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
{
409 unsigned interpolation
;
416 check_location_aliasing(struct explicit_location_info explicit_locations
[][4],
420 unsigned location_limit
,
421 const glsl_type
*type
,
422 unsigned interpolation
,
426 gl_shader_program
*prog
,
427 gl_shader_stage stage
)
430 if (type
->without_array()->is_record()) {
431 /* The component qualifier can't be used on structs so just treat
432 * all component slots as used.
436 unsigned dmul
= type
->without_array()->is_64bit() ? 2 : 1;
437 last_comp
= component
+ type
->without_array()->vector_elements
* dmul
;
440 while (location
< location_limit
) {
441 unsigned i
= component
;
443 /* If there are other outputs assigned to the same location
444 * they must have the same interpolation
448 /* Skip the components used by this output, we only care about
449 * other outputs in the same location
456 struct explicit_location_info
*info
=
457 &explicit_locations
[location
][comp
];
460 if (info
->interpolation
!= interpolation
) {
462 "%s shader has multiple outputs at explicit "
463 "location %u with different interpolation "
465 _mesa_shader_stage_to_string(stage
), location
);
469 if (info
->centroid
!= centroid
||
470 info
->sample
!= sample
||
471 info
->patch
!= patch
) {
473 "%s shader has multiple outputs at explicit "
474 "location %u with different aux storage\n",
475 _mesa_shader_stage_to_string(stage
), location
);
483 /* Component aliasing is not allowed */
484 while (i
< last_comp
) {
485 if (explicit_locations
[location
][i
].var
!= NULL
) {
487 "%s shader has multiple outputs explicitly "
488 "assigned to location %d and component %d\n",
489 _mesa_shader_stage_to_string(stage
),
490 location
, component
);
494 /* Make sure all component at this location have the same type.
496 for (unsigned j
= 0; j
< 4; j
++) {
497 if (explicit_locations
[location
][j
].var
&&
498 explicit_locations
[location
][j
].base_type
!=
499 type
->without_array()->base_type
) {
501 "Varyings sharing the same location must "
502 "have the same underlying numerical type. "
503 "Location %u component %u\n", location
, component
);
508 explicit_locations
[location
][i
].var
= var
;
509 explicit_locations
[location
][i
].base_type
=
510 type
->without_array()->base_type
;
511 explicit_locations
[location
][i
].interpolation
= interpolation
;
512 explicit_locations
[location
][i
].centroid
= centroid
;
513 explicit_locations
[location
][i
].sample
= sample
;
514 explicit_locations
[location
][i
].patch
= patch
;
517 /* We need to do some special handling for doubles as dvec3 and
518 * dvec4 consume two consecutive locations. We don't need to
519 * worry about components beginning at anything other than 0 as
520 * the spec does not allow this for dvec3 and dvec4.
522 if (i
== 4 && last_comp
> 4) {
523 last_comp
= last_comp
- 4;
524 /* Bump location index and reset the component index */
537 * Validate that outputs from one stage match inputs of another
540 cross_validate_outputs_to_inputs(struct gl_context
*ctx
,
541 struct gl_shader_program
*prog
,
542 gl_linked_shader
*producer
,
543 gl_linked_shader
*consumer
)
545 glsl_symbol_table parameters
;
546 struct explicit_location_info explicit_locations
[MAX_VARYING
][4] = { 0 };
548 /* Find all shader outputs in the "producer" stage.
550 foreach_in_list(ir_instruction
, node
, producer
->ir
) {
551 ir_variable
*const var
= node
->as_variable();
553 if (var
== NULL
|| var
->data
.mode
!= ir_var_shader_out
)
556 if (!var
->data
.explicit_location
557 || var
->data
.location
< VARYING_SLOT_VAR0
)
558 parameters
.add_variable(var
);
560 /* User-defined varyings with explicit locations are handled
561 * differently because they do not need to have matching names.
563 const glsl_type
*type
= get_varying_type(var
, producer
->Stage
);
564 unsigned num_elements
= type
->count_attribute_slots(false);
565 unsigned idx
= compute_variable_location_slot(var
, producer
->Stage
);
566 unsigned slot_limit
= idx
+ num_elements
;
569 ctx
->Const
.Program
[producer
->Stage
].MaxOutputComponents
/ 4;
570 if (slot_limit
> slot_max
) {
572 "Invalid location %u in %s shader\n",
573 idx
, _mesa_shader_stage_to_string(producer
->Stage
));
577 if (type
->without_array()->is_interface()) {
578 for (unsigned i
= 0; i
< type
->without_array()->length
; i
++) {
579 glsl_struct_field
*field
= &type
->fields
.structure
[i
];
580 unsigned field_location
= field
->location
-
581 (field
->patch
? VARYING_SLOT_PATCH0
: VARYING_SLOT_VAR0
);
582 if (!check_location_aliasing(explicit_locations
, var
,
584 0, field_location
+ 1,
586 field
->interpolation
,
590 prog
, producer
->Stage
)) {
594 } else if (!check_location_aliasing(explicit_locations
, var
,
595 idx
, var
->data
.location_frac
,
597 var
->data
.interpolation
,
601 prog
, producer
->Stage
)) {
608 /* Find all shader inputs in the "consumer" stage. Any variables that have
609 * matching outputs already in the symbol table must have the same type and
612 * Exception: if the consumer is the geometry shader, then the inputs
613 * should be arrays and the type of the array element should match the type
614 * of the corresponding producer output.
616 foreach_in_list(ir_instruction
, node
, consumer
->ir
) {
617 ir_variable
*const input
= node
->as_variable();
619 if (input
== NULL
|| input
->data
.mode
!= ir_var_shader_in
)
622 if (strcmp(input
->name
, "gl_Color") == 0 && input
->data
.used
) {
623 const ir_variable
*const front_color
=
624 parameters
.get_variable("gl_FrontColor");
626 const ir_variable
*const back_color
=
627 parameters
.get_variable("gl_BackColor");
629 cross_validate_front_and_back_color(prog
, input
,
630 front_color
, back_color
,
631 consumer
->Stage
, producer
->Stage
);
632 } else if (strcmp(input
->name
, "gl_SecondaryColor") == 0 && input
->data
.used
) {
633 const ir_variable
*const front_color
=
634 parameters
.get_variable("gl_FrontSecondaryColor");
636 const ir_variable
*const back_color
=
637 parameters
.get_variable("gl_BackSecondaryColor");
639 cross_validate_front_and_back_color(prog
, input
,
640 front_color
, back_color
,
641 consumer
->Stage
, producer
->Stage
);
643 /* The rules for connecting inputs and outputs change in the presence
644 * of explicit locations. In this case, we no longer care about the
645 * names of the variables. Instead, we care only about the
646 * explicitly assigned location.
