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();
169 /* Find the member type before it was altered by lowering */
170 const glsl_type
*type_wa
= type
->without_array();
172 type_wa
->fields
.structure
[type_wa
->field_index(var
->name
)].type
;
173 name
= ralloc_strdup(NULL
, type_wa
->name
);
177 name
= ralloc_strdup(NULL
, var
->name
);
179 create_xfb_varying_names(mem_ctx
, type
, &name
, strlen(name
), &i
,
180 var
->name
, member_type
, varying_names
);
185 assert(i
== *num_tfeedback_decls
);
186 return has_xfb_qualifiers
;
190 * Validate the types and qualifiers of an output from one stage against the
191 * matching input to another stage.
194 cross_validate_types_and_qualifiers(struct gl_context
*ctx
,
195 struct gl_shader_program
*prog
,
196 const ir_variable
*input
,
197 const ir_variable
*output
,
198 gl_shader_stage consumer_stage
,
199 gl_shader_stage producer_stage
)
201 /* Check that the types match between stages.
203 const glsl_type
*type_to_match
= input
->type
;
205 /* VS -> GS, VS -> TCS, VS -> TES, TES -> GS */
206 const bool extra_array_level
= (producer_stage
== MESA_SHADER_VERTEX
&&
207 consumer_stage
!= MESA_SHADER_FRAGMENT
) ||
208 consumer_stage
== MESA_SHADER_GEOMETRY
;
209 if (extra_array_level
) {
210 assert(type_to_match
->is_array());
211 type_to_match
= type_to_match
->fields
.array
;
214 if (type_to_match
!= output
->type
) {
215 /* There is a bit of a special case for gl_TexCoord. This
216 * built-in is unsized by default. Applications that variable
217 * access it must redeclare it with a size. There is some
218 * language in the GLSL spec that implies the fragment shader
219 * and vertex shader do not have to agree on this size. Other
220 * driver behave this way, and one or two applications seem to
223 * Neither declaration needs to be modified here because the array
224 * sizes are fixed later when update_array_sizes is called.
226 * From page 48 (page 54 of the PDF) of the GLSL 1.10 spec:
228 * "Unlike user-defined varying variables, the built-in
229 * varying variables don't have a strict one-to-one
230 * correspondence between the vertex language and the
231 * fragment language."
233 if (!output
->type
->is_array() || !is_gl_identifier(output
->name
)) {
235 "%s shader output `%s' declared as type `%s', "
236 "but %s shader input declared as type `%s'\n",
237 _mesa_shader_stage_to_string(producer_stage
),
240 _mesa_shader_stage_to_string(consumer_stage
),
246 /* Check that all of the qualifiers match between stages.
249 /* According to the OpenGL and OpenGLES GLSL specs, the centroid qualifier
250 * should match until OpenGL 4.3 and OpenGLES 3.1. The OpenGLES 3.0
251 * conformance test suite does not verify that the qualifiers must match.
252 * The deqp test suite expects the opposite (OpenGLES 3.1) behavior for
253 * OpenGLES 3.0 drivers, so we relax the checking in all cases.
255 if (false /* always skip the centroid check */ &&
256 prog
->data
->Version
< (prog
->IsES
? 310 : 430) &&
257 input
->data
.centroid
!= output
->data
.centroid
) {
259 "%s shader output `%s' %s centroid qualifier, "
260 "but %s shader input %s centroid qualifier\n",
261 _mesa_shader_stage_to_string(producer_stage
),
263 (output
->data
.centroid
) ? "has" : "lacks",
264 _mesa_shader_stage_to_string(consumer_stage
),
265 (input
->data
.centroid
) ? "has" : "lacks");
269 if (input
->data
.sample
!= output
->data
.sample
) {
271 "%s shader output `%s' %s sample qualifier, "
272 "but %s shader input %s sample qualifier\n",
273 _mesa_shader_stage_to_string(producer_stage
),
275 (output
->data
.sample
) ? "has" : "lacks",
276 _mesa_shader_stage_to_string(consumer_stage
),
277 (input
->data
.sample
) ? "has" : "lacks");
281 if (input
->data
.patch
!= output
->data
.patch
) {
283 "%s shader output `%s' %s patch qualifier, "
284 "but %s shader input %s patch qualifier\n",
285 _mesa_shader_stage_to_string(producer_stage
),
287 (output
->data
.patch
) ? "has" : "lacks",
288 _mesa_shader_stage_to_string(consumer_stage
),
289 (input
->data
.patch
) ? "has" : "lacks");
293 /* The GLSL 4.30 and GLSL ES 3.00 specifications say:
295 * "As only outputs need be declared with invariant, an output from
296 * one shader stage will still match an input of a subsequent stage
297 * without the input being declared as invariant."
299 * while GLSL 4.20 says:
301 * "For variables leaving one shader and coming into another shader,
302 * the invariant keyword has to be used in both shaders, or a link
303 * error will result."
305 * and GLSL ES 1.00 section 4.6.4 "Invariance and Linking" says:
307 * "The invariance of varyings that are declared in both the vertex
308 * and fragment shaders must match."
310 if (input
->data
.invariant
!= output
->data
.invariant
&&
311 prog
->data
->Version
< (prog
->IsES
? 300 : 430)) {
313 "%s shader output `%s' %s invariant qualifier, "
314 "but %s shader input %s invariant qualifier\n",
315 _mesa_shader_stage_to_string(producer_stage
),
317 (output
->data
.invariant
) ? "has" : "lacks",
318 _mesa_shader_stage_to_string(consumer_stage
),
319 (input
->data
.invariant
) ? "has" : "lacks");
323 /* GLSL >= 4.40 removes text requiring interpolation qualifiers
324 * to match cross stage, they must only match within the same stage.
326 * From page 84 (page 90 of the PDF) of the GLSL 4.40 spec:
328 * "It is a link-time error if, within the same stage, the interpolation
329 * qualifiers of variables of the same name do not match.
331 * Section 4.3.9 (Interpolation) of the GLSL ES 3.00 spec says:
333 * "When no interpolation qualifier is present, smooth interpolation
336 * So we match variables where one is smooth and the other has no explicit
339 unsigned input_interpolation
= input
->data
.interpolation
;
340 unsigned output_interpolation
= output
->data
.interpolation
;
342 if (input_interpolation
== INTERP_MODE_NONE
)
343 input_interpolation
= INTERP_MODE_SMOOTH
;
344 if (output_interpolation
== INTERP_MODE_NONE
)
345 output_interpolation
= INTERP_MODE_SMOOTH
;
347 if (input_interpolation
!= output_interpolation
&&
348 prog
->data
->Version
< 440) {
349 if (!ctx
->Const
.AllowGLSLCrossStageInterpolationMismatch
) {
351 "%s shader output `%s' specifies %s "
352 "interpolation qualifier, "
353 "but %s shader input specifies %s "
354 "interpolation qualifier\n",
355 _mesa_shader_stage_to_string(producer_stage
),
357 interpolation_string(output
->data
.interpolation
),
358 _mesa_shader_stage_to_string(consumer_stage
),
359 interpolation_string(input
->data
.interpolation
));
363 "%s shader output `%s' specifies %s "
364 "interpolation qualifier, "
365 "but %s shader input specifies %s "
366 "interpolation qualifier\n",
367 _mesa_shader_stage_to_string(producer_stage
),
369 interpolation_string(output
->data
.interpolation
),
370 _mesa_shader_stage_to_string(consumer_stage
),
371 interpolation_string(input
->data
.interpolation
));
377 * Validate front and back color outputs against single color input
380 cross_validate_front_and_back_color(struct gl_context
*ctx
,
381 struct gl_shader_program
*prog
,
382 const ir_variable
*input
,
383 const ir_variable
*front_color
,
384 const ir_variable
*back_color
,
385 gl_shader_stage consumer_stage
,
386 gl_shader_stage producer_stage
)
388 if (front_color
!= NULL
&& front_color
->data
.assigned
)
389 cross_validate_types_and_qualifiers(ctx
, prog
, input
, front_color
,
390 consumer_stage
, producer_stage
);
392 if (back_color
!= NULL
&& back_color
->data
.assigned
)
393 cross_validate_types_and_qualifiers(ctx
, prog
, input
, back_color
,
394 consumer_stage
, producer_stage
);
398 compute_variable_location_slot(ir_variable
*var
, gl_shader_stage stage
)
400 unsigned location_start
= VARYING_SLOT_VAR0
;
403 case MESA_SHADER_VERTEX
:
404 if (var
->data
.mode
== ir_var_shader_in
)
405 location_start
= VERT_ATTRIB_GENERIC0
;
407 case MESA_SHADER_TESS_CTRL
:
408 case MESA_SHADER_TESS_EVAL
:
410 location_start
= VARYING_SLOT_PATCH0
;
412 case MESA_SHADER_FRAGMENT
:
413 if (var
->data
.mode
== ir_var_shader_out
)
414 location_start
= FRAG_RESULT_DATA0
;
420 return var
->data
.location
- location_start
;
423 struct explicit_location_info
{
425 unsigned numerical_type
;
426 unsigned interpolation
;
432 static inline unsigned
433 get_numerical_type(const glsl_type
*type
)
435 /* From the OpenGL 4.6 spec, section 4.4.1 Input Layout Qualifiers, Page 68,
436 * (Location aliasing):
438 * "Further, when location aliasing, the aliases sharing the location
439 * must have the same underlying numerical type (floating-point or
442 if (type
->is_float() || type
->is_double())
443 return GLSL_TYPE_FLOAT
;
444 return GLSL_TYPE_INT
;
448 check_location_aliasing(struct explicit_location_info explicit_locations
[][4],
452 unsigned location_limit
,
453 const glsl_type
*type
,
454 unsigned interpolation
,
458 gl_shader_program
*prog
,
459 gl_shader_stage stage
)
462 if (type
->without_array()->is_record()) {
463 /* The component qualifier can't be used on structs so just treat
464 * all component slots as used.
468 unsigned dmul
= type
->without_array()->is_64bit() ? 2 : 1;
469 last_comp
= component
+ type
->without_array()->vector_elements
* dmul
;
472 while (location
< location_limit
) {
475 struct explicit_location_info
*info
=
476 &explicit_locations
[location
][comp
];
479 /* Component aliasing is not alloed */
480 if (comp
>= component
&& comp
< last_comp
) {
482 "%s shader has multiple outputs explicitly "
483 "assigned to location %d and component %d\n",
484 _mesa_shader_stage_to_string(stage
),
488 /* For all other used components we need to have matching
489 * types, interpolation and auxiliary storage
491 if (info
->numerical_type
!=
492 get_numerical_type(type
->without_array())) {
494 "Varyings sharing the same location must "
495 "have the same underlying numerical type. "
496 "Location %u component %u\n",
501 if (info
->interpolation
!= interpolation
) {
503 "%s shader has multiple outputs at explicit "
504 "location %u with different interpolation "
506 _mesa_shader_stage_to_string(stage
), location
);
510 if (info
->centroid
!= centroid
||
511 info
->sample
!= sample
||
512 info
->patch
!= patch
) {
514 "%s shader has multiple outputs at explicit "
515 "location %u with different aux storage\n",
516 _mesa_shader_stage_to_string(stage
), location
);
520 } else if (comp
>= component
&& comp
< last_comp
) {
522 info
->numerical_type
= get_numerical_type(type
->without_array());
523 info
->interpolation
= interpolation
;
524 info
->centroid
= centroid
;
525 info
->sample
= sample
;
531 /* We need to do some special handling for doubles as dvec3 and
532 * dvec4 consume two consecutive locations. We don't need to
533 * worry about components beginning at anything other than 0 as
534 * the spec does not allow this for dvec3 and dvec4.
