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 * Validate the types and qualifiers of an output from one stage against the
189 * matching input to another stage.
192 cross_validate_types_and_qualifiers(struct gl_context
*ctx
,
193 struct gl_shader_program
*prog
,
194 const ir_variable
*input
,
195 const ir_variable
*output
,
196 gl_shader_stage consumer_stage
,
197 gl_shader_stage producer_stage
)
199 /* Check that the types match between stages.
201 const glsl_type
*type_to_match
= input
->type
;
203 /* VS -> GS, VS -> TCS, VS -> TES, TES -> GS */
204 const bool extra_array_level
= (producer_stage
== MESA_SHADER_VERTEX
&&
205 consumer_stage
!= MESA_SHADER_FRAGMENT
) ||
206 consumer_stage
== MESA_SHADER_GEOMETRY
;
207 if (extra_array_level
) {
208 assert(type_to_match
->is_array());
209 type_to_match
= type_to_match
->fields
.array
;
212 if (type_to_match
!= output
->type
) {
213 /* There is a bit of a special case for gl_TexCoord. This
214 * built-in is unsized by default. Applications that variable
215 * access it must redeclare it with a size. There is some
216 * language in the GLSL spec that implies the fragment shader
217 * and vertex shader do not have to agree on this size. Other
218 * driver behave this way, and one or two applications seem to
221 * Neither declaration needs to be modified here because the array
222 * sizes are fixed later when update_array_sizes is called.
224 * From page 48 (page 54 of the PDF) of the GLSL 1.10 spec:
226 * "Unlike user-defined varying variables, the built-in
227 * varying variables don't have a strict one-to-one
228 * correspondence between the vertex language and the
229 * fragment language."
231 if (!output
->type
->is_array() || !is_gl_identifier(output
->name
)) {
233 "%s shader output `%s' declared as type `%s', "
234 "but %s shader input declared as type `%s'\n",
235 _mesa_shader_stage_to_string(producer_stage
),
238 _mesa_shader_stage_to_string(consumer_stage
),
244 /* Check that all of the qualifiers match between stages.
247 /* According to the OpenGL and OpenGLES GLSL specs, the centroid qualifier
248 * should match until OpenGL 4.3 and OpenGLES 3.1. The OpenGLES 3.0
249 * conformance test suite does not verify that the qualifiers must match.
250 * The deqp test suite expects the opposite (OpenGLES 3.1) behavior for
251 * OpenGLES 3.0 drivers, so we relax the checking in all cases.
253 if (false /* always skip the centroid check */ &&
254 prog
->data
->Version
< (prog
->IsES
? 310 : 430) &&
255 input
->data
.centroid
!= output
->data
.centroid
) {
257 "%s shader output `%s' %s centroid qualifier, "
258 "but %s shader input %s centroid qualifier\n",
259 _mesa_shader_stage_to_string(producer_stage
),
261 (output
->data
.centroid
) ? "has" : "lacks",
262 _mesa_shader_stage_to_string(consumer_stage
),
263 (input
->data
.centroid
) ? "has" : "lacks");
267 if (input
->data
.sample
!= output
->data
.sample
) {
269 "%s shader output `%s' %s sample qualifier, "
270 "but %s shader input %s sample qualifier\n",
271 _mesa_shader_stage_to_string(producer_stage
),
273 (output
->data
.sample
) ? "has" : "lacks",
274 _mesa_shader_stage_to_string(consumer_stage
),
275 (input
->data
.sample
) ? "has" : "lacks");
279 if (input
->data
.patch
!= output
->data
.patch
) {
281 "%s shader output `%s' %s patch qualifier, "
282 "but %s shader input %s patch qualifier\n",
283 _mesa_shader_stage_to_string(producer_stage
),
285 (output
->data
.patch
) ? "has" : "lacks",
286 _mesa_shader_stage_to_string(consumer_stage
),
287 (input
->data
.patch
) ? "has" : "lacks");
291 /* The GLSL 4.30 and GLSL ES 3.00 specifications say:
293 * "As only outputs need be declared with invariant, an output from
294 * one shader stage will still match an input of a subsequent stage
295 * without the input being declared as invariant."
297 * while GLSL 4.20 says:
299 * "For variables leaving one shader and coming into another shader,
300 * the invariant keyword has to be used in both shaders, or a link
301 * error will result."
303 * and GLSL ES 1.00 section 4.6.4 "Invariance and Linking" says:
305 * "The invariance of varyings that are declared in both the vertex
306 * and fragment shaders must match."
308 if (input
->data
.invariant
!= output
->data
.invariant
&&
309 prog
->data
->Version
< (prog
->IsES
? 300 : 430)) {
311 "%s shader output `%s' %s invariant qualifier, "
312 "but %s shader input %s invariant qualifier\n",
313 _mesa_shader_stage_to_string(producer_stage
),
315 (output
->data
.invariant
) ? "has" : "lacks",
316 _mesa_shader_stage_to_string(consumer_stage
),
317 (input
->data
.invariant
) ? "has" : "lacks");
321 /* GLSL >= 4.40 removes text requiring interpolation qualifiers
322 * to match cross stage, they must only match within the same stage.
324 * From page 84 (page 90 of the PDF) of the GLSL 4.40 spec:
326 * "It is a link-time error if, within the same stage, the interpolation
327 * qualifiers of variables of the same name do not match.
329 * Section 4.3.9 (Interpolation) of the GLSL ES 3.00 spec says:
331 * "When no interpolation qualifier is present, smooth interpolation
334 * So we match variables where one is smooth and the other has no explicit
337 unsigned input_interpolation
= input
->data
.interpolation
;
338 unsigned output_interpolation
= output
->data
.interpolation
;
340 if (input_interpolation
== INTERP_MODE_NONE
)
341 input_interpolation
= INTERP_MODE_SMOOTH
;
342 if (output_interpolation
== INTERP_MODE_NONE
)
343 output_interpolation
= INTERP_MODE_SMOOTH
;
345 if (input_interpolation
!= output_interpolation
&&
346 prog
->data
->Version
< 440) {
347 if (!ctx
->Const
.AllowGLSLCrossStageInterpolationMismatch
) {
349 "%s shader output `%s' specifies %s "
350 "interpolation qualifier, "
351 "but %s shader input specifies %s "
352 "interpolation qualifier\n",
353 _mesa_shader_stage_to_string(producer_stage
),
355 interpolation_string(output
->data
.interpolation
),
356 _mesa_shader_stage_to_string(consumer_stage
),
357 interpolation_string(input
->data
.interpolation
));
361 "%s shader output `%s' specifies %s "
362 "interpolation qualifier, "
363 "but %s shader input specifies %s "
364 "interpolation qualifier\n",
365 _mesa_shader_stage_to_string(producer_stage
),
367 interpolation_string(output
->data
.interpolation
),
368 _mesa_shader_stage_to_string(consumer_stage
),
369 interpolation_string(input
->data
.interpolation
));
375 * Validate front and back color outputs against single color input
378 cross_validate_front_and_back_color(struct gl_context
*ctx
,
379 struct gl_shader_program
*prog
,
380 const ir_variable
*input
,
381 const ir_variable
*front_color
,
382 const ir_variable
*back_color
,
383 gl_shader_stage consumer_stage
,
384 gl_shader_stage producer_stage
)
386 if (front_color
!= NULL
&& front_color
->data
.assigned
)
387 cross_validate_types_and_qualifiers(ctx
, prog
, input
, front_color
,
388 consumer_stage
, producer_stage
);
390 if (back_color
!= NULL
&& back_color
->data
.assigned
)
391 cross_validate_types_and_qualifiers(ctx
, prog
, input
, back_color
,
392 consumer_stage
, producer_stage
);
396 compute_variable_location_slot(ir_variable
*var
, gl_shader_stage stage
)
398 unsigned location_start
= VARYING_SLOT_VAR0
;
401 case MESA_SHADER_VERTEX
:
402 if (var
->data
.mode
== ir_var_shader_in
)
403 location_start
= VERT_ATTRIB_GENERIC0
;
405 case MESA_SHADER_TESS_CTRL
:
406 case MESA_SHADER_TESS_EVAL
:
408 location_start
= VARYING_SLOT_PATCH0
;
410 case MESA_SHADER_FRAGMENT
:
411 if (var
->data
.mode
== ir_var_shader_out
)
412 location_start
= FRAG_RESULT_DATA0
;
418 return var
->data
.location
- location_start
;
421 struct explicit_location_info
{
423 unsigned numerical_type
;
424 unsigned interpolation
;
430 static inline unsigned
431 get_numerical_type(const glsl_type
*type
)
433 /* From the OpenGL 4.6 spec, section 4.4.1 Input Layout Qualifiers, Page 68,
434 * (Location aliasing):
436 * "Further, when location aliasing, the aliases sharing the location
437 * must have the same underlying numerical type (floating-point or
440 if (type
->is_float() || type
->is_double())
441 return GLSL_TYPE_FLOAT
;
442 return GLSL_TYPE_INT
;
446 check_location_aliasing(struct explicit_location_info explicit_locations
[][4],
450 unsigned location_limit
,
451 const glsl_type
*type
,
452 unsigned interpolation
,
456 gl_shader_program
*prog
,
457 gl_shader_stage stage
)
460 if (type
->without_array()->is_record()) {
461 /* The component qualifier can't be used on structs so just treat
462 * all component slots as used.
466 unsigned dmul
= type
->without_array()->is_64bit() ? 2 : 1;
467 last_comp
= component
+ type
->without_array()->vector_elements
* dmul
;
470 while (location
< location_limit
) {
473 struct explicit_location_info
*info
=
474 &explicit_locations
[location
][comp
];
477 /* Component aliasing is not alloed */
478 if (comp
>= component
&& comp
< last_comp
) {
480 "%s shader has multiple outputs explicitly "
481 "assigned to location %d and component %d\n",
482 _mesa_shader_stage_to_string(stage
),
486 /* For all other used components we need to have matching
487 * types, interpolation and auxiliary storage
489 if (info
->numerical_type
!=
490 get_numerical_type(type
->without_array())) {
492 "Varyings sharing the same location must "
493 "have the same underlying numerical type. "
494 "Location %u component %u\n",
499 if (info
->interpolation
!= interpolation
) {
501 "%s shader has multiple outputs at explicit "
502 "location %u with different interpolation "
504 _mesa_shader_stage_to_string(stage
), location
);
508 if (info
->centroid
!= centroid
||
509 info
->sample
!= sample
||
510 info
->patch
!= patch
) {
512 "%s shader has multiple outputs at explicit "
513 "location %u with different aux storage\n",
514 _mesa_shader_stage_to_string(stage
), location
);
518 } else if (comp
>= component
&& comp
< last_comp
) {
520 info
->numerical_type
= get_numerical_type(type
->without_array());
521 info
->interpolation
= interpolation
;
522 info
->centroid
= centroid
;
523 info
->sample
= sample
;
529 /* We need to do some special handling for doubles as dvec3 and
530 * dvec4 consume two consecutive locations. We don't need to
531 * worry about components beginning at anything other than 0 as
532 * the spec does not allow this for dvec3 and dvec4.
534 if (comp
== 4 && last_comp
> 4) {
535 last_comp
= last_comp
- 4;
536 /* Bump location index and reset the component index */
550 validate_explicit_variable_location(struct gl_context
*ctx
,
551 struct explicit_location_info explicit_locations
[][4],
553 gl_shader_program
*prog
,
554 gl_linked_shader
*sh
)
556 const glsl_type
*type
= get_varying_type(var
, sh
->Stage
);
557 unsigned num_elements
= type
->count_attribute_slots(false);
558 unsigned idx
= compute_variable_location_slot(var
, sh
->Stage
);
559 unsigned slot_limit
= idx
+ num_elements
;
561 /* Vertex shader inputs and fragment shader outputs are validated in
562 * assign_attribute_or_color_locations() so we should not attempt to
563 * validate them again here.
