe in SSO or versions of OpenGL where interpolation qualifiers are not
1240 * guaranteed to match across stages.
1242 const bool disable_varying_packing
;
1245 * If true, this driver has transform feedback enabled. The transform
1246 * feedback code requires at least some packing be done even when varying
1247 * packing is disabled, fortunately where transform feedback requires
1248 * packing it's safe to override the disabled setting. See
1249 * is_varying_packing_safe().
1251 const bool xfb_enabled
;
1254 * Enum representing the order in which varyings are packed within a
1257 * Currently we pack vec4's first, then vec2's, then scalar values, then
1258 * vec3's. This order ensures that the only vectors that are at risk of
1259 * having to be "double parked" (split between two adjacent varying slots)
1262 enum packing_order_enum
{
1265 PACKING_ORDER_SCALAR
,
1269 static unsigned compute_packing_class(const ir_variable
*var
);
1270 static packing_order_enum
compute_packing_order(const ir_variable
*var
);
1271 static int match_comparator(const void *x_generic
, const void *y_generic
);
1272 static int xfb_comparator(const void *x_generic
, const void *y_generic
);
1275 * Structure recording the relationship between a single producer output
1276 * and a single consumer input.
1280 * Packing class for this varying, computed by compute_packing_class().
1282 unsigned packing_class
;
1285 * Packing order for this varying, computed by compute_packing_order().
1287 packing_order_enum packing_order
;
1288 unsigned num_components
;
1291 * The output variable in the producer stage.
1293 ir_variable
*producer_var
;
1296 * The input variable in the consumer stage.
1298 ir_variable
*consumer_var
;
1301 * The location which has been assigned for this varying. This is
1302 * expressed in multiples of a float, with the first generic varying
1303 * (i.e. the one referred to by VARYING_SLOT_VAR0) represented by the
1306 unsigned generic_location
;
1310 * The number of elements in the \c matches array that are currently in
1313 unsigned num_matches
;
1316 * The number of elements that were set aside for the \c matches array when
1319 unsigned matches_capacity
;
1321 gl_shader_stage producer_stage
;
1322 gl_shader_stage consumer_stage
;
1325 } /* anonymous namespace */
1327 varying_matches::varying_matches(bool disable_varying_packing
,
1329 gl_shader_stage producer_stage
,
1330 gl_shader_stage consumer_stage
)
1331 : disable_varying_packing(disable_varying_packing
),
1332 xfb_enabled(xfb_enabled
),
1333 producer_stage(producer_stage
),
1334 consumer_stage(consumer_stage
)
1336 /* Note: this initial capacity is rather arbitrarily chosen to be large
1337 * enough for many cases without wasting an unreasonable amount of space.
1338 * varying_matches::record() will resize the array if there are more than
1339 * this number of varyings.
1341 this->matches_capacity
= 8;
1342 this->matches
= (match
*)
1343 malloc(sizeof(*this->matches
) * this->matches_capacity
);
1344 this->num_matches
= 0;
1348 varying_matches::~varying_matches()
1350 free(this->matches
);
1355 * Packing is always safe on individual arrays, structures, and matrices. It
1356 * is also safe if the varying is only used for transform feedback.
1359 varying_matches::is_varying_packing_safe(const glsl_type
*type
,
1360 const ir_variable
*var
)
1362 if (consumer_stage
== MESA_SHADER_TESS_EVAL
||
1363 consumer_stage
== MESA_SHADER_TESS_CTRL
||
1364 producer_stage
== MESA_SHADER_TESS_CTRL
)
1367 return xfb_enabled
&& (type
->is_array() || type
->is_record() ||
1368 type
->is_matrix() || var
->data
.is_xfb_only
);
1373 * Record the given producer/consumer variable pair in the list of variables
1374 * that should later be assigned locations.
1376 * It is permissible for \c consumer_var to be NULL (this happens if a
1377 * variable is output by the producer and consumed by transform feedback, but
1378 * not consumed by the consumer).
1380 * If \c producer_var has already been paired up with a consumer_var, or
1381 * producer_var is part of fixed pipeline functionality (and hence already has
1382 * a location assigned), this function has no effect.
1384 * Note: as a side effect this function may change the interpolation type of
1385 * \c producer_var, but only when the change couldn't possibly affect
1389 varying_matches::record(ir_variable
*producer_var
, ir_variable
*consumer_var
)
1391 assert(producer_var
!= NULL
|| consumer_var
!= NULL
);
1393 if ((producer_var
&& (!producer_var
->data
.is_unmatched_generic_inout
||
1394 producer_var
->data
.explicit_location
)) ||
1395 (consumer_var
&& (!consumer_var
->data
.is_unmatched_generic_inout
||
1396 consumer_var
->data
.explicit_location
))) {
1397 /* Either a location already exists for this variable (since it is part
1398 * of fixed functionality), or it has already been recorded as part of a
1404 bool needs_flat_qualifier
= consumer_var
== NULL
&&
1405 (producer_var
->type
->contains_integer() ||
1406 producer_var
->type
->contains_double());
1408 if (!disable_varying_packing
&&
1409 (needs_flat_qualifier
||
1410 (consumer_stage
!= -1 && consumer_stage
!= MESA_SHADER_FRAGMENT
))) {
1411 /* Since this varying is not being consumed by the fragment shader, its
1412 * interpolation type varying cannot possibly affect rendering.
1413 * Also, this variable is non-flat and is (or contains) an integer
1415 * If the consumer stage is unknown, don't modify the interpolation
1416 * type as it could affect rendering later with separate shaders.
1418 * lower_packed_varyings requires all integer varyings to flat,
1419 * regardless of where they appear. We can trivially satisfy that
1420 * requirement by changing the interpolation type to flat here.