648 ir_variable
*output
= NULL
;
649 if (input
->data
.explicit_location
650 && input
->data
.location
>= VARYING_SLOT_VAR0
) {
652 const glsl_type
*type
= get_varying_type(input
, consumer
->Stage
);
653 unsigned num_elements
= type
->count_attribute_slots(false);
655 compute_variable_location_slot(input
, consumer
->Stage
);
656 unsigned slot_limit
= idx
+ num_elements
;
658 while (idx
< slot_limit
) {
659 if (idx
>= MAX_VARYING
) {
661 "Invalid location %u in %s shader\n", idx
,
662 _mesa_shader_stage_to_string(consumer
->Stage
));
666 output
= explicit_locations
[idx
][input
->data
.location_frac
].var
;
668 if (output
== NULL
||
669 input
->data
.location
!= output
->data
.location
) {
671 "%s shader input `%s' with explicit location "
672 "has no matching output\n",
673 _mesa_shader_stage_to_string(consumer
->Stage
),
680 output
= parameters
.get_variable(input
->name
);
683 if (output
!= NULL
) {
684 /* Interface blocks have their own validation elsewhere so don't
685 * try validating them here.
687 if (!(input
->get_interface_type() &&
688 output
->get_interface_type()))
689 cross_validate_types_and_qualifiers(prog
, input
, output
,
693 /* Check for input vars with unmatched output vars in prev stage
694 * taking into account that interface blocks could have a matching
695 * output but with different name, so we ignore them.
697 assert(!input
->data
.assigned
);
698 if (input
->data
.used
&& !input
->get_interface_type() &&
699 !input
->data
.explicit_location
&& !prog
->SeparateShader
)
701 "%s shader input `%s' "
702 "has no matching output in the previous stage\n",
703 _mesa_shader_stage_to_string(consumer
->Stage
),
711 * Demote shader inputs and outputs that are not used in other stages, and
712 * remove them via dead code elimination.
715 remove_unused_shader_inputs_and_outputs(bool is_separate_shader_object
,
716 gl_linked_shader
*sh
,
717 enum ir_variable_mode mode
)
719 if (is_separate_shader_object
)
722 foreach_in_list(ir_instruction
, node
, sh
->ir
) {
723 ir_variable
*const var
= node
->as_variable();
725 if (var
== NULL
|| var
->data
.mode
!= int(mode
))
728 /* A shader 'in' or 'out' variable is only really an input or output if
729 * its value is used by other shader stages. This will cause the
730 * variable to have a location assigned.
732 if (var
->data
.is_unmatched_generic_inout
&& !var
->data
.is_xfb_only
) {
733 assert(var
->data
.mode
!= ir_var_temporary
);
735 /* Assign zeros to demoted inputs to allow more optimizations. */
736 if (var
->data
.mode
== ir_var_shader_in
&& !var
->constant_value
)
737 var
->constant_value
= ir_constant::zero(var
, var
->type
);
739 var
->data
.mode
= ir_var_auto
;
743 /* Eliminate code that is now dead due to unused inputs/outputs being
746 while (do_dead_code(sh
->ir
, false))
752 * Initialize this object based on a string that was passed to
753 * glTransformFeedbackVaryings.
755 * If the input is mal-formed, this call still succeeds, but it sets
756 * this->var_name to a mal-formed input, so tfeedback_decl::find_output_var()
757 * will fail to find any matching variable.
760 tfeedback_decl::init(struct gl_context
*ctx
, const void *mem_ctx
,
763 /* We don't have to be pedantic about what is a valid GLSL variable name,
764 * because any variable with an invalid name can't exist in the IR anyway.
768 this->orig_name
= input
;
769 this->lowered_builtin_array_variable
= none
;
770 this->skip_components
= 0;
771 this->next_buffer_separator
= false;
772 this->matched_candidate
= NULL
;
777 if (ctx
->Extensions
.ARB_transform_feedback3
) {
778 /* Parse gl_NextBuffer. */
779 if (strcmp(input
, "gl_NextBuffer") == 0) {
780 this->next_buffer_separator
= true;
784 /* Parse gl_SkipComponents. */
785 if (strcmp(input
, "gl_SkipComponents1") == 0)
786 this->skip_components
= 1;
787 else if (strcmp(input
, "gl_SkipComponents2") == 0)
788 this->skip_components
= 2;
789 else if (strcmp(input
, "gl_SkipComponents3") == 0)
790 this->skip_components
= 3;
791 else if (strcmp(input
, "gl_SkipComponents4") == 0)
792 this->skip_components
= 4;
794 if (this->skip_components
)
798 /* Parse a declaration. */
799 const char *base_name_end
;
800 long subscript
= parse_program_resource_name(input
, &base_name_end
);
801 this->var_name
= ralloc_strndup(mem_ctx
, input
, base_name_end
- input
);
802 if (this->var_name
== NULL
) {
803 _mesa_error_no_memory(__func__
);
807 if (subscript
>= 0) {
808 this->array_subscript
= subscript
;
809 this->is_subscripted
= true;
811 this->is_subscripted
= false;
814 /* For drivers that lower gl_ClipDistance to gl_ClipDistanceMESA, this
815 * class must behave specially to account for the fact that gl_ClipDistance
816 * is converted from a float[8] to a vec4[2].
818 if (ctx
->Const
.ShaderCompilerOptions
[MESA_SHADER_VERTEX
].LowerCombinedClipCullDistance
&&
819 strcmp(this->var_name
, "gl_ClipDistance") == 0) {
820 this->lowered_builtin_array_variable
= clip_distance
;
822 if (ctx
->Const
.ShaderCompilerOptions
[MESA_SHADER_VERTEX
].LowerCombinedClipCullDistance
&&
823 strcmp(this->var_name
, "gl_CullDistance") == 0) {
824 this->lowered_builtin_array_variable
= cull_distance
;
827 if (ctx
->Const
.LowerTessLevel
&&
828 (strcmp(this->var_name
, "gl_TessLevelOuter") == 0))
829 this->lowered_builtin_array_variable
= tess_level_outer
;
830 if (ctx
->Const
.LowerTessLevel
&&
831 (strcmp(this->var_name
, "gl_TessLevelInner") == 0))
832 this->lowered_builtin_array_variable
= tess_level_inner
;
837 * Determine whether two tfeedback_decl objects refer to the same variable and
838 * array index (if applicable).
