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_shader_program
*prog
,
193 const ir_variable
*input
,
194 const ir_variable
*output
,
195 gl_shader_stage consumer_stage
,
196 gl_shader_stage producer_stage
)
198 /* Check that the types match between stages.
200 const glsl_type
*type_to_match
= input
->type
;
202 /* VS -> GS, VS -> TCS, VS -> TES, TES -> GS */
203 const bool extra_array_level
= (producer_stage
== MESA_SHADER_VERTEX
&&
204 consumer_stage
!= MESA_SHADER_FRAGMENT
) ||
205 consumer_stage
== MESA_SHADER_GEOMETRY
;
206 if (extra_array_level
) {
207 assert(type_to_match
->is_array());
208 type_to_match
= type_to_match
->fields
.array
;
211 if (type_to_match
!= output
->type
) {
212 /* There is a bit of a special case for gl_TexCoord. This
213 * built-in is unsized by default. Applications that variable
214 * access it must redeclare it with a size. There is some
215 * language in the GLSL spec that implies the fragment shader
216 * and vertex shader do not have to agree on this size. Other
217 * driver behave this way, and one or two applications seem to
220 * Neither declaration needs to be modified here because the array
221 * sizes are fixed later when update_array_sizes is called.
223 * From page 48 (page 54 of the PDF) of the GLSL 1.10 spec:
225 * "Unlike user-defined varying variables, the built-in
226 * varying variables don't have a strict one-to-one
227 * correspondence between the vertex language and the
228 * fragment language."
230 if (!output
->type
->is_array() || !is_gl_identifier(output
->name
)) {
232 "%s shader output `%s' declared as type `%s', "
233 "but %s shader input declared as type `%s'\n",
234 _mesa_shader_stage_to_string(producer_stage
),
237 _mesa_shader_stage_to_string(consumer_stage
),
243 /* Check that all of the qualifiers match between stages.
246 /* According to the OpenGL and OpenGLES GLSL specs, the centroid qualifier
247 * should match until OpenGL 4.3 and OpenGLES 3.1. The OpenGLES 3.0
248 * conformance test suite does not verify that the qualifiers must match.
249 * The deqp test suite expects the opposite (OpenGLES 3.1) behavior for
250 * OpenGLES 3.0 drivers, so we relax the checking in all cases.
252 if (false /* always skip the centroid check */ &&
253 prog
->data
->Version
< (prog
->IsES
? 310 : 430) &&
254 input
->data
.centroid
!= output
->data
.centroid
) {
256 "%s shader output `%s' %s centroid qualifier, "
257 "but %s shader input %s centroid qualifier\n",
258 _mesa_shader_stage_to_string(producer_stage
),
260 (output
->data
.centroid
) ? "has" : "lacks",
261 _mesa_shader_stage_to_string(consumer_stage
),
262 (input
->data
.centroid
) ? "has" : "lacks");
266 if (input
->data
.sample
!= output
->data
.sample
) {
268 "%s shader output `%s' %s sample qualifier, "
269 "but %s shader input %s sample qualifier\n",
270 _mesa_shader_stage_to_string(producer_stage
),
272 (output
->data
.sample
) ? "has" : "lacks",
273 _mesa_shader_stage_to_string(consumer_stage
),
274 (input
->data
.sample
) ? "has" : "lacks");
278 if (input
->data
.patch
!= output
->data
.patch
) {
280 "%s shader output `%s' %s patch qualifier, "
281 "but %s shader input %s patch qualifier\n",
282 _mesa_shader_stage_to_string(producer_stage
),
284 (output
->data
.patch
) ? "has" : "lacks",
285 _mesa_shader_stage_to_string(consumer_stage
),
286 (input
->data
.patch
) ? "has" : "lacks");
290 /* The GLSL 4.30 and GLSL ES 3.00 specifications say:
292 * "As only outputs need be declared with invariant, an output from
293 * one shader stage will still match an input of a subsequent stage
294 * without the input being declared as invariant."
296 * while GLSL 4.20 says:
298 * "For variables leaving one shader and coming into another shader,
299 * the invariant keyword has to be used in both shaders, or a link
300 * error will result."
302 * and GLSL ES 1.00 section 4.6.4 "Invariance and Linking" says:
304 * "The invariance of varyings that are declared in both the vertex
305 * and fragment shaders must match."
307 if (input
->data
.invariant
!= output
->data
.invariant
&&
308 prog
->data
->Version
< (prog
->IsES
? 300 : 430)) {
310 "%s shader output `%s' %s invariant qualifier, "
311 "but %s shader input %s invariant qualifier\n",
312 _mesa_shader_stage_to_string(producer_stage
),
314 (output
->data
.invariant
) ? "has" : "lacks",
315 _mesa_shader_stage_to_string(consumer_stage
),
316 (input
->data
.invariant
) ? "has" : "lacks");
320 /* GLSL >= 4.40 removes text requiring interpolation qualifiers
321 * to match cross stage, they must only match within the same stage.
323 * From page 84 (page 90 of the PDF) of the GLSL 4.40 spec:
325 * "It is a link-time error if, within the same stage, the interpolation
326 * qualifiers of variables of the same name do not match.
328 * Section 4.3.9 (Interpolation) of the GLSL ES 3.00 spec says:
330 * "When no interpolation qualifier is present, smooth interpolation
333 * So we match variables where one is smooth and the other has no explicit
336 unsigned input_interpolation
= input
->data
.interpolation
;
337 unsigned output_interpolation
= output
->data
.interpolation
;
339 if (input_interpolation
== INTERP_MODE_NONE
)
340 input_interpolation
= INTERP_MODE_SMOOTH
;
341 if (output_interpolation
== INTERP_MODE_NONE
)
342 output_interpolation
= INTERP_MODE_SMOOTH
;
344 if (input_interpolation
!= output_interpolation
&&
345 prog
->data
->Version
< 440) {
347 "%s shader output `%s' specifies %s "
348 "interpolation qualifier, "
349 "but %s shader input specifies %s "
350 "interpolation qualifier\n",
351 _mesa_shader_stage_to_string(producer_stage
),
353 interpolation_string(output
->data
.interpolation
),
354 _mesa_shader_stage_to_string(consumer_stage
),
355 interpolation_string(input
->data
.interpolation
));
361 * Validate front and back color outputs against single color input
364 cross_validate_front_and_back_color(struct gl_shader_program
*prog
,
365 const ir_variable
*input
,
366 const ir_variable
*front_color
,
367 const ir_variable
*back_color
,
368 gl_shader_stage consumer_stage
,
369 gl_shader_stage producer_stage
)
371 if (front_color
!= NULL
&& front_color
->data
.assigned
)
372 cross_validate_types_and_qualifiers(prog
, input
, front_color
,
373 consumer_stage
, producer_stage
);
375 if (back_color
!= NULL
&& back_color
->data
.assigned
)
376 cross_validate_types_and_qualifiers(prog
, input
, back_color
,
377 consumer_stage
, producer_stage
);
381 compute_variable_location_slot(ir_variable
*var
, gl_shader_stage stage
)
383 unsigned location_start
= VARYING_SLOT_VAR0
;
386 case MESA_SHADER_VERTEX
:
387 if (var
->data
.mode
== ir_var_shader_in
)
388 location_start
= VERT_ATTRIB_GENERIC0
;
390 case MESA_SHADER_TESS_CTRL
:
391 case MESA_SHADER_TESS_EVAL
:
393 location_start
= VARYING_SLOT_PATCH0
;
395 case MESA_SHADER_FRAGMENT
:
396 if (var
->data
.mode
== ir_var_shader_out
)
397 location_start
= FRAG_RESULT_DATA0
;
403 return var
->data
.location
- location_start
;
406 struct explicit_location_info
{
408 unsigned numerical_type
;
409 unsigned interpolation
;
415 static inline unsigned
416 get_numerical_type(const glsl_type
*type
)
418 /* From the OpenGL 4.6 spec, section 4.4.1 Input Layout Qualifiers, Page 68,
419 * (Location aliasing):
421 * "Further, when location aliasing, the aliases sharing the location
422 * must have the same underlying numerical type (floating-point or
425 if (type
->is_float() || type
->is_double())
426 return GLSL_TYPE_FLOAT
;
427 return GLSL_TYPE_INT
;
431 check_location_aliasing(struct explicit_location_info explicit_locations
[][4],
435 unsigned location_limit
,
436 const glsl_type
*type
,
437 unsigned interpolation
,
441 gl_shader_program
*prog
,
442 gl_shader_stage stage
)
445 if (type
->without_array()->is_record()) {
446 /* The component qualifier can't be used on structs so just treat
447 * all component slots as used.
451 unsigned dmul
= type
->without_array()->is_64bit() ? 2 : 1;
452 last_comp
= component
+ type
->without_array()->vector_elements
* dmul
;
455 while (location
< location_limit
) {
458 struct explicit_location_info
*info
=
459 &explicit_locations
[location
][comp
];
462 /* Component aliasing is not alloed */
463 if (comp
>= component
&& comp
< last_comp
) {
465 "%s shader has multiple outputs explicitly "
466 "assigned to location %d and component %d\n",
467 _mesa_shader_stage_to_string(stage
),
471 /* For all other used components we need to have matching
472 * types, interpolation and auxiliary storage
474 if (info
->numerical_type
!=
475 get_numerical_type(type
->without_array())) {
477 "Varyings sharing the same location must "
478 "have the same underlying numerical type. "
479 "Location %u component %u\n",
484 if (info
->interpolation
!= interpolation
) {
486 "%s shader has multiple outputs at explicit "
487 "location %u with different interpolation "
489 _mesa_shader_stage_to_string(stage
), location
);
493 if (info
->centroid
!= centroid
||
494 info
->sample
!= sample
||
495 info
->patch
!= patch
) {
497 "%s shader has multiple outputs at explicit "
498 "location %u with different aux storage\n",
499 _mesa_shader_stage_to_string(stage
), location
);
503 } else if (comp
>= component
&& comp
< last_comp
) {
505 info
->numerical_type
= get_numerical_type(type
->without_array());
506 info
->interpolation
= interpolation
;
507 info
->centroid
= centroid
;
508 info
->sample
= sample
;
514 /* We need to do some special handling for doubles as dvec3 and
515 * dvec4 consume two consecutive locations. We don't need to
516 * worry about components beginning at anything other than 0 as
517 * the spec does not allow this for dvec3 and dvec4.
519 if (comp
== 4 && last_comp
> 4) {
520 last_comp
= last_comp
- 4;
521 /* Bump location index and reset the component index */
535 validate_explicit_variable_location(struct gl_context
*ctx
,
536 struct explicit_location_info explicit_locations
[][4],
538 gl_shader_program
*prog
,
539 gl_linked_shader
*sh
)
541 const glsl_type
*type
= get_varying_type(var
, sh
->Stage
);
542 unsigned num_elements
= type
->count_attribute_slots(false);
543 unsigned idx
= compute_variable_location_slot(var
, sh
->Stage
);
544 unsigned slot_limit
= idx
+ num_elements
;
546 /* Vertex shader inputs and fragment shader outputs are validated in
547 * assign_attribute_or_color_locations() so we should not attempt to
548 * validate them again here.