566 if (var
->data
.mode
== ir_var_shader_out
) {
567 assert(sh
->Stage
!= MESA_SHADER_FRAGMENT
);
569 ctx
->Const
.Program
[sh
->Stage
].MaxOutputComponents
/ 4;
571 assert(var
->data
.mode
== ir_var_shader_in
);
572 assert(sh
->Stage
!= MESA_SHADER_VERTEX
);
574 ctx
->Const
.Program
[sh
->Stage
].MaxInputComponents
/ 4;
577 if (slot_limit
> slot_max
) {
579 "Invalid location %u in %s shader\n",
580 idx
, _mesa_shader_stage_to_string(sh
->Stage
));
584 const glsl_type
*type_without_array
= type
->without_array();
585 if (type_without_array
->is_interface()) {
586 for (unsigned i
= 0; i
< type_without_array
->length
; i
++) {
587 glsl_struct_field
*field
= &type_without_array
->fields
.structure
[i
];
588 unsigned field_location
= field
->location
-
589 (field
->patch
? VARYING_SLOT_PATCH0
: VARYING_SLOT_VAR0
);
590 if (!check_location_aliasing(explicit_locations
, var
,
592 0, field_location
+ 1,
594 field
->interpolation
,
602 } else if (!check_location_aliasing(explicit_locations
, var
,
603 idx
, var
->data
.location_frac
,
605 var
->data
.interpolation
,
617 * Validate explicit locations for the inputs to the first stage and the
618 * outputs of the last stage in an SSO program (everything in between is
619 * validated in cross_validate_outputs_to_inputs).
622 validate_sso_explicit_locations(struct gl_context
*ctx
,
623 struct gl_shader_program
*prog
,
624 gl_shader_stage first_stage
,
625 gl_shader_stage last_stage
)
627 assert(prog
->SeparateShader
);
629 /* VS inputs and FS outputs are validated in
630 * assign_attribute_or_color_locations()
632 bool validate_first_stage
= first_stage
!= MESA_SHADER_VERTEX
;
633 bool validate_last_stage
= last_stage
!= MESA_SHADER_FRAGMENT
;
634 if (!validate_first_stage
&& !validate_last_stage
)
637 struct explicit_location_info explicit_locations
[MAX_VARYING
][4];
639 gl_shader_stage stages
[2] = { first_stage
, last_stage
};
640 bool validate_stage
[2] = { validate_first_stage
, validate_last_stage
};
641 ir_variable_mode var_direction
[2] = { ir_var_shader_in
, ir_var_shader_out
};
643 for (unsigned i
= 0; i
< 2; i
++) {
644 if (!validate_stage
[i
])
647 gl_shader_stage stage
= stages
[i
];
649 gl_linked_shader
*sh
= prog
->_LinkedShaders
[stage
];
652 memset(explicit_locations
, 0, sizeof(explicit_locations
));
654 foreach_in_list(ir_instruction
, node
, sh
->ir
) {
655 ir_variable
*const var
= node
->as_variable();
658 !var
->data
.explicit_location
||
659 var
->data
.location
< VARYING_SLOT_VAR0
||
660 var
->data
.mode
!= var_direction
[i
])
663 if (!validate_explicit_variable_location(
664 ctx
, explicit_locations
, var
, prog
, sh
)) {
672 * Validate that outputs from one stage match inputs of another
675 cross_validate_outputs_to_inputs(struct gl_context
*ctx
,
676 struct gl_shader_program
*prog
,
677 gl_linked_shader
*producer
,
678 gl_linked_shader
*consumer
)
680 glsl_symbol_table parameters
;
681 struct explicit_location_info explicit_locations
[MAX_VARYING
][4] = { 0 };
683 /* Find all shader outputs in the "producer" stage.
685 foreach_in_list(ir_instruction
, node
, producer
->ir
) {
686 ir_variable
*const var
= node
->as_variable();
688 if (var
== NULL
|| var
->data
.mode
!= ir_var_shader_out
)
691 if (!var
->data
.explicit_location
692 || var
->data
.location
< VARYING_SLOT_VAR0
)
693 parameters
.add_variable(var
);
695 /* User-defined varyings with explicit locations are handled
696 * differently because they do not need to have matching names.
698 if (!validate_explicit_variable_location(ctx
,
700 var
, prog
, producer
)) {
707 /* Find all shader inputs in the "consumer" stage. Any variables that have
708 * matching outputs already in the symbol table must have the same type and
711 * Exception: if the consumer is the geometry shader, then the inputs
712 * should be arrays and the type of the array element should match the type
713 * of the corresponding producer output.
715 foreach_in_list(ir_instruction
, node
, consumer
->ir
) {
716 ir_variable
*const input
= node
->as_variable();
718 if (input
== NULL
|| input
->data
.mode
!= ir_var_shader_in
)
721 if (strcmp(input
->name
, "gl_Color") == 0 && input
->data
.used
) {
722 const ir_variable
*const front_color
=
723 parameters
.get_variable("gl_FrontColor");
725 const ir_variable
*const back_color
=
726 parameters
.get_variable("gl_BackColor");
728 cross_validate_front_and_back_color(ctx
, prog
, input
,
729 front_color
, back_color
,
730 consumer
->Stage
, producer
->Stage
);
731 } else if (strcmp(input
->name
, "gl_SecondaryColor") == 0 && input
->data
.used
) {
732 const ir_variable
*const front_color
=
733 parameters
.get_variable("gl_FrontSecondaryColor");
735 const ir_variable
*const back_color
=
736 parameters
.get_variable("gl_BackSecondaryColor");
738 cross_validate_front_and_back_color(ctx
, prog
, input
,
739 front_color
, back_color
,
740 consumer
->Stage
, producer
->Stage
);
742 /* The rules for connecting inputs and outputs change in the presence
743 * of explicit locations. In this case, we no longer care about the
744 * names of the variables. Instead, we care only about the
745 * explicitly assigned location.
747 ir_variable
*output
= NULL
;
748 if (input
->data
.explicit_location
749 && input
->data
.location
>= VARYING_SLOT_VAR0
) {
751 const glsl_type
*type
= get_varying_type(input
, consumer
->Stage
);
752 unsigned num_elements
= type
->count_attribute_slots(false);
754 compute_variable_location_slot(input
, consumer
->Stage
);
755 unsigned slot_limit
= idx
+ num_elements
;
757 while (idx
< slot_limit
) {
758 if (idx
>= MAX_VARYING
) {
760 "Invalid location %u in %s shader\n", idx
,
761 _mesa_shader_stage_to_string(consumer
->Stage
));
765 output
= explicit_locations
[idx
][input
->data
.location_frac
].var
;
767 if (output
== NULL
||
768 input
->data
.location
!= output
->data
.location
) {
770 "%s shader input `%s' with explicit location "
771 "has no matching output\n",
772 _mesa_shader_stage_to_string(consumer
->Stage
),
779 output
= parameters
.get_variable(input
->name
);
782 if (output
!= NULL
) {
783 /* Interface blocks have their own validation elsewhere so don't
784 * try validating them here.
786 if (!(input
->get_interface_type() &&
787 output
->get_interface_type()))
788 cross_validate_types_and_qualifiers(ctx
, prog
, input
, output
,
792 /* Check for input vars with unmatched output vars in prev stage
793 * taking into account that interface blocks could have a matching
794 * output but with different name, so we ignore them.
796 assert(!input
->data
.assigned
);
797 if (input
->data
.used
&& !input
->get_interface_type() &&
798 !input
->data
.explicit_location
&& !prog
->SeparateShader
)
800 "%s shader input `%s' "
801 "has no matching output in the previous stage\n",
802 _mesa_shader_stage_to_string(consumer
->Stage
),
810 * Demote shader inputs and outputs that are not used in other stages, and
811 * remove them via dead code elimination.
814 remove_unused_shader_inputs_and_outputs(bool is_separate_shader_object
,
815 gl_linked_shader
*sh
,
816 enum ir_variable_mode mode
)
818 if (is_separate_shader_object
)
821 foreach_in_list(ir_instruction
, node
, sh
->ir
) {
822 ir_variable
*const var
= node
->as_variable();
824 if (var
== NULL
|| var
->data
.mode
!= int(mode
))
827 /* A shader 'in' or 'out' variable is only really an input or output if
828 * its value is used by other shader stages. This will cause the
829 * variable to have a location assigned.
831 if (var
->data
.is_unmatched_generic_inout
&& !var
->data
.is_xfb_only
) {
832 assert(var
->data
.mode
!= ir_var_temporary
);
834 /* Assign zeros to demoted inputs to allow more optimizations. */
835 if (var
->data
.mode
== ir_var_shader_in
&& !var
->constant_value
)
836 var
->constant_value
= ir_constant::zero(var
, var
->type
);
838 var
->data
.mode
= ir_var_auto
;
842 /* Eliminate code that is now dead due to unused inputs/outputs being
845 while (do_dead_code(sh
->ir
, false))
851 * Initialize this object based on a string that was passed to
852 * glTransformFeedbackVaryings.
854 * If the input is mal-formed, this call still succeeds, but it sets
855 * this->var_name to a mal-formed input, so tfeedback_decl::find_output_var()
856 * will fail to find any matching variable.
859 tfeedback_decl::init(struct gl_context
*ctx
, const void *mem_ctx
,
862 /* We don't have to be pedantic about what is a valid GLSL variable name,
863 * because any variable with an invalid name can't exist in the IR anyway.
867 this->orig_name
= input
;
868 this->lowered_builtin_array_variable
= none
;
869 this->skip_components
= 0;
870 this->next_buffer_separator
= false;
871 this->matched_candidate
= NULL
;
876 if (ctx
->Extensions
.ARB_transform_feedback3
) {
877 /* Parse gl_NextBuffer. */
878 if (strcmp(input
, "gl_NextBuffer") == 0) {
879 this->next_buffer_separator
= true;
883 /* Parse gl_SkipComponents. */
884 if (strcmp(input
, "gl_SkipComponents1") == 0)
885 this->skip_components
= 1;
886 else if (strcmp(input
, "gl_SkipComponents2") == 0)
887 this->skip_components
= 2;
888 else if (strcmp(input
, "gl_SkipComponents3") == 0)
889 this->skip_components
= 3;
890 else if (strcmp(input
, "gl_SkipComponents4") == 0)
891 this->skip_components
= 4;
893 if (this->skip_components
)
897 /* Parse a declaration. */
898 const char *base_name_end
;
899 long subscript
= parse_program_resource_name(input
, &base_name_end
);
900 this->var_name
= ralloc_strndup(mem_ctx
, input
, base_name_end
- input
);
901 if (this->var_name
== NULL
) {
902 _mesa_error_no_memory(__func__
);
906 if (subscript
>= 0) {
907 this->array_subscript
= subscript
;
908 this->is_subscripted
= true;
910 this->is_subscripted
= false;
913 /* For drivers that lower gl_ClipDistance to gl_ClipDistanceMESA, this
914 * class must behave specially to account for the fact that gl_ClipDistance
915 * is converted from a float[8] to a vec4[2].