1423 producer_var
->data
.centroid
= false;
1424 producer_var
->data
.sample
= false;
1425 producer_var
->data
.interpolation
= INTERP_MODE_FLAT
;
1429 consumer_var
->data
.centroid
= false;
1430 consumer_var
->data
.sample
= false;
1431 consumer_var
->data
.interpolation
= INTERP_MODE_FLAT
;
1435 if (this->num_matches
== this->matches_capacity
) {
1436 this->matches_capacity
*= 2;
1437 this->matches
= (match
*)
1438 realloc(this->matches
,
1439 sizeof(*this->matches
) * this->matches_capacity
);
1442 /* We must use the consumer to compute the packing class because in GL4.4+
1443 * there is no guarantee interpolation qualifiers will match across stages.
1445 * From Section 4.5 (Interpolation Qualifiers) of the GLSL 4.30 spec:
1447 * "The type and presence of interpolation qualifiers of variables with
1448 * the same name declared in all linked shaders for the same cross-stage
1449 * interface must match, otherwise the link command will fail.
1451 * When comparing an output from one stage to an input of a subsequent
1452 * stage, the input and output don't match if their interpolation
1453 * qualifiers (or lack thereof) are not the same."
1455 * This text was also in at least revison 7 of the 4.40 spec but is no
1456 * longer in revision 9 and not in the 4.50 spec.
1458 const ir_variable
*const var
= (consumer_var
!= NULL
)
1459 ? consumer_var
: producer_var
;
1460 const gl_shader_stage stage
= (consumer_var
!= NULL
)
1461 ? consumer_stage
: producer_stage
;
1462 const glsl_type
*type
= get_varying_type(var
, stage
);
1464 this->matches
[this->num_matches
].packing_class
1465 = this->compute_packing_class(var
);
1466 this->matches
[this->num_matches
].packing_order
1467 = this->compute_packing_order(var
);
1468 if (this->disable_varying_packing
&& !is_varying_packing_safe(type
, var
)) {
1469 unsigned slots
= type
->count_attribute_slots(false);
1470 this->matches
[this->num_matches
].num_components
= slots
* 4;
1472 this->matches
[this->num_matches
].num_components
1473 = type
->component_slots();
1475 this->matches
[this->num_matches
].producer_var
= producer_var
;
1476 this->matches
[this->num_matches
].consumer_var
= consumer_var
;
1477 this->num_matches
++;
1479 producer_var
->data
.is_unmatched_generic_inout
= 0;
1481 consumer_var
->data
.is_unmatched_generic_inout
= 0;
1486 * Choose locations for all of the variable matches that were previously
1487 * passed to varying_matches::record().
1490 varying_matches::assign_locations(struct gl_shader_program
*prog
,
1491 uint8_t *components
,
1492 uint64_t reserved_slots
)
1494 /* If packing has been disabled then we cannot safely sort the varyings by
1495 * class as it may mean we are using a version of OpenGL where
1496 * interpolation qualifiers are not guaranteed to be matching across
1497 * shaders, sorting in this case could result in mismatching shader
1499 * When packing is disabled the sort orders varyings used by transform
1500 * feedback first, but also depends on *undefined behaviour* of qsort to
1501 * reverse the order of the varyings. See: xfb_comparator().
1503 if (!this->disable_varying_packing
) {
1504 /* Sort varying matches into an order that makes them easy to pack. */
1505 qsort(this->matches
, this->num_matches
, sizeof(*this->matches
),
1506 &varying_matches::match_comparator
);
1508 /* Only sort varyings that are only used by transform feedback. */
1509 qsort(this->matches
, this->num_matches
, sizeof(*this->matches
),
1510 &varying_matches::xfb_comparator
);
1513 unsigned generic_location
= 0;
1514 unsigned generic_patch_location
= MAX_VARYING
*4;
1515 bool previous_var_xfb_only
= false;
1517 for (unsigned i
= 0; i
< this->num_matches
; i
++) {
1518 unsigned *location
= &generic_location
;
1520 const ir_variable
*var
;
1521 const glsl_type
*type
;
1522 bool is_vertex_input
= false;
1523 if (matches
[i
].consumer_var
) {
1524 var
= matches
[i
].consumer_var
;
1525 type
= get_varying_type(var
, consumer_stage
);
1526 if (consumer_stage
== MESA_SHADER_VERTEX
)
1527 is_vertex_input
= true;
1529 var
= matches
[i
].producer_var
;
1530 type
= get_varying_type(var
, producer_stage
);
1533 if (var
->data
.patch
)
1534 location
= &generic_patch_location
;
1536 /* Advance to the next slot if this varying has a different packing
1537 * class than the previous one, and we're not already on a slot
1540 * Also advance to the next slot if packing is disabled. This makes sure
1541 * we don't assign varyings the same locations which is possible
1542 * because we still pack individual arrays, records and matrices even
1543 * when packing is disabled. Note we don't advance to the next slot if
1544 * we can pack varyings together that are only used for transform
1547 if ((this->disable_varying_packing
&&
1548 !(previous_var_xfb_only
&& var
->data
.is_xfb_only
)) ||
1549 (i
> 0 && this->matches
[i
- 1].packing_class
1550 != this->matches
[i
].packing_class
)) {
1551 *location
= ALIGN(*location
, 4);
1554 previous_var_xfb_only
= var
->data
.is_xfb_only
;
1556 /* The number of components taken up by this variable. For vertex shader
1557 * inputs, we use the number of slots * 4, as they have different
1560 unsigned num_components
= is_vertex_input
?
1561 type
->count_attribute_slots(is_vertex_input
) * 4 :
1562 this->matches
[i
].num_components
;
1564 /* The last slot for this variable, inclusive. */
1565 unsigned slot_end
= *location
+ num_components
- 1;
1567 /* FIXME: We could be smarter in the below code and loop back over
1568 * trying to fill any locations that we skipped because we couldn't pack
1569 * the varying between an explicit location. For now just let the user
1570 * hit the linking error if we run out of room and suggest they use
1571 * explicit locations.