536 if (comp
== 4 && last_comp
> 4) {
537 last_comp
= last_comp
- 4;
538 /* Bump location index and reset the component index */
552 validate_explicit_variable_location(struct gl_context
*ctx
,
553 struct explicit_location_info explicit_locations
[][4],
555 gl_shader_program
*prog
,
556 gl_linked_shader
*sh
)
558 const glsl_type
*type
= get_varying_type(var
, sh
->Stage
);
559 unsigned num_elements
= type
->count_attribute_slots(false);
560 unsigned idx
= compute_variable_location_slot(var
, sh
->Stage
);
561 unsigned slot_limit
= idx
+ num_elements
;
563 /* Vertex shader inputs and fragment shader outputs are validated in
564 * assign_attribute_or_color_locations() so we should not attempt to
565 * validate them again here.
568 if (var
->data
.mode
== ir_var_shader_out
) {
569 assert(sh
->Stage
!= MESA_SHADER_FRAGMENT
);
571 ctx
->Const
.Program
[sh
->Stage
].MaxOutputComponents
/ 4;
573 assert(var
->data
.mode
== ir_var_shader_in
);
574 assert(sh
->Stage
!= MESA_SHADER_VERTEX
);
576 ctx
->Const
.Program
[sh
->Stage
].MaxInputComponents
/ 4;
579 if (slot_limit
> slot_max
) {
581 "Invalid location %u in %s shader\n",
582 idx
, _mesa_shader_stage_to_string(sh
->Stage
));
586 const glsl_type
*type_without_array
= type
->without_array();
587 if (type_without_array
->is_interface()) {
588 for (unsigned i
= 0; i
< type_without_array
->length
; i
++) {
589 glsl_struct_field
*field
= &type_without_array
->fields
.structure
[i
];
590 unsigned field_location
= field
->location
-
591 (field
->patch
? VARYING_SLOT_PATCH0
: VARYING_SLOT_VAR0
);
592 if (!check_location_aliasing(explicit_locations
, var
,
594 0, field_location
+ 1,
596 field
->interpolation
,
604 } else if (!check_location_aliasing(explicit_locations
, var
,
605 idx
, var
->data
.location_frac
,
607 var
->data
.interpolation
,
619 * Validate explicit locations for the inputs to the first stage and the
620 * outputs of the last stage in an SSO program (everything in between is
621 * validated in cross_validate_outputs_to_inputs).
624 validate_sso_explicit_locations(struct gl_context
*ctx
,
625 struct gl_shader_program
*prog
,
626 gl_shader_stage first_stage
,
627 gl_shader_stage last_stage
)
629 assert(prog
->SeparateShader
);
631 /* VS inputs and FS outputs are validated in
632 * assign_attribute_or_color_locations()
634 bool validate_first_stage
= first_stage
!= MESA_SHADER_VERTEX
;
635 bool validate_last_stage
= last_stage
!= MESA_SHADER_FRAGMENT
;
636 if (!validate_first_stage
&& !validate_last_stage
)
639 struct explicit_location_info explicit_locations
[MAX_VARYING
][4];
641 gl_shader_stage stages
[2] = { first_stage
, last_stage
};
642 bool validate_stage
[2] = { validate_first_stage
, validate_last_stage
};
643 ir_variable_mode var_direction
[2] = { ir_var_shader_in
, ir_var_shader_out
};
645 for (unsigned i
= 0; i
< 2; i
++) {
646 if (!validate_stage
[i
])
649 gl_shader_stage stage
= stages
[i
];
651 gl_linked_shader
*sh
= prog
->_LinkedShaders
[stage
];
654 memset(explicit_locations
, 0, sizeof(explicit_locations
));
656 foreach_in_list(ir_instruction
, node
, sh
->ir
) {
657 ir_variable
*const var
= node
->as_variable();
660 !var
->data
.explicit_location
||
661 var
->data
.location
< VARYING_SLOT_VAR0
||
662 var
->data
.mode
!= var_direction
[i
])
665 if (!validate_explicit_variable_location(
666 ctx
, explicit_locations
, var
, prog
, sh
)) {
674 * Validate that outputs from one stage match inputs of another
677 cross_validate_outputs_to_inputs(struct gl_context
*ctx
,
678 struct gl_shader_program
*prog
,
679 gl_linked_shader
*producer
,
680 gl_linked_shader
*consumer
)
682 glsl_symbol_table parameters
;
683 struct explicit_location_info explicit_locations
[MAX_VARYING
][4] = { 0 };
685 /* Find all shader outputs in the "producer" stage.
687 foreach_in_list(ir_instruction
, node
, producer
->ir
) {
688 ir_variable
*const var
= node
->as_variable();
690 if (var
== NULL
|| var
->data
.mode
!= ir_var_shader_out
)
693 if (!var
->data
.explicit_location
694 || var
->data
.location
< VARYING_SLOT_VAR0
)
695 parameters
.add_variable(var
);
697 /* User-defined varyings with explicit locations are handled
698 * differently because they do not need to have matching names.
700 if (!validate_explicit_variable_location(ctx
,
702 var
, prog
, producer
)) {
709 /* Find all shader inputs in the "consumer" stage. Any variables that have
710 * matching outputs already in the symbol table must have the same type and
713 * Exception: if the consumer is the geometry shader, then the inputs
714 * should be arrays and the type of the array element should match the type
715 * of the corresponding producer output.
717 foreach_in_list(ir_instruction
, node
, consumer
->ir
) {
718 ir_variable
*const input
= node
->as_variable();
720 if (input
== NULL
|| input
->data
.mode
!= ir_var_shader_in
)
723 if (strcmp(input
->name
, "gl_Color") == 0 && input
->data
.used
) {
724 const ir_variable
*const front_color
=
725 parameters
.get_variable("gl_FrontColor");
727 const ir_variable
*const back_color
=
728 parameters
.get_variable("gl_BackColor");
730 cross_validate_front_and_back_color(ctx
, prog
, input
,
731 front_color
, back_color
,
732 consumer
->Stage
, producer
->Stage
);
733 } else if (strcmp(input
->name
, "gl_SecondaryColor") == 0 && input
->data
.used
) {
734 const ir_variable
*const front_color
=
735 parameters
.get_variable("gl_FrontSecondaryColor");
737 const ir_variable
*const back_color
=
738 parameters
.get_variable("gl_BackSecondaryColor");
740 cross_validate_front_and_back_color(ctx
, prog
, input
,
741 front_color
, back_color
,
742 consumer
->Stage
, producer
->Stage
);
744 /* The rules for connecting inputs and outputs change in the presence
745 * of explicit locations. In this case, we no longer care about the
746 * names of the variables. Instead, we care only about the
747 * explicitly assigned location.
749 ir_variable
*output
= NULL
;
750 if (input
->data
.explicit_location
751 && input
->data
.location
>= VARYING_SLOT_VAR0
) {
753 const glsl_type
*type
= get_varying_type(input
, consumer
->Stage
);
754 unsigned num_elements
= type
->count_attribute_slots(false);
756 compute_variable_location_slot(input
, consumer
->Stage
);
757 unsigned slot_limit
= idx
+ num_elements
;
759 while (idx
< slot_limit
) {
760 if (idx
>= MAX_VARYING
) {
762 "Invalid location %u in %s shader\n", idx
,
763 _mesa_shader_stage_to_string(consumer
->Stage
));
767 output
= explicit_locations
[idx
][input
->data
.location_frac
].var
;
769 if (output
== NULL
||
770 input
->data
.location
!= output
->data
.location
) {
772 "%s shader input `%s' with explicit location "
773 "has no matching output\n",
774 _mesa_shader_stage_to_string(consumer
->Stage
),
781 output
= parameters
.get_variable(input
->name
);
784 if (output
!= NULL
) {
785 /* Interface blocks have their own validation elsewhere so don't
786 * try validating them here.
788 if (!(input
->get_interface_type() &&
789 output
->get_interface_type()))
790 cross_validate_types_and_qualifiers(ctx
, prog
, input
, output
,
794 /* Check for input vars with unmatched output vars in prev stage
795 * taking into account that interface blocks could have a matching
796 * output but with different name, so we ignore them.
798 assert(!input
->data
.assigned
);
799 if (input
->data
.used
&& !input
->get_interface_type() &&
800 !input
->data
.explicit_location
&& !prog
->SeparateShader
)
802 "%s shader input `%s' "
803 "has no matching output in the previous stage\n",
804 _mesa_shader_stage_to_string(consumer
->Stage
),
812 * Demote shader inputs and outputs that are not used in other stages, and
813 * remove them via dead code elimination.
816 remove_unused_shader_inputs_and_outputs(bool is_separate_shader_object
,
817 gl_linked_shader
*sh
,
818 enum ir_variable_mode mode
)
820 if (is_separate_shader_object
)
823 foreach_in_list(ir_instruction
, node
, sh
->ir
) {
824 ir_variable
*const var
= node
->as_variable();
826 if (var
== NULL
|| var
->data
.mode
!= int(mode
))
829 /* A shader 'in' or 'out' variable is only really an input or output if
830 * its value is used by other shader stages. This will cause the
831 * variable to have a location assigned.
833 if (var
->data
.is_unmatched_generic_inout
&& !var
->data
.is_xfb_only
) {
834 assert(var
->data
.mode
!= ir_var_temporary
);
836 /* Assign zeros to demoted inputs to allow more optimizations. */
837 if (var
->data
.mode
== ir_var_shader_in
&& !var
->constant_value
)
838 var
->constant_value
= ir_constant::zero(var
, var
->type
);
840 var
->data
.mode
= ir_var_auto
;
844 /* Eliminate code that is now dead due to unused inputs/outputs being
847 while (do_dead_code(sh
->ir
, false))
853 * Initialize this object based on a string that was passed to
854 * glTransformFeedbackVaryings.
856 * If the input is mal-formed, this call still succeeds, but it sets
857 * this->var_name to a mal-formed input, so tfeedback_decl::find_output_var()
858 * will fail to find any matching variable.
861 tfeedback_decl::init(struct gl_context
*ctx
, const void *mem_ctx
,
864 /* We don't have to be pedantic about what is a valid GLSL variable name,
865 * because any variable with an invalid name can't exist in the IR anyway.
869 this->orig_name
= input
;
870 this->lowered_builtin_array_variable
= none
;
871 this->skip_components
= 0;
872 this->next_buffer_separator
= false;
873 this->matched_candidate
= NULL
;
878 if (ctx
->Extensions
.ARB_transform_feedback3
) {
879 /* Parse gl_NextBuffer. */
880 if (strcmp(input
, "gl_NextBuffer") == 0) {
881 this->next_buffer_separator
= true;
885 /* Parse gl_SkipComponents. */
886 if (strcmp(input
, "gl_SkipComponents1") == 0)
887 this->skip_components
= 1;
888 else if (strcmp(input
, "gl_SkipComponents2") == 0)
889 this->skip_components
= 2;
890 else if (strcmp(input
, "gl_SkipComponents3") == 0)
891 this->skip_components
= 3;
892 else if (strcmp(input
, "gl_SkipComponents4") == 0)
893 this->skip_components
= 4;
895 if (this->skip_components
)
899 /* Parse a declaration. */
900 const char *base_name_end
;
901 long subscript
= parse_program_resource_name(input
, &base_name_end
);
902 this->var_name
= ralloc_strndup(mem_ctx
, input
, base_name_end
- input
);
903 if (this->var_name
== NULL
) {
904 _mesa_error_no_memory(__func__
);
908 if (subscript
>= 0) {
909 this->array_subscript
= subscript
;
910 this->is_subscripted
= true;
912 this->is_subscripted
= false;
915 /* For drivers that lower gl_ClipDistance to gl_ClipDistanceMESA, this
916 * class must behave specially to account for the fact that gl_ClipDistance
917 * is converted from a float[8] to a vec4[2].