551 if (var
->data
.mode
== ir_var_shader_out
) {
552 assert(sh
->Stage
!= MESA_SHADER_FRAGMENT
);
554 ctx
->Const
.Program
[sh
->Stage
].MaxOutputComponents
/ 4;
556 assert(var
->data
.mode
== ir_var_shader_in
);
557 assert(sh
->Stage
!= MESA_SHADER_VERTEX
);
559 ctx
->Const
.Program
[sh
->Stage
].MaxInputComponents
/ 4;
562 if (slot_limit
> slot_max
) {
564 "Invalid location %u in %s shader\n",
565 idx
, _mesa_shader_stage_to_string(sh
->Stage
));
569 if (type
->without_array()->is_interface()) {
570 for (unsigned i
= 0; i
< type
->without_array()->length
; i
++) {
571 glsl_struct_field
*field
= &type
->fields
.structure
[i
];
572 unsigned field_location
= field
->location
-
573 (field
->patch
? VARYING_SLOT_PATCH0
: VARYING_SLOT_VAR0
);
574 if (!check_location_aliasing(explicit_locations
, var
,
576 0, field_location
+ 1,
578 field
->interpolation
,
586 } else if (!check_location_aliasing(explicit_locations
, var
,
587 idx
, var
->data
.location_frac
,
589 var
->data
.interpolation
,
601 * Validate explicit locations for the inputs to the first stage and the
602 * outputs of the last stage in an SSO program (everything in between is
603 * validated in cross_validate_outputs_to_inputs).
606 validate_sso_explicit_locations(struct gl_context
*ctx
,
607 struct gl_shader_program
*prog
,
608 gl_shader_stage first_stage
,
609 gl_shader_stage last_stage
)
611 assert(prog
->SeparateShader
);
613 /* VS inputs and FS outputs are validated in
614 * assign_attribute_or_color_locations()
616 bool validate_first_stage
= first_stage
!= MESA_SHADER_VERTEX
;
617 bool validate_last_stage
= last_stage
!= MESA_SHADER_FRAGMENT
;
618 if (!validate_first_stage
&& !validate_last_stage
)
621 struct explicit_location_info explicit_locations
[MAX_VARYING
][4];
623 gl_shader_stage stages
[2] = { first_stage
, last_stage
};
624 bool validate_stage
[2] = { validate_first_stage
, validate_last_stage
};
625 ir_variable_mode var_direction
[2] = { ir_var_shader_in
, ir_var_shader_out
};
627 for (unsigned i
= 0; i
< 2; i
++) {
628 if (!validate_stage
[i
])
631 gl_shader_stage stage
= stages
[i
];
633 gl_linked_shader
*sh
= prog
->_LinkedShaders
[stage
];
636 memset(explicit_locations
, 0, sizeof(explicit_locations
));
638 foreach_in_list(ir_instruction
, node
, sh
->ir
) {
639 ir_variable
*const var
= node
->as_variable();
642 !var
->data
.explicit_location
||
643 var
->data
.location
< VARYING_SLOT_VAR0
||
644 var
->data
.mode
!= var_direction
[i
])
647 if (!validate_explicit_variable_location(
648 ctx
, explicit_locations
, var
, prog
, sh
)) {
656 * Validate that outputs from one stage match inputs of another
659 cross_validate_outputs_to_inputs(struct gl_context
*ctx
,
660 struct gl_shader_program
*prog
,
661 gl_linked_shader
*producer
,
662 gl_linked_shader
*consumer
)
664 glsl_symbol_table parameters
;
665 struct explicit_location_info explicit_locations
[MAX_VARYING
][4] = { 0 };
667 /* Find all shader outputs in the "producer" stage.
669 foreach_in_list(ir_instruction
, node
, producer
->ir
) {
670 ir_variable
*const var
= node
->as_variable();
672 if (var
== NULL
|| var
->data
.mode
!= ir_var_shader_out
)
675 if (!var
->data
.explicit_location
676 || var
->data
.location
< VARYING_SLOT_VAR0
)
677 parameters
.add_variable(var
);
679 /* User-defined varyings with explicit locations are handled
680 * differently because they do not need to have matching names.
682 if (!validate_explicit_variable_location(ctx
,
684 var
, prog
, producer
)) {
691 /* Find all shader inputs in the "consumer" stage. Any variables that have
692 * matching outputs already in the symbol table must have the same type and
695 * Exception: if the consumer is the geometry shader, then the inputs
696 * should be arrays and the type of the array element should match the type
697 * of the corresponding producer output.
699 foreach_in_list(ir_instruction
, node
, consumer
->ir
) {
700 ir_variable
*const input
= node
->as_variable();
702 if (input
== NULL
|| input
->data
.mode
!= ir_var_shader_in
)
705 if (strcmp(input
->name
, "gl_Color") == 0 && input
->data
.used
) {
706 const ir_variable
*const front_color
=
707 parameters
.get_variable("gl_FrontColor");
709 const ir_variable
*const back_color
=
710 parameters
.get_variable("gl_BackColor");
712 cross_validate_front_and_back_color(prog
, input
,
713 front_color
, back_color
,
714 consumer
->Stage
, producer
->Stage
);
715 } else if (strcmp(input
->name
, "gl_SecondaryColor") == 0 && input
->data
.used
) {
716 const ir_variable
*const front_color
=
717 parameters
.get_variable("gl_FrontSecondaryColor");
719 const ir_variable
*const back_color
=
720 parameters
.get_variable("gl_BackSecondaryColor");
722 cross_validate_front_and_back_color(prog
, input
,
723 front_color
, back_color
,
724 consumer
->Stage
, producer
->Stage
);
726 /* The rules for connecting inputs and outputs change in the presence
727 * of explicit locations. In this case, we no longer care about the
728 * names of the variables. Instead, we care only about the
729 * explicitly assigned location.
731 ir_variable
*output
= NULL
;
732 if (input
->data
.explicit_location
733 && input
->data
.location
>= VARYING_SLOT_VAR0
) {
735 const glsl_type
*type
= get_varying_type(input
, consumer
->Stage
);
736 unsigned num_elements
= type
->count_attribute_slots(false);
738 compute_variable_location_slot(input
, consumer
->Stage
);
739 unsigned slot_limit
= idx
+ num_elements
;
741 while (idx
< slot_limit
) {
742 if (idx
>= MAX_VARYING
) {
744 "Invalid location %u in %s shader\n", idx
,
745 _mesa_shader_stage_to_string(consumer
->Stage
));
749 output
= explicit_locations
[idx
][input
->data
.location_frac
].var
;
751 if (output
== NULL
||
752 input
->data
.location
!= output
->data
.location
) {
754 "%s shader input `%s' with explicit location "
755 "has no matching output\n",
756 _mesa_shader_stage_to_string(consumer
->Stage
),
763 output
= parameters
.get_variable(input
->name
);
766 if (output
!= NULL
) {
767 /* Interface blocks have their own validation elsewhere so don't
768 * try validating them here.
770 if (!(input
->get_interface_type() &&
771 output
->get_interface_type()))
772 cross_validate_types_and_qualifiers(prog
, input
, output
,
776 /* Check for input vars with unmatched output vars in prev stage
777 * taking into account that interface blocks could have a matching
778 * output but with different name, so we ignore them.
780 assert(!input
->data
.assigned
);
781 if (input
->data
.used
&& !input
->get_interface_type() &&
782 !input
->data
.explicit_location
&& !prog
->SeparateShader
)
784 "%s shader input `%s' "
785 "has no matching output in the previous stage\n",
786 _mesa_shader_stage_to_string(consumer
->Stage
),
794 * Demote shader inputs and outputs that are not used in other stages, and
795 * remove them via dead code elimination.
798 remove_unused_shader_inputs_and_outputs(bool is_separate_shader_object
,
799 gl_linked_shader
*sh
,
800 enum ir_variable_mode mode
)
802 if (is_separate_shader_object
)
805 foreach_in_list(ir_instruction
, node
, sh
->ir
) {
806 ir_variable
*const var
= node
->as_variable();
808 if (var
== NULL
|| var
->data
.mode
!= int(mode
))
811 /* A shader 'in' or 'out' variable is only really an input or output if
812 * its value is used by other shader stages. This will cause the
813 * variable to have a location assigned.
815 if (var
->data
.is_unmatched_generic_inout
&& !var
->data
.is_xfb_only
) {
816 assert(var
->data
.mode
!= ir_var_temporary
);
818 /* Assign zeros to demoted inputs to allow more optimizations. */
819 if (var
->data
.mode
== ir_var_shader_in
&& !var
->constant_value
)
820 var
->constant_value
= ir_constant::zero(var
, var
->type
);
822 var
->data
.mode
= ir_var_auto
;
826 /* Eliminate code that is now dead due to unused inputs/outputs being
829 while (do_dead_code(sh
->ir
, false))
835 * Initialize this object based on a string that was passed to
836 * glTransformFeedbackVaryings.
838 * If the input is mal-formed, this call still succeeds, but it sets
839 * this->var_name to a mal-formed input, so tfeedback_decl::find_output_var()
840 * will fail to find any matching variable.
843 tfeedback_decl::init(struct gl_context
*ctx
, const void *mem_ctx
,
846 /* We don't have to be pedantic about what is a valid GLSL variable name,
847 * because any variable with an invalid name can't exist in the IR anyway.
851 this->orig_name
= input
;
852 this->lowered_builtin_array_variable
= none
;
853 this->skip_components
= 0;
854 this->next_buffer_separator
= false;
855 this->matched_candidate
= NULL
;
860 if (ctx
->Extensions
.ARB_transform_feedback3
) {
861 /* Parse gl_NextBuffer. */
862 if (strcmp(input
, "gl_NextBuffer") == 0) {
863 this->next_buffer_separator
= true;
867 /* Parse gl_SkipComponents. */
868 if (strcmp(input
, "gl_SkipComponents1") == 0)
869 this->skip_components
= 1;
870 else if (strcmp(input
, "gl_SkipComponents2") == 0)
871 this->skip_components
= 2;
872 else if (strcmp(input
, "gl_SkipComponents3") == 0)
873 this->skip_components
= 3;
874 else if (strcmp(input
, "gl_SkipComponents4") == 0)
875 this->skip_components
= 4;
877 if (this->skip_components
)
881 /* Parse a declaration. */
882 const char *base_name_end
;
883 long subscript
= parse_program_resource_name(input
, &base_name_end
);
884 this->var_name
= ralloc_strndup(mem_ctx
, input
, base_name_end
- input
);
885 if (this->var_name
== NULL
) {
886 _mesa_error_no_memory(__func__
);
890 if (subscript
>= 0) {
891 this->array_subscript
= subscript
;
892 this->is_subscripted
= true;
894 this->is_subscripted
= false;
897 /* For drivers that lower gl_ClipDistance to gl_ClipDistanceMESA, this
898 * class must behave specially to account for the fact that gl_ClipDistance
899 * is converted from a float[8] to a vec4[2].
901 if (ctx
->Const
.ShaderCompilerOptions
[MESA_SHADER_VERTEX
].LowerCombinedClipCullDistance
&&
902 strcmp(this->var_name
, "gl_ClipDistance") == 0) {
903 this->lowered_builtin_array_variable
= clip_distance
;
905 if (ctx
->Const
.ShaderCompilerOptions
[MESA_SHADER_VERTEX
].LowerCombinedClipCullDistance
&&
906 strcmp(this->var_name
, "gl_CullDistance") == 0) {
907 this->lowered_builtin_array_variable
= cull_distance
;
910 if (ctx
->Const
.LowerTessLevel
&&
911 (strcmp(this->var_name
, "gl_TessLevelOuter") == 0))
912 this->lowered_builtin_array_variable
= tess_level_outer
;
913 if (ctx
->Const
.LowerTessLevel
&&
914 (strcmp(this->var_name
, "gl_TessLevelInner") == 0))
915 this->lowered_builtin_array_variable
= tess_level_inner
;
920 * Determine whether two tfeedback_decl objects refer to the same variable and
921 * array index (if applicable).