1573 while (slot_end
< MAX_VARYING
* 4u) {
1574 const unsigned slots
= (slot_end
/ 4u) - (*location
/ 4u) + 1;
1575 const uint64_t slot_mask
= ((1ull << slots
) - 1) << (*location
/ 4u);
1578 if (reserved_slots
& slot_mask
) {
1579 *location
= ALIGN(*location
+ 1, 4);
1580 slot_end
= *location
+ num_components
- 1;
1587 if (!var
->data
.patch
&& slot_end
>= MAX_VARYING
* 4u) {
1588 linker_error(prog
, "insufficient contiguous locations available for "
1589 "%s it is possible an array or struct could not be "
1590 "packed between varyings with explicit locations. Try "
1591 "using an explicit location for arrays and structs.",
1595 if (slot_end
< MAX_VARYINGS_INCL_PATCH
* 4u) {
1596 for (unsigned j
= *location
/ 4u; j
< slot_end
/ 4u; j
++)
1598 components
[slot_end
/ 4u] = (slot_end
& 3) + 1;
1601 this->matches
[i
].generic_location
= *location
;
1603 *location
= slot_end
+ 1;
1606 return (generic_location
+ 3) / 4;
1611 * Update the producer and consumer shaders to reflect the locations
1612 * assignments that were made by varying_matches::assign_locations().
1615 varying_matches::store_locations() const
1617 for (unsigned i
= 0; i
< this->num_matches
; i
++) {
1618 ir_variable
*producer_var
= this->matches
[i
].producer_var
;
1619 ir_variable
*consumer_var
= this->matches
[i
].consumer_var
;
1620 unsigned generic_location
= this->matches
[i
].generic_location
;
1621 unsigned slot
= generic_location
/ 4;
1622 unsigned offset
= generic_location
% 4;
1625 producer_var
->data
.location
= VARYING_SLOT_VAR0
+ slot
;
1626 producer_var
->data
.location_frac
= offset
;
1630 assert(consumer_var
->data
.location
== -1);
1631 consumer_var
->data
.location
= VARYING_SLOT_VAR0
+ slot
;
1632 consumer_var
->data
.location_frac
= offset
;
1639 * Compute the "packing class" of the given varying. This is an unsigned
1640 * integer with the property that two variables in the same packing class can
1641 * be safely backed into the same vec4.
1644 varying_matches::compute_packing_class(const ir_variable
*var
)
1646 /* Without help from the back-end, there is no way to pack together
1647 * variables with different interpolation types, because
1648 * lower_packed_varyings must choose exactly one interpolation type for
1649 * each packed varying it creates.
1651 * However, we can safely pack together floats, ints, and uints, because:
1653 * - varyings of base type "int" and "uint" must use the "flat"
1654 * interpolation type, which can only occur in GLSL 1.30 and above.
1656 * - On platforms that support GLSL 1.30 and above, lower_packed_varyings
1657 * can store flat floats as ints without losing any information (using
1658 * the ir_unop_bitcast_* opcodes).
1660 * Therefore, the packing class depends only on the interpolation type.
1662 unsigned packing_class
= var
->data
.centroid
| (var
->data
.sample
<< 1) |
1663 (var
->data
.patch
<< 2);
1665 packing_class
+= var
->is_interpolation_flat()
1666 ? unsigned(INTERP_MODE_FLAT
) : var
->data
.interpolation
;
1667 return packing_class
;
1672 * Compute the "packing order" of the given varying. This is a sort key we
1673 * use to determine when to attempt to pack the given varying relative to
1674 * other varyings in the same packing class.
1676 varying_matches::packing_order_enum
1677 varying_matches::compute_packing_order(const ir_variable
*var
)
1679 const glsl_type
*element_type
= var
->type
;
1681 while (element_type
->is_array()) {
1682 element_type
= element_type
->fields
.array
;
1685 switch (element_type
->component_slots() % 4) {
1686 case 1: return PACKING_ORDER_SCALAR
;
1687 case 2: return PACKING_ORDER_VEC2
;
1688 case 3: return PACKING_ORDER_VEC3
;
1689 case 0: return PACKING_ORDER_VEC4
;
1691 assert(!"Unexpected value of vector_elements");
1692 return PACKING_ORDER_VEC4
;
1698 * Comparison function passed to qsort() to sort varyings by packing_class and
1699 * then by packing_order.
1702 varying_matches::match_comparator(const void *x_generic
, const void *y_generic
)
1704 const match
*x
= (const match
*) x_generic
;
1705 const match
*y
= (const match
*) y_generic
;
1707 if (x
->packing_class
!= y
->packing_class
)
1708 return x
->packing_class
- y
->packing_class
;
1709 return x
->packing_order
- y
->packing_order
;
1714 * Comparison function passed to qsort() to sort varyings used only by
1715 * transform feedback when packing of other varyings is disabled.
1718 varying_matches::xfb_comparator(const void *x_generic
, const void *y_generic
)
1720 const match
*x
= (const match
*) x_generic
;
1722 if (x
->producer_var
!= NULL
&& x
->producer_var
->data
.is_xfb_only
)
1723 return match_comparator(x_generic
, y_generic
);
1725 /* FIXME: When the comparator returns 0 it means the elements being
1726 * compared are equivalent. However the qsort documentation says:
1728 * "The order of equivalent elements is undefined."
1730 * In practice the sort ends up reversing the order of the varyings which
1731 * means locations are also assigned in this reversed order and happens to
1732 * be what we want. This is also whats happening in
1733 * varying_matches::match_comparator().
1740 * Is the given variable a varying variable to be counted against the
1741 * limit in ctx->Const.MaxVarying?