919 if (ctx
->Const
.ShaderCompilerOptions
[MESA_SHADER_VERTEX
].LowerCombinedClipCullDistance
&&
920 strcmp(this->var_name
, "gl_ClipDistance") == 0) {
921 this->lowered_builtin_array_variable
= clip_distance
;
923 if (ctx
->Const
.ShaderCompilerOptions
[MESA_SHADER_VERTEX
].LowerCombinedClipCullDistance
&&
924 strcmp(this->var_name
, "gl_CullDistance") == 0) {
925 this->lowered_builtin_array_variable
= cull_distance
;
928 if (ctx
->Const
.LowerTessLevel
&&
929 (strcmp(this->var_name
, "gl_TessLevelOuter") == 0))
930 this->lowered_builtin_array_variable
= tess_level_outer
;
931 if (ctx
->Const
.LowerTessLevel
&&
932 (strcmp(this->var_name
, "gl_TessLevelInner") == 0))
933 this->lowered_builtin_array_variable
= tess_level_inner
;
938 * Determine whether two tfeedback_decl objects refer to the same variable and
939 * array index (if applicable).
942 tfeedback_decl::is_same(const tfeedback_decl
&x
, const tfeedback_decl
&y
)
944 assert(x
.is_varying() && y
.is_varying());
946 if (strcmp(x
.var_name
, y
.var_name
) != 0)
948 if (x
.is_subscripted
!= y
.is_subscripted
)
950 if (x
.is_subscripted
&& x
.array_subscript
!= y
.array_subscript
)
957 * Assign a location and stream ID for this tfeedback_decl object based on the
958 * transform feedback candidate found by find_candidate.
960 * If an error occurs, the error is reported through linker_error() and false
964 tfeedback_decl::assign_location(struct gl_context
*ctx
,
965 struct gl_shader_program
*prog
)
967 assert(this->is_varying());
969 unsigned fine_location
970 = this->matched_candidate
->toplevel_var
->data
.location
* 4
971 + this->matched_candidate
->toplevel_var
->data
.location_frac
972 + this->matched_candidate
->offset
;
973 const unsigned dmul
=
974 this->matched_candidate
->type
->without_array()->is_64bit() ? 2 : 1;
976 if (this->matched_candidate
->type
->is_array()) {
978 const unsigned matrix_cols
=
979 this->matched_candidate
->type
->fields
.array
->matrix_columns
;
980 const unsigned vector_elements
=
981 this->matched_candidate
->type
->fields
.array
->vector_elements
;
982 unsigned actual_array_size
;
983 switch (this->lowered_builtin_array_variable
) {
985 actual_array_size
= prog
->last_vert_prog
?
986 prog
->last_vert_prog
->info
.clip_distance_array_size
: 0;
989 actual_array_size
= prog
->last_vert_prog
?
990 prog
->last_vert_prog
->info
.cull_distance_array_size
: 0;
992 case tess_level_outer
:
993 actual_array_size
= 4;
995 case tess_level_inner
:
996 actual_array_size
= 2;
1000 actual_array_size
= this->matched_candidate
->type
->array_size();
1004 if (this->is_subscripted
) {
1005 /* Check array bounds. */
1006 if (this->array_subscript
>= actual_array_size
) {
1007 linker_error(prog
, "Transform feedback varying %s has index "
1008 "%i, but the array size is %u.",
1009 this->orig_name
, this->array_subscript
,
1013 unsigned array_elem_size
= this->lowered_builtin_array_variable
?
1014 1 : vector_elements
* matrix_cols
* dmul
;
1015 fine_location
+= array_elem_size
* this->array_subscript
;
1018 this->size
= actual_array_size
;
1020 this->vector_elements
= vector_elements
;
1021 this->matrix_columns
= matrix_cols
;
1022 if (this->lowered_builtin_array_variable
)
1023 this->type
= GL_FLOAT
;
1025 this->type
= this->matched_candidate
->type
->fields
.array
->gl_type
;
1027 /* Regular variable (scalar, vector, or matrix) */
1028 if (this->is_subscripted
) {
1029 linker_error(prog
, "Transform feedback varying %s requested, "
1030 "but %s is not an array.",
1031 this->orig_name
, this->var_name
);
1035 this->vector_elements
= this->matched_candidate
->type
->vector_elements
;
1036 this->matrix_columns
= this->matched_candidate
->type
->matrix_columns
;
1037 this->type
= this->matched_candidate
->type
->gl_type
;
1039 this->location
= fine_location
/ 4;
1040 this->location_frac
= fine_location
% 4;
1042 /* From GL_EXT_transform_feedback:
1043 * A program will fail to link if:
1045 * * the total number of components to capture in any varying
1046 * variable in <varyings> is greater than the constant
1047 * MAX_TRANSFORM_FEEDBACK_SEPARATE_COMPONENTS_EXT and the
1048 * buffer mode is SEPARATE_ATTRIBS_EXT;
1050 if (prog
->TransformFeedback
.BufferMode
== GL_SEPARATE_ATTRIBS
&&
1051 this->num_components() >
1052 ctx
->Const
.MaxTransformFeedbackSeparateComponents
) {
1053 linker_error(prog
, "Transform feedback varying %s exceeds "
1054 "MAX_TRANSFORM_FEEDBACK_SEPARATE_COMPONENTS.",
1059 /* Only transform feedback varyings can be assigned to non-zero streams,
1060 * so assign the stream id here.
1062 this->stream_id
= this->matched_candidate
->toplevel_var
->data
.stream
;
1064 unsigned array_offset
= this->array_subscript
* 4 * dmul
;
1065 unsigned struct_offset
= this->matched_candidate
->offset
* 4 * dmul
;
1066 this->buffer
= this->matched_candidate
->toplevel_var
->data
.xfb_buffer
;
1067 this->offset
= this->matched_candidate
->toplevel_var
->data
.offset
+
1068 array_offset
+ struct_offset
;
1075 tfeedback_decl::get_num_outputs() const
1077 if (!this->is_varying()) {
1080 return (this->num_components() + this->location_frac
+ 3)/4;
1085 * Update gl_transform_feedback_info to reflect this tfeedback_decl.
1087 * If an error occurs, the error is reported through linker_error() and false
1091 tfeedback_decl::store(struct gl_context
*ctx
, struct gl_shader_program
*prog
,
1092 struct gl_transform_feedback_info
*info
,
1093 unsigned buffer
, unsigned buffer_index
,
1094 const unsigned max_outputs
, bool *explicit_stride
,
1095 bool has_xfb_qualifiers
) const
1097 unsigned xfb_offset
= 0;
1098 unsigned size
= this->size
;
1099 /* Handle gl_SkipComponents. */
1100 if (this->skip_components
) {
1101 info
->Buffers
[buffer
].Stride
+= this->skip_components
;
1102 size
= this->skip_components
;
1106 if (this->next_buffer_separator
) {
1111 if (has_xfb_qualifiers
) {
1112 xfb_offset
= this->offset
/ 4;
1114 xfb_offset
= info
->Buffers
[buffer
].Stride
;
1116 info
->Varyings
[info
->NumVarying
].Offset
= xfb_offset
* 4;
1119 unsigned location
= this->location
;
1120 unsigned location_frac
= this->location_frac
;
1121 unsigned num_components
= this->num_components();
1122 while (num_components
> 0) {
1123 unsigned output_size
= MIN2(num_components
, 4 - location_frac
);
1124 assert((info
->NumOutputs
== 0 && max_outputs
== 0) ||
1125 info
->NumOutputs
< max_outputs
);
1127 /* From the ARB_enhanced_layouts spec:
1129 * "If such a block member or variable is not written during a shader
1130 * invocation, the buffer contents at the assigned offset will be
1131 * undefined. Even if there are no static writes to a variable or
1132 * member that is assigned a transform feedback offset, the space is
1133 * still allocated in the buffer and still affects the stride."
1135 if (this->is_varying_written()) {
1136 info
->Outputs
[info
->NumOutputs
].ComponentOffset
= location_frac
;
1137 info
->Outputs
[info
->NumOutputs
].OutputRegister
= location
;
1138 info
->Outputs
[info
->NumOutputs
].NumComponents
= output_size
;
1139 info
->Outputs
[info
->NumOutputs
].StreamId
= stream_id
;
1140 info
->Outputs
[info
->NumOutputs
].OutputBuffer
= buffer
;
1141 info
->Outputs
[info
->NumOutputs
].DstOffset
= xfb_offset
;
1144 info
->Buffers
[buffer
].Stream
= this->stream_id
;
1145 xfb_offset
+= output_size
;
1147 num_components
-= output_size
;
1153 if (explicit_stride
&& explicit_stride
[buffer
]) {
1154 if (this->is_64bit() && info
->Buffers
[buffer
].Stride
% 2) {
1155 linker_error(prog
, "invalid qualifier xfb_stride=%d must be a "
1156 "multiple of 8 as its applied to a type that is or "
1157 "contains a double.",
1158 info
->Buffers
[buffer
].Stride
* 4);
1162 if ((this->offset
/ 4) / info
->Buffers
[buffer
].Stride
!=
1163 (xfb_offset
- 1) / info
->Buffers
[buffer
].Stride
) {
1164 linker_error(prog
, "xfb_offset (%d) overflows xfb_stride (%d) for "
1165 "buffer (%d)", xfb_offset
* 4,
1166 info
->Buffers
[buffer
].Stride
* 4, buffer
);
1170 info
->Buffers
[buffer
].Stride
= xfb_offset
;
1173 /* From GL_EXT_transform_feedback:
1174 * A program will fail to link if:
1176 * * the total number of components to capture is greater than
1177 * the constant MAX_TRANSFORM_FEEDBACK_INTERLEAVED_COMPONENTS_EXT
1178 * and the buffer mode is INTERLEAVED_ATTRIBS_EXT.
1180 * From GL_ARB_enhanced_layouts:
1182 * "The resulting stride (implicit or explicit) must be less than or
1183 * equal to the implementation-dependent constant
1184 * gl_MaxTransformFeedbackInterleavedComponents."
1186 if ((prog
->TransformFeedback
.BufferMode
== GL_INTERLEAVED_ATTRIBS
||
1187 has_xfb_qualifiers
) &&
1188 info
->Buffers
[buffer
].Stride
>
1189 ctx
->Const
.MaxTransformFeedbackInterleavedComponents
) {
1190 linker_error(prog
, "The MAX_TRANSFORM_FEEDBACK_INTERLEAVED_COMPONENTS "
1191 "limit has been exceeded.");
1196 info
->Varyings
[info
->NumVarying
].Name
= ralloc_strdup(prog
,
1198 info
->Varyings
[info
->NumVarying
].Type
= this->type
;
1199 info
->Varyings
[info
->NumVarying
].Size
= size
;
1200 info
->Varyings
[info
->NumVarying
].BufferIndex
= buffer_index
;
1202 info
->Buffers
[buffer
].NumVaryings
++;
1208 const tfeedback_candidate
*
1209 tfeedback_decl::find_candidate(gl_shader_program
*prog
,
1210 hash_table
*tfeedback_candidates
)
1212 const char *name
= this->var_name
;
1213 switch (this->lowered_builtin_array_variable
) {
1215 name
= this->var_name
;
1218 name
= "gl_ClipDistanceMESA";
1221 name
= "gl_CullDistanceMESA";
1223 case tess_level_outer
:
1224 name
= "gl_TessLevelOuterMESA";
1226 case tess_level_inner
:
1227 name
= "gl_TessLevelInnerMESA";
1230 hash_entry
*entry
= _mesa_hash_table_search(tfeedback_candidates
, name
);
1232 this->matched_candidate
= entry
?