924 tfeedback_decl::is_same(const tfeedback_decl
&x
, const tfeedback_decl
&y
)
926 assert(x
.is_varying() && y
.is_varying());
928 if (strcmp(x
.var_name
, y
.var_name
) != 0)
930 if (x
.is_subscripted
!= y
.is_subscripted
)
932 if (x
.is_subscripted
&& x
.array_subscript
!= y
.array_subscript
)
939 * Assign a location and stream ID for this tfeedback_decl object based on the
940 * transform feedback candidate found by find_candidate.
942 * If an error occurs, the error is reported through linker_error() and false
946 tfeedback_decl::assign_location(struct gl_context
*ctx
,
947 struct gl_shader_program
*prog
)
949 assert(this->is_varying());
951 unsigned fine_location
952 = this->matched_candidate
->toplevel_var
->data
.location
* 4
953 + this->matched_candidate
->toplevel_var
->data
.location_frac
954 + this->matched_candidate
->offset
;
955 const unsigned dmul
=
956 this->matched_candidate
->type
->without_array()->is_64bit() ? 2 : 1;
958 if (this->matched_candidate
->type
->is_array()) {
960 const unsigned matrix_cols
=
961 this->matched_candidate
->type
->fields
.array
->matrix_columns
;
962 const unsigned vector_elements
=
963 this->matched_candidate
->type
->fields
.array
->vector_elements
;
964 unsigned actual_array_size
;
965 switch (this->lowered_builtin_array_variable
) {
967 actual_array_size
= prog
->last_vert_prog
?
968 prog
->last_vert_prog
->info
.clip_distance_array_size
: 0;
971 actual_array_size
= prog
->last_vert_prog
?
972 prog
->last_vert_prog
->info
.cull_distance_array_size
: 0;
974 case tess_level_outer
:
975 actual_array_size
= 4;
977 case tess_level_inner
:
978 actual_array_size
= 2;
982 actual_array_size
= this->matched_candidate
->type
->array_size();
986 if (this->is_subscripted
) {
987 /* Check array bounds. */
988 if (this->array_subscript
>= actual_array_size
) {
989 linker_error(prog
, "Transform feedback varying %s has index "
990 "%i, but the array size is %u.",
991 this->orig_name
, this->array_subscript
,
995 unsigned array_elem_size
= this->lowered_builtin_array_variable
?
996 1 : vector_elements
* matrix_cols
* dmul
;
997 fine_location
+= array_elem_size
* this->array_subscript
;
1000 this->size
= actual_array_size
;
1002 this->vector_elements
= vector_elements
;
1003 this->matrix_columns
= matrix_cols
;
1004 if (this->lowered_builtin_array_variable
)
1005 this->type
= GL_FLOAT
;
1007 this->type
= this->matched_candidate
->type
->fields
.array
->gl_type
;
1009 /* Regular variable (scalar, vector, or matrix) */
1010 if (this->is_subscripted
) {
1011 linker_error(prog
, "Transform feedback varying %s requested, "
1012 "but %s is not an array.",
1013 this->orig_name
, this->var_name
);
1017 this->vector_elements
= this->matched_candidate
->type
->vector_elements
;
1018 this->matrix_columns
= this->matched_candidate
->type
->matrix_columns
;
1019 this->type
= this->matched_candidate
->type
->gl_type
;
1021 this->location
= fine_location
/ 4;
1022 this->location_frac
= fine_location
% 4;
1024 /* From GL_EXT_transform_feedback:
1025 * A program will fail to link if:
1027 * * the total number of components to capture in any varying
1028 * variable in <varyings> is greater than the constant
1029 * MAX_TRANSFORM_FEEDBACK_SEPARATE_COMPONENTS_EXT and the
1030 * buffer mode is SEPARATE_ATTRIBS_EXT;
1032 if (prog
->TransformFeedback
.BufferMode
== GL_SEPARATE_ATTRIBS
&&
1033 this->num_components() >
1034 ctx
->Const
.MaxTransformFeedbackSeparateComponents
) {
1035 linker_error(prog
, "Transform feedback varying %s exceeds "
1036 "MAX_TRANSFORM_FEEDBACK_SEPARATE_COMPONENTS.",
1041 /* Only transform feedback varyings can be assigned to non-zero streams,
1042 * so assign the stream id here.
1044 this->stream_id
= this->matched_candidate
->toplevel_var
->data
.stream
;
1046 unsigned array_offset
= this->array_subscript
* 4 * dmul
;
1047 unsigned struct_offset
= this->matched_candidate
->offset
* 4 * dmul
;
1048 this->buffer
= this->matched_candidate
->toplevel_var
->data
.xfb_buffer
;
1049 this->offset
= this->matched_candidate
->toplevel_var
->data
.offset
+
1050 array_offset
+ struct_offset
;
1057 tfeedback_decl::get_num_outputs() const
1059 if (!this->is_varying()) {
1062 return (this->num_components() + this->location_frac
+ 3)/4;
1067 * Update gl_transform_feedback_info to reflect this tfeedback_decl.
1069 * If an error occurs, the error is reported through linker_error() and false
1073 tfeedback_decl::store(struct gl_context
*ctx
, struct gl_shader_program
*prog
,
1074 struct gl_transform_feedback_info
*info
,
1075 unsigned buffer
, unsigned buffer_index
,
1076 const unsigned max_outputs
, bool *explicit_stride
,
1077 bool has_xfb_qualifiers
) const
1079 unsigned xfb_offset
= 0;
1080 unsigned size
= this->size
;
1081 /* Handle gl_SkipComponents. */
1082 if (this->skip_components
) {
1083 info
->Buffers
[buffer
].Stride
+= this->skip_components
;
1084 size
= this->skip_components
;
1088 if (this->next_buffer_separator
) {
1093 if (has_xfb_qualifiers
) {
1094 xfb_offset
= this->offset
/ 4;
1096 xfb_offset
= info
->Buffers
[buffer
].Stride
;
1098 info
->Varyings
[info
->NumVarying
].Offset
= xfb_offset
* 4;
1101 unsigned location
= this->location
;
1102 unsigned location_frac
= this->location_frac
;
1103 unsigned num_components
= this->num_components();
1104 while (num_components
> 0) {
1105 unsigned output_size
= MIN2(num_components
, 4 - location_frac
);
1106 assert((info
->NumOutputs
== 0 && max_outputs
== 0) ||
1107 info
->NumOutputs
< max_outputs
);
1109 /* From the ARB_enhanced_layouts spec:
1111 * "If such a block member or variable is not written during a shader
1112 * invocation, the buffer contents at the assigned offset will be
1113 * undefined. Even if there are no static writes to a variable or
1114 * member that is assigned a transform feedback offset, the space is
1115 * still allocated in the buffer and still affects the stride."
1117 if (this->is_varying_written()) {
1118 info
->Outputs
[info
->NumOutputs
].ComponentOffset
= location_frac
;
1119 info
->Outputs
[info
->NumOutputs
].OutputRegister
= location
;
1120 info
->Outputs
[info
->NumOutputs
].NumComponents
= output_size
;
1121 info
->Outputs
[info
->NumOutputs
].StreamId
= stream_id
;
1122 info
->Outputs
[info
->NumOutputs
].OutputBuffer
= buffer
;
1123 info
->Outputs
[info
->NumOutputs
].DstOffset
= xfb_offset
;
1126 info
->Buffers
[buffer
].Stream
= this->stream_id
;
1127 xfb_offset
+= output_size
;
1129 num_components
-= output_size
;
1135 if (explicit_stride
&& explicit_stride
[buffer
]) {
1136 if (this->is_64bit() && info
->Buffers
[buffer
].Stride
% 2) {
1137 linker_error(prog
, "invalid qualifier xfb_stride=%d must be a "
1138 "multiple of 8 as its applied to a type that is or "
1139 "contains a double.",
1140 info
->Buffers
[buffer
].Stride
* 4);
1144 if ((this->offset
/ 4) / info
->Buffers
[buffer
].Stride
!=
1145 (xfb_offset
- 1) / info
->Buffers
[buffer
].Stride
) {
1146 linker_error(prog
, "xfb_offset (%d) overflows xfb_stride (%d) for "
1147 "buffer (%d)", xfb_offset
* 4,
1148 info
->Buffers
[buffer
].Stride
* 4, buffer
);
1152 info
->Buffers
[buffer
].Stride
= xfb_offset
;
1155 /* From GL_EXT_transform_feedback:
1156 * A program will fail to link if:
1158 * * the total number of components to capture is greater than
1159 * the constant MAX_TRANSFORM_FEEDBACK_INTERLEAVED_COMPONENTS_EXT
1160 * and the buffer mode is INTERLEAVED_ATTRIBS_EXT.
1162 * From GL_ARB_enhanced_layouts:
1164 * "The resulting stride (implicit or explicit) must be less than or
1165 * equal to the implementation-dependent constant
1166 * gl_MaxTransformFeedbackInterleavedComponents."
1168 if ((prog
->TransformFeedback
.BufferMode
== GL_INTERLEAVED_ATTRIBS
||
1169 has_xfb_qualifiers
) &&
1170 info
->Buffers
[buffer
].Stride
>
1171 ctx
->Const
.MaxTransformFeedbackInterleavedComponents
) {
1172 linker_error(prog
, "The MAX_TRANSFORM_FEEDBACK_INTERLEAVED_COMPONENTS "
1173 "limit has been exceeded.");
1178 info
->Varyings
[info
->NumVarying
].Name
= ralloc_strdup(prog
,
1180 info
->Varyings
[info
->NumVarying
].Type
= this->type
;
1181 info
->Varyings
[info
->NumVarying
].Size
= size
;
1182 info
->Varyings
[info
->NumVarying
].BufferIndex
= buffer_index
;
1184 info
->Buffers
[buffer
].NumVaryings
++;
1190 const tfeedback_candidate
*
1191 tfeedback_decl::find_candidate(gl_shader_program
*prog
,
1192 hash_table
*tfeedback_candidates
)
1194 const char *name
= this->var_name
;
1195 switch (this->lowered_builtin_array_variable
) {
1197 name
= this->var_name
;
1200 name
= "gl_ClipDistanceMESA";
1203 name
= "gl_CullDistanceMESA";
1205 case tess_level_outer
:
1206 name
= "gl_TessLevelOuterMESA";
1208 case tess_level_inner
:
1209 name
= "gl_TessLevelInnerMESA";
1212 hash_entry
*entry
= _mesa_hash_table_search(tfeedback_candidates
, name
);
1214 this->matched_candidate
= entry
?
1215 (const tfeedback_candidate
*) entry
->data
: NULL
;
1217 if (!this->matched_candidate
) {
1218 /* From GL_EXT_transform_feedback:
1219 * A program will fail to link if:
1221 * * any variable name specified in the <varyings> array is not
1222 * declared as an output in the geometry shader (if present) or
1223 * the vertex shader (if no geometry shader is present);
1225 linker_error(prog
, "Transform feedback varying %s undeclared.",
1229 return this->matched_candidate
;
1234 * Parse all the transform feedback declarations that were passed to
1235 * glTransformFeedbackVaryings() and store them in tfeedback_decl objects.