841 tfeedback_decl::is_same(const tfeedback_decl
&x
, const tfeedback_decl
&y
)
843 assert(x
.is_varying() && y
.is_varying());
845 if (strcmp(x
.var_name
, y
.var_name
) != 0)
847 if (x
.is_subscripted
!= y
.is_subscripted
)
849 if (x
.is_subscripted
&& x
.array_subscript
!= y
.array_subscript
)
856 * Assign a location and stream ID for this tfeedback_decl object based on the
857 * transform feedback candidate found by find_candidate.
859 * If an error occurs, the error is reported through linker_error() and false
863 tfeedback_decl::assign_location(struct gl_context
*ctx
,
864 struct gl_shader_program
*prog
)
866 assert(this->is_varying());
868 unsigned fine_location
869 = this->matched_candidate
->toplevel_var
->data
.location
* 4
870 + this->matched_candidate
->toplevel_var
->data
.location_frac
871 + this->matched_candidate
->offset
;
872 const unsigned dmul
=
873 this->matched_candidate
->type
->without_array()->is_64bit() ? 2 : 1;
875 if (this->matched_candidate
->type
->is_array()) {
877 const unsigned matrix_cols
=
878 this->matched_candidate
->type
->fields
.array
->matrix_columns
;
879 const unsigned vector_elements
=
880 this->matched_candidate
->type
->fields
.array
->vector_elements
;
881 unsigned actual_array_size
;
882 switch (this->lowered_builtin_array_variable
) {
884 actual_array_size
= prog
->last_vert_prog
?
885 prog
->last_vert_prog
->info
.clip_distance_array_size
: 0;
888 actual_array_size
= prog
->last_vert_prog
?
889 prog
->last_vert_prog
->info
.cull_distance_array_size
: 0;
891 case tess_level_outer
:
892 actual_array_size
= 4;
894 case tess_level_inner
:
895 actual_array_size
= 2;
899 actual_array_size
= this->matched_candidate
->type
->array_size();
903 if (this->is_subscripted
) {
904 /* Check array bounds. */
905 if (this->array_subscript
>= actual_array_size
) {
906 linker_error(prog
, "Transform feedback varying %s has index "
907 "%i, but the array size is %u.",
908 this->orig_name
, this->array_subscript
,
912 unsigned array_elem_size
= this->lowered_builtin_array_variable
?
913 1 : vector_elements
* matrix_cols
* dmul
;
914 fine_location
+= array_elem_size
* this->array_subscript
;
917 this->size
= actual_array_size
;
919 this->vector_elements
= vector_elements
;
920 this->matrix_columns
= matrix_cols
;
921 if (this->lowered_builtin_array_variable
)
922 this->type
= GL_FLOAT
;
924 this->type
= this->matched_candidate
->type
->fields
.array
->gl_type
;
926 /* Regular variable (scalar, vector, or matrix) */
927 if (this->is_subscripted
) {
928 linker_error(prog
, "Transform feedback varying %s requested, "
929 "but %s is not an array.",
930 this->orig_name
, this->var_name
);
934 this->vector_elements
= this->matched_candidate
->type
->vector_elements
;
935 this->matrix_columns
= this->matched_candidate
->type
->matrix_columns
;
936 this->type
= this->matched_candidate
->type
->gl_type
;
938 this->location
= fine_location
/ 4;
939 this->location_frac
= fine_location
% 4;
941 /* From GL_EXT_transform_feedback:
942 * A program will fail to link if:
944 * * the total number of components to capture in any varying
945 * variable in <varyings> is greater than the constant
946 * MAX_TRANSFORM_FEEDBACK_SEPARATE_COMPONENTS_EXT and the
947 * buffer mode is SEPARATE_ATTRIBS_EXT;
949 if (prog
->TransformFeedback
.BufferMode
== GL_SEPARATE_ATTRIBS
&&
950 this->num_components() >
951 ctx
->Const
.MaxTransformFeedbackSeparateComponents
) {
952 linker_error(prog
, "Transform feedback varying %s exceeds "
953 "MAX_TRANSFORM_FEEDBACK_SEPARATE_COMPONENTS.",
958 /* Only transform feedback varyings can be assigned to non-zero streams,
959 * so assign the stream id here.
961 this->stream_id
= this->matched_candidate
->toplevel_var
->data
.stream
;
963 unsigned array_offset
= this->array_subscript
* 4 * dmul
;
964 unsigned struct_offset
= this->matched_candidate
->offset
* 4 * dmul
;
965 this->buffer
= this->matched_candidate
->toplevel_var
->data
.xfb_buffer
;
966 this->offset
= this->matched_candidate
->toplevel_var
->data
.offset
+
967 array_offset
+ struct_offset
;
974 tfeedback_decl::get_num_outputs() const
976 if (!this->is_varying()) {
979 return (this->num_components() + this->location_frac
+ 3)/4;
984 * Update gl_transform_feedback_info to reflect this tfeedback_decl.
986 * If an error occurs, the error is reported through linker_error() and false
990 tfeedback_decl::store(struct gl_context
*ctx
, struct gl_shader_program
*prog
,
991 struct gl_transform_feedback_info
*info
,
992 unsigned buffer
, unsigned buffer_index
,
993 const unsigned max_outputs
, bool *explicit_stride
,
994 bool has_xfb_qualifiers
) const
996 unsigned xfb_offset
= 0;
997 unsigned size
= this->size
;
998 /* Handle gl_SkipComponents. */
999 if (this->skip_components
) {
1000 info
->Buffers
[buffer
].Stride
+= this->skip_components
;
1001 size
= this->skip_components
;
1005 if (this->next_buffer_separator
) {
1010 if (has_xfb_qualifiers
) {
1011 xfb_offset
= this->offset
/ 4;
1013 xfb_offset
= info
->Buffers
[buffer
].Stride
;
1015 info
->Varyings
[info
->NumVarying
].Offset
= xfb_offset
* 4;
1018 unsigned location
= this->location
;
1019 unsigned location_frac
= this->location_frac
;
1020 unsigned num_components
= this->num_components();
1021 while (num_components
> 0) {
1022 unsigned output_size
= MIN2(num_components
, 4 - location_frac
);
1023 assert((info
->NumOutputs
== 0 && max_outputs
== 0) ||
1024 info
->NumOutputs
< max_outputs
);
1026 /* From the ARB_enhanced_layouts spec:
1028 * "If such a block member or variable is not written during a shader
1029 * invocation, the buffer contents at the assigned offset will be
1030 * undefined. Even if there are no static writes to a variable or
1031 * member that is assigned a transform feedback offset, the space is
1032 * still allocated in the buffer and still affects the stride."