1742 * This includes variables such as texcoords, colors and generic
1743 * varyings, but excludes variables such as gl_FrontFacing and gl_FragCoord.
1746 var_counts_against_varying_limit(gl_shader_stage stage
, const ir_variable
*var
)
1748 /* Only fragment shaders will take a varying variable as an input */
1749 if (stage
== MESA_SHADER_FRAGMENT
&&
1750 var
->data
.mode
== ir_var_shader_in
) {
1751 switch (var
->data
.location
) {
1752 case VARYING_SLOT_POS
:
1753 case VARYING_SLOT_FACE
:
1754 case VARYING_SLOT_PNTC
:
1765 * Visitor class that generates tfeedback_candidate structs describing all
1766 * possible targets of transform feedback.
1768 * tfeedback_candidate structs are stored in the hash table
1769 * tfeedback_candidates, which is passed to the constructor. This hash table
1770 * maps varying names to instances of the tfeedback_candidate struct.
1772 class tfeedback_candidate_generator
: public program_resource_visitor
1775 tfeedback_candidate_generator(void *mem_ctx
,
1776 hash_table
*tfeedback_candidates
)
1778 tfeedback_candidates(tfeedback_candidates
),
1784 void process(ir_variable
*var
)
1786 /* All named varying interface blocks should be flattened by now */
1787 assert(!var
->is_interface_instance());
1789 this->toplevel_var
= var
;
1790 this->varying_floats
= 0;
1791 program_resource_visitor::process(var
);
1795 virtual void visit_field(const glsl_type
*type
, const char *name
,
1796 bool /* row_major */,
1797 const glsl_type
* /* record_type */,
1798 const enum glsl_interface_packing
,
1799 bool /* last_field */)
1801 assert(!type
->without_array()->is_record());
1802 assert(!type
->without_array()->is_interface());
1804 tfeedback_candidate
*candidate
1805 = rzalloc(this->mem_ctx
, tfeedback_candidate
);
1806 candidate
->toplevel_var
= this->toplevel_var
;
1807 candidate
->type
= type
;
1808 candidate
->offset
= this->varying_floats
;
1809 _mesa_hash_table_insert(this->tfeedback_candidates
,
1810 ralloc_strdup(this->mem_ctx
, name
),
1812 this->varying_floats
+= type
->component_slots();
1816 * Memory context used to allocate hash table keys and values.
1818 void * const mem_ctx
;
1821 * Hash table in which tfeedback_candidate objects should be stored.
1823 hash_table
* const tfeedback_candidates
;
1826 * Pointer to the toplevel variable that is being traversed.
1828 ir_variable
*toplevel_var
;
1831 * Total number of varying floats that have been visited so far. This is
1832 * used to determine the offset to each varying within the toplevel
1835 unsigned varying_floats
;
1842 populate_consumer_input_sets(void *mem_ctx
, exec_list
*ir
,
1843 hash_table
*consumer_inputs
,
1844 hash_table
*consumer_interface_inputs
,
1845 ir_variable
*consumer_inputs_with_locations
[VARYING_SLOT_TESS_MAX
])
1847 memset(consumer_inputs_with_locations
,
1849 sizeof(consumer_inputs_with_locations
[0]) * VARYING_SLOT_TESS_MAX
);
1851 foreach_in_list(ir_instruction
, node
, ir
) {
1852 ir_variable
*const input_var
= node
->as_variable();
1854 if (input_var
!= NULL
&& input_var
->data
.mode
== ir_var_shader_in
) {
1855 /* All interface blocks should have been lowered by this point */
1856 assert(!input_var
->type
->is_interface());
1858 if (input_var
->data
.explicit_location
) {
1859 /* assign_varying_locations only cares about finding the
1860 * ir_variable at the start of a contiguous location block.
1862 * - For !producer, consumer_inputs_with_locations isn't used.
1864 * - For !consumer, consumer_inputs_with_locations is empty.
1866 * For consumer && producer, if you were trying to set some
1867 * ir_variable to the middle of a location block on the other side
1868 * of producer/consumer, cross_validate_outputs_to_inputs() should
1869 * be link-erroring due to either type mismatch or location
1870 * overlaps. If the variables do match up, then they've got a
1871 * matching data.location and you only looked at
1872 * consumer_inputs_with_locations[var->data.location], not any
1873 * following entries for the array/structure.
1875 consumer_inputs_with_locations
[input_var
->data
.location
] =
1877 } else if (input_var
->get_interface_type() != NULL
) {
1878 char *const iface_field_name
=
1879 ralloc_asprintf(mem_ctx
, "%s.%s",
1880 input_var
->get_interface_type()->without_array()->name
,
1882 _mesa_hash_table_insert(consumer_interface_inputs
,
1883 iface_field_name
, input_var
);
1885 _mesa_hash_table_insert(consumer_inputs
,
1886 ralloc_strdup(mem_ctx
, input_var
->name
),
1894 * Find a variable from the consumer that "matches" the specified variable
1896 * This function only finds inputs with names that match. There is no
1897 * validation (here) that the types, etc. are compatible.