1233 (const tfeedback_candidate
*) entry
->data
: NULL
;
1235 if (!this->matched_candidate
) {
1236 /* From GL_EXT_transform_feedback:
1237 * A program will fail to link if:
1239 * * any variable name specified in the <varyings> array is not
1240 * declared as an output in the geometry shader (if present) or
1241 * the vertex shader (if no geometry shader is present);
1243 linker_error(prog
, "Transform feedback varying %s undeclared.",
1247 return this->matched_candidate
;
1252 * Parse all the transform feedback declarations that were passed to
1253 * glTransformFeedbackVaryings() and store them in tfeedback_decl objects.
1255 * If an error occurs, the error is reported through linker_error() and false
1259 parse_tfeedback_decls(struct gl_context
*ctx
, struct gl_shader_program
*prog
,
1260 const void *mem_ctx
, unsigned num_names
,
1261 char **varying_names
, tfeedback_decl
*decls
)
1263 for (unsigned i
= 0; i
< num_names
; ++i
) {
1264 decls
[i
].init(ctx
, mem_ctx
, varying_names
[i
]);
1266 if (!decls
[i
].is_varying())
1269 /* From GL_EXT_transform_feedback:
1270 * A program will fail to link if:
1272 * * any two entries in the <varyings> array specify the same varying
1275 * We interpret this to mean "any two entries in the <varyings> array
1276 * specify the same varying variable and array index", since transform
1277 * feedback of arrays would be useless otherwise.
1279 for (unsigned j
= 0; j
< i
; ++j
) {
1280 if (!decls
[j
].is_varying())
1283 if (tfeedback_decl::is_same(decls
[i
], decls
[j
])) {
1284 linker_error(prog
, "Transform feedback varying %s specified "
1285 "more than once.", varying_names
[i
]);
1295 cmp_xfb_offset(const void * x_generic
, const void * y_generic
)
1297 tfeedback_decl
*x
= (tfeedback_decl
*) x_generic
;
1298 tfeedback_decl
*y
= (tfeedback_decl
*) y_generic
;
1300 if (x
->get_buffer() != y
->get_buffer())
1301 return x
->get_buffer() - y
->get_buffer();
1302 return x
->get_offset() - y
->get_offset();
1306 * Store transform feedback location assignments into
1307 * prog->sh.LinkedTransformFeedback based on the data stored in
1310 * If an error occurs, the error is reported through linker_error() and false
1314 store_tfeedback_info(struct gl_context
*ctx
, struct gl_shader_program
*prog
,
1315 unsigned num_tfeedback_decls
,
1316 tfeedback_decl
*tfeedback_decls
, bool has_xfb_qualifiers
)
1318 if (!prog
->last_vert_prog
)
1321 /* Make sure MaxTransformFeedbackBuffers is less than 32 so the bitmask for
1322 * tracking the number of buffers doesn't overflow.
1324 assert(ctx
->Const
.MaxTransformFeedbackBuffers
< 32);
1326 bool separate_attribs_mode
=
1327 prog
->TransformFeedback
.BufferMode
== GL_SEPARATE_ATTRIBS
;
1329 struct gl_program
*xfb_prog
= prog
->last_vert_prog
;
1330 xfb_prog
->sh
.LinkedTransformFeedback
=
1331 rzalloc(xfb_prog
, struct gl_transform_feedback_info
);
1333 /* The xfb_offset qualifier does not have to be used in increasing order
1334 * however some drivers expect to receive the list of transform feedback
1335 * declarations in order so sort it now for convenience.
1337 if (has_xfb_qualifiers
) {
1338 qsort(tfeedback_decls
, num_tfeedback_decls
, sizeof(*tfeedback_decls
),
1341 xfb_prog
->sh
.LinkedTransformFeedback
->api_enabled
= true;
1344 xfb_prog
->sh
.LinkedTransformFeedback
->Varyings
=
1345 rzalloc_array(xfb_prog
, struct gl_transform_feedback_varying_info
,
1346 num_tfeedback_decls
);
1348 unsigned num_outputs
= 0;
1349 for (unsigned i
= 0; i
< num_tfeedback_decls
; ++i
) {
1350 if (tfeedback_decls
[i
].is_varying_written())
1351 num_outputs
+= tfeedback_decls
[i
].get_num_outputs();
1354 xfb_prog
->sh
.LinkedTransformFeedback
->Outputs
=
1355 rzalloc_array(xfb_prog
, struct gl_transform_feedback_output
,
1358 unsigned num_buffers
= 0;
1359 unsigned buffers
= 0;
1361 if (!has_xfb_qualifiers
&& separate_attribs_mode
) {
1362 /* GL_SEPARATE_ATTRIBS */
1363 for (unsigned i
= 0; i
< num_tfeedback_decls
; ++i
) {
1364 if (!tfeedback_decls
[i
].store(ctx
, prog
,
1365 xfb_prog
->sh
.LinkedTransformFeedback
,
1366 num_buffers
, num_buffers
, num_outputs
,
1367 NULL
, has_xfb_qualifiers
))
1370 buffers
|= 1 << num_buffers
;
1375 /* GL_INVERLEAVED_ATTRIBS */
1376 int buffer_stream_id
= -1;
1378 num_tfeedback_decls
? tfeedback_decls
[0].get_buffer() : 0;
1379 bool explicit_stride
[MAX_FEEDBACK_BUFFERS
] = { false };
1381 /* Apply any xfb_stride global qualifiers */
1382 if (has_xfb_qualifiers
) {
1383 for (unsigned j
= 0; j
< MAX_FEEDBACK_BUFFERS
; j
++) {
1384 if (prog
->TransformFeedback
.BufferStride
[j
]) {
1385 explicit_stride
[j
] = true;
1386 xfb_prog
->sh
.LinkedTransformFeedback
->Buffers
[j
].Stride
=
1387 prog
->TransformFeedback
.BufferStride
[j
] / 4;
1392 for (unsigned i
= 0; i
< num_tfeedback_decls
; ++i
) {
1393 if (has_xfb_qualifiers
&&
1394 buffer
!= tfeedback_decls
[i
].get_buffer()) {
1395 /* we have moved to the next buffer so reset stream id */
1396 buffer_stream_id
= -1;
1400 if (tfeedback_decls
[i
].is_next_buffer_separator()) {
1401 if (!tfeedback_decls
[i
].store(ctx
, prog
,
1402 xfb_prog
->sh
.LinkedTransformFeedback
,
1403 buffer
, num_buffers
, num_outputs
,
1404 explicit_stride
, has_xfb_qualifiers
))
1407 buffer_stream_id
= -1;
1411 if (has_xfb_qualifiers
) {
1412 buffer
= tfeedback_decls
[i
].get_buffer();
1414 buffer
= num_buffers
;
1417 if (tfeedback_decls
[i
].is_varying()) {
1418 if (buffer_stream_id
== -1) {
1419 /* First varying writing to this buffer: remember its stream */
1420 buffer_stream_id
= (int) tfeedback_decls
[i
].get_stream_id();
1422 /* Only mark a buffer as active when there is a varying
1423 * attached to it. This behaviour is based on a revised version
1424 * of section 13.2.2 of the GL 4.6 spec.
1426 buffers
|= 1 << buffer
;
1427 } else if (buffer_stream_id
!=
1428 (int) tfeedback_decls
[i
].get_stream_id()) {
1429 /* Varying writes to the same buffer from a different stream */
1431 "Transform feedback can't capture varyings belonging "
1432 "to different vertex streams in a single buffer. "
1433 "Varying %s writes to buffer from stream %u, other "
1434 "varyings in the same buffer write from stream %u.",
1435 tfeedback_decls
[i
].name(),
1436 tfeedback_decls
[i
].get_stream_id(),
1442 if (!tfeedback_decls
[i
].store(ctx
, prog
,
1443 xfb_prog
->sh
.LinkedTransformFeedback
,
1444 buffer
, num_buffers
, num_outputs
,
1445 explicit_stride
, has_xfb_qualifiers
))
1450 assert(xfb_prog
->sh
.LinkedTransformFeedback
->NumOutputs
== num_outputs
);
1452 xfb_prog
->sh
.LinkedTransformFeedback
->ActiveBuffers
= buffers
;
1459 * Data structure recording the relationship between outputs of one shader
1460 * stage (the "producer") and inputs of another (the "consumer").
1462 class varying_matches
1465 varying_matches(bool disable_varying_packing
, bool xfb_enabled
,
1466 bool enhanced_layouts_enabled
,
1467 gl_shader_stage producer_stage
,
1468 gl_shader_stage consumer_stage
);
1470 void record(ir_variable
*producer_var
, ir_variable
*consumer_var
);
1471 unsigned assign_locations(struct gl_shader_program
*prog
,
1472 uint8_t *components
,
1473 uint64_t reserved_slots
);
1474 void store_locations() const;
1477 bool is_varying_packing_safe(const glsl_type
*type
,
1478 const ir_variable
*var
);
1481 * If true, this driver disables varying packing, so all varyings need to
1482 * be aligned on slot boundaries, and take up a number of slots equal to
1483 * their number of matrix columns times their array size.
1485 * Packing may also be disabled because our current packing method is not
1486 * safe in SSO or versions of OpenGL where interpolation qualifiers are not
1487 * guaranteed to match across stages.
1489 const bool disable_varying_packing
;
1492 * If true, this driver has transform feedback enabled. The transform
1493 * feedback code requires at least some packing be done even when varying
1494 * packing is disabled, fortunately where transform feedback requires
1495 * packing it's safe to override the disabled setting. See
1496 * is_varying_packing_safe().
1498 const bool xfb_enabled
;
1500 const bool enhanced_layouts_enabled
;
1503 * Enum representing the order in which varyings are packed within a
1506 * Currently we pack vec4's first, then vec2's, then scalar values, then
1507 * vec3's. This order ensures that the only vectors that are at risk of
1508 * having to be "double parked" (split between two adjacent varying slots)
1511 enum packing_order_enum
{
1514 PACKING_ORDER_SCALAR
,
1518 static unsigned compute_packing_class(const ir_variable
*var
);
1519 static packing_order_enum
compute_packing_order(const ir_variable
*var
);
1520 static int match_comparator(const void *x_generic
, const void *y_generic
);
1521 static int xfb_comparator(const void *x_generic
, const void *y_generic
);
1524 * Structure recording the relationship between a single producer output
1525 * and a single consumer input.
1529 * Packing class for this varying, computed by compute_packing_class().
1531 unsigned packing_class
;
1534 * Packing order for this varying, computed by compute_packing_order().
1536 packing_order_enum packing_order
;
1537 unsigned num_components
;
1540 * The output variable in the producer stage.
1542 ir_variable
*producer_var
;
1545 * The input variable in the consumer stage.
1547 ir_variable
*consumer_var
;
1550 * The location which has been assigned for this varying. This is
1551 * expressed in multiples of a float, with the first generic varying
1552 * (i.e. the one referred to by VARYING_SLOT_VAR0) represented by the
1555 unsigned generic_location
;
1559 * The number of elements in the \c matches array that are currently in
1562 unsigned num_matches
;
1565 * The number of elements that were set aside for the \c matches array when
1568 unsigned matches_capacity
;
1570 gl_shader_stage producer_stage
;
1571 gl_shader_stage consumer_stage
;
1574 } /* anonymous namespace */
1576 varying_matches::varying_matches(bool disable_varying_packing
,
1578 bool enhanced_layouts_enabled
,
1579 gl_shader_stage producer_stage
,
1580 gl_shader_stage consumer_stage
)
1581 : disable_varying_packing(disable_varying_packing
),
1582 xfb_enabled(xfb_enabled
),
1583 enhanced_layouts_enabled(enhanced_layouts_enabled
),
1584 producer_stage(producer_stage
),
1585 consumer_stage(consumer_stage
)
1587 /* Note: this initial capacity is rather arbitrarily chosen to be large
1588 * enough for many cases without wasting an unreasonable amount of space.