1034 if (this->is_varying_written()) {
1035 info
->Outputs
[info
->NumOutputs
].ComponentOffset
= location_frac
;
1036 info
->Outputs
[info
->NumOutputs
].OutputRegister
= location
;
1037 info
->Outputs
[info
->NumOutputs
].NumComponents
= output_size
;
1038 info
->Outputs
[info
->NumOutputs
].StreamId
= stream_id
;
1039 info
->Outputs
[info
->NumOutputs
].OutputBuffer
= buffer
;
1040 info
->Outputs
[info
->NumOutputs
].DstOffset
= xfb_offset
;
1043 info
->Buffers
[buffer
].Stream
= this->stream_id
;
1044 xfb_offset
+= output_size
;
1046 num_components
-= output_size
;
1052 if (explicit_stride
&& explicit_stride
[buffer
]) {
1053 if (this->is_64bit() && info
->Buffers
[buffer
].Stride
% 2) {
1054 linker_error(prog
, "invalid qualifier xfb_stride=%d must be a "
1055 "multiple of 8 as its applied to a type that is or "
1056 "contains a double.",
1057 info
->Buffers
[buffer
].Stride
* 4);
1061 if ((this->offset
/ 4) / info
->Buffers
[buffer
].Stride
!=
1062 (xfb_offset
- 1) / info
->Buffers
[buffer
].Stride
) {
1063 linker_error(prog
, "xfb_offset (%d) overflows xfb_stride (%d) for "
1064 "buffer (%d)", xfb_offset
* 4,
1065 info
->Buffers
[buffer
].Stride
* 4, buffer
);
1069 info
->Buffers
[buffer
].Stride
= xfb_offset
;
1072 /* From GL_EXT_transform_feedback:
1073 * A program will fail to link if:
1075 * * the total number of components to capture is greater than
1076 * the constant MAX_TRANSFORM_FEEDBACK_INTERLEAVED_COMPONENTS_EXT
1077 * and the buffer mode is INTERLEAVED_ATTRIBS_EXT.
1079 * From GL_ARB_enhanced_layouts:
1081 * "The resulting stride (implicit or explicit) must be less than or
1082 * equal to the implementation-dependent constant
1083 * gl_MaxTransformFeedbackInterleavedComponents."
1085 if ((prog
->TransformFeedback
.BufferMode
== GL_INTERLEAVED_ATTRIBS
||
1086 has_xfb_qualifiers
) &&
1087 info
->Buffers
[buffer
].Stride
>
1088 ctx
->Const
.MaxTransformFeedbackInterleavedComponents
) {
1089 linker_error(prog
, "The MAX_TRANSFORM_FEEDBACK_INTERLEAVED_COMPONENTS "
1090 "limit has been exceeded.");
1095 info
->Varyings
[info
->NumVarying
].Name
= ralloc_strdup(prog
,
1097 info
->Varyings
[info
->NumVarying
].Type
= this->type
;
1098 info
->Varyings
[info
->NumVarying
].Size
= size
;
1099 info
->Varyings
[info
->NumVarying
].BufferIndex
= buffer_index
;
1101 info
->Buffers
[buffer
].NumVaryings
++;
1107 const tfeedback_candidate
*
1108 tfeedback_decl::find_candidate(gl_shader_program
*prog
,
1109 hash_table
*tfeedback_candidates
)
1111 const char *name
= this->var_name
;
1112 switch (this->lowered_builtin_array_variable
) {
1114 name
= this->var_name
;
1117 name
= "gl_ClipDistanceMESA";
1120 name
= "gl_CullDistanceMESA";
1122 case tess_level_outer
:
1123 name
= "gl_TessLevelOuterMESA";
1125 case tess_level_inner
:
1126 name
= "gl_TessLevelInnerMESA";
1129 hash_entry
*entry
= _mesa_hash_table_search(tfeedback_candidates
, name
);
1131 this->matched_candidate
= entry
?
1132 (const tfeedback_candidate
*) entry
->data
: NULL
;
1134 if (!this->matched_candidate
) {
1135 /* From GL_EXT_transform_feedback:
1136 * A program will fail to link if:
1138 * * any variable name specified in the <varyings> array is not
1139 * declared as an output in the geometry shader (if present) or
1140 * the vertex shader (if no geometry shader is present);
1142 linker_error(prog
, "Transform feedback varying %s undeclared.",
1146 return this->matched_candidate
;
1151 * Parse all the transform feedback declarations that were passed to
1152 * glTransformFeedbackVaryings() and store them in tfeedback_decl objects.
1154 * If an error occurs, the error is reported through linker_error() and false
1158 parse_tfeedback_decls(struct gl_context
*ctx
, struct gl_shader_program
*prog
,
1159 const void *mem_ctx
, unsigned num_names
,
1160 char **varying_names
, tfeedback_decl
*decls
)
1162 for (unsigned i
= 0; i
< num_names
; ++i
) {
1163 decls
[i
].init(ctx
, mem_ctx
, varying_names
[i
]);
1165 if (!decls
[i
].is_varying())
1168 /* From GL_EXT_transform_feedback:
1169 * A program will fail to link if:
1171 * * any two entries in the <varyings> array specify the same varying
1174 * We interpret this to mean "any two entries in the <varyings> array
1175 * specify the same varying variable and array index", since transform
1176 * feedback of arrays would be useless otherwise.
1178 for (unsigned j
= 0; j
< i
; ++j
) {
1179 if (!decls
[j
].is_varying())
1182 if (tfeedback_decl::is_same(decls
[i
], decls
[j
])) {
1183 linker_error(prog
, "Transform feedback varying %s specified "
1184 "more than once.", varying_names
[i
]);
1194 cmp_xfb_offset(const void * x_generic
, const void * y_generic
)
1196 tfeedback_decl
*x
= (tfeedback_decl
*) x_generic
;
1197 tfeedback_decl
*y
= (tfeedback_decl
*) y_generic
;
1199 if (x
->get_buffer() != y
->get_buffer())
1200 return x
->get_buffer() - y
->get_buffer();
1201 return x
->get_offset() - y
->get_offset();
1205 * Store transform feedback location assignments into
1206 * prog->sh.LinkedTransformFeedback based on the data stored in
1209 * If an error occurs, the error is reported through linker_error() and false
1213 store_tfeedback_info(struct gl_context
*ctx
, struct gl_shader_program
*prog
,
1214 unsigned num_tfeedback_decls
,
1215 tfeedback_decl
*tfeedback_decls
, bool has_xfb_qualifiers
)
1217 if (!prog
->last_vert_prog
)
1220 /* Make sure MaxTransformFeedbackBuffers is less than 32 so the bitmask for
1221 * tracking the number of buffers doesn't overflow.