917 if (ctx
->Const
.ShaderCompilerOptions
[MESA_SHADER_VERTEX
].LowerCombinedClipCullDistance
&&
918 strcmp(this->var_name
, "gl_ClipDistance") == 0) {
919 this->lowered_builtin_array_variable
= clip_distance
;
921 if (ctx
->Const
.ShaderCompilerOptions
[MESA_SHADER_VERTEX
].LowerCombinedClipCullDistance
&&
922 strcmp(this->var_name
, "gl_CullDistance") == 0) {
923 this->lowered_builtin_array_variable
= cull_distance
;
926 if (ctx
->Const
.LowerTessLevel
&&
927 (strcmp(this->var_name
, "gl_TessLevelOuter") == 0))
928 this->lowered_builtin_array_variable
= tess_level_outer
;
929 if (ctx
->Const
.LowerTessLevel
&&
930 (strcmp(this->var_name
, "gl_TessLevelInner") == 0))
931 this->lowered_builtin_array_variable
= tess_level_inner
;
936 * Determine whether two tfeedback_decl objects refer to the same variable and
937 * array index (if applicable).
940 tfeedback_decl::is_same(const tfeedback_decl
&x
, const tfeedback_decl
&y
)
942 assert(x
.is_varying() && y
.is_varying());
944 if (strcmp(x
.var_name
, y
.var_name
) != 0)
946 if (x
.is_subscripted
!= y
.is_subscripted
)
948 if (x
.is_subscripted
&& x
.array_subscript
!= y
.array_subscript
)
955 * Assign a location and stream ID for this tfeedback_decl object based on the
956 * transform feedback candidate found by find_candidate.
958 * If an error occurs, the error is reported through linker_error() and false
962 tfeedback_decl::assign_location(struct gl_context
*ctx
,
963 struct gl_shader_program
*prog
)
965 assert(this->is_varying());
967 unsigned fine_location
968 = this->matched_candidate
->toplevel_var
->data
.location
* 4
969 + this->matched_candidate
->toplevel_var
->data
.location_frac
970 + this->matched_candidate
->offset
;
971 const unsigned dmul
=
972 this->matched_candidate
->type
->without_array()->is_64bit() ? 2 : 1;
974 if (this->matched_candidate
->type
->is_array()) {
976 const unsigned matrix_cols
=
977 this->matched_candidate
->type
->fields
.array
->matrix_columns
;
978 const unsigned vector_elements
=
979 this->matched_candidate
->type
->fields
.array
->vector_elements
;
980 unsigned actual_array_size
;
981 switch (this->lowered_builtin_array_variable
) {
983 actual_array_size
= prog
->last_vert_prog
?
984 prog
->last_vert_prog
->info
.clip_distance_array_size
: 0;
987 actual_array_size
= prog
->last_vert_prog
?
988 prog
->last_vert_prog
->info
.cull_distance_array_size
: 0;
990 case tess_level_outer
:
991 actual_array_size
= 4;
993 case tess_level_inner
:
994 actual_array_size
= 2;
998 actual_array_size
= this->matched_candidate
->type
->array_size();
1002 if (this->is_subscripted
) {
1003 /* Check array bounds. */
1004 if (this->array_subscript
>= actual_array_size
) {
1005 linker_error(prog
, "Transform feedback varying %s has index "
1006 "%i, but the array size is %u.",
1007 this->orig_name
, this->array_subscript
,
1011 unsigned array_elem_size
= this->lowered_builtin_array_variable
?
1012 1 : vector_elements
* matrix_cols
* dmul
;
1013 fine_location
+= array_elem_size
* this->array_subscript
;
1016 this->size
= actual_array_size
;
1018 this->vector_elements
= vector_elements
;
1019 this->matrix_columns
= matrix_cols
;
1020 if (this->lowered_builtin_array_variable
)
1021 this->type
= GL_FLOAT
;
1023 this->type
= this->matched_candidate
->type
->fields
.array
->gl_type
;
1025 /* Regular variable (scalar, vector, or matrix) */
1026 if (this->is_subscripted
) {
1027 linker_error(prog
, "Transform feedback varying %s requested, "
1028 "but %s is not an array.",
1029 this->orig_name
, this->var_name
);
1033 this->vector_elements
= this->matched_candidate
->type
->vector_elements
;
1034 this->matrix_columns
= this->matched_candidate
->type
->matrix_columns
;
1035 this->type
= this->matched_candidate
->type
->gl_type
;
1037 this->location
= fine_location
/ 4;
1038 this->location_frac
= fine_location
% 4;
1040 /* From GL_EXT_transform_feedback:
1041 * A program will fail to link if:
1043 * * the total number of components to capture in any varying
1044 * variable in <varyings> is greater than the constant
1045 * MAX_TRANSFORM_FEEDBACK_SEPARATE_COMPONENTS_EXT and the
1046 * buffer mode is SEPARATE_ATTRIBS_EXT;
1048 if (prog
->TransformFeedback
.BufferMode
== GL_SEPARATE_ATTRIBS
&&
1049 this->num_components() >
1050 ctx
->Const
.MaxTransformFeedbackSeparateComponents
) {
1051 linker_error(prog
, "Transform feedback varying %s exceeds "
1052 "MAX_TRANSFORM_FEEDBACK_SEPARATE_COMPONENTS.",
1057 /* Only transform feedback varyings can be assigned to non-zero streams,
1058 * so assign the stream id here.
1060 this->stream_id
= this->matched_candidate
->toplevel_var
->data
.stream
;
1062 unsigned array_offset
= this->array_subscript
* 4 * dmul
;
1063 unsigned struct_offset
= this->matched_candidate
->offset
* 4 * dmul
;
1064 this->buffer
= this->matched_candidate
->toplevel_var
->data
.xfb_buffer
;
1065 this->offset
= this->matched_candidate
->toplevel_var
->data
.offset
+
1066 array_offset
+ struct_offset
;
1073 tfeedback_decl::get_num_outputs() const
1075 if (!this->is_varying()) {
1078 return (this->num_components() + this->location_frac
+ 3)/4;
1083 * Update gl_transform_feedback_info to reflect this tfeedback_decl.
1085 * If an error occurs, the error is reported through linker_error() and false
1089 tfeedback_decl::store(struct gl_context
*ctx
, struct gl_shader_program
*prog
,
1090 struct gl_transform_feedback_info
*info
,
1091 unsigned buffer
, unsigned buffer_index
,
1092 const unsigned max_outputs
, bool *explicit_stride
,
1093 bool has_xfb_qualifiers
) const
1095 unsigned xfb_offset
= 0;
1096 unsigned size
= this->size
;
1097 /* Handle gl_SkipComponents. */
1098 if (this->skip_components
) {
1099 info
->Buffers
[buffer
].Stride
+= this->skip_components
;
1100 size
= this->skip_components
;
1104 if (this->next_buffer_separator
) {
1109 if (has_xfb_qualifiers
) {
1110 xfb_offset
= this->offset
/ 4;
1112 xfb_offset
= info
->Buffers
[buffer
].Stride
;
1114 info
->Varyings
[info
->NumVarying
].Offset
= xfb_offset
* 4;
1117 unsigned location
= this->location
;
1118 unsigned location_frac
= this->location_frac
;
1119 unsigned num_components
= this->num_components();
1120 while (num_components
> 0) {
1121 unsigned output_size
= MIN2(num_components
, 4 - location_frac
);
1122 assert((info
->NumOutputs
== 0 && max_outputs
== 0) ||
1123 info
->NumOutputs
< max_outputs
);
1125 /* From the ARB_enhanced_layouts spec:
1127 * "If such a block member or variable is not written during a shader
1128 * invocation, the buffer contents at the assigned offset will be
1129 * undefined. Even if there are no static writes to a variable or
1130 * member that is assigned a transform feedback offset, the space is
1131 * still allocated in the buffer and still affects the stride."
1133 if (this->is_varying_written()) {
1134 info
->Outputs
[info
->NumOutputs
].ComponentOffset
= location_frac
;
1135 info
->Outputs
[info
->NumOutputs
].OutputRegister
= location
;
1136 info
->Outputs
[info
->NumOutputs
].NumComponents
= output_size
;
1137 info
->Outputs
[info
->NumOutputs
].StreamId
= stream_id
;
1138 info
->Outputs
[info
->NumOutputs
].OutputBuffer
= buffer
;
1139 info
->Outputs
[info
->NumOutputs
].DstOffset
= xfb_offset
;
1142 info
->Buffers
[buffer
].Stream
= this->stream_id
;
1143 xfb_offset
+= output_size
;
1145 num_components
-= output_size
;
1151 if (explicit_stride
&& explicit_stride
[buffer
]) {
1152 if (this->is_64bit() && info
->Buffers
[buffer
].Stride
% 2) {
1153 linker_error(prog
, "invalid qualifier xfb_stride=%d must be a "
1154 "multiple of 8 as its applied to a type that is or "
1155 "contains a double.",
1156 info
->Buffers
[buffer
].Stride
* 4);
1160 if ((this->offset
/ 4) / info
->Buffers
[buffer
].Stride
!=
1161 (xfb_offset
- 1) / info
->Buffers
[buffer
].Stride
) {
1162 linker_error(prog
, "xfb_offset (%d) overflows xfb_stride (%d) for "
1163 "buffer (%d)", xfb_offset
* 4,
1164 info
->Buffers
[buffer
].Stride
* 4, buffer
);
1168 info
->Buffers
[buffer
].Stride
= xfb_offset
;
1171 /* From GL_EXT_transform_feedback:
1172 * A program will fail to link if:
1174 * * the total number of components to capture is greater than
1175 * the constant MAX_TRANSFORM_FEEDBACK_INTERLEAVED_COMPONENTS_EXT
1176 * and the buffer mode is INTERLEAVED_ATTRIBS_EXT.
1178 * From GL_ARB_enhanced_layouts:
1180 * "The resulting stride (implicit or explicit) must be less than or
1181 * equal to the implementation-dependent constant
1182 * gl_MaxTransformFeedbackInterleavedComponents."
1184 if ((prog
->TransformFeedback
.BufferMode
== GL_INTERLEAVED_ATTRIBS
||
1185 has_xfb_qualifiers
) &&
1186 info
->Buffers
[buffer
].Stride
>
1187 ctx
->Const
.MaxTransformFeedbackInterleavedComponents
) {
1188 linker_error(prog
, "The MAX_TRANSFORM_FEEDBACK_INTERLEAVED_COMPONENTS "
1189 "limit has been exceeded.");
1194 info
->Varyings
[info
->NumVarying
].Name
= ralloc_strdup(prog
,
1196 info
->Varyings
[info
->NumVarying
].Type
= this->type
;
1197 info
->Varyings
[info
->NumVarying
].Size
= size
;
1198 info
->Varyings
[info
->NumVarying
].BufferIndex
= buffer_index
;
1200 info
->Buffers
[buffer
].NumVaryings
++;
1206 const tfeedback_candidate
*
1207 tfeedback_decl::find_candidate(gl_shader_program
*prog
,
1208 hash_table
*tfeedback_candidates
)
1210 const char *name
= this->var_name
;
1211 switch (this->lowered_builtin_array_variable
) {
1213 name
= this->var_name
;
1216 name
= "gl_ClipDistanceMESA";
1219 name
= "gl_CullDistanceMESA";
1221 case tess_level_outer
:
1222 name
= "gl_TessLevelOuterMESA";
1224 case tess_level_inner
:
1225 name
= "gl_TessLevelInnerMESA";
1228 hash_entry
*entry
= _mesa_hash_table_search(tfeedback_candidates
, name
);
1230 this->matched_candidate
= entry
?
1231 (const tfeedback_candidate
*) entry
->data
: NULL
;
1233 if (!this->matched_candidate
) {
1234 /* From GL_EXT_transform_feedback:
1235 * A program will fail to link if:
1237 * * any variable name specified in the <varyings> array is not
1238 * declared as an output in the geometry shader (if present) or
1239 * the vertex shader (if no geometry shader is present);
1241 linker_error(prog
, "Transform feedback varying %s undeclared.",
1245 return this->matched_candidate
;
1250 * Parse all the transform feedback declarations that were passed to
1251 * glTransformFeedbackVaryings() and store them in tfeedback_decl objects.