1237 * If an error occurs, the error is reported through linker_error() and false
1241 parse_tfeedback_decls(struct gl_context
*ctx
, struct gl_shader_program
*prog
,
1242 const void *mem_ctx
, unsigned num_names
,
1243 char **varying_names
, tfeedback_decl
*decls
)
1245 for (unsigned i
= 0; i
< num_names
; ++i
) {
1246 decls
[i
].init(ctx
, mem_ctx
, varying_names
[i
]);
1248 if (!decls
[i
].is_varying())
1251 /* From GL_EXT_transform_feedback:
1252 * A program will fail to link if:
1254 * * any two entries in the <varyings> array specify the same varying
1257 * We interpret this to mean "any two entries in the <varyings> array
1258 * specify the same varying variable and array index", since transform
1259 * feedback of arrays would be useless otherwise.
1261 for (unsigned j
= 0; j
< i
; ++j
) {
1262 if (!decls
[j
].is_varying())
1265 if (tfeedback_decl::is_same(decls
[i
], decls
[j
])) {
1266 linker_error(prog
, "Transform feedback varying %s specified "
1267 "more than once.", varying_names
[i
]);
1277 cmp_xfb_offset(const void * x_generic
, const void * y_generic
)
1279 tfeedback_decl
*x
= (tfeedback_decl
*) x_generic
;
1280 tfeedback_decl
*y
= (tfeedback_decl
*) y_generic
;
1282 if (x
->get_buffer() != y
->get_buffer())
1283 return x
->get_buffer() - y
->get_buffer();
1284 return x
->get_offset() - y
->get_offset();
1288 * Store transform feedback location assignments into
1289 * prog->sh.LinkedTransformFeedback based on the data stored in
1292 * If an error occurs, the error is reported through linker_error() and false
1296 store_tfeedback_info(struct gl_context
*ctx
, struct gl_shader_program
*prog
,
1297 unsigned num_tfeedback_decls
,
1298 tfeedback_decl
*tfeedback_decls
, bool has_xfb_qualifiers
)
1300 if (!prog
->last_vert_prog
)
1303 /* Make sure MaxTransformFeedbackBuffers is less than 32 so the bitmask for
1304 * tracking the number of buffers doesn't overflow.
1306 assert(ctx
->Const
.MaxTransformFeedbackBuffers
< 32);
1308 bool separate_attribs_mode
=
1309 prog
->TransformFeedback
.BufferMode
== GL_SEPARATE_ATTRIBS
;
1311 struct gl_program
*xfb_prog
= prog
->last_vert_prog
;
1312 xfb_prog
->sh
.LinkedTransformFeedback
=
1313 rzalloc(xfb_prog
, struct gl_transform_feedback_info
);
1315 /* The xfb_offset qualifier does not have to be used in increasing order
1316 * however some drivers expect to receive the list of transform feedback
1317 * declarations in order so sort it now for convenience.
1319 if (has_xfb_qualifiers
) {
1320 qsort(tfeedback_decls
, num_tfeedback_decls
, sizeof(*tfeedback_decls
),
1323 xfb_prog
->sh
.LinkedTransformFeedback
->api_enabled
= true;
1326 xfb_prog
->sh
.LinkedTransformFeedback
->Varyings
=
1327 rzalloc_array(xfb_prog
, struct gl_transform_feedback_varying_info
,
1328 num_tfeedback_decls
);
1330 unsigned num_outputs
= 0;
1331 for (unsigned i
= 0; i
< num_tfeedback_decls
; ++i
) {
1332 if (tfeedback_decls
[i
].is_varying_written())
1333 num_outputs
+= tfeedback_decls
[i
].get_num_outputs();
1336 xfb_prog
->sh
.LinkedTransformFeedback
->Outputs
=
1337 rzalloc_array(xfb_prog
, struct gl_transform_feedback_output
,
1340 unsigned num_buffers
= 0;
1341 unsigned buffers
= 0;
1343 if (!has_xfb_qualifiers
&& separate_attribs_mode
) {
1344 /* GL_SEPARATE_ATTRIBS */
1345 for (unsigned i
= 0; i
< num_tfeedback_decls
; ++i
) {
1346 if (!tfeedback_decls
[i
].store(ctx
, prog
,
1347 xfb_prog
->sh
.LinkedTransformFeedback
,
1348 num_buffers
, num_buffers
, num_outputs
,
1349 NULL
, has_xfb_qualifiers
))
1352 buffers
|= 1 << num_buffers
;
1357 /* GL_INVERLEAVED_ATTRIBS */
1358 int buffer_stream_id
= -1;
1360 num_tfeedback_decls
? tfeedback_decls
[0].get_buffer() : 0;
1361 bool explicit_stride
[MAX_FEEDBACK_BUFFERS
] = { false };
1363 /* Apply any xfb_stride global qualifiers */
1364 if (has_xfb_qualifiers
) {
1365 for (unsigned j
= 0; j
< MAX_FEEDBACK_BUFFERS
; j
++) {
1366 if (prog
->TransformFeedback
.BufferStride
[j
]) {
1368 explicit_stride
[j
] = true;
1369 xfb_prog
->sh
.LinkedTransformFeedback
->Buffers
[j
].Stride
=
1370 prog
->TransformFeedback
.BufferStride
[j
] / 4;
1375 for (unsigned i
= 0; i
< num_tfeedback_decls
; ++i
) {
1376 if (has_xfb_qualifiers
&&
1377 buffer
!= tfeedback_decls
[i
].get_buffer()) {
1378 /* we have moved to the next buffer so reset stream id */
1379 buffer_stream_id
= -1;
1383 if (tfeedback_decls
[i
].is_next_buffer_separator()) {
1384 if (!tfeedback_decls
[i
].store(ctx
, prog
,
1385 xfb_prog
->sh
.LinkedTransformFeedback
,
1386 buffer
, num_buffers
, num_outputs
,
1387 explicit_stride
, has_xfb_qualifiers
))
1390 buffer_stream_id
= -1;
1392 } else if (tfeedback_decls
[i
].is_varying()) {
1393 if (buffer_stream_id
== -1) {
1394 /* First varying writing to this buffer: remember its stream */
1395 buffer_stream_id
= (int) tfeedback_decls
[i
].get_stream_id();
1396 } else if (buffer_stream_id
!=
1397 (int) tfeedback_decls
[i
].get_stream_id()) {
1398 /* Varying writes to the same buffer from a different stream */
1400 "Transform feedback can't capture varyings belonging "
1401 "to different vertex streams in a single buffer. "
1402 "Varying %s writes to buffer from stream %u, other "
1403 "varyings in the same buffer write from stream %u.",
1404 tfeedback_decls
[i
].name(),
1405 tfeedback_decls
[i
].get_stream_id(),
1411 if (has_xfb_qualifiers
) {
1412 buffer
= tfeedback_decls
[i
].get_buffer();
1414 buffer
= num_buffers
;
1416 buffers
|= 1 << buffer
;
1418 if (!tfeedback_decls
[i
].store(ctx
, prog
,
1419 xfb_prog
->sh
.LinkedTransformFeedback
,
1420 buffer
, num_buffers
, num_outputs
,
1421 explicit_stride
, has_xfb_qualifiers
))
1426 assert(xfb_prog
->sh
.LinkedTransformFeedback
->NumOutputs
== num_outputs
);
1428 xfb_prog
->sh
.LinkedTransformFeedback
->ActiveBuffers
= buffers
;
1435 * Data structure recording the relationship between outputs of one shader
1436 * stage (the "producer") and inputs of another (the "consumer").
1438 class varying_matches
1441 varying_matches(bool disable_varying_packing
, bool xfb_enabled
,
1442 bool enhanced_layouts_enabled
,
1443 gl_shader_stage producer_stage
,
1444 gl_shader_stage consumer_stage
);
1446 void record(ir_variable
*producer_var
, ir_variable
*consumer_var
);
1447 unsigned assign_locations(struct gl_shader_program
*prog
,
1448 uint8_t *components
,
1449 uint64_t reserved_slots
);
1450 void store_locations() const;
1453 bool is_varying_packing_safe(const glsl_type
*type
,
1454 const ir_variable
*var
);
1457 * If true, this driver disables varying packing, so all varyings need to
1458 * be aligned on slot boundaries, and take up a number of slots equal to
1459 * their number of matrix columns times their array size.
1461 * Packing may also be disabled because our current packing method is not
1462 * safe in SSO or versions of OpenGL where interpolation qualifiers are not
1463 * guaranteed to match across stages.
1465 const bool disable_varying_packing
;
1468 * If true, this driver has transform feedback enabled. The transform
1469 * feedback code requires at least some packing be done even when varying
1470 * packing is disabled, fortunately where transform feedback requires
1471 * packing it's safe to override the disabled setting. See
1472 * is_varying_packing_safe().
1474 const bool xfb_enabled
;
1476 const bool enhanced_layouts_enabled
;
1479 * Enum representing the order in which varyings are packed within a
1482 * Currently we pack vec4's first, then vec2's, then scalar values, then
1483 * vec3's. This order ensures that the only vectors that are at risk of
1484 * having to be "double parked" (split between two adjacent varying slots)
1487 enum packing_order_enum
{
1490 PACKING_ORDER_SCALAR
,
1494 static unsigned compute_packing_class(const ir_variable
*var
);
1495 static packing_order_enum
compute_packing_order(const ir_variable
*var
);
1496 static int match_comparator(const void *x_generic
, const void *y_generic
);
1497 static int xfb_comparator(const void *x_generic
, const void *y_generic
);
1500 * Structure recording the relationship between a single producer output
1501 * and a single consumer input.
1505 * Packing class for this varying, computed by compute_packing_class().
1507 unsigned packing_class
;
1510 * Packing order for this varying, computed by compute_packing_order().
1512 packing_order_enum packing_order
;
1513 unsigned num_components
;
1516 * The output variable in the producer stage.
1518 ir_variable
*producer_var
;
1521 * The input variable in the consumer stage.
1523 ir_variable
*consumer_var
;
1526 * The location which has been assigned for this varying. This is
1527 * expressed in multiples of a float, with the first generic varying
1528 * (i.e. the one referred to by VARYING_SLOT_VAR0) represented by the
1531 unsigned generic_location
;
1535 * The number of elements in the \c matches array that are currently in
1538 unsigned num_matches
;
1541 * The number of elements that were set aside for the \c matches array when
1544 unsigned matches_capacity
;
1546 gl_shader_stage producer_stage
;
1547 gl_shader_stage consumer_stage
;
1550 } /* anonymous namespace */
1552 varying_matches::varying_matches(bool disable_varying_packing
,
1554 bool enhanced_layouts_enabled
,
1555 gl_shader_stage producer_stage
,
1556 gl_shader_stage consumer_stage
)
1557 : disable_varying_packing(disable_varying_packing
),
1558 xfb_enabled(xfb_enabled
),
1559 enhanced_layouts_enabled(enhanced_layouts_enabled
),
1560 producer_stage(producer_stage
),
1561 consumer_stage(consumer_stage
)
1563 /* Note: this initial capacity is rather arbitrarily chosen to be large
1564 * enough for many cases without wasting an unreasonable amount of space.
1565 * varying_matches::record() will resize the array if there are more than
1566 * this number of varyings.
1568 this->matches_capacity
= 8;
1569 this->matches
= (match
*)
1570 malloc(sizeof(*this->matches
) * this->matches_capacity
);
1571 this->num_matches
= 0;
1575 varying_matches::~varying_matches()
1577 free(this->matches
);
1582 * Packing is always safe on individual arrays, structures, and matrices. It
1583 * is also safe if the varying is only used for transform feedback.