1223 assert(ctx
->Const
.MaxTransformFeedbackBuffers
< 32);
1225 bool separate_attribs_mode
=
1226 prog
->TransformFeedback
.BufferMode
== GL_SEPARATE_ATTRIBS
;
1228 struct gl_program
*xfb_prog
= prog
->last_vert_prog
;
1229 xfb_prog
->sh
.LinkedTransformFeedback
=
1230 rzalloc(xfb_prog
, struct gl_transform_feedback_info
);
1232 /* The xfb_offset qualifier does not have to be used in increasing order
1233 * however some drivers expect to receive the list of transform feedback
1234 * declarations in order so sort it now for convenience.
1236 if (has_xfb_qualifiers
)
1237 qsort(tfeedback_decls
, num_tfeedback_decls
, sizeof(*tfeedback_decls
),
1240 xfb_prog
->sh
.LinkedTransformFeedback
->Varyings
=
1241 rzalloc_array(xfb_prog
, struct gl_transform_feedback_varying_info
,
1242 num_tfeedback_decls
);
1244 unsigned num_outputs
= 0;
1245 for (unsigned i
= 0; i
< num_tfeedback_decls
; ++i
) {
1246 if (tfeedback_decls
[i
].is_varying_written())
1247 num_outputs
+= tfeedback_decls
[i
].get_num_outputs();
1250 xfb_prog
->sh
.LinkedTransformFeedback
->Outputs
=
1251 rzalloc_array(xfb_prog
, struct gl_transform_feedback_output
,
1254 unsigned num_buffers
= 0;
1255 unsigned buffers
= 0;
1257 if (!has_xfb_qualifiers
&& separate_attribs_mode
) {
1258 /* GL_SEPARATE_ATTRIBS */
1259 for (unsigned i
= 0; i
< num_tfeedback_decls
; ++i
) {
1260 if (!tfeedback_decls
[i
].store(ctx
, prog
,
1261 xfb_prog
->sh
.LinkedTransformFeedback
,
1262 num_buffers
, num_buffers
, num_outputs
,
1263 NULL
, has_xfb_qualifiers
))
1266 buffers
|= 1 << num_buffers
;
1271 /* GL_INVERLEAVED_ATTRIBS */
1272 int buffer_stream_id
= -1;
1274 num_tfeedback_decls
? tfeedback_decls
[0].get_buffer() : 0;
1275 bool explicit_stride
[MAX_FEEDBACK_BUFFERS
] = { false };
1277 /* Apply any xfb_stride global qualifiers */
1278 if (has_xfb_qualifiers
) {
1279 for (unsigned j
= 0; j
< MAX_FEEDBACK_BUFFERS
; j
++) {
1280 if (prog
->TransformFeedback
.BufferStride
[j
]) {
1282 explicit_stride
[j
] = true;
1283 xfb_prog
->sh
.LinkedTransformFeedback
->Buffers
[j
].Stride
=
1284 prog
->TransformFeedback
.BufferStride
[j
] / 4;
1289 for (unsigned i
= 0; i
< num_tfeedback_decls
; ++i
) {
1290 if (has_xfb_qualifiers
&&
1291 buffer
!= tfeedback_decls
[i
].get_buffer()) {
1292 /* we have moved to the next buffer so reset stream id */
1293 buffer_stream_id
= -1;
1297 if (tfeedback_decls
[i
].is_next_buffer_separator()) {
1298 if (!tfeedback_decls
[i
].store(ctx
, prog
,
1299 xfb_prog
->sh
.LinkedTransformFeedback
,
1300 buffer
, num_buffers
, num_outputs
,
1301 explicit_stride
, has_xfb_qualifiers
))
1304 buffer_stream_id
= -1;
1306 } else if (tfeedback_decls
[i
].is_varying()) {
1307 if (buffer_stream_id
== -1) {
1308 /* First varying writing to this buffer: remember its stream */
1309 buffer_stream_id
= (int) tfeedback_decls
[i
].get_stream_id();
1310 } else if (buffer_stream_id
!=
1311 (int) tfeedback_decls
[i
].get_stream_id()) {
1312 /* Varying writes to the same buffer from a different stream */
1314 "Transform feedback can't capture varyings belonging "
1315 "to different vertex streams in a single buffer. "
1316 "Varying %s writes to buffer from stream %u, other "
1317 "varyings in the same buffer write from stream %u.",
1318 tfeedback_decls
[i
].name(),
1319 tfeedback_decls
[i
].get_stream_id(),
1325 if (has_xfb_qualifiers
) {
1326 buffer
= tfeedback_decls
[i
].get_buffer();
1328 buffer
= num_buffers
;
1330 buffers
|= 1 << buffer
;
1332 if (!tfeedback_decls
[i
].store(ctx
, prog
,
1333 xfb_prog
->sh
.LinkedTransformFeedback
,
1334 buffer
, num_buffers
, num_outputs
,
1335 explicit_stride
, has_xfb_qualifiers
))
1340 assert(xfb_prog
->sh
.LinkedTransformFeedback
->NumOutputs
== num_outputs
);
1342 xfb_prog
->sh
.LinkedTransformFeedback
->ActiveBuffers
= buffers
;
1349 * Data structure recording the relationship between outputs of one shader
1350 * stage (the "producer") and inputs of another (the "consumer").
1352 class varying_matches
1355 varying_matches(bool disable_varying_packing
, bool xfb_enabled
,
1356 bool enhanced_layouts_enabled
,
1357 gl_shader_stage producer_stage
,
1358 gl_shader_stage consumer_stage
);
1360 void record(ir_variable
*producer_var
, ir_variable
*consumer_var
);
1361 unsigned assign_locations(struct gl_shader_program
*prog
,
1362 uint8_t *components
,
1363 uint64_t reserved_slots
);
1364 void store_locations() const;
1367 bool is_varying_packing_safe(const glsl_type
*type
,
1368 const ir_variable
*var
);
1371 * If true, this driver disables varying packing, so all varyings need to
1372 * be aligned on slot boundaries, and take up a number of slots equal to
1373 * their number of matrix columns times their array size.
1375 * Packing may also be disabled because our current packing method is not
1376 * safe in SSO or versions of OpenGL where interpolation qualifiers are not
1377 * guaranteed to match across stages.
1379 const bool disable_varying_packing
;
1382 * If true, this driver has transform feedback enabled. The transform
1383 * feedback code requires at least some packing be done even when varying
1384 * packing is disabled, fortunately where transform feedback requires
1385 * packing it's safe to override the disabled setting. See
1386 * is_varying_packing_safe().