1253 * If an error occurs, the error is reported through linker_error() and false
1257 parse_tfeedback_decls(struct gl_context
*ctx
, struct gl_shader_program
*prog
,
1258 const void *mem_ctx
, unsigned num_names
,
1259 char **varying_names
, tfeedback_decl
*decls
)
1261 for (unsigned i
= 0; i
< num_names
; ++i
) {
1262 decls
[i
].init(ctx
, mem_ctx
, varying_names
[i
]);
1264 if (!decls
[i
].is_varying())
1267 /* From GL_EXT_transform_feedback:
1268 * A program will fail to link if:
1270 * * any two entries in the <varyings> array specify the same varying
1273 * We interpret this to mean "any two entries in the <varyings> array
1274 * specify the same varying variable and array index", since transform
1275 * feedback of arrays would be useless otherwise.
1277 for (unsigned j
= 0; j
< i
; ++j
) {
1278 if (!decls
[j
].is_varying())
1281 if (tfeedback_decl::is_same(decls
[i
], decls
[j
])) {
1282 linker_error(prog
, "Transform feedback varying %s specified "
1283 "more than once.", varying_names
[i
]);
1293 cmp_xfb_offset(const void * x_generic
, const void * y_generic
)
1295 tfeedback_decl
*x
= (tfeedback_decl
*) x_generic
;
1296 tfeedback_decl
*y
= (tfeedback_decl
*) y_generic
;
1298 if (x
->get_buffer() != y
->get_buffer())
1299 return x
->get_buffer() - y
->get_buffer();
1300 return x
->get_offset() - y
->get_offset();
1304 * Store transform feedback location assignments into
1305 * prog->sh.LinkedTransformFeedback based on the data stored in
1308 * If an error occurs, the error is reported through linker_error() and false
1312 store_tfeedback_info(struct gl_context
*ctx
, struct gl_shader_program
*prog
,
1313 unsigned num_tfeedback_decls
,
1314 tfeedback_decl
*tfeedback_decls
, bool has_xfb_qualifiers
)
1316 if (!prog
->last_vert_prog
)
1319 /* Make sure MaxTransformFeedbackBuffers is less than 32 so the bitmask for
1320 * tracking the number of buffers doesn't overflow.
1322 assert(ctx
->Const
.MaxTransformFeedbackBuffers
< 32);
1324 bool separate_attribs_mode
=
1325 prog
->TransformFeedback
.BufferMode
== GL_SEPARATE_ATTRIBS
;
1327 struct gl_program
*xfb_prog
= prog
->last_vert_prog
;
1328 xfb_prog
->sh
.LinkedTransformFeedback
=
1329 rzalloc(xfb_prog
, struct gl_transform_feedback_info
);
1331 /* The xfb_offset qualifier does not have to be used in increasing order
1332 * however some drivers expect to receive the list of transform feedback
1333 * declarations in order so sort it now for convenience.
1335 if (has_xfb_qualifiers
) {
1336 qsort(tfeedback_decls
, num_tfeedback_decls
, sizeof(*tfeedback_decls
),
1339 xfb_prog
->sh
.LinkedTransformFeedback
->api_enabled
= true;
1342 xfb_prog
->sh
.LinkedTransformFeedback
->Varyings
=
1343 rzalloc_array(xfb_prog
, struct gl_transform_feedback_varying_info
,
1344 num_tfeedback_decls
);
1346 unsigned num_outputs
= 0;
1347 for (unsigned i
= 0; i
< num_tfeedback_decls
; ++i
) {
1348 if (tfeedback_decls
[i
].is_varying_written())
1349 num_outputs
+= tfeedback_decls
[i
].get_num_outputs();
1352 xfb_prog
->sh
.LinkedTransformFeedback
->Outputs
=
1353 rzalloc_array(xfb_prog
, struct gl_transform_feedback_output
,
1356 unsigned num_buffers
= 0;
1357 unsigned buffers
= 0;
1359 if (!has_xfb_qualifiers
&& separate_attribs_mode
) {
1360 /* GL_SEPARATE_ATTRIBS */
1361 for (unsigned i
= 0; i
< num_tfeedback_decls
; ++i
) {
1362 if (!tfeedback_decls
[i
].store(ctx
, prog
,
1363 xfb_prog
->sh
.LinkedTransformFeedback
,
1364 num_buffers
, num_buffers
, num_outputs
,
1365 NULL
, has_xfb_qualifiers
))
1368 buffers
|= 1 << num_buffers
;
1373 /* GL_INVERLEAVED_ATTRIBS */
1374 int buffer_stream_id
= -1;
1376 num_tfeedback_decls
? tfeedback_decls
[0].get_buffer() : 0;
1377 bool explicit_stride
[MAX_FEEDBACK_BUFFERS
] = { false };
1379 /* Apply any xfb_stride global qualifiers */
1380 if (has_xfb_qualifiers
) {
1381 for (unsigned j
= 0; j
< MAX_FEEDBACK_BUFFERS
; j
++) {
1382 if (prog
->TransformFeedback
.BufferStride
[j
]) {
1383 explicit_stride
[j
] = true;
1384 xfb_prog
->sh
.LinkedTransformFeedback
->Buffers
[j
].Stride
=
1385 prog
->TransformFeedback
.BufferStride
[j
] / 4;
1390 for (unsigned i
= 0; i
< num_tfeedback_decls
; ++i
) {
1391 if (has_xfb_qualifiers
&&
1392 buffer
!= tfeedback_decls
[i
].get_buffer()) {
1393 /* we have moved to the next buffer so reset stream id */
1394 buffer_stream_id
= -1;
1398 if (tfeedback_decls
[i
].is_next_buffer_separator()) {
1399 if (!tfeedback_decls
[i
].store(ctx
, prog
,
1400 xfb_prog
->sh
.LinkedTransformFeedback
,
1401 buffer
, num_buffers
, num_outputs
,
1402 explicit_stride
, has_xfb_qualifiers
))
1405 buffer_stream_id
= -1;
1409 if (has_xfb_qualifiers
) {
1410 buffer
= tfeedback_decls
[i
].get_buffer();
1412 buffer
= num_buffers
;
1415 if (tfeedback_decls
[i
].is_varying()) {
1416 if (buffer_stream_id
== -1) {
1417 /* First varying writing to this buffer: remember its stream */
1418 buffer_stream_id
= (int) tfeedback_decls
[i
].get_stream_id();
1420 /* Only mark a buffer as active when there is a varying
1421 * attached to it. This behaviour is based on a revised version
1422 * of section 13.2.2 of the GL 4.6 spec.
1424 buffers
|= 1 << buffer
;
1425 } else if (buffer_stream_id
!=
1426 (int) tfeedback_decls
[i
].get_stream_id()) {
1427 /* Varying writes to the same buffer from a different stream */
1429 "Transform feedback can't capture varyings belonging "
1430 "to different vertex streams in a single buffer. "
1431 "Varying %s writes to buffer from stream %u, other "
1432 "varyings in the same buffer write from stream %u.",
1433 tfeedback_decls
[i
].name(),
1434 tfeedback_decls
[i
].get_stream_id(),
1440 if (!tfeedback_decls
[i
].store(ctx
, prog
,
1441 xfb_prog
->sh
.LinkedTransformFeedback
,
1442 buffer
, num_buffers
, num_outputs
,
1443 explicit_stride
, has_xfb_qualifiers
))
1448 assert(xfb_prog
->sh
.LinkedTransformFeedback
->NumOutputs
== num_outputs
);
1450 xfb_prog
->sh
.LinkedTransformFeedback
->ActiveBuffers
= buffers
;
1457 * Data structure recording the relationship between outputs of one shader
1458 * stage (the "producer") and inputs of another (the "consumer").
1460 class varying_matches
1463 varying_matches(bool disable_varying_packing
, bool xfb_enabled
,
1464 bool enhanced_layouts_enabled
,
1465 gl_shader_stage producer_stage
,
1466 gl_shader_stage consumer_stage
);
1468 void record(ir_variable
*producer_var
, ir_variable
*consumer_var
);
1469 unsigned assign_locations(struct gl_shader_program
*prog
,
1470 uint8_t *components
,
1471 uint64_t reserved_slots
);
1472 void store_locations() const;
1475 bool is_varying_packing_safe(const glsl_type
*type
,
1476 const ir_variable
*var
);
1479 * If true, this driver disables varying packing, so all varyings need to
1480 * be aligned on slot boundaries, and take up a number of slots equal to
1481 * their number of matrix columns times their array size.
1483 * Packing may also be disabled because our current packing method is not
1484 * safe in SSO or versions of OpenGL where interpolation qualifiers are not
1485 * guaranteed to match across stages.
1487 const bool disable_varying_packing
;
1490 * If true, this driver has transform feedback enabled. The transform
1491 * feedback code requires at least some packing be done even when varying
1492 * packing is disabled, fortunately where transform feedback requires
1493 * packing it's safe to override the disabled setting. See
1494 * is_varying_packing_safe().
1496 const bool xfb_enabled
;
1498 const bool enhanced_layouts_enabled
;
1501 * Enum representing the order in which varyings are packed within a
1504 * Currently we pack vec4's first, then vec2's, then scalar values, then
1505 * vec3's. This order ensures that the only vectors that are at risk of
1506 * having to be "double parked" (split between two adjacent varying slots)
1509 enum packing_order_enum
{
1512 PACKING_ORDER_SCALAR
,
1516 static unsigned compute_packing_class(const ir_variable
*var
);
1517 static packing_order_enum
compute_packing_order(const ir_variable
*var
);
1518 static int match_comparator(const void *x_generic
, const void *y_generic
);
1519 static int xfb_comparator(const void *x_generic
, const void *y_generic
);
1522 * Structure recording the relationship between a single producer output
1523 * and a single consumer input.
1527 * Packing class for this varying, computed by compute_packing_class().
1529 unsigned packing_class
;
1532 * Packing order for this varying, computed by compute_packing_order().
1534 packing_order_enum packing_order
;
1535 unsigned num_components
;
1538 * The output variable in the producer stage.
1540 ir_variable
*producer_var
;
1543 * The input variable in the consumer stage.
1545 ir_variable
*consumer_var
;
1548 * The location which has been assigned for this varying. This is
1549 * expressed in multiples of a float, with the first generic varying
1550 * (i.e. the one referred to by VARYING_SLOT_VAR0) represented by the
1553 unsigned generic_location
;
1557 * The number of elements in the \c matches array that are currently in
1560 unsigned num_matches
;
1563 * The number of elements that were set aside for the \c matches array when
1566 unsigned matches_capacity
;
1568 gl_shader_stage producer_stage
;
1569 gl_shader_stage consumer_stage
;
1572 } /* anonymous namespace */
1574 varying_matches::varying_matches(bool disable_varying_packing
,
1576 bool enhanced_layouts_enabled
,
1577 gl_shader_stage producer_stage
,
1578 gl_shader_stage consumer_stage
)
1579 : disable_varying_packing(disable_varying_packing
),
1580 xfb_enabled(xfb_enabled
),
1581 enhanced_layouts_enabled(enhanced_layouts_enabled
),
1582 producer_stage(producer_stage
),
1583 consumer_stage(consumer_stage
)
1585 /* Note: this initial capacity is rather arbitrarily chosen to be large
1586 * enough for many cases without wasting an unreasonable amount of space.
1587 * varying_matches::record() will resize the array if there are more than
1588 * this number of varyings.