1586 varying_matches::is_varying_packing_safe(const glsl_type
*type
,
1587 const ir_variable
*var
)
1589 if (consumer_stage
== MESA_SHADER_TESS_EVAL
||
1590 consumer_stage
== MESA_SHADER_TESS_CTRL
||
1591 producer_stage
== MESA_SHADER_TESS_CTRL
)
1594 return xfb_enabled
&& (type
->is_array() || type
->is_record() ||
1595 type
->is_matrix() || var
->data
.is_xfb_only
);
1600 * Record the given producer/consumer variable pair in the list of variables
1601 * that should later be assigned locations.
1603 * It is permissible for \c consumer_var to be NULL (this happens if a
1604 * variable is output by the producer and consumed by transform feedback, but
1605 * not consumed by the consumer).
1607 * If \c producer_var has already been paired up with a consumer_var, or
1608 * producer_var is part of fixed pipeline functionality (and hence already has
1609 * a location assigned), this function has no effect.
1611 * Note: as a side effect this function may change the interpolation type of
1612 * \c producer_var, but only when the change couldn't possibly affect
1616 varying_matches::record(ir_variable
*producer_var
, ir_variable
*consumer_var
)
1618 assert(producer_var
!= NULL
|| consumer_var
!= NULL
);
1620 if ((producer_var
&& (!producer_var
->data
.is_unmatched_generic_inout
||
1621 producer_var
->data
.explicit_location
)) ||
1622 (consumer_var
&& (!consumer_var
->data
.is_unmatched_generic_inout
||
1623 consumer_var
->data
.explicit_location
))) {
1624 /* Either a location already exists for this variable (since it is part
1625 * of fixed functionality), or it has already been recorded as part of a
1631 bool needs_flat_qualifier
= consumer_var
== NULL
&&
1632 (producer_var
->type
->contains_integer() ||
1633 producer_var
->type
->contains_double());
1635 if (!disable_varying_packing
&&
1636 (needs_flat_qualifier
||
1637 (consumer_stage
!= MESA_SHADER_NONE
&& consumer_stage
!= MESA_SHADER_FRAGMENT
))) {
1638 /* Since this varying is not being consumed by the fragment shader, its
1639 * interpolation type varying cannot possibly affect rendering.
1640 * Also, this variable is non-flat and is (or contains) an integer
1642 * If the consumer stage is unknown, don't modify the interpolation
1643 * type as it could affect rendering later with separate shaders.
1645 * lower_packed_varyings requires all integer varyings to flat,
1646 * regardless of where they appear. We can trivially satisfy that
1647 * requirement by changing the interpolation type to flat here.
1650 producer_var
->data
.centroid
= false;
1651 producer_var
->data
.sample
= false;
1652 producer_var
->data
.interpolation
= INTERP_MODE_FLAT
;
1656 consumer_var
->data
.centroid
= false;
1657 consumer_var
->data
.sample
= false;
1658 consumer_var
->data
.interpolation
= INTERP_MODE_FLAT
;
1662 if (this->num_matches
== this->matches_capacity
) {
1663 this->matches_capacity
*= 2;
1664 this->matches
= (match
*)
1665 realloc(this->matches
,
1666 sizeof(*this->matches
) * this->matches_capacity
);
1669 /* We must use the consumer to compute the packing class because in GL4.4+
1670 * there is no guarantee interpolation qualifiers will match across stages.
1672 * From Section 4.5 (Interpolation Qualifiers) of the GLSL 4.30 spec:
1674 * "The type and presence of interpolation qualifiers of variables with
1675 * the same name declared in all linked shaders for the same cross-stage
1676 * interface must match, otherwise the link command will fail.
1678 * When comparing an output from one stage to an input of a subsequent
1679 * stage, the input and output don't match if their interpolation
1680 * qualifiers (or lack thereof) are not the same."
1682 * This text was also in at least revison 7 of the 4.40 spec but is no
1683 * longer in revision 9 and not in the 4.50 spec.
1685 const ir_variable
*const var
= (consumer_var
!= NULL
)
1686 ? consumer_var
: producer_var
;
1687 const gl_shader_stage stage
= (consumer_var
!= NULL
)
1688 ? consumer_stage
: producer_stage
;
1689 const glsl_type
*type
= get_varying_type(var
, stage
);
1691 if (producer_var
&& consumer_var
&&
1692 consumer_var
->data
.must_be_shader_input
) {
1693 producer_var
->data
.must_be_shader_input
= 1;
1696 this->matches
[this->num_matches
].packing_class
1697 = this->compute_packing_class(var
);
1698 this->matches
[this->num_matches
].packing_order
1699 = this->compute_packing_order(var
);
1700 if ((this->disable_varying_packing
&& !is_varying_packing_safe(type
, var
)) ||
1701 var
->data
.must_be_shader_input
) {
1702 unsigned slots
= type
->count_attribute_slots(false);
1703 this->matches
[this->num_matches
].num_components
= slots
* 4;
1705 this->matches
[this->num_matches
].num_components
1706 = type
->component_slots();
1709 this->matches
[this->num_matches
].producer_var
= producer_var
;
1710 this->matches
[this->num_matches
].consumer_var
= consumer_var
;
1711 this->num_matches
++;
1713 producer_var
->data
.is_unmatched_generic_inout
= 0;
1715 consumer_var
->data
.is_unmatched_generic_inout
= 0;
1720 * Choose locations for all of the variable matches that were previously
1721 * passed to varying_matches::record().
1724 varying_matches::assign_locations(struct gl_shader_program
*prog
,
1725 uint8_t *components
,
1726 uint64_t reserved_slots
)
1728 /* If packing has been disabled then we cannot safely sort the varyings by
1729 * class as it may mean we are using a version of OpenGL where
1730 * interpolation qualifiers are not guaranteed to be matching across
1731 * shaders, sorting in this case could result in mismatching shader
1733 * When packing is disabled the sort orders varyings used by transform
1734 * feedback first, but also depends on *undefined behaviour* of qsort to
1735 * reverse the order of the varyings. See: xfb_comparator().
1737 if (!this->disable_varying_packing
) {
1738 /* Sort varying matches into an order that makes them easy to pack. */
1739 qsort(this->matches
, this->num_matches
, sizeof(*this->matches
),
1740 &varying_matches::match_comparator
);
1742 /* Only sort varyings that are only used by transform feedback. */
1743 qsort(this->matches
, this->num_matches
, sizeof(*this->matches
),
1744 &varying_matches::xfb_comparator
);
1747 unsigned generic_location
= 0;
1748 unsigned generic_patch_location
= MAX_VARYING
*4;
1749 bool previous_var_xfb_only
= false;
1751 for (unsigned i
= 0; i
< this->num_matches
; i
++) {
1752 unsigned *location
= &generic_location
;
1754 const ir_variable
*var
;
1755 const glsl_type
*type
;
1756 bool is_vertex_input
= false;
1757 if (matches
[i
].consumer_var
) {
1758 var
= matches
[i
].consumer_var
;
1759 type
= get_varying_type(var
, consumer_stage
);
1760 if (consumer_stage
== MESA_SHADER_VERTEX
)
1761 is_vertex_input
= true;
1763 var
= matches
[i
].producer_var
;
1764 type
= get_varying_type(var
, producer_stage
);
1767 if (var
->data
.patch
)
1768 location
= &generic_patch_location
;
1770 /* Advance to the next slot if this varying has a different packing
1771 * class than the previous one, and we're not already on a slot
1774 * Also advance to the next slot if packing is disabled. This makes sure
1775 * we don't assign varyings the same locations which is possible
1776 * because we still pack individual arrays, records and matrices even
1777 * when packing is disabled. Note we don't advance to the next slot if
1778 * we can pack varyings together that are only used for transform
1781 if (var
->data
.must_be_shader_input
||
1782 (this->disable_varying_packing
&&
1783 !(previous_var_xfb_only
&& var
->data
.is_xfb_only
)) ||
1784 (i
> 0 && this->matches
[i
- 1].packing_class
1785 != this->matches
[i
].packing_class
)) {
1786 *location
= ALIGN(*location
, 4);
1789 previous_var_xfb_only
= var
->data
.is_xfb_only
;
1791 /* The number of components taken up by this variable. For vertex shader
1792 * inputs, we use the number of slots * 4, as they have different
1795 unsigned num_components
= is_vertex_input
?
1796 type
->count_attribute_slots(is_vertex_input
) * 4 :
1797 this->matches
[i
].num_components
;
1799 /* The last slot for this variable, inclusive. */
1800 unsigned slot_end
= *location
+ num_components
- 1;
1802 /* FIXME: We could be smarter in the below code and loop back over
1803 * trying to fill any locations that we skipped because we couldn't pack
1804 * the varying between an explicit location. For now just let the user
1805 * hit the linking error if we run out of room and suggest they use
1806 * explicit locations.
1808 while (slot_end
< MAX_VARYING
* 4u) {
1809 const unsigned slots
= (slot_end
/ 4u) - (*location
/ 4u) + 1;
1810 const uint64_t slot_mask
= ((1ull << slots
) - 1) << (*location
/ 4u);
1813 if (reserved_slots
& slot_mask
) {
1814 *location
= ALIGN(*location
+ 1, 4);
1815 slot_end
= *location
+ num_components
- 1;
1822 if (!var
->data
.patch
&& slot_end
>= MAX_VARYING
* 4u) {
1823 linker_error(prog
, "insufficient contiguous locations available for "
1824 "%s it is possible an array or struct could not be "
1825 "packed between varyings with explicit locations. Try "
1826 "using an explicit location for arrays and structs.",
1830 if (slot_end
< MAX_VARYINGS_INCL_PATCH
* 4u) {
1831 for (unsigned j
= *location
/ 4u; j
< slot_end
/ 4u; j
++)
1833 components
[slot_end
/ 4u] = (slot_end
& 3) + 1;
1836 this->matches
[i
].generic_location
= *location
;
1838 *location
= slot_end
+ 1;
1841 return (generic_location
+ 3) / 4;
1846 * Update the producer and consumer shaders to reflect the locations
1847 * assignments that were made by varying_matches::assign_locations().
1850 varying_matches::store_locations() const
1852 /* Check is location needs to be packed with lower_packed_varyings() or if
1853 * we can just use ARB_enhanced_layouts packing.
1855 bool pack_loc
[MAX_VARYINGS_INCL_PATCH
] = { 0 };
1856 const glsl_type
*loc_type
[MAX_VARYINGS_INCL_PATCH
][4] = { {NULL
, NULL
} };
1858 for (unsigned i
= 0; i
< this->num_matches
; i
++) {
1859 ir_variable
*producer_var
= this->matches
[i
].producer_var
;
1860 ir_variable
*consumer_var
= this->matches
[i
].consumer_var
;
1861 unsigned generic_location
= this->matches
[i
].generic_location
;
1862 unsigned slot
= generic_location
/ 4;
1863 unsigned offset
= generic_location
% 4;
1866 producer_var
->data
.location
= VARYING_SLOT_VAR0
+ slot
;
1867 producer_var
->data
.location_frac
= offset
;
1871 assert(consumer_var
->data
.location
== -1);
1872 consumer_var
->data
.location
= VARYING_SLOT_VAR0
+ slot
;
1873 consumer_var
->data
.location_frac
= offset
;
1876 /* Find locations suitable for native packing via
1877 * ARB_enhanced_layouts.