1388 const bool xfb_enabled
;
1390 const bool enhanced_layouts_enabled
;
1393 * Enum representing the order in which varyings are packed within a
1396 * Currently we pack vec4's first, then vec2's, then scalar values, then
1397 * vec3's. This order ensures that the only vectors that are at risk of
1398 * having to be "double parked" (split between two adjacent varying slots)
1401 enum packing_order_enum
{
1404 PACKING_ORDER_SCALAR
,
1408 static unsigned compute_packing_class(const ir_variable
*var
);
1409 static packing_order_enum
compute_packing_order(const ir_variable
*var
);
1410 static int match_comparator(const void *x_generic
, const void *y_generic
);
1411 static int xfb_comparator(const void *x_generic
, const void *y_generic
);
1414 * Structure recording the relationship between a single producer output
1415 * and a single consumer input.
1419 * Packing class for this varying, computed by compute_packing_class().
1421 unsigned packing_class
;
1424 * Packing order for this varying, computed by compute_packing_order().
1426 packing_order_enum packing_order
;
1427 unsigned num_components
;
1430 * The output variable in the producer stage.
1432 ir_variable
*producer_var
;
1435 * The input variable in the consumer stage.
1437 ir_variable
*consumer_var
;
1440 * The location which has been assigned for this varying. This is
1441 * expressed in multiples of a float, with the first generic varying
1442 * (i.e. the one referred to by VARYING_SLOT_VAR0) represented by the
1445 unsigned generic_location
;
1449 * The number of elements in the \c matches array that are currently in
1452 unsigned num_matches
;
1455 * The number of elements that were set aside for the \c matches array when
1458 unsigned matches_capacity
;
1460 gl_shader_stage producer_stage
;
1461 gl_shader_stage consumer_stage
;
1464 } /* anonymous namespace */
1466 varying_matches::varying_matches(bool disable_varying_packing
,
1468 bool enhanced_layouts_enabled
,
1469 gl_shader_stage producer_stage
,
1470 gl_shader_stage consumer_stage
)
1471 : disable_varying_packing(disable_varying_packing
),
1472 xfb_enabled(xfb_enabled
),
1473 enhanced_layouts_enabled(enhanced_layouts_enabled
),
1474 producer_stage(producer_stage
),
1475 consumer_stage(consumer_stage
)
1477 /* Note: this initial capacity is rather arbitrarily chosen to be large
1478 * enough for many cases without wasting an unreasonable amount of space.
1479 * varying_matches::record() will resize the array if there are more than
1480 * this number of varyings.
1482 this->matches_capacity
= 8;
1483 this->matches
= (match
*)
1484 malloc(sizeof(*this->matches
) * this->matches_capacity
);
1485 this->num_matches
= 0;
1489 varying_matches::~varying_matches()
1491 free(this->matches
);
1496 * Packing is always safe on individual arrays, structures, and matrices. It
1497 * is also safe if the varying is only used for transform feedback.
1500 varying_matches::is_varying_packing_safe(const glsl_type
*type
,
1501 const ir_variable
*var
)
1503 if (consumer_stage
== MESA_SHADER_TESS_EVAL
||
1504 consumer_stage
== MESA_SHADER_TESS_CTRL
||
1505 producer_stage
== MESA_SHADER_TESS_CTRL
)
1508 return xfb_enabled
&& (type
->is_array() || type
->is_record() ||
1509 type
->is_matrix() || var
->data
.is_xfb_only
);
1514 * Record the given producer/consumer variable pair in the list of variables
1515 * that should later be assigned locations.
1517 * It is permissible for \c consumer_var to be NULL (this happens if a
1518 * variable is output by the producer and consumed by transform feedback, but
1519 * not consumed by the consumer).
1521 * If \c producer_var has already been paired up with a consumer_var, or
1522 * producer_var is part of fixed pipeline functionality (and hence already has
1523 * a location assigned), this function has no effect.
1525 * Note: as a side effect this function may change the interpolation type of
1526 * \c producer_var, but only when the change couldn't possibly affect
1530 varying_matches::record(ir_variable
*producer_var
, ir_variable
*consumer_var
)
1532 assert(producer_var
!= NULL
|| consumer_var
!= NULL
);
1534 if ((producer_var
&& (!producer_var
->data
.is_unmatched_generic_inout
||
1535 producer_var
->data
.explicit_location
)) ||
1536 (consumer_var
&& (!consumer_var
->data
.is_unmatched_generic_inout
||
1537 consumer_var
->data
.explicit_location
))) {
1538 /* Either a location already exists for this variable (since it is part
1539 * of fixed functionality), or it has already been recorded as part of a
1545 bool needs_flat_qualifier
= consumer_var
== NULL
&&
1546 (producer_var
->type
->contains_integer() ||
1547 producer_var
->type
->contains_double());
1549 if (!disable_varying_packing
&&
1550 (needs_flat_qualifier
||
1551 (consumer_stage
!= MESA_SHADER_NONE
&& consumer_stage
!= MESA_SHADER_FRAGMENT
))) {
1552 /* Since this varying is not being consumed by the fragment shader, its
1553 * interpolation type varying cannot possibly affect rendering.
1554 * Also, this variable is non-flat and is (or contains) an integer
1556 * If the consumer stage is unknown, don't modify the interpolation
1557 * type as it could affect rendering later with separate shaders.
1559 * lower_packed_varyings requires all integer varyings to flat,
1560 * regardless of where they appear. We can trivially satisfy that
1561 * requirement by changing the interpolation type to flat here.
1564 producer_var
->data
.centroid
= false;
1565 producer_var
->data
.sample
= false;
1566 producer_var
->data
.interpolation
= INTERP_MODE_FLAT
;
1570 consumer_var
->data
.centroid
= false;
1571 consumer_var
->data
.sample
= false;
1572 consumer_var
->data
.interpolation
= INTERP_MODE_FLAT
;
1576 if (this->num_matches
== this->matches_capacity
) {
1577 this->matches_capacity
*= 2;
1578 this->matches
= (match
*)
1579 realloc(this->matches
,
1580 sizeof(*this->matches
) * this->matches_capacity
);
1583 /* We must use the consumer to compute the packing class because in GL4.4+
1584 * there is no guarantee interpolation qualifiers will match across stages.
1586 * From Section 4.5 (Interpolation Qualifiers) of the GLSL 4.30 spec:
1588 * "The type and presence of interpolation qualifiers of variables with
1589 * the same name declared in all linked shaders for the same cross-stage
1590 * interface must match, otherwise the link command will fail.