1900 get_matching_input(void *mem_ctx
,
1901 const ir_variable
*output_var
,
1902 hash_table
*consumer_inputs
,
1903 hash_table
*consumer_interface_inputs
,
1904 ir_variable
*consumer_inputs_with_locations
[VARYING_SLOT_TESS_MAX
])
1906 ir_variable
*input_var
;
1908 if (output_var
->data
.explicit_location
) {
1909 input_var
= consumer_inputs_with_locations
[output_var
->data
.location
];
1910 } else if (output_var
->get_interface_type() != NULL
) {
1911 char *const iface_field_name
=
1912 ralloc_asprintf(mem_ctx
, "%s.%s",
1913 output_var
->get_interface_type()->without_array()->name
,
1915 hash_entry
*entry
= _mesa_hash_table_search(consumer_interface_inputs
, iface_field_name
);
1916 input_var
= entry
? (ir_variable
*) entry
->data
: NULL
;
1918 hash_entry
*entry
= _mesa_hash_table_search(consumer_inputs
, output_var
->name
);
1919 input_var
= entry
? (ir_variable
*) entry
->data
: NULL
;
1922 return (input_var
== NULL
|| input_var
->data
.mode
!= ir_var_shader_in
)
1929 io_variable_cmp(const void *_a
, const void *_b
)
1931 const ir_variable
*const a
= *(const ir_variable
**) _a
;
1932 const ir_variable
*const b
= *(const ir_variable
**) _b
;
1934 if (a
->data
.explicit_location
&& b
->data
.explicit_location
)
1935 return b
->data
.location
- a
->data
.location
;
1937 if (a
->data
.explicit_location
&& !b
->data
.explicit_location
)
1940 if (!a
->data
.explicit_location
&& b
->data
.explicit_location
)
1943 return -strcmp(a
->name
, b
->name
);
1947 * Sort the shader IO variables into canonical order
1950 canonicalize_shader_io(exec_list
*ir
, enum ir_variable_mode io_mode
)
1952 ir_variable
*var_table
[MAX_PROGRAM_OUTPUTS
* 4];
1953 unsigned num_variables
= 0;
1955 foreach_in_list(ir_instruction
, node
, ir
) {
1956 ir_variable
*const var
= node
->as_variable();
1958 if (var
== NULL
|| var
->data
.mode
!= io_mode
)
1961 /* If we have already encountered more I/O variables that could
1962 * successfully link, bail.
1964 if (num_variables
== ARRAY_SIZE(var_table
))
1967 var_table
[num_variables
++] = var
;
1970 if (num_variables
== 0)
1973 /* Sort the list in reverse order (io_variable_cmp handles this). Later
1974 * we're going to push the variables on to the IR list as a stack, so we
1975 * want the last variable (in canonical order) to be first in the list.
1977 qsort(var_table
, num_variables
, sizeof(var_table
[0]), io_variable_cmp
);
1979 /* Remove the variable from it's current location in the IR, and put it at
1982 for (unsigned i
= 0; i
< num_variables
; i
++) {
1983 var_table
[i
]->remove();
1984 ir
->push_head(var_table
[i
]);
1989 * Generate a bitfield map of the explicit locations for shader varyings.
1991 * Note: For Tessellation shaders we are sitting right on the limits of the
1992 * 64 bit map. Per-vertex and per-patch both have separate location domains
1993 * with a max of MAX_VARYING.
1996 reserved_varying_slot(struct gl_linked_shader
*stage
,
1997 ir_variable_mode io_mode
)
1999 assert(io_mode
== ir_var_shader_in
|| io_mode
== ir_var_shader_out
);
2000 /* Avoid an overflow of the returned value */
2001 assert(MAX_VARYINGS_INCL_PATCH
<= 64);
2009 foreach_in_list(ir_instruction
, node
, stage
->ir
) {
2010 ir_variable
*const var
= node
->as_variable();
2012 if (var
== NULL
|| var
->data
.mode
!= io_mode
||
2013 !var
->data
.explicit_location
||
2014 var
->data
.location
< VARYING_SLOT_VAR0
)
2017 var_slot
= var
->data
.location
- VARYING_SLOT_VAR0
;
2019 unsigned num_elements
= get_varying_type(var
, stage
->Stage
)
2020 ->count_attribute_slots(stage
->Stage
== MESA_SHADER_VERTEX
);
2021 for (unsigned i
= 0; i
< num_elements
; i
++) {
2022 if (var_slot
>= 0 && var_slot
< MAX_VARYINGS_INCL_PATCH
)
2023 slots
|= UINT64_C(1) << var_slot
;
2033 * Assign locations for all variables that are produced in one pipeline stage
2034 * (the "producer") and consumed in the next stage (the "consumer").
2036 * Variables produced by the producer may also be consumed by transform
2039 * \param num_tfeedback_decls is the number of declarations indicating
2040 * variables that may be consumed by transform feedback.
2042 * \param tfeedback_decls is a pointer to an array of tfeedback_decl objects
2043 * representing the result of parsing the strings passed to
2044 * glTransformFeedbackVaryings(). assign_location() will be called for
2045 * each of these objects that matches one of the outputs of the
2048 * When num_tfeedback_decls is nonzero, it is permissible for the consumer to
2049 * be NULL. In this case, varying locations are assigned solely based on the
2050 * requirements of transform feedback.
2053 assign_varying_locations(struct gl_context
*ctx
,
2055 struct gl_shader_program
*prog
,
2056 gl_linked_shader
*producer
,
2057 gl_linked_shader
*consumer
,
2058 unsigned num_tfeedback_decls
,
2059 tfeedback_decl
*tfeedback_decls
,
2060 const uint64_t reserved_slots
)
2062 /* Tessellation shaders treat inputs and outputs as shared memory and can
2063 * access inputs and outputs of other invocations.
2064 * Therefore, they can't be lowered to temps easily (and definitely not
2067 bool unpackable_tess
=
2068 (consumer
&& consumer
->Stage
== MESA_SHADER_TESS_EVAL
) ||
2069 (consumer
&& consumer
->Stage
== MESA_SHADER_TESS_CTRL
) ||
2070 (producer
&& producer
->Stage
== MESA_SHADER_TESS_CTRL
);
2072 /* Transform feedback code assumes varying arrays are packed, so if the
2073 * driver has disabled varying packing, make sure to at least enable
2074 * packing required by transform feedback.