1879 if (producer_var
&& consumer_var
) {
1880 if (enhanced_layouts_enabled
) {
1881 const glsl_type
*type
=
1882 get_varying_type(producer_var
, producer_stage
);
1883 if (type
->is_array() || type
->is_matrix() || type
->is_record() ||
1884 type
->is_double()) {
1885 unsigned comp_slots
= type
->component_slots() + offset
;
1886 unsigned slots
= comp_slots
/ 4;
1890 for (unsigned j
= 0; j
< slots
; j
++) {
1891 pack_loc
[slot
+ j
] = true;
1893 } else if (offset
+ type
->vector_elements
> 4) {
1894 pack_loc
[slot
] = true;
1895 pack_loc
[slot
+ 1] = true;
1897 loc_type
[slot
][offset
] = type
;
1903 /* Attempt to use ARB_enhanced_layouts for more efficient packing if
1906 if (enhanced_layouts_enabled
) {
1907 for (unsigned i
= 0; i
< this->num_matches
; i
++) {
1908 ir_variable
*producer_var
= this->matches
[i
].producer_var
;
1909 ir_variable
*consumer_var
= this->matches
[i
].consumer_var
;
1910 unsigned generic_location
= this->matches
[i
].generic_location
;
1911 unsigned slot
= generic_location
/ 4;
1913 if (pack_loc
[slot
] || !producer_var
|| !consumer_var
)
1916 const glsl_type
*type
=
1917 get_varying_type(producer_var
, producer_stage
);
1918 bool type_match
= true;
1919 for (unsigned j
= 0; j
< 4; j
++) {
1920 if (loc_type
[slot
][j
]) {
1921 if (type
->base_type
!= loc_type
[slot
][j
]->base_type
)
1927 producer_var
->data
.explicit_location
= 1;
1928 consumer_var
->data
.explicit_location
= 1;
1929 producer_var
->data
.explicit_component
= 1;
1930 consumer_var
->data
.explicit_component
= 1;
1938 * Compute the "packing class" of the given varying. This is an unsigned
1939 * integer with the property that two variables in the same packing class can
1940 * be safely backed into the same vec4.
1943 varying_matches::compute_packing_class(const ir_variable
*var
)
1945 /* Without help from the back-end, there is no way to pack together
1946 * variables with different interpolation types, because
1947 * lower_packed_varyings must choose exactly one interpolation type for
1948 * each packed varying it creates.
1950 * However, we can safely pack together floats, ints, and uints, because:
1952 * - varyings of base type "int" and "uint" must use the "flat"
1953 * interpolation type, which can only occur in GLSL 1.30 and above.
1955 * - On platforms that support GLSL 1.30 and above, lower_packed_varyings
1956 * can store flat floats as ints without losing any information (using
1957 * the ir_unop_bitcast_* opcodes).
1959 * Therefore, the packing class depends only on the interpolation type.
1961 unsigned packing_class
= var
->data
.centroid
| (var
->data
.sample
<< 1) |
1962 (var
->data
.patch
<< 2) |
1963 (var
->data
.must_be_shader_input
<< 3);
1965 packing_class
+= var
->is_interpolation_flat()
1966 ? unsigned(INTERP_MODE_FLAT
) : var
->data
.interpolation
;
1967 return packing_class
;
1972 * Compute the "packing order" of the given varying. This is a sort key we
1973 * use to determine when to attempt to pack the given varying relative to
1974 * other varyings in the same packing class.
1976 varying_matches::packing_order_enum
1977 varying_matches::compute_packing_order(const ir_variable
*var
)
1979 const glsl_type
*element_type
= var
->type
;
1981 while (element_type
->is_array()) {
1982 element_type
= element_type
->fields
.array
;
1985 switch (element_type
->component_slots() % 4) {
1986 case 1: return PACKING_ORDER_SCALAR
;
1987 case 2: return PACKING_ORDER_VEC2
;
1988 case 3: return PACKING_ORDER_VEC3
;
1989 case 0: return PACKING_ORDER_VEC4
;
1991 assert(!"Unexpected value of vector_elements");
1992 return PACKING_ORDER_VEC4
;
1998 * Comparison function passed to qsort() to sort varyings by packing_class and
1999 * then by packing_order.
2002 varying_matches::match_comparator(const void *x_generic
, const void *y_generic
)
2004 const match
*x
= (const match
*) x_generic
;
2005 const match
*y
= (const match
*) y_generic
;
2007 if (x
->packing_class
!= y
->packing_class
)
2008 return x
->packing_class
- y
->packing_class
;
2009 return x
->packing_order
- y
->packing_order
;
2014 * Comparison function passed to qsort() to sort varyings used only by
2015 * transform feedback when packing of other varyings is disabled.
2018 varying_matches::xfb_comparator(const void *x_generic
, const void *y_generic
)
2020 const match
*x
= (const match
*) x_generic
;
2022 if (x
->producer_var
!= NULL
&& x
->producer_var
->data
.is_xfb_only
)
2023 return match_comparator(x_generic
, y_generic
);
2025 /* FIXME: When the comparator returns 0 it means the elements being
2026 * compared are equivalent. However the qsort documentation says:
2028 * "The order of equivalent elements is undefined."
2030 * In practice the sort ends up reversing the order of the varyings which
2031 * means locations are also assigned in this reversed order and happens to
2032 * be what we want. This is also whats happening in
2033 * varying_matches::match_comparator().
2040 * Is the given variable a varying variable to be counted against the
2041 * limit in ctx->Const.MaxVarying?
2042 * This includes variables such as texcoords, colors and generic
2043 * varyings, but excludes variables such as gl_FrontFacing and gl_FragCoord.
2046 var_counts_against_varying_limit(gl_shader_stage stage
, const ir_variable
*var
)
2048 /* Only fragment shaders will take a varying variable as an input */
2049 if (stage
== MESA_SHADER_FRAGMENT
&&
2050 var
->data
.mode
== ir_var_shader_in
) {
2051 switch (var
->data
.location
) {
2052 case VARYING_SLOT_POS
:
2053 case VARYING_SLOT_FACE
:
2054 case VARYING_SLOT_PNTC
:
2065 * Visitor class that generates tfeedback_candidate structs describing all
2066 * possible targets of transform feedback.
2068 * tfeedback_candidate structs are stored in the hash table
2069 * tfeedback_candidates, which is passed to the constructor. This hash table
2070 * maps varying names to instances of the tfeedback_candidate struct.
2072 class tfeedback_candidate_generator
: public program_resource_visitor
2075 tfeedback_candidate_generator(void *mem_ctx
,
2076 hash_table
*tfeedback_candidates
)
2078 tfeedback_candidates(tfeedback_candidates
),
2084 void process(ir_variable
*var
)
2086 /* All named varying interface blocks should be flattened by now */
2087 assert(!var
->is_interface_instance());
2089 this->toplevel_var
= var
;
2090 this->varying_floats
= 0;
2091 program_resource_visitor::process(var
, false);
2095 virtual void visit_field(const glsl_type
*type
, const char *name
,
2096 bool /* row_major */,
2097 const glsl_type
* /* record_type */,
2098 const enum glsl_interface_packing
,
2099 bool /* last_field */)
2101 assert(!type
->without_array()->is_record());
2102 assert(!type
->without_array()->is_interface());
2104 tfeedback_candidate
*candidate
2105 = rzalloc(this->mem_ctx
, tfeedback_candidate
);
2106 candidate
->toplevel_var
= this->toplevel_var
;
2107 candidate
->type
= type
;
2108 candidate
->offset
= this->varying_floats
;
2109 _mesa_hash_table_insert(this->tfeedback_candidates
,
2110 ralloc_strdup(this->mem_ctx
, name
),
2112 this->varying_floats
+= type
->component_slots();
2116 * Memory context used to allocate hash table keys and values.
2118 void * const mem_ctx
;
2121 * Hash table in which tfeedback_candidate objects should be stored.
2123 hash_table
* const tfeedback_candidates
;
2126 * Pointer to the toplevel variable that is being traversed.
2128 ir_variable
*toplevel_var
;
2131 * Total number of varying floats that have been visited so far. This is
2132 * used to determine the offset to each varying within the toplevel
2135 unsigned varying_floats
;
2142 populate_consumer_input_sets(void *mem_ctx
, exec_list
*ir
,
2143 hash_table
*consumer_inputs
,
2144 hash_table
*consumer_interface_inputs
,
2145 ir_variable
*consumer_inputs_with_locations
[VARYING_SLOT_TESS_MAX
])
2147 memset(consumer_inputs_with_locations
,
2149 sizeof(consumer_inputs_with_locations
[0]) * VARYING_SLOT_TESS_MAX
);
2151 foreach_in_list(ir_instruction
, node
, ir
) {
2152 ir_variable
*const input_var
= node
->as_variable();
2154 if (input_var
!= NULL
&& input_var
->data
.mode
== ir_var_shader_in
) {
2155 /* All interface blocks should have been lowered by this point */
2156 assert(!input_var
->type
->is_interface());
2158 if (input_var
->data
.explicit_location
) {
2159 /* assign_varying_locations only cares about finding the
2160 * ir_variable at the start of a contiguous location block.
2162 * - For !producer, consumer_inputs_with_locations isn't used.
2164 * - For !consumer, consumer_inputs_with_locations is empty.
2166 * For consumer && producer, if you were trying to set some
2167 * ir_variable to the middle of a location block on the other side
2168 * of producer/consumer, cross_validate_outputs_to_inputs() should
2169 * be link-erroring due to either type mismatch or location
2170 * overlaps. If the variables do match up, then they've got a
2171 * matching data.location and you only looked at
2172 * consumer_inputs_with_locations[var->data.location], not any
2173 * following entries for the array/structure.
2175 consumer_inputs_with_locations
[input_var
->data
.location
] =
2177 } else if (input_var
->get_interface_type() != NULL
) {
2178 char *const iface_field_name
=
2179 ralloc_asprintf(mem_ctx
, "%s.%s",
2180 input_var
->get_interface_type()->without_array()->name
,
2182 _mesa_hash_table_insert(consumer_interface_inputs
,
2183 iface_field_name
, input_var
);
2185 _mesa_hash_table_insert(consumer_inputs
,
2186 ralloc_strdup(mem_ctx
, input_var
->name
),
2194 * Find a variable from the consumer that "matches" the specified variable
2196 * This function only finds inputs with names that match. There is no
2197 * validation (here) that the types, etc. are compatible.
2200 get_matching_input(void *mem_ctx
,
2201 const ir_variable
*output_var
,
2202 hash_table
*consumer_inputs
,
2203 hash_table
*consumer_interface_inputs
,
2204 ir_variable
*consumer_inputs_with_locations
[VARYING_SLOT_TESS_MAX
])
2206 ir_variable
*input_var
;
2208 if (output_var
->data
.explicit_location
) {
2209 input_var
= consumer_inputs_with_locations
[output_var
->data
.location
];
2210 } else if (output_var
->get_interface_type() != NULL
) {
2211 char *const iface_field_name
=
2212 ralloc_asprintf(mem_ctx
, "%s.%s",
2213 output_var
->get_interface_type()->without_array()->name
,
2215 hash_entry
*entry
= _mesa_hash_table_search(consumer_interface_inputs
, iface_field_name
);
2216 input_var
= entry
? (ir_variable
*) entry
->data
: NULL
;
2218 hash_entry
*entry
= _mesa_hash_table_search(consumer_inputs
, output_var
->name
);
2219 input_var
= entry
? (ir_variable
*) entry
->data
: NULL
;
2222 return (input_var
== NULL
|| input_var
->data
.mode
!= ir_var_shader_in
)
2229 io_variable_cmp(const void *_a
, const void *_b
)
2231 const ir_variable
*const a
= *(const ir_variable
**) _a
;
2232 const ir_variable
*const b
= *(const ir_variable
**) _b
;
2234 if (a
->data
.explicit_location
&& b
->data
.explicit_location
)
2235 return b
->data
.location
- a
->data
.location
;
2237 if (a
->data
.explicit_location
&& !b
->data
.explicit_location
)
2240 if (!a
->data
.explicit_location
&& b
->data
.explicit_location
)
2243 return -strcmp(a
->name
, b
->name
);
2247 * Sort the shader IO variables into canonical order
2250 canonicalize_shader_io(exec_list
*ir
, enum ir_variable_mode io_mode
)
2252 ir_variable
*var_table
[MAX_PROGRAM_OUTPUTS
* 4];
2253 unsigned num_variables
= 0;
2255 foreach_in_list(ir_instruction
, node
, ir
) {
2256 ir_variable
*const var
= node
->as_variable();
2258 if (var
== NULL
|| var
->data
.mode
!= io_mode
)
2261 /* If we have already encountered more I/O variables that could
2262 * successfully link, bail.