1590 this->matches_capacity
= 8;
1591 this->matches
= (match
*)
1592 malloc(sizeof(*this->matches
) * this->matches_capacity
);
1593 this->num_matches
= 0;
1597 varying_matches::~varying_matches()
1599 free(this->matches
);
1604 * Packing is always safe on individual arrays, structures, and matrices. It
1605 * is also safe if the varying is only used for transform feedback.
1608 varying_matches::is_varying_packing_safe(const glsl_type
*type
,
1609 const ir_variable
*var
)
1611 if (consumer_stage
== MESA_SHADER_TESS_EVAL
||
1612 consumer_stage
== MESA_SHADER_TESS_CTRL
||
1613 producer_stage
== MESA_SHADER_TESS_CTRL
)
1616 return xfb_enabled
&& (type
->is_array() || type
->is_record() ||
1617 type
->is_matrix() || var
->data
.is_xfb_only
);
1622 * Record the given producer/consumer variable pair in the list of variables
1623 * that should later be assigned locations.
1625 * It is permissible for \c consumer_var to be NULL (this happens if a
1626 * variable is output by the producer and consumed by transform feedback, but
1627 * not consumed by the consumer).
1629 * If \c producer_var has already been paired up with a consumer_var, or
1630 * producer_var is part of fixed pipeline functionality (and hence already has
1631 * a location assigned), this function has no effect.
1633 * Note: as a side effect this function may change the interpolation type of
1634 * \c producer_var, but only when the change couldn't possibly affect
1638 varying_matches::record(ir_variable
*producer_var
, ir_variable
*consumer_var
)
1640 assert(producer_var
!= NULL
|| consumer_var
!= NULL
);
1642 if ((producer_var
&& (!producer_var
->data
.is_unmatched_generic_inout
||
1643 producer_var
->data
.explicit_location
)) ||
1644 (consumer_var
&& (!consumer_var
->data
.is_unmatched_generic_inout
||
1645 consumer_var
->data
.explicit_location
))) {
1646 /* Either a location already exists for this variable (since it is part
1647 * of fixed functionality), or it has already been recorded as part of a
1653 bool needs_flat_qualifier
= consumer_var
== NULL
&&
1654 (producer_var
->type
->contains_integer() ||
1655 producer_var
->type
->contains_double());
1657 if (!disable_varying_packing
&&
1658 (needs_flat_qualifier
||
1659 (consumer_stage
!= MESA_SHADER_NONE
&& consumer_stage
!= MESA_SHADER_FRAGMENT
))) {
1660 /* Since this varying is not being consumed by the fragment shader, its
1661 * interpolation type varying cannot possibly affect rendering.
1662 * Also, this variable is non-flat and is (or contains) an integer
1664 * If the consumer stage is unknown, don't modify the interpolation
1665 * type as it could affect rendering later with separate shaders.
1667 * lower_packed_varyings requires all integer varyings to flat,
1668 * regardless of where they appear. We can trivially satisfy that
1669 * requirement by changing the interpolation type to flat here.
1672 producer_var
->data
.centroid
= false;
1673 producer_var
->data
.sample
= false;
1674 producer_var
->data
.interpolation
= INTERP_MODE_FLAT
;
1678 consumer_var
->data
.centroid
= false;
1679 consumer_var
->data
.sample
= false;
1680 consumer_var
->data
.interpolation
= INTERP_MODE_FLAT
;
1684 if (this->num_matches
== this->matches_capacity
) {
1685 this->matches_capacity
*= 2;
1686 this->matches
= (match
*)
1687 realloc(this->matches
,
1688 sizeof(*this->matches
) * this->matches_capacity
);
1691 /* We must use the consumer to compute the packing class because in GL4.4+
1692 * there is no guarantee interpolation qualifiers will match across stages.
1694 * From Section 4.5 (Interpolation Qualifiers) of the GLSL 4.30 spec:
1696 * "The type and presence of interpolation qualifiers of variables with
1697 * the same name declared in all linked shaders for the same cross-stage
1698 * interface must match, otherwise the link command will fail.
1700 * When comparing an output from one stage to an input of a subsequent
1701 * stage, the input and output don't match if their interpolation
1702 * qualifiers (or lack thereof) are not the same."
1704 * This text was also in at least revison 7 of the 4.40 spec but is no
1705 * longer in revision 9 and not in the 4.50 spec.
1707 const ir_variable
*const var
= (consumer_var
!= NULL
)
1708 ? consumer_var
: producer_var
;
1709 const gl_shader_stage stage
= (consumer_var
!= NULL
)
1710 ? consumer_stage
: producer_stage
;
1711 const glsl_type
*type
= get_varying_type(var
, stage
);
1713 if (producer_var
&& consumer_var
&&
1714 consumer_var
->data
.must_be_shader_input
) {
1715 producer_var
->data
.must_be_shader_input
= 1;
1718 this->matches
[this->num_matches
].packing_class
1719 = this->compute_packing_class(var
);
1720 this->matches
[this->num_matches
].packing_order
1721 = this->compute_packing_order(var
);
1722 if ((this->disable_varying_packing
&& !is_varying_packing_safe(type
, var
)) ||
1723 var
->data
.must_be_shader_input
) {
1724 unsigned slots
= type
->count_attribute_slots(false);
1725 this->matches
[this->num_matches
].num_components
= slots
* 4;
1727 this->matches
[this->num_matches
].num_components
1728 = type
->component_slots();
1731 this->matches
[this->num_matches
].producer_var
= producer_var
;
1732 this->matches
[this->num_matches
].consumer_var
= consumer_var
;
1733 this->num_matches
++;
1735 producer_var
->data
.is_unmatched_generic_inout
= 0;
1737 consumer_var
->data
.is_unmatched_generic_inout
= 0;
1742 * Choose locations for all of the variable matches that were previously
1743 * passed to varying_matches::record().
1746 varying_matches::assign_locations(struct gl_shader_program
*prog
,
1747 uint8_t *components
,
1748 uint64_t reserved_slots
)
1750 /* If packing has been disabled then we cannot safely sort the varyings by
1751 * class as it may mean we are using a version of OpenGL where
1752 * interpolation qualifiers are not guaranteed to be matching across
1753 * shaders, sorting in this case could result in mismatching shader
1755 * When packing is disabled the sort orders varyings used by transform
1756 * feedback first, but also depends on *undefined behaviour* of qsort to
1757 * reverse the order of the varyings. See: xfb_comparator().
1759 if (!this->disable_varying_packing
) {
1760 /* Sort varying matches into an order that makes them easy to pack. */
1761 qsort(this->matches
, this->num_matches
, sizeof(*this->matches
),
1762 &varying_matches::match_comparator
);
1764 /* Only sort varyings that are only used by transform feedback. */
1765 qsort(this->matches
, this->num_matches
, sizeof(*this->matches
),
1766 &varying_matches::xfb_comparator
);
1769 unsigned generic_location
= 0;
1770 unsigned generic_patch_location
= MAX_VARYING
*4;
1771 bool previous_var_xfb_only
= false;
1773 for (unsigned i
= 0; i
< this->num_matches
; i
++) {
1774 unsigned *location
= &generic_location
;
1776 const ir_variable
*var
;
1777 const glsl_type
*type
;
1778 bool is_vertex_input
= false;
1779 if (matches
[i
].consumer_var
) {
1780 var
= matches
[i
].consumer_var
;
1781 type
= get_varying_type(var
, consumer_stage
);
1782 if (consumer_stage
== MESA_SHADER_VERTEX
)
1783 is_vertex_input
= true;
1785 var
= matches
[i
].producer_var
;
1786 type
= get_varying_type(var
, producer_stage
);
1789 if (var
->data
.patch
)
1790 location
= &generic_patch_location
;
1792 /* Advance to the next slot if this varying has a different packing
1793 * class than the previous one, and we're not already on a slot
1796 * Also advance to the next slot if packing is disabled. This makes sure
1797 * we don't assign varyings the same locations which is possible
1798 * because we still pack individual arrays, records and matrices even
1799 * when packing is disabled. Note we don't advance to the next slot if
1800 * we can pack varyings together that are only used for transform
1803 if (var
->data
.must_be_shader_input
||
1804 (this->disable_varying_packing
&&
1805 !(previous_var_xfb_only
&& var
->data
.is_xfb_only
)) ||
1806 (i
> 0 && this->matches
[i
- 1].packing_class
1807 != this->matches
[i
].packing_class
)) {
1808 *location
= ALIGN(*location
, 4);
1811 previous_var_xfb_only
= var
->data
.is_xfb_only
;
1813 /* The number of components taken up by this variable. For vertex shader
1814 * inputs, we use the number of slots * 4, as they have different
1817 unsigned num_components
= is_vertex_input
?
1818 type
->count_attribute_slots(is_vertex_input
) * 4 :
1819 this->matches
[i
].num_components
;
1821 /* The last slot for this variable, inclusive. */
1822 unsigned slot_end
= *location
+ num_components
- 1;
1824 /* FIXME: We could be smarter in the below code and loop back over
1825 * trying to fill any locations that we skipped because we couldn't pack
1826 * the varying between an explicit location. For now just let the user
1827 * hit the linking error if we run out of room and suggest they use
1828 * explicit locations.
1830 while (slot_end
< MAX_VARYING
* 4u) {
1831 const unsigned slots
= (slot_end
/ 4u) - (*location
/ 4u) + 1;
1832 const uint64_t slot_mask
= ((1ull << slots
) - 1) << (*location
/ 4u);
1835 if (reserved_slots
& slot_mask
) {
1836 *location
= ALIGN(*location
+ 1, 4);
1837 slot_end
= *location
+ num_components
- 1;
1844 if (!var
->data
.patch
&& slot_end
>= MAX_VARYING
* 4u) {
1845 linker_error(prog
, "insufficient contiguous locations available for "
1846 "%s it is possible an array or struct could not be "
1847 "packed between varyings with explicit locations. Try "
1848 "using an explicit location for arrays and structs.",
1852 if (slot_end
< MAX_VARYINGS_INCL_PATCH
* 4u) {
1853 for (unsigned j
= *location
/ 4u; j
< slot_end
/ 4u; j
++)
1855 components
[slot_end
/ 4u] = (slot_end
& 3) + 1;
1858 this->matches
[i
].generic_location
= *location
;
1860 *location
= slot_end
+ 1;
1863 return (generic_location
+ 3) / 4;
1868 * Update the producer and consumer shaders to reflect the locations
1869 * assignments that were made by varying_matches::assign_locations().
1872 varying_matches::store_locations() const
1874 /* Check is location needs to be packed with lower_packed_varyings() or if
1875 * we can just use ARB_enhanced_layouts packing.
1877 bool pack_loc
[MAX_VARYINGS_INCL_PATCH
] = { 0 };
1878 const glsl_type
*loc_type
[MAX_VARYINGS_INCL_PATCH
][4] = { {NULL
, NULL
} };
1880 for (unsigned i
= 0; i
< this->num_matches
; i
++) {
1881 ir_variable
*producer_var
= this->matches
[i
].producer_var
;
1882 ir_variable
*consumer_var
= this->matches
[i
].consumer_var
;
1883 unsigned generic_location
= this->matches
[i
].generic_location
;
1884 unsigned slot
= generic_location
/ 4;
1885 unsigned offset
= generic_location
% 4;
1888 producer_var
->data
.location
= VARYING_SLOT_VAR0
+ slot
;
1889 producer_var
->data
.location_frac
= offset
;
1893 assert(consumer_var
->data
.location
== -1);
1894 consumer_var
->data
.location
= VARYING_SLOT_VAR0
+ slot
;
1895 consumer_var
->data
.location_frac
= offset
;
1898 /* Find locations suitable for native packing via
1899 * ARB_enhanced_layouts.