2077 ctx
->Extensions
.EXT_transform_feedback
&& !unpackable_tess
;
2079 /* Disable packing on outward facing interfaces for SSO because in ES we
2080 * need to retain the unpacked varying information for draw time
2083 * Packing is still enabled on individual arrays, structs, and matrices as
2084 * these are required by the transform feedback code and it is still safe
2085 * to do so. We also enable packing when a varying is only used for
2086 * transform feedback and its not a SSO.
2088 bool disable_varying_packing
=
2089 ctx
->Const
.DisableVaryingPacking
|| unpackable_tess
;
2090 if (prog
->SeparateShader
&& (producer
== NULL
|| consumer
== NULL
))
2091 disable_varying_packing
= true;
2093 varying_matches
matches(disable_varying_packing
, xfb_enabled
,
2094 producer
? producer
->Stage
: (gl_shader_stage
)-1,
2095 consumer
? consumer
->Stage
: (gl_shader_stage
)-1);
2096 hash_table
*tfeedback_candidates
=
2097 _mesa_hash_table_create(NULL
, _mesa_key_hash_string
,
2098 _mesa_key_string_equal
);
2099 hash_table
*consumer_inputs
=
2100 _mesa_hash_table_create(NULL
, _mesa_key_hash_string
,
2101 _mesa_key_string_equal
);
2102 hash_table
*consumer_interface_inputs
=
2103 _mesa_hash_table_create(NULL
, _mesa_key_hash_string
,
2104 _mesa_key_string_equal
);
2105 ir_variable
*consumer_inputs_with_locations
[VARYING_SLOT_TESS_MAX
] = {
2109 unsigned consumer_vertices
= 0;
2110 if (consumer
&& consumer
->Stage
== MESA_SHADER_GEOMETRY
)
2111 consumer_vertices
= prog
->Geom
.VerticesIn
;
2113 /* Operate in a total of four passes.
2115 * 1. Sort inputs / outputs into a canonical order. This is necessary so
2116 * that inputs / outputs of separable shaders will be assigned
2117 * predictable locations regardless of the order in which declarations
2118 * appeared in the shader source.
2120 * 2. Assign locations for any matching inputs and outputs.
2122 * 3. Mark output variables in the producer that do not have locations as
2123 * not being outputs. This lets the optimizer eliminate them.
2125 * 4. Mark input variables in the consumer that do not have locations as
2126 * not being inputs. This lets the optimizer eliminate them.
2129 canonicalize_shader_io(consumer
->ir
, ir_var_shader_in
);
2132 canonicalize_shader_io(producer
->ir
, ir_var_shader_out
);
2135 linker::populate_consumer_input_sets(mem_ctx
, consumer
->ir
,
2137 consumer_interface_inputs
,
2138 consumer_inputs_with_locations
);
2141 foreach_in_list(ir_instruction
, node
, producer
->ir
) {
2142 ir_variable
*const output_var
= node
->as_variable();
2144 if (output_var
== NULL
|| output_var
->data
.mode
!= ir_var_shader_out
)
2147 /* Only geometry shaders can use non-zero streams */
2148 assert(output_var
->data
.stream
== 0 ||
2149 (output_var
->data
.stream
< MAX_VERTEX_STREAMS
&&
2150 producer
->Stage
== MESA_SHADER_GEOMETRY
));
2152 if (num_tfeedback_decls
> 0) {
2153 tfeedback_candidate_generator
g(mem_ctx
, tfeedback_candidates
);
2154 g
.process(output_var
);
2157 ir_variable
*const input_var
=
2158 linker::get_matching_input(mem_ctx
, output_var
, consumer_inputs
,
2159 consumer_interface_inputs
,
2160 consumer_inputs_with_locations
);
2162 /* If a matching input variable was found, add this output (and the
2163 * input) to the set. If this is a separable program and there is no
2164 * consumer stage, add the output.
2166 * Always add TCS outputs. They are shared by all invocations
2167 * within a patch and can be used as shared memory.
2169 if (input_var
|| (prog
->SeparateShader
&& consumer
== NULL
) ||
2170 producer
->Stage
== MESA_SHADER_TESS_CTRL
) {
2171 matches
.record(output_var
, input_var
);
2174 /* Only stream 0 outputs can be consumed in the next stage */
2175 if (input_var
&& output_var
->data
.stream
!= 0) {
2176 linker_error(prog
, "output %s is assigned to stream=%d but "
2177 "is linked to an input, which requires stream=0",
2178 output_var
->name
, output_var
->data
.stream
);
2183 /* If there's no producer stage, then this must be a separable program.
2184 * For example, we may have a program that has just a fragment shader.
2185 * Later this program will be used with some arbitrary vertex (or
2186 * geometry) shader program. This means that locations must be assigned
2187 * for all the inputs.