1589 * varying_matches::record() will resize the array if there are more than
1590 * this number of varyings.
1592 this->matches_capacity
= 8;
1593 this->matches
= (match
*)
1594 malloc(sizeof(*this->matches
) * this->matches_capacity
);
1595 this->num_matches
= 0;
1599 varying_matches::~varying_matches()
1601 free(this->matches
);
1606 * Packing is always safe on individual arrays, structures, and matrices. It
1607 * is also safe if the varying is only used for transform feedback.
1610 varying_matches::is_varying_packing_safe(const glsl_type
*type
,
1611 const ir_variable
*var
)
1613 if (consumer_stage
== MESA_SHADER_TESS_EVAL
||
1614 consumer_stage
== MESA_SHADER_TESS_CTRL
||
1615 producer_stage
== MESA_SHADER_TESS_CTRL
)
1618 return xfb_enabled
&& (type
->is_array() || type
->is_record() ||
1619 type
->is_matrix() || var
->data
.is_xfb_only
);
1624 * Record the given producer/consumer variable pair in the list of variables
1625 * that should later be assigned locations.
1627 * It is permissible for \c consumer_var to be NULL (this happens if a
1628 * variable is output by the producer and consumed by transform feedback, but
1629 * not consumed by the consumer).
1631 * If \c producer_var has already been paired up with a consumer_var, or
1632 * producer_var is part of fixed pipeline functionality (and hence already has
1633 * a location assigned), this function has no effect.
1635 * Note: as a side effect this function may change the interpolation type of
1636 * \c producer_var, but only when the change couldn't possibly affect
1640 varying_matches::record(ir_variable
*producer_var
, ir_variable
*consumer_var
)
1642 assert(producer_var
!= NULL
|| consumer_var
!= NULL
);
1644 if ((producer_var
&& (!producer_var
->data
.is_unmatched_generic_inout
||
1645 producer_var
->data
.explicit_location
)) ||
1646 (consumer_var
&& (!consumer_var
->data
.is_unmatched_generic_inout
||
1647 consumer_var
->data
.explicit_location
))) {
1648 /* Either a location already exists for this variable (since it is part
1649 * of fixed functionality), or it has already been recorded as part of a
1655 bool needs_flat_qualifier
= consumer_var
== NULL
&&
1656 (producer_var
->type
->contains_integer() ||
1657 producer_var
->type
->contains_double());
1659 if (!disable_varying_packing
&&
1660 (needs_flat_qualifier
||
1661 (consumer_stage
!= MESA_SHADER_NONE
&& consumer_stage
!= MESA_SHADER_FRAGMENT
))) {
1662 /* Since this varying is not being consumed by the fragment shader, its
1663 * interpolation type varying cannot possibly affect rendering.
1664 * Also, this variable is non-flat and is (or contains) an integer
1666 * If the consumer stage is unknown, don't modify the interpolation
1667 * type as it could affect rendering later with separate shaders.
1669 * lower_packed_varyings requires all integer varyings to flat,
1670 * regardless of where they appear. We can trivially satisfy that
1671 * requirement by changing the interpolation type to flat here.
1674 producer_var
->data
.centroid
= false;
1675 producer_var
->data
.sample
= false;
1676 producer_var
->data
.interpolation
= INTERP_MODE_FLAT
;
1680 consumer_var
->data
.centroid
= false;
1681 consumer_var
->data
.sample
= false;
1682 consumer_var
->data
.interpolation
= INTERP_MODE_FLAT
;
1686 if (this->num_matches
== this->matches_capacity
) {
1687 this->matches_capacity
*= 2;
1688 this->matches
= (match
*)
1689 realloc(this->matches
,
1690 sizeof(*this->matches
) * this->matches_capacity
);
1693 /* We must use the consumer to compute the packing class because in GL4.4+
1694 * there is no guarantee interpolation qualifiers will match across stages.
1696 * From Section 4.5 (Interpolation Qualifiers) of the GLSL 4.30 spec:
1698 * "The type and presence of interpolation qualifiers of variables with
1699 * the same name declared in all linked shaders for the same cross-stage
1700 * interface must match, otherwise the link command will fail.
1702 * When comparing an output from one stage to an input of a subsequent
1703 * stage, the input and output don't match if their interpolation
1704 * qualifiers (or lack thereof) are not the same."
1706 * This text was also in at least revison 7 of the 4.40 spec but is no
1707 * longer in revision 9 and not in the 4.50 spec.
1709 const ir_variable
*const var
= (consumer_var
!= NULL
)
1710 ? consumer_var
: producer_var
;
1711 const gl_shader_stage stage
= (consumer_var
!= NULL
)
1712 ? consumer_stage
: producer_stage
;
1713 const glsl_type
*type
= get_varying_type(var
, stage
);
1715 if (producer_var
&& consumer_var
&&
1716 consumer_var
->data
.must_be_shader_input
) {
1717 producer_var
->data
.must_be_shader_input
= 1;
1720 this->matches
[this->num_matches
].packing_class
1721 = this->compute_packing_class(var
);
1722 this->matches
[this->num_matches
].packing_order
1723 = this->compute_packing_order(var
);
1724 if ((this->disable_varying_packing
&& !is_varying_packing_safe(type
, var
)) ||
1725 var
->data
.must_be_shader_input
) {
1726 unsigned slots
= type
->count_attribute_slots(false);
1727 this->matches
[this->num_matches
].num_components
= slots
* 4;
1729 this->matches
[this->num_matches
].num_components
1730 = type
->component_slots();
1733 this->matches
[this->num_matches
].producer_var
= producer_var
;
1734 this->matches
[this->num_matches
].consumer_var
= consumer_var
;
1735 this->num_matches
++;
1737 producer_var
->data
.is_unmatched_generic_inout
= 0;
1739 consumer_var
->data
.is_unmatched_generic_inout
= 0;
1744 * Choose locations for all of the variable matches that were previously
1745 * passed to varying_matches::record().
1748 varying_matches::assign_locations(struct gl_shader_program
*prog
,
1749 uint8_t *components
,
1750 uint64_t reserved_slots
)
1752 /* If packing has been disabled then we cannot safely sort the varyings by
1753 * class as it may mean we are using a version of OpenGL where
1754 * interpolation qualifiers are not guaranteed to be matching across
1755 * shaders, sorting in this case could result in mismatching shader
1757 * When packing is disabled the sort orders varyings used by transform
1758 * feedback first, but also depends on *undefined behaviour* of qsort to
1759 * reverse the order of the varyings. See: xfb_comparator().
1761 if (!this->disable_varying_packing
) {
1762 /* Sort varying matches into an order that makes them easy to pack. */
1763 qsort(this->matches
, this->num_matches
, sizeof(*this->matches
),
1764 &varying_matches::match_comparator
);
1766 /* Only sort varyings that are only used by transform feedback. */
1767 qsort(this->matches
, this->num_matches
, sizeof(*this->matches
),
1768 &varying_matches::xfb_comparator
);
1771 unsigned generic_location
= 0;
1772 unsigned generic_patch_location
= MAX_VARYING
*4;
1773 bool previous_var_xfb_only
= false;
1775 for (unsigned i
= 0; i
< this->num_matches
; i
++) {
1776 unsigned *location
= &generic_location
;
1777 const ir_variable
*var
;
1778 const glsl_type
*type
;
1779 bool is_vertex_input
= false;
1781 if (matches
[i
].consumer_var
) {
1782 var
= matches
[i
].consumer_var
;
1783 type
= get_varying_type(var
, consumer_stage
);
1784 if (consumer_stage
== MESA_SHADER_VERTEX
)
1785 is_vertex_input
= true;
1787 var
= matches
[i
].producer_var
;
1788 type
= get_varying_type(var
, producer_stage
);
1791 if (var
->data
.patch
)
1792 location
= &generic_patch_location
;
1794 /* Advance to the next slot if this varying has a different packing
1795 * class than the previous one, and we're not already on a slot
1798 * Also advance to the next slot if packing is disabled. This makes sure
1799 * we don't assign varyings the same locations which is possible
1800 * because we still pack individual arrays, records and matrices even
1801 * when packing is disabled. Note we don't advance to the next slot if
1802 * we can pack varyings together that are only used for transform
1805 if (var
->data
.must_be_shader_input
||
1806 (this->disable_varying_packing
&&
1807 !(previous_var_xfb_only
&& var
->data
.is_xfb_only
)) ||
1808 (i
> 0 && this->matches
[i
- 1].packing_class
1809 != this->matches
[i
].packing_class
)) {
1810 *location
= ALIGN(*location
, 4);
1813 previous_var_xfb_only
= var
->data
.is_xfb_only
;
1815 /* The number of components taken up by this variable. For vertex shader
1816 * inputs, we use the number of slots * 4, as they have different
1819 unsigned num_components
= is_vertex_input
?
1820 type
->count_attribute_slots(is_vertex_input
) * 4 :
1821 this->matches
[i
].num_components
;
1823 /* The last slot for this variable, inclusive. */
1824 unsigned slot_end
= *location
+ num_components
- 1;
1826 /* FIXME: We could be smarter in the below code and loop back over
1827 * trying to fill any locations that we skipped because we couldn't pack
1828 * the varying between an explicit location. For now just let the user
1829 * hit the linking error if we run out of room and suggest they use
1830 * explicit locations.
1832 while (slot_end
< MAX_VARYING
* 4u) {
1833 const unsigned slots
= (slot_end
/ 4u) - (*location
/ 4u) + 1;
1834 const uint64_t slot_mask
= ((1ull << slots
) - 1) << (*location
/ 4u);
1837 if (reserved_slots
& slot_mask
) {
1838 *location
= ALIGN(*location
+ 1, 4);
1839 slot_end
= *location
+ num_components
- 1;
1846 if (!var
->data
.patch
&& slot_end
>= MAX_VARYING
* 4u) {
1847 linker_error(prog
, "insufficient contiguous locations available for "
1848 "%s it is possible an array or struct could not be "
1849 "packed between varyings with explicit locations. Try "
1850 "using an explicit location for arrays and structs.",
1854 if (slot_end
< MAX_VARYINGS_INCL_PATCH
* 4u) {
1855 for (unsigned j
= *location
/ 4u; j
< slot_end
/ 4u; j
++)
1857 components
[slot_end
/ 4u] = (slot_end
& 3) + 1;
1860 this->matches
[i
].generic_location
= *location
;
1862 *location
= slot_end
+ 1;
1865 return (generic_location
+ 3) / 4;
1870 * Update the producer and consumer shaders to reflect the locations
1871 * assignments that were made by varying_matches::assign_locations().
1874 varying_matches::store_locations() const
1876 /* Check is location needs to be packed with lower_packed_varyings() or if
1877 * we can just use ARB_enhanced_layouts packing.
1879 bool pack_loc
[MAX_VARYINGS_INCL_PATCH
] = { 0 };
1880 const glsl_type
*loc_type
[MAX_VARYINGS_INCL_PATCH
][4] = { {NULL
, NULL
} };
1882 for (unsigned i
= 0; i
< this->num_matches
; i
++) {
1883 ir_variable
*producer_var
= this->matches
[i
].producer_var
;
1884 ir_variable
*consumer_var
= this->matches
[i
].consumer_var
;
1885 unsigned generic_location
= this->matches
[i
].generic_location
;
1886 unsigned slot
= generic_location
/ 4;
1887 unsigned offset
= generic_location
% 4;
1890 producer_var
->data
.location
= VARYING_SLOT_VAR0
+ slot
;
1891 producer_var
->data
.location_frac
= offset
;
1895 assert(consumer_var
->data
.location
== -1);
1896 consumer_var
->data
.location
= VARYING_SLOT_VAR0
+ slot
;
1897 consumer_var
->data
.location_frac
= offset
;
1900 /* Find locations suitable for native packing via
1901 * ARB_enhanced_layouts.