1901 if (producer_var
&& consumer_var
) {
1902 if (enhanced_layouts_enabled
) {
1903 const glsl_type
*type
=
1904 get_varying_type(producer_var
, producer_stage
);
1905 if (type
->is_array() || type
->is_matrix() || type
->is_record() ||
1906 type
->is_double()) {
1907 unsigned comp_slots
= type
->component_slots() + offset
;
1908 unsigned slots
= comp_slots
/ 4;
1912 for (unsigned j
= 0; j
< slots
; j
++) {
1913 pack_loc
[slot
+ j
] = true;
1915 } else if (offset
+ type
->vector_elements
> 4) {
1916 pack_loc
[slot
] = true;
1917 pack_loc
[slot
+ 1] = true;
1919 loc_type
[slot
][offset
] = type
;
1925 /* Attempt to use ARB_enhanced_layouts for more efficient packing if
1928 if (enhanced_layouts_enabled
) {
1929 for (unsigned i
= 0; i
< this->num_matches
; i
++) {
1930 ir_variable
*producer_var
= this->matches
[i
].producer_var
;
1931 ir_variable
*consumer_var
= this->matches
[i
].consumer_var
;
1932 unsigned generic_location
= this->matches
[i
].generic_location
;
1933 unsigned slot
= generic_location
/ 4;
1935 if (pack_loc
[slot
] || !producer_var
|| !consumer_var
)
1938 const glsl_type
*type
=
1939 get_varying_type(producer_var
, producer_stage
);
1940 bool type_match
= true;
1941 for (unsigned j
= 0; j
< 4; j
++) {
1942 if (loc_type
[slot
][j
]) {
1943 if (type
->base_type
!= loc_type
[slot
][j
]->base_type
)
1949 producer_var
->data
.explicit_location
= 1;
1950 consumer_var
->data
.explicit_location
= 1;
1951 producer_var
->data
.explicit_component
= 1;
1952 consumer_var
->data
.explicit_component
= 1;
1960 * Compute the "packing class" of the given varying. This is an unsigned
1961 * integer with the property that two variables in the same packing class can
1962 * be safely backed into the same vec4.
1965 varying_matches::compute_packing_class(const ir_variable
*var
)
1967 /* Without help from the back-end, there is no way to pack together
1968 * variables with different interpolation types, because
1969 * lower_packed_varyings must choose exactly one interpolation type for
1970 * each packed varying it creates.
1972 * However, we can safely pack together floats, ints, and uints, because:
1974 * - varyings of base type "int" and "uint" must use the "flat"
1975 * interpolation type, which can only occur in GLSL 1.30 and above.
1977 * - On platforms that support GLSL 1.30 and above, lower_packed_varyings
1978 * can store flat floats as ints without losing any information (using
1979 * the ir_unop_bitcast_* opcodes).
1981 * Therefore, the packing class depends only on the interpolation type.
1983 unsigned packing_class
= var
->data
.centroid
| (var
->data
.sample
<< 1) |
1984 (var
->data
.patch
<< 2) |
1985 (var
->data
.must_be_shader_input
<< 3);
1987 packing_class
+= var
->is_interpolation_flat()
1988 ? unsigned(INTERP_MODE_FLAT
) : var
->data
.interpolation
;
1989 return packing_class
;
1994 * Compute the "packing order" of the given varying. This is a sort key we
1995 * use to determine when to attempt to pack the given varying relative to
1996 * other varyings in the same packing class.
1998 varying_matches::packing_order_enum
1999 varying_matches::compute_packing_order(const ir_variable
*var
)
2001 const glsl_type
*element_type
= var
->type
;
2003 while (element_type
->is_array()) {
2004 element_type
= element_type
->fields
.array
;
2007 switch (element_type
->component_slots() % 4) {
2008 case 1: return PACKING_ORDER_SCALAR
;
2009 case 2: return PACKING_ORDER_VEC2
;
2010 case 3: return PACKING_ORDER_VEC3
;
2011 case 0: return PACKING_ORDER_VEC4
;
2013 assert(!"Unexpected value of vector_elements");
2014 return PACKING_ORDER_VEC4
;
2020 * Comparison function passed to qsort() to sort varyings by packing_class and
2021 * then by packing_order.
2024 varying_matches::match_comparator(const void *x_generic
, const void *y_generic
)
2026 const match
*x
= (const match
*) x_generic
;
2027 const match
*y
= (const match
*) y_generic
;
2029 if (x
->packing_class
!= y
->packing_class
)
2030 return x
->packing_class
- y
->packing_class
;
2031 return x
->packing_order
- y
->packing_order
;
2036 * Comparison function passed to qsort() to sort varyings used only by
2037 * transform feedback when packing of other varyings is disabled.
2040 varying_matches::xfb_comparator(const void *x_generic
, const void *y_generic
)
2042 const match
*x
= (const match
*) x_generic
;
2044 if (x
->producer_var
!= NULL
&& x
->producer_var
->data
.is_xfb_only
)
2045 return match_comparator(x_generic
, y_generic
);
2047 /* FIXME: When the comparator returns 0 it means the elements being
2048 * compared are equivalent. However the qsort documentation says:
2050 * "The order of equivalent elements is undefined."
2052 * In practice the sort ends up reversing the order of the varyings which
2053 * means locations are also assigned in this reversed order and happens to
2054 * be what we want. This is also whats happening in
2055 * varying_matches::match_comparator().
2062 * Is the given variable a varying variable to be counted against the
2063 * limit in ctx->Const.MaxVarying?
2064 * This includes variables such as texcoords, colors and generic
2065 * varyings, but excludes variables such as gl_FrontFacing and gl_FragCoord.
2068 var_counts_against_varying_limit(gl_shader_stage stage
, const ir_variable
*var
)
2070 /* Only fragment shaders will take a varying variable as an input */
2071 if (stage
== MESA_SHADER_FRAGMENT
&&
2072 var
->data
.mode
== ir_var_shader_in
) {
2073 switch (var
->data
.location
) {
2074 case VARYING_SLOT_POS
:
2075 case VARYING_SLOT_FACE
:
2076 case VARYING_SLOT_PNTC
:
2087 * Visitor class that generates tfeedback_candidate structs describing all
2088 * possible targets of transform feedback.
2090 * tfeedback_candidate structs are stored in the hash table
2091 * tfeedback_candidates, which is passed to the constructor. This hash table
2092 * maps varying names to instances of the tfeedback_candidate struct.
2094 class tfeedback_candidate_generator
: public program_resource_visitor
2097 tfeedback_candidate_generator(void *mem_ctx
,
2098 hash_table
*tfeedback_candidates
)
2100 tfeedback_candidates(tfeedback_candidates
),
2106 void process(ir_variable
*var
)
2108 /* All named varying interface blocks should be flattened by now */
2109 assert(!var
->is_interface_instance());
2111 this->toplevel_var
= var
;
2112 this->varying_floats
= 0;
2113 program_resource_visitor::process(var
, false);
2117 virtual void visit_field(const glsl_type
*type
, const char *name
,
2118 bool /* row_major */,
2119 const glsl_type
* /* record_type */,
2120 const enum glsl_interface_packing
,
2121 bool /* last_field */)
2123 assert(!type
->without_array()->is_record());
2124 assert(!type
->without_array()->is_interface());
2126 tfeedback_candidate
*candidate
2127 = rzalloc(this->mem_ctx
, tfeedback_candidate
);
2128 candidate
->toplevel_var
= this->toplevel_var
;
2129 candidate
->type
= type
;
2130 candidate
->offset
= this->varying_floats
;
2131 _mesa_hash_table_insert(this->tfeedback_candidates
,
2132 ralloc_strdup(this->mem_ctx
, name
),
2134 this->varying_floats
+= type
->component_slots();
2138 * Memory context used to allocate hash table keys and values.
2140 void * const mem_ctx
;
2143 * Hash table in which tfeedback_candidate objects should be stored.
2145 hash_table
* const tfeedback_candidates
;
2148 * Pointer to the toplevel variable that is being traversed.
2150 ir_variable
*toplevel_var
;
2153 * Total number of varying floats that have been visited so far. This is
2154 * used to determine the offset to each varying within the toplevel
2157 unsigned varying_floats
;
2164 populate_consumer_input_sets(void *mem_ctx
, exec_list
*ir
,
2165 hash_table
*consumer_inputs
,
2166 hash_table
*consumer_interface_inputs
,
2167 ir_variable
*consumer_inputs_with_locations
[VARYING_SLOT_TESS_MAX
])
2169 memset(consumer_inputs_with_locations
,
2171 sizeof(consumer_inputs_with_locations
[0]) * VARYING_SLOT_TESS_MAX
);
2173 foreach_in_list(ir_instruction
, node
, ir
) {
2174 ir_variable
*const input_var
= node
->as_variable();
2176 if (input_var
!= NULL
&& input_var
->data
.mode
== ir_var_shader_in
) {
2177 /* All interface blocks should have been lowered by this point */
2178 assert(!input_var
->type
->is_interface());
2180 if (input_var
->data
.explicit_location
) {
2181 /* assign_varying_locations only cares about finding the
2182 * ir_variable at the start of a contiguous location block.
2184 * - For !producer, consumer_inputs_with_locations isn't used.
2186 * - For !consumer, consumer_inputs_with_locations is empty.
2188 * For consumer && producer, if you were trying to set some
2189 * ir_variable to the middle of a location block on the other side
2190 * of producer/consumer, cross_validate_outputs_to_inputs() should
2191 * be link-erroring due to either type mismatch or location
2192 * overlaps. If the variables do match up, then they've got a
2193 * matching data.location and you only looked at
2194 * consumer_inputs_with_locations[var->data.location], not any
2195 * following entries for the array/structure.
2197 consumer_inputs_with_locations
[input_var
->data
.location
] =
2199 } else if (input_var
->get_interface_type() != NULL
) {
2200 char *const iface_field_name
=
2201 ralloc_asprintf(mem_ctx
, "%s.%s",
2202 input_var
->get_interface_type()->without_array()->name
,
2204 _mesa_hash_table_insert(consumer_interface_inputs
,
2205 iface_field_name
, input_var
);
2207 _mesa_hash_table_insert(consumer_inputs
,
2208 ralloc_strdup(mem_ctx
, input_var
->name
),
2216 * Find a variable from the consumer that "matches" the specified variable
2218 * This function only finds inputs with names that match. There is no
2219 * validation (here) that the types, etc. are compatible.
2222 get_matching_input(void *mem_ctx
,
2223 const ir_variable
*output_var
,
2224 hash_table
*consumer_inputs
,
2225 hash_table
*consumer_interface_inputs
,
2226 ir_variable
*consumer_inputs_with_locations
[VARYING_SLOT_TESS_MAX
])
2228 ir_variable
*input_var
;
2230 if (output_var
->data
.explicit_location
) {
2231 input_var
= consumer_inputs_with_locations
[output_var
->data
.location
];
2232 } else if (output_var
->get_interface_type() != NULL
) {
2233 char *const iface_field_name
=
2234 ralloc_asprintf(mem_ctx
, "%s.%s",
2235 output_var
->get_interface_type()->without_array()->name
,
2237 hash_entry
*entry
= _mesa_hash_table_search(consumer_interface_inputs
, iface_field_name
);
2238 input_var
= entry
? (ir_variable
*) entry
->data
: NULL
;
2240 hash_entry
*entry
= _mesa_hash_table_search(consumer_inputs
, output_var
->name
);
2241 input_var
= entry
? (ir_variable
*) entry
->data
: NULL
;
2244 return (input_var
== NULL
|| input_var
->data
.mode
!= ir_var_shader_in
)
2251 io_variable_cmp(const void *_a
, const void *_b
)
2253 const ir_variable
*const a
= *(const ir_variable
**) _a
;
2254 const ir_variable
*const b
= *(const ir_variable
**) _b
;
2256 if (a
->data
.explicit_location
&& b
->data
.explicit_location
)
2257 return b
->data
.location
- a
->data
.location
;
2259 if (a
->data
.explicit_location
&& !b
->data
.explicit_location
)
2262 if (!a
->data
.explicit_location
&& b
->data
.explicit_location
)
2265 return -strcmp(a
->name
, b
->name
);
2269 * Sort the shader IO variables into canonical order
2272 canonicalize_shader_io(exec_list
*ir
, enum ir_variable_mode io_mode
)
2274 ir_variable
*var_table
[MAX_PROGRAM_OUTPUTS
* 4];
2275 unsigned num_variables
= 0;
2277 foreach_in_list(ir_instruction
, node
, ir
) {
2278 ir_variable
*const var
= node
->as_variable();
2280 if (var
== NULL
|| var
->data
.mode
!= io_mode
)
2283 /* If we have already encountered more I/O variables that could
2284 * successfully link, bail.