2189 foreach_in_list(ir_instruction
, node
, consumer
->ir
) {
2190 ir_variable
*const input_var
= node
->as_variable();
2192 if (input_var
== NULL
|| input_var
->data
.mode
!= ir_var_shader_in
)
2195 matches
.record(NULL
, input_var
);
2199 _mesa_hash_table_destroy(consumer_inputs
, NULL
);
2200 _mesa_hash_table_destroy(consumer_interface_inputs
, NULL
);
2202 for (unsigned i
= 0; i
< num_tfeedback_decls
; ++i
) {
2203 if (!tfeedback_decls
[i
].is_varying())
2206 const tfeedback_candidate
*matched_candidate
2207 = tfeedback_decls
[i
].find_candidate(prog
, tfeedback_candidates
);
2209 if (matched_candidate
== NULL
) {
2210 _mesa_hash_table_destroy(tfeedback_candidates
, NULL
);
2214 if (matched_candidate
->toplevel_var
->data
.is_unmatched_generic_inout
) {
2215 matched_candidate
->toplevel_var
->data
.is_xfb_only
= 1;
2216 matches
.record(matched_candidate
->toplevel_var
, NULL
);
2220 uint8_t components
[MAX_VARYINGS_INCL_PATCH
] = {0};
2221 const unsigned slots_used
= matches
.assign_locations(
2222 prog
, components
, reserved_slots
);
2223 matches
.store_locations();
2225 for (unsigned i
= 0; i
< num_tfeedback_decls
; ++i
) {
2226 if (!tfeedback_decls
[i
].is_varying())
2229 if (!tfeedback_decls
[i
].assign_location(ctx
, prog
)) {
2230 _mesa_hash_table_destroy(tfeedback_candidates
, NULL
);
2234 _mesa_hash_table_destroy(tfeedback_candidates
, NULL
);
2236 if (consumer
&& producer
) {
2237 foreach_in_list(ir_instruction
, node
, consumer
->ir
) {
2238 ir_variable
*const var
= node
->as_variable();
2240 if (var
&& var
->data
.mode
== ir_var_shader_in
&&
2241 var
->data
.is_unmatched_generic_inout
) {
2242 if (!prog
->IsES
&& prog
->data
->Version
<= 120) {
2243 /* On page 25 (page 31 of the PDF) of the GLSL 1.20 spec:
2245 * Only those varying variables used (i.e. read) in
2246 * the fragment shader executable must be written to
2247 * by the vertex shader executable; declaring
2248 * superfluous varying variables in a vertex shader is
2251 * We interpret this text as meaning that the VS must
2252 * write the variable for the FS to read it. See
2253 * "glsl1-varying read but not written" in piglit.
2255 linker_error(prog
, "%s shader varying %s not written "
2257 _mesa_shader_stage_to_string(consumer
->Stage
),
2259 _mesa_shader_stage_to_string(producer
->Stage
));
2261 linker_warning(prog
, "%s shader varying %s not written "
2263 _mesa_shader_stage_to_string(consumer
->Stage
),
2265 _mesa_shader_stage_to_string(producer
->Stage
));
2270 /* Now that validation is done its safe to remove unused varyings. As
2271 * we have both a producer and consumer its safe to remove unused
2272 * varyings even if the program is a SSO because the stages are being
2273 * linked together i.e. we have a multi-stage SSO.
2275 remove_unused_shader_inputs_and_outputs(false, producer
,
2277 remove_unused_shader_inputs_and_outputs(false, consumer
,
2282 lower_packed_varyings(mem_ctx
, slots_used
, components
, ir_var_shader_out
,
2283 0, producer
, disable_varying_packing
,
2288 lower_packed_varyings(mem_ctx
, slots_used
, components
, ir_var_shader_in
,
2289 consumer_vertices
, consumer
,
2290 disable_varying_packing
, xfb_enabled
);
2297 check_against_output_limit(struct gl_context
*ctx
,
2298 struct gl_shader_program
*prog
,
2299 gl_linked_shader
*producer
,
2300 unsigned num_explicit_locations
)
2302 unsigned output_vectors
= num_explicit_locations
;
2304 foreach_in_list(ir_instruction
, node
, producer
->ir
) {
2305 ir_variable
*const var
= node
->as_variable();
2307 if (var
&& !var
->data
.explicit_location
&&
2308 var
->data
.mode
== ir_var_shader_out
&&
2309 var_counts_against_varying_limit(producer
->Stage
, var
)) {
2310 /* outputs for fragment shader can't be doubles */
2311 output_vectors
+= var
->type
->count_attribute_slots(false);
2315 assert(producer
->Stage
!= MESA_SHADER_FRAGMENT
);
2316 unsigned max_output_components
=
2317 ctx
->Const
.Program
[producer
->Stage
].MaxOutputComponents
;
2319 const unsigned output_components
= output_vectors
* 4;
2320 if (output_components
> max_output_components
) {
2321 if (ctx
->API
== API_OPENGLES2
|| prog
->IsES
)
2322 linker_error(prog
, "%s shader uses too many output vectors "
2324 _mesa_shader_stage_to_string(producer
->Stage
),
2326 max_output_components
/ 4);
2328 linker_error(prog
, "%s shader uses too many output components "
2330 _mesa_shader_stage_to_string(producer
->Stage
),
2332 max_output_components
);
2341 check_against_input_limit(struct gl_context
*ctx
,
2342 struct gl_shader_program
*prog
,
2343 gl_linked_shader
*consumer
,
2344 unsigned num_explicit_locations
)
2346 unsigned input_vectors
= num_explicit_locations
;
2348 foreach_in_list(ir_instruction
, node
, consumer
->ir
) {
2349 ir_variable
*const var
= node
->as_variable();
2351 if (var
&& !var
->data
.explicit_location
&&
2352 var
->data
.mode
== ir_var_shader_in
&&
2353 var_counts_against_varying_limit(consumer
->Stage
, var
)) {
2354 /* vertex inputs aren't varying counted */
2355 input_vectors
+= var
->type
->count_attribute_slots(false);
2359 assert(consumer
->Stage
!= MESA_SHADER_VERTEX
);
2360 unsigned max_input_components
=
2361 ctx
->Const
.Program
[consumer
->Stage
].MaxInputComponents
;
2363 const unsigned input_components
= input_vectors
* 4;
2364 if (input_components
> max_input_components
) {
2365 if (ctx
->API
== API_OPENGLES2
|| prog
->IsES
)
2366 linker_error(prog
, "%s shader uses too many input vectors "
2368 _mesa_shader_stage_to_string(consumer
->Stage
),
2370 max_input_components
/ 4);
2372 linker_error(prog
, "%s shader uses too many input components "
2374 _mesa_shader_stage_to_string(consumer
->Stage
),
2376 max_input_components
);
2385 link_varyings(struct gl_shader_program
*prog
, unsigned first
, unsigned last
,
2386 struct gl_context
*ctx
, void *mem_ctx
)
2388 bool has_xfb_qualifiers
= false;
2389 unsigned num_tfeedback_decls
= 0;
2390 char **varying_names
= NULL
;
2391 tfeedback_decl
*tfeedback_decls
= NULL
;
2393 /* From the ARB_enhanced_layouts spec:
2395 * "If the shader used to record output variables for transform feedback
2396 * varyings uses the "xfb_buffer", "xfb_offset", or "xfb_stride" layout
2397 * qualifiers, the values specified by TransformFeedbackVaryings are
2398 * ignored, and the set of variables captured for transform feedback is
2399 * instead derived from the specified layout qualifiers."