1592 * When comparing an output from one stage to an input of a subsequent
1593 * stage, the input and output don't match if their interpolation
1594 * qualifiers (or lack thereof) are not the same."
1596 * This text was also in at least revison 7 of the 4.40 spec but is no
1597 * longer in revision 9 and not in the 4.50 spec.
1599 const ir_variable
*const var
= (consumer_var
!= NULL
)
1600 ? consumer_var
: producer_var
;
1601 const gl_shader_stage stage
= (consumer_var
!= NULL
)
1602 ? consumer_stage
: producer_stage
;
1603 const glsl_type
*type
= get_varying_type(var
, stage
);
1605 if (producer_var
&& consumer_var
&&
1606 consumer_var
->data
.must_be_shader_input
) {
1607 producer_var
->data
.must_be_shader_input
= 1;
1610 this->matches
[this->num_matches
].packing_class
1611 = this->compute_packing_class(var
);
1612 this->matches
[this->num_matches
].packing_order
1613 = this->compute_packing_order(var
);
1614 if ((this->disable_varying_packing
&& !is_varying_packing_safe(type
, var
)) ||
1615 var
->data
.must_be_shader_input
) {
1616 unsigned slots
= type
->count_attribute_slots(false);
1617 this->matches
[this->num_matches
].num_components
= slots
* 4;
1619 this->matches
[this->num_matches
].num_components
1620 = type
->component_slots();
1623 this->matches
[this->num_matches
].producer_var
= producer_var
;
1624 this->matches
[this->num_matches
].consumer_var
= consumer_var
;
1625 this->num_matches
++;
1627 producer_var
->data
.is_unmatched_generic_inout
= 0;
1629 consumer_var
->data
.is_unmatched_generic_inout
= 0;
1634 * Choose locations for all of the variable matches that were previously
1635 * passed to varying_matches::record().
1638 varying_matches::assign_locations(struct gl_shader_program
*prog
,
1639 uint8_t *components
,
1640 uint64_t reserved_slots
)
1642 /* If packing has been disabled then we cannot safely sort the varyings by
1643 * class as it may mean we are using a version of OpenGL where
1644 * interpolation qualifiers are not guaranteed to be matching across
1645 * shaders, sorting in this case could result in mismatching shader
1647 * When packing is disabled the sort orders varyings used by transform
1648 * feedback first, but also depends on *undefined behaviour* of qsort to
1649 * reverse the order of the varyings. See: xfb_comparator().
1651 if (!this->disable_varying_packing
) {
1652 /* Sort varying matches into an order that makes them easy to pack. */
1653 qsort(this->matches
, this->num_matches
, sizeof(*this->matches
),
1654 &varying_matches::match_comparator
);
1656 /* Only sort varyings that are only used by transform feedback. */
1657 qsort(this->matches
, this->num_matches
, sizeof(*this->matches
),
1658 &varying_matches::xfb_comparator
);
1661 unsigned generic_location
= 0;
1662 unsigned generic_patch_location
= MAX_VARYING
*4;
1663 bool previous_var_xfb_only
= false;
1665 for (unsigned i
= 0; i
< this->num_matches
; i
++) {
1666 unsigned *location
= &generic_location
;
1668 const ir_variable
*var
;
1669 const glsl_type
*type
;
1670 bool is_vertex_input
= false;
1671 if (matches
[i
].consumer_var
) {
1672 var
= matches
[i
].consumer_var
;
1673 type
= get_varying_type(var
, consumer_stage
);
1674 if (consumer_stage
== MESA_SHADER_VERTEX
)
1675 is_vertex_input
= true;
1677 var
= matches
[i
].producer_var
;
1678 type
= get_varying_type(var
, producer_stage
);
1681 if (var
->data
.patch
)
1682 location
= &generic_patch_location
;
1684 /* Advance to the next slot if this varying has a different packing
1685 * class than the previous one, and we're not already on a slot
1688 * Also advance to the next slot if packing is disabled. This makes sure
1689 * we don't assign varyings the same locations which is possible
1690 * because we still pack individual arrays, records and matrices even
1691 * when packing is disabled. Note we don't advance to the next slot if
1692 * we can pack varyings together that are only used for transform
1695 if (var
->data
.must_be_shader_input
||
1696 (this->disable_varying_packing
&&
1697 !(previous_var_xfb_only
&& var
->data
.is_xfb_only
)) ||
1698 (i
> 0 && this->matches
[i
- 1].packing_class
1699 != this->matches
[i
].packing_class
)) {
1700 *location
= ALIGN(*location
, 4);
1703 previous_var_xfb_only
= var
->data
.is_xfb_only
;
1705 /* The number of components taken up by this variable. For vertex shader
1706 * inputs, we use the number of slots * 4, as they have different
1709 unsigned num_components
= is_vertex_input
?
1710 type
->count_attribute_slots(is_vertex_input
) * 4 :
1711 this->matches
[i
].num_components
;
1713 /* The last slot for this variable, inclusive. */
1714 unsigned slot_end
= *location
+ num_components
- 1;
1716 /* FIXME: We could be smarter in the below code and loop back over
1717 * trying to fill any locations that we skipped because we couldn't pack
1718 * the varying between an explicit location. For now just let the user
1719 * hit the linking error if we run out of room and suggest they use
1720 * explicit locations.
1722 while (slot_end
< MAX_VARYING
* 4u) {
1723 const unsigned slots
= (slot_end
/ 4u) - (*location
/ 4u) + 1;
1724 const uint64_t slot_mask
= ((1ull << slots
) - 1) << (*location
/ 4u);
1727 if (reserved_slots
& slot_mask
) {
1728 *location
= ALIGN(*location
+ 1, 4);
1729 slot_end
= *location
+ num_components
- 1;
1736 if (!var
->data
.patch
&& slot_end
>= MAX_VARYING
* 4u) {
1737 linker_error(prog
, "insufficient contiguous locations available for "
1738 "%s it is possible an array or struct could not be "
1739 "packed between varyings with explicit locations. Try "
1740 "using an explicit location for arrays and structs.",
1744 if (slot_end
< MAX_VARYINGS_INCL_PATCH
* 4u) {
1745 for (unsigned j
= *location
/ 4u; j
< slot_end
/ 4u; j
++)
1747 components
[slot_end
/ 4u] = (slot_end
& 3) + 1;
1750 this->matches
[i
].generic_location
= *location
;
1752 *location
= slot_end
+ 1;
1755 return (generic_location
+ 3) / 4;
1760 * Update the producer and consumer shaders to reflect the locations
1761 * assignments that were made by varying_matches::assign_locations().