2286 if (num_variables
== ARRAY_SIZE(var_table
))
2289 var_table
[num_variables
++] = var
;
2292 if (num_variables
== 0)
2295 /* Sort the list in reverse order (io_variable_cmp handles this). Later
2296 * we're going to push the variables on to the IR list as a stack, so we
2297 * want the last variable (in canonical order) to be first in the list.
2299 qsort(var_table
, num_variables
, sizeof(var_table
[0]), io_variable_cmp
);
2301 /* Remove the variable from it's current location in the IR, and put it at
2304 for (unsigned i
= 0; i
< num_variables
; i
++) {
2305 var_table
[i
]->remove();
2306 ir
->push_head(var_table
[i
]);
2311 * Generate a bitfield map of the explicit locations for shader varyings.
2313 * Note: For Tessellation shaders we are sitting right on the limits of the
2314 * 64 bit map. Per-vertex and per-patch both have separate location domains
2315 * with a max of MAX_VARYING.
2318 reserved_varying_slot(struct gl_linked_shader
*stage
,
2319 ir_variable_mode io_mode
)
2321 assert(io_mode
== ir_var_shader_in
|| io_mode
== ir_var_shader_out
);
2322 /* Avoid an overflow of the returned value */
2323 assert(MAX_VARYINGS_INCL_PATCH
<= 64);
2331 foreach_in_list(ir_instruction
, node
, stage
->ir
) {
2332 ir_variable
*const var
= node
->as_variable();
2334 if (var
== NULL
|| var
->data
.mode
!= io_mode
||
2335 !var
->data
.explicit_location
||
2336 var
->data
.location
< VARYING_SLOT_VAR0
)
2339 var_slot
= var
->data
.location
- VARYING_SLOT_VAR0
;
2341 unsigned num_elements
= get_varying_type(var
, stage
->Stage
)
2342 ->count_attribute_slots(io_mode
== ir_var_shader_in
&&
2343 stage
->Stage
== MESA_SHADER_VERTEX
);
2344 for (unsigned i
= 0; i
< num_elements
; i
++) {
2345 if (var_slot
>= 0 && var_slot
< MAX_VARYINGS_INCL_PATCH
)
2346 slots
|= UINT64_C(1) << var_slot
;
2356 * Assign locations for all variables that are produced in one pipeline stage
2357 * (the "producer") and consumed in the next stage (the "consumer").
2359 * Variables produced by the producer may also be consumed by transform
2362 * \param num_tfeedback_decls is the number of declarations indicating
2363 * variables that may be consumed by transform feedback.
2365 * \param tfeedback_decls is a pointer to an array of tfeedback_decl objects
2366 * representing the result of parsing the strings passed to
2367 * glTransformFeedbackVaryings(). assign_location() will be called for
2368 * each of these objects that matches one of the outputs of the
2371 * When num_tfeedback_decls is nonzero, it is permissible for the consumer to
2372 * be NULL. In this case, varying locations are assigned solely based on the
2373 * requirements of transform feedback.
2376 assign_varying_locations(struct gl_context
*ctx
,
2378 struct gl_shader_program
*prog
,
2379 gl_linked_shader
*producer
,
2380 gl_linked_shader
*consumer
,
2381 unsigned num_tfeedback_decls
,
2382 tfeedback_decl
*tfeedback_decls
,
2383 const uint64_t reserved_slots
)
2385 /* Tessellation shaders treat inputs and outputs as shared memory and can
2386 * access inputs and outputs of other invocations.
2387 * Therefore, they can't be lowered to temps easily (and definitely not
2390 bool unpackable_tess
=
2391 (consumer
&& consumer
->Stage
== MESA_SHADER_TESS_EVAL
) ||
2392 (consumer
&& consumer
->Stage
== MESA_SHADER_TESS_CTRL
) ||
2393 (producer
&& producer
->Stage
== MESA_SHADER_TESS_CTRL
);
2395 /* Transform feedback code assumes varying arrays are packed, so if the
2396 * driver has disabled varying packing, make sure to at least enable
2397 * packing required by transform feedback.
2400 ctx
->Extensions
.EXT_transform_feedback
&& !unpackable_tess
;
2402 /* Disable packing on outward facing interfaces for SSO because in ES we
2403 * need to retain the unpacked varying information for draw time
2406 * Packing is still enabled on individual arrays, structs, and matrices as
2407 * these are required by the transform feedback code and it is still safe
2408 * to do so. We also enable packing when a varying is only used for
2409 * transform feedback and its not a SSO.
2411 bool disable_varying_packing
=
2412 ctx
->Const
.DisableVaryingPacking
|| unpackable_tess
;
2413 if (prog
->SeparateShader
&& (producer
== NULL
|| consumer
== NULL
))
2414 disable_varying_packing
= true;
2416 varying_matches
matches(disable_varying_packing
, xfb_enabled
,
2417 ctx
->Extensions
.ARB_enhanced_layouts
,
2418 producer
? producer
->Stage
: MESA_SHADER_NONE
,
2419 consumer
? consumer
->Stage
: MESA_SHADER_NONE
);
2420 hash_table
*tfeedback_candidates
=
2421 _mesa_hash_table_create(NULL
, _mesa_key_hash_string
,
2422 _mesa_key_string_equal
);
2423 hash_table
*consumer_inputs
=
2424 _mesa_hash_table_create(NULL
, _mesa_key_hash_string
,
2425 _mesa_key_string_equal
);
2426 hash_table
*consumer_interface_inputs
=
2427 _mesa_hash_table_create(NULL
, _mesa_key_hash_string
,
2428 _mesa_key_string_equal
);
2429 ir_variable
*consumer_inputs_with_locations
[VARYING_SLOT_TESS_MAX
] = {
2433 unsigned consumer_vertices
= 0;
2434 if (consumer
&& consumer
->Stage
== MESA_SHADER_GEOMETRY
)
2435 consumer_vertices
= prog
->Geom
.VerticesIn
;
2437 /* Operate in a total of four passes.
2439 * 1. Sort inputs / outputs into a canonical order. This is necessary so
2440 * that inputs / outputs of separable shaders will be assigned
2441 * predictable locations regardless of the order in which declarations
2442 * appeared in the shader source.
2444 * 2. Assign locations for any matching inputs and outputs.
2446 * 3. Mark output variables in the producer that do not have locations as
2447 * not being outputs. This lets the optimizer eliminate them.
2449 * 4. Mark input variables in the consumer that do not have locations as
2450 * not being inputs. This lets the optimizer eliminate them.
2453 canonicalize_shader_io(consumer
->ir
, ir_var_shader_in
);
2456 canonicalize_shader_io(producer
->ir
, ir_var_shader_out
);
2459 linker::populate_consumer_input_sets(mem_ctx
, consumer
->ir
,
2461 consumer_interface_inputs
,
2462 consumer_inputs_with_locations
);
2465 foreach_in_list(ir_instruction
, node
, producer
->ir
) {
2466 ir_variable
*const output_var
= node
->as_variable();
2468 if (output_var
== NULL
|| output_var
->data
.mode
!= ir_var_shader_out
)
2471 /* Only geometry shaders can use non-zero streams */
2472 assert(output_var
->data
.stream
== 0 ||
2473 (output_var
->data
.stream
< MAX_VERTEX_STREAMS
&&
2474 producer
->Stage
== MESA_SHADER_GEOMETRY
));
2476 if (num_tfeedback_decls
> 0) {
2477 tfeedback_candidate_generator
g(mem_ctx
, tfeedback_candidates
);
2478 g
.process(output_var
);
2481 ir_variable
*const input_var
=
2482 linker::get_matching_input(mem_ctx
, output_var
, consumer_inputs
,
2483 consumer_interface_inputs
,
2484 consumer_inputs_with_locations
);
2486 /* If a matching input variable was found, add this output (and the
2487 * input) to the set. If this is a separable program and there is no
2488 * consumer stage, add the output.
2490 * Always add TCS outputs. They are shared by all invocations
2491 * within a patch and can be used as shared memory.
2493 if (input_var
|| (prog
->SeparateShader
&& consumer
== NULL
) ||
2494 producer
->Stage
== MESA_SHADER_TESS_CTRL
) {
2495 matches
.record(output_var
, input_var
);
2498 /* Only stream 0 outputs can be consumed in the next stage */
2499 if (input_var
&& output_var
->data
.stream
!= 0) {
2500 linker_error(prog
, "output %s is assigned to stream=%d but "
2501 "is linked to an input, which requires stream=0",
2502 output_var
->name
, output_var
->data
.stream
);
2507 /* If there's no producer stage, then this must be a separable program.
2508 * For example, we may have a program that has just a fragment shader.
2509 * Later this program will be used with some arbitrary vertex (or
2510 * geometry) shader program. This means that locations must be assigned
2511 * for all the inputs.
2513 foreach_in_list(ir_instruction
, node
, consumer
->ir
) {
2514 ir_variable
*const input_var
= node
->as_variable();
2516 if (input_var
== NULL
|| input_var
->data
.mode
!= ir_var_shader_in
)
2519 matches
.record(NULL
, input_var
);
2523 for (unsigned i
= 0; i
< num_tfeedback_decls
; ++i
) {
2524 if (!tfeedback_decls
[i
].is_varying())
2527 const tfeedback_candidate
*matched_candidate
2528 = tfeedback_decls
[i
].find_candidate(prog
, tfeedback_candidates
);
2530 if (matched_candidate
== NULL
) {
2531 _mesa_hash_table_destroy(tfeedback_candidates
, NULL
);
2535 /* Mark xfb varyings as always active */
2536 matched_candidate
->toplevel_var
->data
.always_active_io
= 1;
2538 /* Mark any corresponding inputs as always active also. We must do this
2539 * because we have a NIR pass that lowers vectors to scalars and another
2540 * that removes unused varyings.
2541 * We don't split varyings marked as always active because there is no
2542 * point in doing so. This means we need to mark both sides of the
2543 * interface as always active otherwise we will have a mismatch and
2544 * start removing things we shouldn't.