2401 for (int i
= MESA_SHADER_FRAGMENT
- 1; i
>= 0; i
--) {
2402 /* Find last stage before fragment shader */
2403 if (prog
->_LinkedShaders
[i
]) {
2404 has_xfb_qualifiers
=
2405 process_xfb_layout_qualifiers(mem_ctx
, prog
->_LinkedShaders
[i
],
2406 prog
, &num_tfeedback_decls
,
2412 if (!has_xfb_qualifiers
) {
2413 num_tfeedback_decls
= prog
->TransformFeedback
.NumVarying
;
2414 varying_names
= prog
->TransformFeedback
.VaryingNames
;
2417 if (num_tfeedback_decls
!= 0) {
2418 /* From GL_EXT_transform_feedback:
2419 * A program will fail to link if:
2421 * * the <count> specified by TransformFeedbackVaryingsEXT is
2422 * non-zero, but the program object has no vertex or geometry
2425 if (first
>= MESA_SHADER_FRAGMENT
) {
2426 linker_error(prog
, "Transform feedback varyings specified, but "
2427 "no vertex, tessellation, or geometry shader is "
2432 tfeedback_decls
= rzalloc_array(mem_ctx
, tfeedback_decl
,
2433 num_tfeedback_decls
);
2434 if (!parse_tfeedback_decls(ctx
, prog
, mem_ctx
, num_tfeedback_decls
,
2435 varying_names
, tfeedback_decls
))
2439 /* If there is no fragment shader we need to set transform feedback.
2441 * For SSO we also need to assign output locations. We assign them here
2442 * because we need to do it for both single stage programs and multi stage
2445 if (last
< MESA_SHADER_FRAGMENT
&&
2446 (num_tfeedback_decls
!= 0 || prog
->SeparateShader
)) {
2447 const uint64_t reserved_out_slots
=
2448 reserved_varying_slot(prog
->_LinkedShaders
[last
], ir_var_shader_out
);
2449 if (!assign_varying_locations(ctx
, mem_ctx
, prog
,
2450 prog
->_LinkedShaders
[last
], NULL
,
2451 num_tfeedback_decls
, tfeedback_decls
,
2452 reserved_out_slots
))
2456 if (last
<= MESA_SHADER_FRAGMENT
) {
2457 /* Remove unused varyings from the first/last stage unless SSO */
2458 remove_unused_shader_inputs_and_outputs(prog
->SeparateShader
,
2459 prog
->_LinkedShaders
[first
],
2461 remove_unused_shader_inputs_and_outputs(prog
->SeparateShader
,
2462 prog
->_LinkedShaders
[last
],
2465 /* If the program is made up of only a single stage */
2466 if (first
== last
) {
2467 gl_linked_shader
*const sh
= prog
->_LinkedShaders
[last
];
2469 do_dead_builtin_varyings(ctx
, NULL
, sh
, 0, NULL
);
2470 do_dead_builtin_varyings(ctx
, sh
, NULL
, num_tfeedback_decls
,
2473 if (prog
->SeparateShader
) {
2474 const uint64_t reserved_slots
=
2475 reserved_varying_slot(sh
, ir_var_shader_in
);
2477 /* Assign input locations for SSO, output locations are already
2480 if (!assign_varying_locations(ctx
, mem_ctx
, prog
,
2481 NULL
/* producer */,
2483 0 /* num_tfeedback_decls */,
2484 NULL
/* tfeedback_decls */,
2489 /* Linking the stages in the opposite order (from fragment to vertex)
2490 * ensures that inter-shader outputs written to in an earlier stage
2491 * are eliminated if they are (transitively) not used in a later
2495 for (int i
= next
- 1; i
>= 0; i
--) {
2496 if (prog
->_LinkedShaders
[i
] == NULL
&& i
!= 0)
2499 gl_linked_shader
*const sh_i
= prog
->_LinkedShaders
[i
];
2500 gl_linked_shader
*const sh_next
= prog
->_LinkedShaders
[next
];
2502 const uint64_t reserved_out_slots
=
2503 reserved_varying_slot(sh_i
, ir_var_shader_out
);
2504 const uint64_t reserved_in_slots
=
2505 reserved_varying_slot(sh_next
, ir_var_shader_in
);
2507 do_dead_builtin_varyings(ctx
, sh_i
, sh_next
,
2508 next
== MESA_SHADER_FRAGMENT
? num_tfeedback_decls
: 0,
2511 if (!assign_varying_locations(ctx
, mem_ctx
, prog
, sh_i
, sh_next
,
2512 next
== MESA_SHADER_FRAGMENT
? num_tfeedback_decls
: 0,
2514 reserved_out_slots
| reserved_in_slots
))
2517 /* This must be done after all dead varyings are eliminated. */
2519 unsigned slots_used
= _mesa_bitcount_64(reserved_out_slots
);
2520 if (!check_against_output_limit(ctx
, prog
, sh_i
, slots_used
)) {
2525 unsigned slots_used
= _mesa_bitcount_64(reserved_in_slots
);
2526 if (!check_against_input_limit(ctx
, prog
, sh_next
, slots_used
))
2534 if (!store_tfeedback_info(ctx
, prog
, num_tfeedback_decls
, tfeedback_decls
,
2535 has_xfb_qualifiers
))