2264 if (num_variables
== ARRAY_SIZE(var_table
))
2267 var_table
[num_variables
++] = var
;
2270 if (num_variables
== 0)
2273 /* Sort the list in reverse order (io_variable_cmp handles this). Later
2274 * we're going to push the variables on to the IR list as a stack, so we
2275 * want the last variable (in canonical order) to be first in the list.
2277 qsort(var_table
, num_variables
, sizeof(var_table
[0]), io_variable_cmp
);
2279 /* Remove the variable from it's current location in the IR, and put it at
2282 for (unsigned i
= 0; i
< num_variables
; i
++) {
2283 var_table
[i
]->remove();
2284 ir
->push_head(var_table
[i
]);
2289 * Generate a bitfield map of the explicit locations for shader varyings.
2291 * Note: For Tessellation shaders we are sitting right on the limits of the
2292 * 64 bit map. Per-vertex and per-patch both have separate location domains
2293 * with a max of MAX_VARYING.
2296 reserved_varying_slot(struct gl_linked_shader
*stage
,
2297 ir_variable_mode io_mode
)
2299 assert(io_mode
== ir_var_shader_in
|| io_mode
== ir_var_shader_out
);
2300 /* Avoid an overflow of the returned value */
2301 assert(MAX_VARYINGS_INCL_PATCH
<= 64);
2309 foreach_in_list(ir_instruction
, node
, stage
->ir
) {
2310 ir_variable
*const var
= node
->as_variable();
2312 if (var
== NULL
|| var
->data
.mode
!= io_mode
||
2313 !var
->data
.explicit_location
||
2314 var
->data
.location
< VARYING_SLOT_VAR0
)
2317 var_slot
= var
->data
.location
- VARYING_SLOT_VAR0
;
2319 unsigned num_elements
= get_varying_type(var
, stage
->Stage
)
2320 ->count_attribute_slots(io_mode
== ir_var_shader_in
&&
2321 stage
->Stage
== MESA_SHADER_VERTEX
);
2322 for (unsigned i
= 0; i
< num_elements
; i
++) {
2323 if (var_slot
>= 0 && var_slot
< MAX_VARYINGS_INCL_PATCH
)
2324 slots
|= UINT64_C(1) << var_slot
;
2334 * Assign locations for all variables that are produced in one pipeline stage
2335 * (the "producer") and consumed in the next stage (the "consumer").
2337 * Variables produced by the producer may also be consumed by transform
2340 * \param num_tfeedback_decls is the number of declarations indicating
2341 * variables that may be consumed by transform feedback.
2343 * \param tfeedback_decls is a pointer to an array of tfeedback_decl objects
2344 * representing the result of parsing the strings passed to
2345 * glTransformFeedbackVaryings(). assign_location() will be called for
2346 * each of these objects that matches one of the outputs of the
2349 * When num_tfeedback_decls is nonzero, it is permissible for the consumer to
2350 * be NULL. In this case, varying locations are assigned solely based on the
2351 * requirements of transform feedback.
2354 assign_varying_locations(struct gl_context
*ctx
,
2356 struct gl_shader_program
*prog
,
2357 gl_linked_shader
*producer
,
2358 gl_linked_shader
*consumer
,
2359 unsigned num_tfeedback_decls
,
2360 tfeedback_decl
*tfeedback_decls
,
2361 const uint64_t reserved_slots
)
2363 /* Tessellation shaders treat inputs and outputs as shared memory and can
2364 * access inputs and outputs of other invocations.
2365 * Therefore, they can't be lowered to temps easily (and definitely not
2368 bool unpackable_tess
=
2369 (consumer
&& consumer
->Stage
== MESA_SHADER_TESS_EVAL
) ||
2370 (consumer
&& consumer
->Stage
== MESA_SHADER_TESS_CTRL
) ||
2371 (producer
&& producer
->Stage
== MESA_SHADER_TESS_CTRL
);
2373 /* Transform feedback code assumes varying arrays are packed, so if the
2374 * driver has disabled varying packing, make sure to at least enable
2375 * packing required by transform feedback.
2378 ctx
->Extensions
.EXT_transform_feedback
&& !unpackable_tess
;
2380 /* Disable packing on outward facing interfaces for SSO because in ES we
2381 * need to retain the unpacked varying information for draw time
2384 * Packing is still enabled on individual arrays, structs, and matrices as
2385 * these are required by the transform feedback code and it is still safe
2386 * to do so. We also enable packing when a varying is only used for
2387 * transform feedback and its not a SSO.
2389 bool disable_varying_packing
=
2390 ctx
->Const
.DisableVaryingPacking
|| unpackable_tess
;
2391 if (prog
->SeparateShader
&& (producer
== NULL
|| consumer
== NULL
))
2392 disable_varying_packing
= true;
2394 varying_matches
matches(disable_varying_packing
, xfb_enabled
,
2395 ctx
->Extensions
.ARB_enhanced_layouts
,
2396 producer
? producer
->Stage
: MESA_SHADER_NONE
,
2397 consumer
? consumer
->Stage
: MESA_SHADER_NONE
);
2398 hash_table
*tfeedback_candidates
=
2399 _mesa_hash_table_create(NULL
, _mesa_key_hash_string
,
2400 _mesa_key_string_equal
);
2401 hash_table
*consumer_inputs
=
2402 _mesa_hash_table_create(NULL
, _mesa_key_hash_string
,
2403 _mesa_key_string_equal
);
2404 hash_table
*consumer_interface_inputs
=
2405 _mesa_hash_table_create(NULL
, _mesa_key_hash_string
,
2406 _mesa_key_string_equal
);
2407 ir_variable
*consumer_inputs_with_locations
[VARYING_SLOT_TESS_MAX
] = {
2411 unsigned consumer_vertices
= 0;
2412 if (consumer
&& consumer
->Stage
== MESA_SHADER_GEOMETRY
)
2413 consumer_vertices
= prog
->Geom
.VerticesIn
;
2415 /* Operate in a total of four passes.
2417 * 1. Sort inputs / outputs into a canonical order. This is necessary so
2418 * that inputs / outputs of separable shaders will be assigned
2419 * predictable locations regardless of the order in which declarations
2420 * appeared in the shader source.
2422 * 2. Assign locations for any matching inputs and outputs.
2424 * 3. Mark output variables in the producer that do not have locations as
2425 * not being outputs. This lets the optimizer eliminate them.
2427 * 4. Mark input variables in the consumer that do not have locations as
2428 * not being inputs. This lets the optimizer eliminate them.
2431 canonicalize_shader_io(consumer
->ir
, ir_var_shader_in
);
2434 canonicalize_shader_io(producer
->ir
, ir_var_shader_out
);
2437 linker::populate_consumer_input_sets(mem_ctx
, consumer
->ir
,
2439 consumer_interface_inputs
,
2440 consumer_inputs_with_locations
);
2443 foreach_in_list(ir_instruction
, node
, producer
->ir
) {
2444 ir_variable
*const output_var
= node
->as_variable();
2446 if (output_var
== NULL
|| output_var
->data
.mode
!= ir_var_shader_out
)
2449 /* Only geometry shaders can use non-zero streams */
2450 assert(output_var
->data
.stream
== 0 ||
2451 (output_var
->data
.stream
< MAX_VERTEX_STREAMS
&&
2452 producer
->Stage
== MESA_SHADER_GEOMETRY
));
2454 if (num_tfeedback_decls
> 0) {
2455 tfeedback_candidate_generator
g(mem_ctx
, tfeedback_candidates
);
2456 g
.process(output_var
);
2459 ir_variable
*const input_var
=
2460 linker::get_matching_input(mem_ctx
, output_var
, consumer_inputs
,
2461 consumer_interface_inputs
,
2462 consumer_inputs_with_locations
);
2464 /* If a matching input variable was found, add this output (and the
2465 * input) to the set. If this is a separable program and there is no
2466 * consumer stage, add the output.
2468 * Always add TCS outputs. They are shared by all invocations
2469 * within a patch and can be used as shared memory.
2471 if (input_var
|| (prog
->SeparateShader
&& consumer
== NULL
) ||
2472 producer
->Stage
== MESA_SHADER_TESS_CTRL
) {
2473 matches
.record(output_var
, input_var
);
2476 /* Only stream 0 outputs can be consumed in the next stage */
2477 if (input_var
&& output_var
->data
.stream
!= 0) {
2478 linker_error(prog
, "output %s is assigned to stream=%d but "
2479 "is linked to an input, which requires stream=0",
2480 output_var
->name
, output_var
->data
.stream
);
2485 /* If there's no producer stage, then this must be a separable program.
2486 * For example, we may have a program that has just a fragment shader.
2487 * Later this program will be used with some arbitrary vertex (or
2488 * geometry) shader program. This means that locations must be assigned
2489 * for all the inputs.
2491 foreach_in_list(ir_instruction
, node
, consumer
->ir
) {
2492 ir_variable
*const input_var
= node
->as_variable();
2494 if (input_var
== NULL
|| input_var
->data
.mode
!= ir_var_shader_in
)
2497 matches
.record(NULL
, input_var
);
2501 for (unsigned i
= 0; i
< num_tfeedback_decls
; ++i
) {
2502 if (!tfeedback_decls
[i
].is_varying())
2505 const tfeedback_candidate
*matched_candidate
2506 = tfeedback_decls
[i
].find_candidate(prog
, tfeedback_candidates
);
2508 if (matched_candidate
== NULL
) {
2509 _mesa_hash_table_destroy(tfeedback_candidates
, NULL
);
2513 /* Mark xfb varyings as always active */
2514 matched_candidate
->toplevel_var
->data
.always_active_io
= 1;
2516 /* Mark any corresponding inputs as always active also. We must do this
2517 * because we have a NIR pass that lowers vectors to scalars and another
2518 * that removes unused varyings.
2519 * We don't split varyings marked as always active because there is no
2520 * point in doing so. This means we need to mark both sides of the
2521 * interface as always active otherwise we will have a mismatch and
2522 * start removing things we shouldn't.