1903 if (producer_var
&& consumer_var
) {
1904 if (enhanced_layouts_enabled
) {
1905 const glsl_type
*type
=
1906 get_varying_type(producer_var
, producer_stage
);
1907 if (type
->is_array() || type
->is_matrix() || type
->is_record() ||
1908 type
->is_double()) {
1909 unsigned comp_slots
= type
->component_slots() + offset
;
1910 unsigned slots
= comp_slots
/ 4;
1914 for (unsigned j
= 0; j
< slots
; j
++) {
1915 pack_loc
[slot
+ j
] = true;
1917 } else if (offset
+ type
->vector_elements
> 4) {
1918 pack_loc
[slot
] = true;
1919 pack_loc
[slot
+ 1] = true;
1921 loc_type
[slot
][offset
] = type
;
1927 /* Attempt to use ARB_enhanced_layouts for more efficient packing if
1930 if (enhanced_layouts_enabled
) {
1931 for (unsigned i
= 0; i
< this->num_matches
; i
++) {
1932 ir_variable
*producer_var
= this->matches
[i
].producer_var
;
1933 ir_variable
*consumer_var
= this->matches
[i
].consumer_var
;
1934 unsigned generic_location
= this->matches
[i
].generic_location
;
1935 unsigned slot
= generic_location
/ 4;
1937 if (pack_loc
[slot
] || !producer_var
|| !consumer_var
)
1940 const glsl_type
*type
=
1941 get_varying_type(producer_var
, producer_stage
);
1942 bool type_match
= true;
1943 for (unsigned j
= 0; j
< 4; j
++) {
1944 if (loc_type
[slot
][j
]) {
1945 if (type
->base_type
!= loc_type
[slot
][j
]->base_type
)
1951 producer_var
->data
.explicit_location
= 1;
1952 consumer_var
->data
.explicit_location
= 1;
1953 producer_var
->data
.explicit_component
= 1;
1954 consumer_var
->data
.explicit_component
= 1;
1962 * Compute the "packing class" of the given varying. This is an unsigned
1963 * integer with the property that two variables in the same packing class can
1964 * be safely backed into the same vec4.
1967 varying_matches::compute_packing_class(const ir_variable
*var
)
1969 /* Without help from the back-end, there is no way to pack together
1970 * variables with different interpolation types, because
1971 * lower_packed_varyings must choose exactly one interpolation type for
1972 * each packed varying it creates.
1974 * However, we can safely pack together floats, ints, and uints, because:
1976 * - varyings of base type "int" and "uint" must use the "flat"
1977 * interpolation type, which can only occur in GLSL 1.30 and above.
1979 * - On platforms that support GLSL 1.30 and above, lower_packed_varyings
1980 * can store flat floats as ints without losing any information (using
1981 * the ir_unop_bitcast_* opcodes).
1983 * Therefore, the packing class depends only on the interpolation type.
1985 unsigned packing_class
= var
->data
.centroid
| (var
->data
.sample
<< 1) |
1986 (var
->data
.patch
<< 2) |
1987 (var
->data
.must_be_shader_input
<< 3);
1989 packing_class
+= var
->is_interpolation_flat()
1990 ? unsigned(INTERP_MODE_FLAT
) : var
->data
.interpolation
;
1991 return packing_class
;
1996 * Compute the "packing order" of the given varying. This is a sort key we
1997 * use to determine when to attempt to pack the given varying relative to
1998 * other varyings in the same packing class.
2000 varying_matches::packing_order_enum
2001 varying_matches::compute_packing_order(const ir_variable
*var
)
2003 const glsl_type
*element_type
= var
->type
;
2005 while (element_type
->is_array()) {
2006 element_type
= element_type
->fields
.array
;
2009 switch (element_type
->component_slots() % 4) {
2010 case 1: return PACKING_ORDER_SCALAR
;
2011 case 2: return PACKING_ORDER_VEC2
;
2012 case 3: return PACKING_ORDER_VEC3
;
2013 case 0: return PACKING_ORDER_VEC4
;
2015 assert(!"Unexpected value of vector_elements");
2016 return PACKING_ORDER_VEC4
;
2022 * Comparison function passed to qsort() to sort varyings by packing_class and
2023 * then by packing_order.
2026 varying_matches::match_comparator(const void *x_generic
, const void *y_generic
)
2028 const match
*x
= (const match
*) x_generic
;
2029 const match
*y
= (const match
*) y_generic
;
2031 if (x
->packing_class
!= y
->packing_class
)
2032 return x
->packing_class
- y
->packing_class
;
2033 return x
->packing_order
- y
->packing_order
;
2038 * Comparison function passed to qsort() to sort varyings used only by
2039 * transform feedback when packing of other varyings is disabled.
2042 varying_matches::xfb_comparator(const void *x_generic
, const void *y_generic
)
2044 const match
*x
= (const match
*) x_generic
;
2046 if (x
->producer_var
!= NULL
&& x
->producer_var
->data
.is_xfb_only
)
2047 return match_comparator(x_generic
, y_generic
);
2049 /* FIXME: When the comparator returns 0 it means the elements being
2050 * compared are equivalent. However the qsort documentation says:
2052 * "The order of equivalent elements is undefined."
2054 * In practice the sort ends up reversing the order of the varyings which
2055 * means locations are also assigned in this reversed order and happens to
2056 * be what we want. This is also whats happening in
2057 * varying_matches::match_comparator().
2064 * Is the given variable a varying variable to be counted against the
2065 * limit in ctx->Const.MaxVarying?
2066 * This includes variables such as texcoords, colors and generic
2067 * varyings, but excludes variables such as gl_FrontFacing and gl_FragCoord.
2070 var_counts_against_varying_limit(gl_shader_stage stage
, const ir_variable
*var
)
2072 /* Only fragment shaders will take a varying variable as an input */
2073 if (stage
== MESA_SHADER_FRAGMENT
&&
2074 var
->data
.mode
== ir_var_shader_in
) {
2075 switch (var
->data
.location
) {
2076 case VARYING_SLOT_POS
:
2077 case VARYING_SLOT_FACE
:
2078 case VARYING_SLOT_PNTC
:
2089 * Visitor class that generates tfeedback_candidate structs describing all
2090 * possible targets of transform feedback.
2092 * tfeedback_candidate structs are stored in the hash table
2093 * tfeedback_candidates, which is passed to the constructor. This hash table
2094 * maps varying names to instances of the tfeedback_candidate struct.
2096 class tfeedback_candidate_generator
: public program_resource_visitor
2099 tfeedback_candidate_generator(void *mem_ctx
,
2100 hash_table
*tfeedback_candidates
)
2102 tfeedback_candidates(tfeedback_candidates
),
2108 void process(ir_variable
*var
)
2110 /* All named varying interface blocks should be flattened by now */
2111 assert(!var
->is_interface_instance());
2113 this->toplevel_var
= var
;
2114 this->varying_floats
= 0;
2115 program_resource_visitor::process(var
, false);
2119 virtual void visit_field(const glsl_type
*type
, const char *name
,
2120 bool /* row_major */,
2121 const glsl_type
* /* record_type */,
2122 const enum glsl_interface_packing
,
2123 bool /* last_field */)
2125 assert(!type
->without_array()->is_record());
2126 assert(!type
->without_array()->is_interface());
2128 tfeedback_candidate
*candidate
2129 = rzalloc(this->mem_ctx
, tfeedback_candidate
);
2130 candidate
->toplevel_var
= this->toplevel_var
;
2131 candidate
->type
= type
;
2132 candidate
->offset
= this->varying_floats
;
2133 _mesa_hash_table_insert(this->tfeedback_candidates
,
2134 ralloc_strdup(this->mem_ctx
, name
),
2136 this->varying_floats
+= type
->component_slots();
2140 * Memory context used to allocate hash table keys and values.
2142 void * const mem_ctx
;
2145 * Hash table in which tfeedback_candidate objects should be stored.
2147 hash_table
* const tfeedback_candidates
;
2150 * Pointer to the toplevel variable that is being traversed.
2152 ir_variable
*toplevel_var
;
2155 * Total number of varying floats that have been visited so far. This is
2156 * used to determine the offset to each varying within the toplevel
2159 unsigned varying_floats
;
2166 populate_consumer_input_sets(void *mem_ctx
, exec_list
*ir
,
2167 hash_table
*consumer_inputs
,
2168 hash_table
*consumer_interface_inputs
,
2169 ir_variable
*consumer_inputs_with_locations
[VARYING_SLOT_TESS_MAX
])
2171 memset(consumer_inputs_with_locations
,
2173 sizeof(consumer_inputs_with_locations
[0]) * VARYING_SLOT_TESS_MAX
);
2175 foreach_in_list(ir_instruction
, node
, ir
) {
2176 ir_variable
*const input_var
= node
->as_variable();
2178 if (input_var
!= NULL
&& input_var
->data
.mode
== ir_var_shader_in
) {
2179 /* All interface blocks should have been lowered by this point */
2180 assert(!input_var
->type
->is_interface());
2182 if (input_var
->data
.explicit_location
) {
2183 /* assign_varying_locations only cares about finding the
2184 * ir_variable at the start of a contiguous location block.
2186 * - For !producer, consumer_inputs_with_locations isn't used.
2188 * - For !consumer, consumer_inputs_with_locations is empty.
2190 * For consumer && producer, if you were trying to set some
2191 * ir_variable to the middle of a location block on the other side
2192 * of producer/consumer, cross_validate_outputs_to_inputs() should
2193 * be link-erroring due to either type mismatch or location
2194 * overlaps. If the variables do match up, then they've got a
2195 * matching data.location and you only looked at
2196 * consumer_inputs_with_locations[var->data.location], not any
2197 * following entries for the array/structure.
2199 consumer_inputs_with_locations
[input_var
->data
.location
] =
2201 } else if (input_var
->get_interface_type() != NULL
) {
2202 char *const iface_field_name
=
2203 ralloc_asprintf(mem_ctx
, "%s.%s",
2204 input_var
->get_interface_type()->without_array()->name
,
2206 _mesa_hash_table_insert(consumer_interface_inputs
,
2207 iface_field_name
, input_var
);
2209 _mesa_hash_table_insert(consumer_inputs
,
2210 ralloc_strdup(mem_ctx
, input_var
->name
),
2218 * Find a variable from the consumer that "matches" the specified variable
2220 * This function only finds inputs with names that match. There is no
2221 * validation (here) that the types, etc. are compatible.