2546 ir_variable
*const input_var
=
2547 linker::get_matching_input(mem_ctx
, matched_candidate
->toplevel_var
,
2549 consumer_interface_inputs
,
2550 consumer_inputs_with_locations
);
2552 input_var
->data
.always_active_io
= 1;
2554 if (matched_candidate
->toplevel_var
->data
.is_unmatched_generic_inout
) {
2555 matched_candidate
->toplevel_var
->data
.is_xfb_only
= 1;
2556 matches
.record(matched_candidate
->toplevel_var
, NULL
);
2560 _mesa_hash_table_destroy(consumer_inputs
, NULL
);
2561 _mesa_hash_table_destroy(consumer_interface_inputs
, NULL
);
2563 uint8_t components
[MAX_VARYINGS_INCL_PATCH
] = {0};
2564 const unsigned slots_used
= matches
.assign_locations(
2565 prog
, components
, reserved_slots
);
2566 matches
.store_locations();
2568 for (unsigned i
= 0; i
< num_tfeedback_decls
; ++i
) {
2569 if (!tfeedback_decls
[i
].is_varying())
2572 if (!tfeedback_decls
[i
].assign_location(ctx
, prog
)) {
2573 _mesa_hash_table_destroy(tfeedback_candidates
, NULL
);
2577 _mesa_hash_table_destroy(tfeedback_candidates
, NULL
);
2579 if (consumer
&& producer
) {
2580 foreach_in_list(ir_instruction
, node
, consumer
->ir
) {
2581 ir_variable
*const var
= node
->as_variable();
2583 if (var
&& var
->data
.mode
== ir_var_shader_in
&&
2584 var
->data
.is_unmatched_generic_inout
) {
2585 if (!prog
->IsES
&& prog
->data
->Version
<= 120) {
2586 /* On page 25 (page 31 of the PDF) of the GLSL 1.20 spec:
2588 * Only those varying variables used (i.e. read) in
2589 * the fragment shader executable must be written to
2590 * by the vertex shader executable; declaring
2591 * superfluous varying variables in a vertex shader is
2594 * We interpret this text as meaning that the VS must
2595 * write the variable for the FS to read it. See
2596 * "glsl1-varying read but not written" in piglit.
2598 linker_error(prog
, "%s shader varying %s not written "
2600 _mesa_shader_stage_to_string(consumer
->Stage
),
2602 _mesa_shader_stage_to_string(producer
->Stage
));
2604 linker_warning(prog
, "%s shader varying %s not written "
2606 _mesa_shader_stage_to_string(consumer
->Stage
),
2608 _mesa_shader_stage_to_string(producer
->Stage
));
2613 /* Now that validation is done its safe to remove unused varyings. As
2614 * we have both a producer and consumer its safe to remove unused
2615 * varyings even if the program is a SSO because the stages are being
2616 * linked together i.e. we have a multi-stage SSO.
2618 remove_unused_shader_inputs_and_outputs(false, producer
,
2620 remove_unused_shader_inputs_and_outputs(false, consumer
,
2625 lower_packed_varyings(mem_ctx
, slots_used
, components
, ir_var_shader_out
,
2626 0, producer
, disable_varying_packing
,
2631 lower_packed_varyings(mem_ctx
, slots_used
, components
, ir_var_shader_in
,
2632 consumer_vertices
, consumer
,
2633 disable_varying_packing
, xfb_enabled
);
2640 check_against_output_limit(struct gl_context
*ctx
,
2641 struct gl_shader_program
*prog
,
2642 gl_linked_shader
*producer
,
2643 unsigned num_explicit_locations
)
2645 unsigned output_vectors
= num_explicit_locations
;
2647 foreach_in_list(ir_instruction
, node
, producer
->ir
) {
2648 ir_variable
*const var
= node
->as_variable();
2650 if (var
&& !var
->data
.explicit_location
&&
2651 var
->data
.mode
== ir_var_shader_out
&&
2652 var_counts_against_varying_limit(producer
->Stage
, var
)) {
2653 /* outputs for fragment shader can't be doubles */
2654 output_vectors
+= var
->type
->count_attribute_slots(false);
2658 assert(producer
->Stage
!= MESA_SHADER_FRAGMENT
);
2659 unsigned max_output_components
=
2660 ctx
->Const
.Program
[producer
->Stage
].MaxOutputComponents
;
2662 const unsigned output_components
= output_vectors
* 4;
2663 if (output_components
> max_output_components
) {
2664 if (ctx
->API
== API_OPENGLES2
|| prog
->IsES
)
2665 linker_error(prog
, "%s shader uses too many output vectors "
2667 _mesa_shader_stage_to_string(producer
->Stage
),
2669 max_output_components
/ 4);
2671 linker_error(prog
, "%s shader uses too many output components "
2673 _mesa_shader_stage_to_string(producer
->Stage
),
2675 max_output_components
);
2684 check_against_input_limit(struct gl_context
*ctx
,
2685 struct gl_shader_program
*prog
,
2686 gl_linked_shader
*consumer
,
2687 unsigned num_explicit_locations
)
2689 unsigned input_vectors
= num_explicit_locations
;
2691 foreach_in_list(ir_instruction
, node
, consumer
->ir
) {
2692 ir_variable
*const var
= node
->as_variable();
2694 if (var
&& !var
->data
.explicit_location
&&
2695 var
->data
.mode
== ir_var_shader_in
&&
2696 var_counts_against_varying_limit(consumer
->Stage
, var
)) {
2697 /* vertex inputs aren't varying counted */
2698 input_vectors
+= var
->type
->count_attribute_slots(false);
2702 assert(consumer
->Stage
!= MESA_SHADER_VERTEX
);
2703 unsigned max_input_components
=
2704 ctx
->Const
.Program
[consumer
->Stage
].MaxInputComponents
;
2706 const unsigned input_components
= input_vectors
* 4;
2707 if (input_components
> max_input_components
) {
2708 if (ctx
->API
== API_OPENGLES2
|| prog
->IsES
)
2709 linker_error(prog
, "%s shader uses too many input vectors "
2711 _mesa_shader_stage_to_string(consumer
->Stage
),
2713 max_input_components
/ 4);
2715 linker_error(prog
, "%s shader uses too many input components "
2717 _mesa_shader_stage_to_string(consumer
->Stage
),
2719 max_input_components
);
2728 link_varyings(struct gl_shader_program
*prog
, unsigned first
, unsigned last
,
2729 struct gl_context
*ctx
, void *mem_ctx
)
2731 bool has_xfb_qualifiers
= false;
2732 unsigned num_tfeedback_decls
= 0;
2733 char **varying_names
= NULL
;
2734 tfeedback_decl
*tfeedback_decls
= NULL
;
2736 /* From the ARB_enhanced_layouts spec:
2738 * "If the shader used to record output variables for transform feedback
2739 * varyings uses the "xfb_buffer", "xfb_offset", or "xfb_stride" layout
2740 * qualifiers, the values specified by TransformFeedbackVaryings are
2741 * ignored, and the set of variables captured for transform feedback is
2742 * instead derived from the specified layout qualifiers."
2744 for (int i
= MESA_SHADER_FRAGMENT
- 1; i
>= 0; i
--) {
2745 /* Find last stage before fragment shader */
2746 if (prog
->_LinkedShaders
[i
]) {
2747 has_xfb_qualifiers
=
2748 process_xfb_layout_qualifiers(mem_ctx
, prog
->_LinkedShaders
[i
],
2749 prog
, &num_tfeedback_decls
,
2755 if (!has_xfb_qualifiers
) {
2756 num_tfeedback_decls
= prog
->TransformFeedback
.NumVarying
;
2757 varying_names
= prog
->TransformFeedback
.VaryingNames
;
2760 if (num_tfeedback_decls
!= 0) {
2761 /* From GL_EXT_transform_feedback:
2762 * A program will fail to link if:
2764 * * the <count> specified by TransformFeedbackVaryingsEXT is
2765 * non-zero, but the program object has no vertex or geometry
2768 if (first
>= MESA_SHADER_FRAGMENT
) {
2769 linker_error(prog
, "Transform feedback varyings specified, but "
2770 "no vertex, tessellation, or geometry shader is "
2775 tfeedback_decls
= rzalloc_array(mem_ctx
, tfeedback_decl
,
2776 num_tfeedback_decls
);
2777 if (!parse_tfeedback_decls(ctx
, prog
, mem_ctx
, num_tfeedback_decls
,
2778 varying_names
, tfeedback_decls
))
2782 /* If there is no fragment shader we need to set transform feedback.
2784 * For SSO we also need to assign output locations. We assign them here
2785 * because we need to do it for both single stage programs and multi stage
2788 if (last
< MESA_SHADER_FRAGMENT
&&
2789 (num_tfeedback_decls
!= 0 || prog
->SeparateShader
)) {
2790 const uint64_t reserved_out_slots
=
2791 reserved_varying_slot(prog
->_LinkedShaders
[last
], ir_var_shader_out
);
2792 if (!assign_varying_locations(ctx
, mem_ctx
, prog
,
2793 prog
->_LinkedShaders
[last
], NULL
,
2794 num_tfeedback_decls
, tfeedback_decls
,
2795 reserved_out_slots
))
2799 if (last
<= MESA_SHADER_FRAGMENT
) {
2800 /* Remove unused varyings from the first/last stage unless SSO */
2801 remove_unused_shader_inputs_and_outputs(prog
->SeparateShader
,
2802 prog
->_LinkedShaders
[first
],
2804 remove_unused_shader_inputs_and_outputs(prog
->SeparateShader
,
2805 prog
->_LinkedShaders
[last
],
2808 /* If the program is made up of only a single stage */
2809 if (first
== last
) {
2810 gl_linked_shader
*const sh
= prog
->_LinkedShaders
[last
];
2812 do_dead_builtin_varyings(ctx
, NULL
, sh
, 0, NULL
);
2813 do_dead_builtin_varyings(ctx
, sh
, NULL
, num_tfeedback_decls
,
2816 if (prog
->SeparateShader
) {
2817 const uint64_t reserved_slots
=
2818 reserved_varying_slot(sh
, ir_var_shader_in
);
2820 /* Assign input locations for SSO, output locations are already
2823 if (!assign_varying_locations(ctx
, mem_ctx
, prog
,
2824 NULL
/* producer */,
2826 0 /* num_tfeedback_decls */,
2827 NULL
/* tfeedback_decls */,
2832 /* Linking the stages in the opposite order (from fragment to vertex)
2833 * ensures that inter-shader outputs written to in an earlier stage
2834 * are eliminated if they are (transitively) not used in a later
2838 for (int i
= next
- 1; i
>= 0; i
--) {
2839 if (prog
->_LinkedShaders
[i
] == NULL
&& i
!= 0)
2842 gl_linked_shader
*const sh_i
= prog
->_LinkedShaders
[i
];
2843 gl_linked_shader
*const sh_next
= prog
->_LinkedShaders
[next
];
2845 const uint64_t reserved_out_slots
=
2846 reserved_varying_slot(sh_i
, ir_var_shader_out
);
2847 const uint64_t reserved_in_slots
=
2848 reserved_varying_slot(sh_next
, ir_var_shader_in
);
2850 do_dead_builtin_varyings(ctx
, sh_i
, sh_next
,
2851 next
== MESA_SHADER_FRAGMENT
? num_tfeedback_decls
: 0,
2854 if (!assign_varying_locations(ctx
, mem_ctx
, prog
, sh_i
, sh_next
,
2855 next
== MESA_SHADER_FRAGMENT
? num_tfeedback_decls
: 0,
2857 reserved_out_slots
| reserved_in_slots
))
2860 /* This must be done after all dead varyings are eliminated. */
2862 unsigned slots_used
= _mesa_bitcount_64(reserved_out_slots
);
2863 if (!check_against_output_limit(ctx
, prog
, sh_i
, slots_used
)) {
2868 unsigned slots_used
= _mesa_bitcount_64(reserved_in_slots
);
2869 if (!check_against_input_limit(ctx
, prog
, sh_next
, slots_used
))
2877 if (!store_tfeedback_info(ctx
, prog
, num_tfeedback_decls
, tfeedback_decls
,
2878 has_xfb_qualifiers
))