2524 ir_variable
*const input_var
=
2525 linker::get_matching_input(mem_ctx
, matched_candidate
->toplevel_var
,
2527 consumer_interface_inputs
,
2528 consumer_inputs_with_locations
);
2530 input_var
->data
.always_active_io
= 1;
2532 if (matched_candidate
->toplevel_var
->data
.is_unmatched_generic_inout
) {
2533 matched_candidate
->toplevel_var
->data
.is_xfb_only
= 1;
2534 matches
.record(matched_candidate
->toplevel_var
, NULL
);
2538 _mesa_hash_table_destroy(consumer_inputs
, NULL
);
2539 _mesa_hash_table_destroy(consumer_interface_inputs
, NULL
);
2541 uint8_t components
[MAX_VARYINGS_INCL_PATCH
] = {0};
2542 const unsigned slots_used
= matches
.assign_locations(
2543 prog
, components
, reserved_slots
);
2544 matches
.store_locations();
2546 for (unsigned i
= 0; i
< num_tfeedback_decls
; ++i
) {
2547 if (!tfeedback_decls
[i
].is_varying())
2550 if (!tfeedback_decls
[i
].assign_location(ctx
, prog
)) {
2551 _mesa_hash_table_destroy(tfeedback_candidates
, NULL
);
2555 _mesa_hash_table_destroy(tfeedback_candidates
, NULL
);
2557 if (consumer
&& producer
) {
2558 foreach_in_list(ir_instruction
, node
, consumer
->ir
) {
2559 ir_variable
*const var
= node
->as_variable();
2561 if (var
&& var
->data
.mode
== ir_var_shader_in
&&
2562 var
->data
.is_unmatched_generic_inout
) {
2563 if (!prog
->IsES
&& prog
->data
->Version
<= 120) {
2564 /* On page 25 (page 31 of the PDF) of the GLSL 1.20 spec:
2566 * Only those varying variables used (i.e. read) in
2567 * the fragment shader executable must be written to
2568 * by the vertex shader executable; declaring
2569 * superfluous varying variables in a vertex shader is
2572 * We interpret this text as meaning that the VS must
2573 * write the variable for the FS to read it. See
2574 * "glsl1-varying read but not written" in piglit.
2576 linker_error(prog
, "%s shader varying %s not written "
2578 _mesa_shader_stage_to_string(consumer
->Stage
),
2580 _mesa_shader_stage_to_string(producer
->Stage
));
2582 linker_warning(prog
, "%s shader varying %s not written "
2584 _mesa_shader_stage_to_string(consumer
->Stage
),
2586 _mesa_shader_stage_to_string(producer
->Stage
));
2591 /* Now that validation is done its safe to remove unused varyings. As
2592 * we have both a producer and consumer its safe to remove unused
2593 * varyings even if the program is a SSO because the stages are being
2594 * linked together i.e. we have a multi-stage SSO.
2596 remove_unused_shader_inputs_and_outputs(false, producer
,
2598 remove_unused_shader_inputs_and_outputs(false, consumer
,
2603 lower_packed_varyings(mem_ctx
, slots_used
, components
, ir_var_shader_out
,
2604 0, producer
, disable_varying_packing
,
2609 lower_packed_varyings(mem_ctx
, slots_used
, components
, ir_var_shader_in
,
2610 consumer_vertices
, consumer
,
2611 disable_varying_packing
, xfb_enabled
);
2618 check_against_output_limit(struct gl_context
*ctx
,
2619 struct gl_shader_program
*prog
,
2620 gl_linked_shader
*producer
,
2621 unsigned num_explicit_locations
)
2623 unsigned output_vectors
= num_explicit_locations
;
2625 foreach_in_list(ir_instruction
, node
, producer
->ir
) {
2626 ir_variable
*const var
= node
->as_variable();
2628 if (var
&& !var
->data
.explicit_location
&&
2629 var
->data
.mode
== ir_var_shader_out
&&
2630 var_counts_against_varying_limit(producer
->Stage
, var
)) {
2631 /* outputs for fragment shader can't be doubles */
2632 output_vectors
+= var
->type
->count_attribute_slots(false);
2636 assert(producer
->Stage
!= MESA_SHADER_FRAGMENT
);
2637 unsigned max_output_components
=
2638 ctx
->Const
.Program
[producer
->Stage
].MaxOutputComponents
;
2640 const unsigned output_components
= output_vectors
* 4;
2641 if (output_components
> max_output_components
) {
2642 if (ctx
->API
== API_OPENGLES2
|| prog
->IsES
)
2643 linker_error(prog
, "%s shader uses too many output vectors "
2645 _mesa_shader_stage_to_string(producer
->Stage
),
2647 max_output_components
/ 4);
2649 linker_error(prog
, "%s shader uses too many output components "
2651 _mesa_shader_stage_to_string(producer
->Stage
),
2653 max_output_components
);
2662 check_against_input_limit(struct gl_context
*ctx
,
2663 struct gl_shader_program
*prog
,
2664 gl_linked_shader
*consumer
,
2665 unsigned num_explicit_locations
)
2667 unsigned input_vectors
= num_explicit_locations
;
2669 foreach_in_list(ir_instruction
, node
, consumer
->ir
) {
2670 ir_variable
*const var
= node
->as_variable();
2672 if (var
&& !var
->data
.explicit_location
&&
2673 var
->data
.mode
== ir_var_shader_in
&&
2674 var_counts_against_varying_limit(consumer
->Stage
, var
)) {
2675 /* vertex inputs aren't varying counted */
2676 input_vectors
+= var
->type
->count_attribute_slots(false);
2680 assert(consumer
->Stage
!= MESA_SHADER_VERTEX
);
2681 unsigned max_input_components
=
2682 ctx
->Const
.Program
[consumer
->Stage
].MaxInputComponents
;
2684 const unsigned input_components
= input_vectors
* 4;
2685 if (input_components
> max_input_components
) {
2686 if (ctx
->API
== API_OPENGLES2
|| prog
->IsES
)
2687 linker_error(prog
, "%s shader uses too many input vectors "
2689 _mesa_shader_stage_to_string(consumer
->Stage
),
2691 max_input_components
/ 4);
2693 linker_error(prog
, "%s shader uses too many input components "
2695 _mesa_shader_stage_to_string(consumer
->Stage
),
2697 max_input_components
);
2706 link_varyings(struct gl_shader_program
*prog
, unsigned first
, unsigned last
,
2707 struct gl_context
*ctx
, void *mem_ctx
)
2709 bool has_xfb_qualifiers
= false;
2710 unsigned num_tfeedback_decls
= 0;
2711 char **varying_names
= NULL
;
2712 tfeedback_decl
*tfeedback_decls
= NULL
;
2714 /* From the ARB_enhanced_layouts spec:
2716 * "If the shader used to record output variables for transform feedback
2717 * varyings uses the "xfb_buffer", "xfb_offset", or "xfb_stride" layout
2718 * qualifiers, the values specified by TransformFeedbackVaryings are
2719 * ignored, and the set of variables captured for transform feedback is
2720 * instead derived from the specified layout qualifiers."
2722 for (int i
= MESA_SHADER_FRAGMENT
- 1; i
>= 0; i
--) {
2723 /* Find last stage before fragment shader */
2724 if (prog
->_LinkedShaders
[i
]) {
2725 has_xfb_qualifiers
=
2726 process_xfb_layout_qualifiers(mem_ctx
, prog
->_LinkedShaders
[i
],
2727 prog
, &num_tfeedback_decls
,
2733 if (!has_xfb_qualifiers
) {
2734 num_tfeedback_decls
= prog
->TransformFeedback
.NumVarying
;
2735 varying_names
= prog
->TransformFeedback
.VaryingNames
;
2738 if (num_tfeedback_decls
!= 0) {
2739 /* From GL_EXT_transform_feedback:
2740 * A program will fail to link if:
2742 * * the <count> specified by TransformFeedbackVaryingsEXT is
2743 * non-zero, but the program object has no vertex or geometry
2746 if (first
>= MESA_SHADER_FRAGMENT
) {
2747 linker_error(prog
, "Transform feedback varyings specified, but "
2748 "no vertex, tessellation, or geometry shader is "
2753 tfeedback_decls
= rzalloc_array(mem_ctx
, tfeedback_decl
,
2754 num_tfeedback_decls
);
2755 if (!parse_tfeedback_decls(ctx
, prog
, mem_ctx
, num_tfeedback_decls
,
2756 varying_names
, tfeedback_decls
))
2760 /* If there is no fragment shader we need to set transform feedback.
2762 * For SSO we also need to assign output locations. We assign them here
2763 * because we need to do it for both single stage programs and multi stage
2766 if (last
< MESA_SHADER_FRAGMENT
&&
2767 (num_tfeedback_decls
!= 0 || prog
->SeparateShader
)) {
2768 const uint64_t reserved_out_slots
=
2769 reserved_varying_slot(prog
->_LinkedShaders
[last
], ir_var_shader_out
);
2770 if (!assign_varying_locations(ctx
, mem_ctx
, prog
,
2771 prog
->_LinkedShaders
[last
], NULL
,
2772 num_tfeedback_decls
, tfeedback_decls
,
2773 reserved_out_slots
))
2777 if (last
<= MESA_SHADER_FRAGMENT
) {
2778 /* Remove unused varyings from the first/last stage unless SSO */
2779 remove_unused_shader_inputs_and_outputs(prog
->SeparateShader
,
2780 prog
->_LinkedShaders
[first
],
2782 remove_unused_shader_inputs_and_outputs(prog
->SeparateShader
,
2783 prog
->_LinkedShaders
[last
],
2786 /* If the program is made up of only a single stage */
2787 if (first
== last
) {
2788 gl_linked_shader
*const sh
= prog
->_LinkedShaders
[last
];
2790 do_dead_builtin_varyings(ctx
, NULL
, sh
, 0, NULL
);
2791 do_dead_builtin_varyings(ctx
, sh
, NULL
, num_tfeedback_decls
,
2794 if (prog
->SeparateShader
) {
2795 const uint64_t reserved_slots
=
2796 reserved_varying_slot(sh
, ir_var_shader_in
);
2798 /* Assign input locations for SSO, output locations are already
2801 if (!assign_varying_locations(ctx
, mem_ctx
, prog
,
2802 NULL
/* producer */,
2804 0 /* num_tfeedback_decls */,
2805 NULL
/* tfeedback_decls */,
2810 /* Linking the stages in the opposite order (from fragment to vertex)
2811 * ensures that inter-shader outputs written to in an earlier stage
2812 * are eliminated if they are (transitively) not used in a later
2816 for (int i
= next
- 1; i
>= 0; i
--) {
2817 if (prog
->_LinkedShaders
[i
] == NULL
&& i
!= 0)
2820 gl_linked_shader
*const sh_i
= prog
->_LinkedShaders
[i
];
2821 gl_linked_shader
*const sh_next
= prog
->_LinkedShaders
[next
];
2823 const uint64_t reserved_out_slots
=
2824 reserved_varying_slot(sh_i
, ir_var_shader_out
);
2825 const uint64_t reserved_in_slots
=
2826 reserved_varying_slot(sh_next
, ir_var_shader_in
);
2828 do_dead_builtin_varyings(ctx
, sh_i
, sh_next
,
2829 next
== MESA_SHADER_FRAGMENT
? num_tfeedback_decls
: 0,
2832 if (!assign_varying_locations(ctx
, mem_ctx
, prog
, sh_i
, sh_next
,
2833 next
== MESA_SHADER_FRAGMENT
? num_tfeedback_decls
: 0,
2835 reserved_out_slots
| reserved_in_slots
))
2838 /* This must be done after all dead varyings are eliminated. */
2840 unsigned slots_used
= _mesa_bitcount_64(reserved_out_slots
);
2841 if (!check_against_output_limit(ctx
, prog
, sh_i
, slots_used
)) {
2846 unsigned slots_used
= _mesa_bitcount_64(reserved_in_slots
);
2847 if (!check_against_input_limit(ctx
, prog
, sh_next
, slots_used
))
2855 if (!store_tfeedback_info(ctx
, prog
, num_tfeedback_decls
, tfeedback_decls
,
2856 has_xfb_qualifiers
))