2224 get_matching_input(void *mem_ctx
,
2225 const ir_variable
*output_var
,
2226 hash_table
*consumer_inputs
,
2227 hash_table
*consumer_interface_inputs
,
2228 ir_variable
*consumer_inputs_with_locations
[VARYING_SLOT_TESS_MAX
])
2230 ir_variable
*input_var
;
2232 if (output_var
->data
.explicit_location
) {
2233 input_var
= consumer_inputs_with_locations
[output_var
->data
.location
];
2234 } else if (output_var
->get_interface_type() != NULL
) {
2235 char *const iface_field_name
=
2236 ralloc_asprintf(mem_ctx
, "%s.%s",
2237 output_var
->get_interface_type()->without_array()->name
,
2239 hash_entry
*entry
= _mesa_hash_table_search(consumer_interface_inputs
, iface_field_name
);
2240 input_var
= entry
? (ir_variable
*) entry
->data
: NULL
;
2242 hash_entry
*entry
= _mesa_hash_table_search(consumer_inputs
, output_var
->name
);
2243 input_var
= entry
? (ir_variable
*) entry
->data
: NULL
;
2246 return (input_var
== NULL
|| input_var
->data
.mode
!= ir_var_shader_in
)
2253 io_variable_cmp(const void *_a
, const void *_b
)
2255 const ir_variable
*const a
= *(const ir_variable
**) _a
;
2256 const ir_variable
*const b
= *(const ir_variable
**) _b
;
2258 if (a
->data
.explicit_location
&& b
->data
.explicit_location
)
2259 return b
->data
.location
- a
->data
.location
;
2261 if (a
->data
.explicit_location
&& !b
->data
.explicit_location
)
2264 if (!a
->data
.explicit_location
&& b
->data
.explicit_location
)
2267 return -strcmp(a
->name
, b
->name
);
2271 * Sort the shader IO variables into canonical order
2274 canonicalize_shader_io(exec_list
*ir
, enum ir_variable_mode io_mode
)
2276 ir_variable
*var_table
[MAX_PROGRAM_OUTPUTS
* 4];
2277 unsigned num_variables
= 0;
2279 foreach_in_list(ir_instruction
, node
, ir
) {
2280 ir_variable
*const var
= node
->as_variable();
2282 if (var
== NULL
|| var
->data
.mode
!= io_mode
)
2285 /* If we have already encountered more I/O variables that could
2286 * successfully link, bail.
2288 if (num_variables
== ARRAY_SIZE(var_table
))
2291 var_table
[num_variables
++] = var
;
2294 if (num_variables
== 0)
2297 /* Sort the list in reverse order (io_variable_cmp handles this). Later
2298 * we're going to push the variables on to the IR list as a stack, so we
2299 * want the last variable (in canonical order) to be first in the list.
2301 qsort(var_table
, num_variables
, sizeof(var_table
[0]), io_variable_cmp
);
2303 /* Remove the variable from it's current location in the IR, and put it at
2306 for (unsigned i
= 0; i
< num_variables
; i
++) {
2307 var_table
[i
]->remove();
2308 ir
->push_head(var_table
[i
]);
2313 * Generate a bitfield map of the explicit locations for shader varyings.
2315 * Note: For Tessellation shaders we are sitting right on the limits of the
2316 * 64 bit map. Per-vertex and per-patch both have separate location domains
2317 * with a max of MAX_VARYING.
2320 reserved_varying_slot(struct gl_linked_shader
*stage
,
2321 ir_variable_mode io_mode
)
2323 assert(io_mode
== ir_var_shader_in
|| io_mode
== ir_var_shader_out
);
2324 /* Avoid an overflow of the returned value */
2325 assert(MAX_VARYINGS_INCL_PATCH
<= 64);
2333 foreach_in_list(ir_instruction
, node
, stage
->ir
) {
2334 ir_variable
*const var
= node
->as_variable();
2336 if (var
== NULL
|| var
->data
.mode
!= io_mode
||
2337 !var
->data
.explicit_location
||
2338 var
->data
.location
< VARYING_SLOT_VAR0
)
2341 var_slot
= var
->data
.location
- VARYING_SLOT_VAR0
;
2343 unsigned num_elements
= get_varying_type(var
, stage
->Stage
)
2344 ->count_attribute_slots(io_mode
== ir_var_shader_in
&&
2345 stage
->Stage
== MESA_SHADER_VERTEX
);
2346 for (unsigned i
= 0; i
< num_elements
; i
++) {
2347 if (var_slot
>= 0 && var_slot
< MAX_VARYINGS_INCL_PATCH
)
2348 slots
|= UINT64_C(1) << var_slot
;
2358 * Assign locations for all variables that are produced in one pipeline stage
2359 * (the "producer") and consumed in the next stage (the "consumer").
2361 * Variables produced by the producer may also be consumed by transform
2364 * \param num_tfeedback_decls is the number of declarations indicating
2365 * variables that may be consumed by transform feedback.
2367 * \param tfeedback_decls is a pointer to an array of tfeedback_decl objects
2368 * representing the result of parsing the strings passed to
2369 * glTransformFeedbackVaryings(). assign_location() will be called for
2370 * each of these objects that matches one of the outputs of the
2373 * When num_tfeedback_decls is nonzero, it is permissible for the consumer to
2374 * be NULL. In this case, varying locations are assigned solely based on the
2375 * requirements of transform feedback.
2378 assign_varying_locations(struct gl_context
*ctx
,
2380 struct gl_shader_program
*prog
,
2381 gl_linked_shader
*producer
,
2382 gl_linked_shader
*consumer
,
2383 unsigned num_tfeedback_decls
,
2384 tfeedback_decl
*tfeedback_decls
,
2385 const uint64_t reserved_slots
)
2387 /* Tessellation shaders treat inputs and outputs as shared memory and can
2388 * access inputs and outputs of other invocations.
2389 * Therefore, they can't be lowered to temps easily (and definitely not
2392 bool unpackable_tess
=
2393 (consumer
&& consumer
->Stage
== MESA_SHADER_TESS_EVAL
) ||
2394 (consumer
&& consumer
->Stage
== MESA_SHADER_TESS_CTRL
) ||
2395 (producer
&& producer
->Stage
== MESA_SHADER_TESS_CTRL
);
2397 /* Transform feedback code assumes varying arrays are packed, so if the
2398 * driver has disabled varying packing, make sure to at least enable
2399 * packing required by transform feedback.
2402 ctx
->Extensions
.EXT_transform_feedback
&& !unpackable_tess
;
2404 /* Disable packing on outward facing interfaces for SSO because in ES we
2405 * need to retain the unpacked varying information for draw time
2408 * Packing is still enabled on individual arrays, structs, and matrices as
2409 * these are required by the transform feedback code and it is still safe
2410 * to do so. We also enable packing when a varying is only used for
2411 * transform feedback and its not a SSO.
2413 bool disable_varying_packing
=
2414 ctx
->Const
.DisableVaryingPacking
|| unpackable_tess
;
2415 if (prog
->SeparateShader
&& (producer
== NULL
|| consumer
== NULL
))
2416 disable_varying_packing
= true;
2418 varying_matches
matches(disable_varying_packing
, xfb_enabled
,
2419 ctx
->Extensions
.ARB_enhanced_layouts
,
2420 producer
? producer
->Stage
: MESA_SHADER_NONE
,
2421 consumer
? consumer
->Stage
: MESA_SHADER_NONE
);
2422 hash_table
*tfeedback_candidates
=
2423 _mesa_hash_table_create(NULL
, _mesa_key_hash_string
,
2424 _mesa_key_string_equal
);
2425 hash_table
*consumer_inputs
=
2426 _mesa_hash_table_create(NULL
, _mesa_key_hash_string
,
2427 _mesa_key_string_equal
);
2428 hash_table
*consumer_interface_inputs
=
2429 _mesa_hash_table_create(NULL
, _mesa_key_hash_string
,
2430 _mesa_key_string_equal
);
2431 ir_variable
*consumer_inputs_with_locations
[VARYING_SLOT_TESS_MAX
] = {
2435 unsigned consumer_vertices
= 0;
2436 if (consumer
&& consumer
->Stage
== MESA_SHADER_GEOMETRY
)
2437 consumer_vertices
= prog
->Geom
.VerticesIn
;
2439 /* Operate in a total of four passes.
2441 * 1. Sort inputs / outputs into a canonical order. This is necessary so
2442 * that inputs / outputs of separable shaders will be assigned
2443 * predictable locations regardless of the order in which declarations
2444 * appeared in the shader source.
2446 * 2. Assign locations for any matching inputs and outputs.
2448 * 3. Mark output variables in the producer that do not have locations as
2449 * not being outputs. This lets the optimizer eliminate them.
2451 * 4. Mark input variables in the consumer that do not have locations as
2452 * not being inputs. This lets the optimizer eliminate them.
2455 canonicalize_shader_io(consumer
->ir
, ir_var_shader_in
);
2458 canonicalize_shader_io(producer
->ir
, ir_var_shader_out
);
2461 linker::populate_consumer_input_sets(mem_ctx
, consumer
->ir
,
2463 consumer_interface_inputs
,
2464 consumer_inputs_with_locations
);
2467 foreach_in_list(ir_instruction
, node
, producer
->ir
) {
2468 ir_variable
*const output_var
= node
->as_variable();
2470 if (output_var
== NULL
|| output_var
->data
.mode
!= ir_var_shader_out
)
2473 /* Only geometry shaders can use non-zero streams */
2474 assert(output_var
->data
.stream
== 0 ||
2475 (output_var
->data
.stream
< MAX_VERTEX_STREAMS
&&
2476 producer
->Stage
== MESA_SHADER_GEOMETRY
));
2478 if (num_tfeedback_decls
> 0) {
2479 tfeedback_candidate_generator
g(mem_ctx
, tfeedback_candidates
);
2480 g
.process(output_var
);
2483 ir_variable
*const input_var
=
2484 linker::get_matching_input(mem_ctx
, output_var
, consumer_inputs
,
2485 consumer_interface_inputs
,
2486 consumer_inputs_with_locations
);
2488 /* If a matching input variable was found, add this output (and the
2489 * input) to the set. If this is a separable program and there is no
2490 * consumer stage, add the output.
2492 * Always add TCS outputs. They are shared by all invocations
2493 * within a patch and can be used as shared memory.
2495 if (input_var
|| (prog
->SeparateShader
&& consumer
== NULL
) ||
2496 producer
->Stage
== MESA_SHADER_TESS_CTRL
) {
2497 matches
.record(output_var
, input_var
);
2500 /* Only stream 0 outputs can be consumed in the next stage */
2501 if (input_var
&& output_var
->data
.stream
!= 0) {
2502 linker_error(prog
, "output %s is assigned to stream=%d but "
2503 "is linked to an input, which requires stream=0",
2504 output_var
->name
, output_var
->data
.stream
);
2509 /* If there's no producer stage, then this must be a separable program.
2510 * For example, we may have a program that has just a fragment shader.
2511 * Later this program will be used with some arbitrary vertex (or
2512 * geometry) shader program. This means that locations must be assigned
2513 * for all the inputs.
2515 foreach_in_list(ir_instruction
, node
, consumer
->ir
) {
2516 ir_variable
*const input_var
= node
->as_variable();
2518 if (input_var
== NULL
|| input_var
->data
.mode
!= ir_var_shader_in
)
2521 matches
.record(NULL
, input_var
);
2525 for (unsigned i
= 0; i
< num_tfeedback_decls
; ++i
) {
2526 if (!tfeedback_decls
[i
].is_varying())
2529 const tfeedback_candidate
*matched_candidate
2530 = tfeedback_decls
[i
].find_candidate(prog
, tfeedback_candidates
);
2532 if (matched_candidate
== NULL
) {
2533 _mesa_hash_table_destroy(tfeedback_candidates
, NULL
);
2537 /* Mark xfb varyings as always active */
2538 matched_candidate
->toplevel_var
->data
.always_active_io
= 1;
2540 /* Mark any corresponding inputs as always active also. We must do this
2541 * because we have a NIR pass that lowers vectors to scalars and another
2542 * that removes unused varyings.
2543 * We don't split varyings marked as always active because there is no
2544 * point in doing so. This means we need to mark both sides of the
2545 * interface as always active otherwise we will have a mismatch and
2546 * start removing things we shouldn't.
2548 ir_variable
*const input_var
=
2549 linker::get_matching_input(mem_ctx
, matched_candidate
->toplevel_var
,
2551 consumer_interface_inputs
,
2552 consumer_inputs_with_locations
);
2554 input_var
->data
.always_active_io
= 1;
2556 if (matched_candidate
->toplevel_var
->data
.is_unmatched_generic_inout
) {
2557 matched_candidate
->toplevel_var
->data
.is_xfb_only
= 1;
2558 matches
.record(matched_candidate
->toplevel_var
, NULL
);
2562 _mesa_hash_table_destroy(consumer_inputs
, NULL
);
2563 _mesa_hash_table_destroy(consumer_interface_inputs
, NULL
);
2565 uint8_t components
[MAX_VARYINGS_INCL_PATCH
] = {0};
2566 const unsigned slots_used
= matches
.assign_locations(
2567 prog
, components
, reserved_slots
);
2568 matches
.store_locations();
2570 for (unsigned i
= 0; i
< num_tfeedback_decls
; ++i
) {
2571 if (!tfeedback_decls
[i
].is_varying())
2574 if (!tfeedback_decls
[i
].assign_location(ctx
, prog
)) {
2575 _mesa_hash_table_destroy(tfeedback_candidates
, NULL
);
2579 _mesa_hash_table_destroy(tfeedback_candidates
, NULL
);
2581 if (consumer
&& producer
) {
2582 foreach_in_list(ir_instruction
, node
, consumer
->ir
) {
2583 ir_variable
*const var
= node
->as_variable();
2585 if (var
&& var
->data
.mode
== ir_var_shader_in
&&
2586 var
->data
.is_unmatched_generic_inout
) {
2587 if (!prog
->IsES
&& prog
->data
->Version
<= 120) {
2588 /* On page 25 (page 31 of the PDF) of the GLSL 1.20 spec:
2590 * Only those varying variables used (i.e. read) in
2591 * the fragment shader executable must be written to
2592 * by the vertex shader executable; declaring
2593 * superfluous varying variables in a vertex shader is
2596 * We interpret this text as meaning that the VS must
2597 * write the variable for the FS to read it. See
2598 * "glsl1-varying read but not written" in piglit.
2600 linker_error(prog
, "%s shader varying %s not written "
2602 _mesa_shader_stage_to_string(consumer
->Stage
),
2604 _mesa_shader_stage_to_string(producer
->Stage
));
2606 linker_warning(prog
, "%s shader varying %s not written "
2608 _mesa_shader_stage_to_string(consumer
->Stage
),
2610 _mesa_shader_stage_to_string(producer
->Stage
));
2615 /* Now that validation is done its safe to remove unused varyings. As
2616 * we have both a producer and consumer its safe to remove unused
2617 * varyings even if the program is a SSO because the stages are being
2618 * linked together i.e. we have a multi-stage SSO.
2620 remove_unused_shader_inputs_and_outputs(false, producer
,
2622 remove_unused_shader_inputs_and_outputs(false, consumer
,
2627 lower_packed_varyings(mem_ctx
, slots_used
, components
, ir_var_shader_out
,
2628 0, producer
, disable_varying_packing
,
2633 lower_packed_varyings(mem_ctx
, slots_used
, components
, ir_var_shader_in
,
2634 consumer_vertices
, consumer
,
2635 disable_varying_packing
, xfb_enabled
);
2642 check_against_output_limit(struct gl_context
*ctx
,
2643 struct gl_shader_program
*prog
,
2644 gl_linked_shader
*producer
,
2645 unsigned num_explicit_locations
)
2647 unsigned output_vectors
= num_explicit_locations
;
2649 foreach_in_list(ir_instruction
, node
, producer
->ir
) {
2650 ir_variable
*const var
= node
->as_variable();
2652 if (var
&& !var
->data
.explicit_location
&&
2653 var
->data
.mode
== ir_var_shader_out
&&
2654 var_counts_against_varying_limit(producer
->Stage
, var
)) {
2655 /* outputs for fragment shader can't be doubles */
2656 output_vectors
+= var
->type
->count_attribute_slots(false);
2660 assert(producer
->Stage
!= MESA_SHADER_FRAGMENT
);
2661 unsigned max_output_components
=
2662 ctx
->Const
.Program
[producer
->Stage
].MaxOutputComponents
;
2664 const unsigned output_components
= output_vectors
* 4;
2665 if (output_components
> max_output_components
) {
2666 if (ctx
->API
== API_OPENGLES2
|| prog
->IsES
)
2667 linker_error(prog
, "%s shader uses too many output vectors "
2669 _mesa_shader_stage_to_string(producer
->Stage
),
2671 max_output_components
/ 4);
2673 linker_error(prog
, "%s shader uses too many output components "
2675 _mesa_shader_stage_to_string(producer
->Stage
),
2677 max_output_components
);
2686 check_against_input_limit(struct gl_context
*ctx
,
2687 struct gl_shader_program
*prog
,
2688 gl_linked_shader
*consumer
,
2689 unsigned num_explicit_locations
)
2691 unsigned input_vectors
= num_explicit_locations
;
2693 foreach_in_list(ir_instruction
, node
, consumer
->ir
) {
2694 ir_variable
*const var
= node
->as_variable();
2696 if (var
&& !var
->data
.explicit_location
&&
2697 var
->data
.mode
== ir_var_shader_in
&&
2698 var_counts_against_varying_limit(consumer
->Stage
, var
)) {
2699 /* vertex inputs aren't varying counted */
2700 input_vectors
+= var
->type
->count_attribute_slots(false);
2704 assert(consumer
->Stage
!= MESA_SHADER_VERTEX
);
2705 unsigned max_input_components
=
2706 ctx
->Const
.Program
[consumer
->Stage
].MaxInputComponents
;
2708 const unsigned input_components
= input_vectors
* 4;
2709 if (input_components
> max_input_components
) {
2710 if (ctx
->API
== API_OPENGLES2
|| prog
->IsES
)
2711 linker_error(prog
, "%s shader uses too many input vectors "
2713 _mesa_shader_stage_to_string(consumer
->Stage
),
2715 max_input_components
/ 4);
2717 linker_error(prog
, "%s shader uses too many input components "
2719 _mesa_shader_stage_to_string(consumer
->Stage
),
2721 max_input_components
);
2730 link_varyings(struct gl_shader_program
*prog
, unsigned first
, unsigned last
,
2731 struct gl_context
*ctx
, void *mem_ctx
)
2733 bool has_xfb_qualifiers
= false;
2734 unsigned num_tfeedback_decls
= 0;
2735 char **varying_names
= NULL
;
2736 tfeedback_decl
*tfeedback_decls
= NULL
;
2738 /* From the ARB_enhanced_layouts spec:
2740 * "If the shader used to record output variables for transform feedback
2741 * varyings uses the "xfb_buffer", "xfb_offset", or "xfb_stride" layout
2742 * qualifiers, the values specified by TransformFeedbackVaryings are
2743 * ignored, and the set of variables captured for transform feedback is
2744 * instead derived from the specified layout qualifiers."
2746 for (int i
= MESA_SHADER_FRAGMENT
- 1; i
>= 0; i
--) {
2747 /* Find last stage before fragment shader */
2748 if (prog
->_LinkedShaders
[i
]) {
2749 has_xfb_qualifiers
=
2750 process_xfb_layout_qualifiers(mem_ctx
, prog
->_LinkedShaders
[i
],
2751 prog
, &num_tfeedback_decls
,
2757 if (!has_xfb_qualifiers
) {
2758 num_tfeedback_decls
= prog
->TransformFeedback
.NumVarying
;
2759 varying_names
= prog
->TransformFeedback
.VaryingNames
;
2762 if (num_tfeedback_decls
!= 0) {
2763 /* From GL_EXT_transform_feedback:
2764 * A program will fail to link if:
2766 * * the <count> specified by TransformFeedbackVaryingsEXT is
2767 * non-zero, but the program object has no vertex or geometry
2770 if (first
>= MESA_SHADER_FRAGMENT
) {
2771 linker_error(prog
, "Transform feedback varyings specified, but "
2772 "no vertex, tessellation, or geometry shader is "
2777 tfeedback_decls
= rzalloc_array(mem_ctx
, tfeedback_decl
,
2778 num_tfeedback_decls
);
2779 if (!parse_tfeedback_decls(ctx
, prog
, mem_ctx
, num_tfeedback_decls
,
2780 varying_names
, tfeedback_decls
))
2784 /* If there is no fragment shader we need to set transform feedback.
2786 * For SSO we also need to assign output locations. We assign them here
2787 * because we need to do it for both single stage programs and multi stage
2790 if (last
< MESA_SHADER_FRAGMENT
&&
2791 (num_tfeedback_decls
!= 0 || prog
->SeparateShader
)) {
2792 const uint64_t reserved_out_slots
=
2793 reserved_varying_slot(prog
->_LinkedShaders
[last
], ir_var_shader_out
);
2794 if (!assign_varying_locations(ctx
, mem_ctx
, prog
,
2795 prog
->_LinkedShaders
[last
], NULL
,
2796 num_tfeedback_decls
, tfeedback_decls
,
2797 reserved_out_slots
))
2801 if (last
<= MESA_SHADER_FRAGMENT
) {
2802 /* Remove unused varyings from the first/last stage unless SSO */
2803 remove_unused_shader_inputs_and_outputs(prog
->SeparateShader
,
2804 prog
->_LinkedShaders
[first
],
2806 remove_unused_shader_inputs_and_outputs(prog
->SeparateShader
,
2807 prog
->_LinkedShaders
[last
],
2810 /* If the program is made up of only a single stage */
2811 if (first
== last
) {
2812 gl_linked_shader
*const sh
= prog
->_LinkedShaders
[last
];
2814 do_dead_builtin_varyings(ctx
, NULL
, sh
, 0, NULL
);
2815 do_dead_builtin_varyings(ctx
, sh
, NULL
, num_tfeedback_decls
,
2818 if (prog
->SeparateShader
) {
2819 const uint64_t reserved_slots
=
2820 reserved_varying_slot(sh
, ir_var_shader_in
);
2822 /* Assign input locations for SSO, output locations are already
2825 if (!assign_varying_locations(ctx
, mem_ctx
, prog
,
2826 NULL
/* producer */,
2828 0 /* num_tfeedback_decls */,
2829 NULL
/* tfeedback_decls */,
2834 /* Linking the stages in the opposite order (from fragment to vertex)
2835 * ensures that inter-shader outputs written to in an earlier stage
2836 * are eliminated if they are (transitively) not used in a later
2840 for (int i
= next
- 1; i
>= 0; i
--) {
2841 if (prog
->_LinkedShaders
[i
] == NULL
&& i
!= 0)
2844 gl_linked_shader
*const sh_i
= prog
->_LinkedShaders
[i
];
2845 gl_linked_shader
*const sh_next
= prog
->_LinkedShaders
[next
];
2847 const uint64_t reserved_out_slots
=
2848 reserved_varying_slot(sh_i
, ir_var_shader_out
);
2849 const uint64_t reserved_in_slots
=
2850 reserved_varying_slot(sh_next
, ir_var_shader_in
);
2852 do_dead_builtin_varyings(ctx
, sh_i
, sh_next
,
2853 next
== MESA_SHADER_FRAGMENT
? num_tfeedback_decls
: 0,
2856 if (!assign_varying_locations(ctx
, mem_ctx
, prog
, sh_i
, sh_next
,
2857 next
== MESA_SHADER_FRAGMENT
? num_tfeedback_decls
: 0,
2859 reserved_out_slots
| reserved_in_slots
))
2862 /* This must be done after all dead varyings are eliminated. */
2864 unsigned slots_used
= _mesa_bitcount_64(reserved_out_slots
);
2865 if (!check_against_output_limit(ctx
, prog
, sh_i
, slots_used
)) {
2870 unsigned slots_used
= _mesa_bitcount_64(reserved_in_slots
);
2871 if (!check_against_input_limit(ctx
, prog
, sh_next
, slots_used
))
2879 if (!store_tfeedback_info(ctx
, prog
, num_tfeedback_decls
, tfeedback_decls
,
2880 has_xfb_qualifiers
))