1764 varying_matches::store_locations() const
1766 /* Check is location needs to be packed with lower_packed_varyings() or if
1767 * we can just use ARB_enhanced_layouts packing.
1769 bool pack_loc
[MAX_VARYINGS_INCL_PATCH
] = { 0 };
1770 const glsl_type
*loc_type
[MAX_VARYINGS_INCL_PATCH
][4] = { {NULL
, NULL
} };
1772 for (unsigned i
= 0; i
< this->num_matches
; i
++) {
1773 ir_variable
*producer_var
= this->matches
[i
].producer_var
;
1774 ir_variable
*consumer_var
= this->matches
[i
].consumer_var
;
1775 unsigned generic_location
= this->matches
[i
].generic_location
;
1776 unsigned slot
= generic_location
/ 4;
1777 unsigned offset
= generic_location
% 4;
1780 producer_var
->data
.location
= VARYING_SLOT_VAR0
+ slot
;
1781 producer_var
->data
.location_frac
= offset
;
1785 assert(consumer_var
->data
.location
== -1);
1786 consumer_var
->data
.location
= VARYING_SLOT_VAR0
+ slot
;
1787 consumer_var
->data
.location_frac
= offset
;
1790 /* Find locations suitable for native packing via
1791 * ARB_enhanced_layouts.
1793 if (producer_var
&& consumer_var
) {
1794 if (enhanced_layouts_enabled
) {
1795 const glsl_type
*type
=
1796 get_varying_type(producer_var
, producer_stage
);
1797 if (type
->is_array() || type
->is_matrix() || type
->is_record() ||
1798 type
->is_double()) {
1799 unsigned comp_slots
= type
->component_slots() + offset
;
1800 unsigned slots
= comp_slots
/ 4;
1804 for (unsigned j
= 0; j
< slots
; j
++) {
1805 pack_loc
[slot
+ j
] = true;
1807 } else if (offset
+ type
->vector_elements
> 4) {
1808 pack_loc
[slot
] = true;
1809 pack_loc
[slot
+ 1] = true;
1811 loc_type
[slot
][offset
] = type
;
1817 /* Attempt to use ARB_enhanced_layouts for more efficient packing if
1820 if (enhanced_layouts_enabled
) {
1821 for (unsigned i
= 0; i
< this->num_matches
; i
++) {
1822 ir_variable
*producer_var
= this->matches
[i
].producer_var
;
1823 ir_variable
*consumer_var
= this->matches
[i
].consumer_var
;
1824 unsigned generic_location
= this->matches
[i
].generic_location
;
1825 unsigned slot
= generic_location
/ 4;
1827 if (pack_loc
[slot
] || !producer_var
|| !consumer_var
)
1830 const glsl_type
*type
=
1831 get_varying_type(producer_var
, producer_stage
);
1832 bool type_match
= true;
1833 for (unsigned j
= 0; j
< 4; j
++) {
1834 if (loc_type
[slot
][j
]) {
1835 if (type
->base_type
!= loc_type
[slot
][j
]->base_type
)
1841 producer_var
->data
.explicit_location
= 1;
1842 consumer_var
->data
.explicit_location
= 1;
1843 producer_var
->data
.explicit_component
= 1;
1844 consumer_var
->data
.explicit_component
= 1;
1852 * Compute the "packing class" of the given varying. This is an unsigned
1853 * integer with the property that two variables in the same packing class can
1854 * be safely backed into the same vec4.
1857 varying_matches::compute_packing_class(const ir_variable
*var
)
1859 /* Without help from the back-end, there is no way to pack together
1860 * variables with different interpolation types, because
1861 * lower_packed_varyings must choose exactly one interpolation type for
1862 * each packed varying it creates.
1864 * However, we can safely pack together floats, ints, and uints, because:
1866 * - varyings of base type "int" and "uint" must use the "flat"
1867 * interpolation type, which can only occur in GLSL 1.30 and above.
1869 * - On platforms that support GLSL 1.30 and above, lower_packed_varyings
1870 * can store flat floats as ints without losing any information (using
1871 * the ir_unop_bitcast_* opcodes).
1873 * Therefore, the packing class depends only on the interpolation type.
1875 unsigned packing_class
= var
->data
.centroid
| (var
->data
.sample
<< 1) |
1876 (var
->data
.patch
<< 2) |
1877 (var
->data
.must_be_shader_input
<< 3);
1879 packing_class
+= var
->is_interpolation_flat()
1880 ? unsigned(INTERP_MODE_FLAT
) : var
->data
.interpolation
;
1881 return packing_class
;
1886 * Compute the "packing order" of the given varying. This is a sort key we
1887 * use to determine when to attempt to pack the given varying relative to
1888 * other varyings in the same packing class.
1890 varying_matches::packing_order_enum
1891 varying_matches::compute_packing_order(const ir_variable
*var
)
1893 const glsl_type
*element_type
= var
->type
;
1895 while (element_type
->is_array()) {
1896 element_type
= element_type
->fields
.array
;
1899 switch (element_type
->component_slots() % 4) {
1900 case 1: return PACKING_ORDER_SCALAR
;
1901 case 2: return PACKING_ORDER_VEC2
;
1902 case 3: return PACKING_ORDER_VEC3
;
1903 case 0: return PACKING_ORDER_VEC4
;
1905 assert(!"Unexpected value of vector_elements");
1906 return PACKING_ORDER_VEC4
;
1912 * Comparison function passed to qsort() to sort varyings by packing_class and
1913 * then by packing_order.
1916 varying_matches::match_comparator(const void *x_generic
, const void *y_generic
)
1918 const match
*x
= (const match
*) x_generic
;
1919 const match
*y
= (const match
*) y_generic
;
1921 if (x
->packing_class
!= y
->packing_class
)
1922 return x
->packing_class
- y
->packing_class
;
1923 return x
->packing_order
- y
->packing_order
;
1928 * Comparison function passed to qsort() to sort varyings used only by
1929 * transform feedback when packing of other varyings is disabled.
1932 varying_matches::xfb_comparator(const void *x_generic
, const void *y_generic
)
1934 const match
*x
= (const match
*) x_generic
;
1936 if (x
->producer_var
!= NULL
&& x
->producer_var
->data
.is_xfb_only
)
1937 return match_comparator(x_generic
, y_generic
);
1939 /* FIXME: When the comparator returns 0 it means the elements being
1940 * compared are equivalent. However the qsort documentation says:
1942 * "The order of equivalent elements is undefined."
1944 * In practice the sort ends up reversing the order of the varyings which
1945 * means locations are also assigned in this reversed order and happens to
1946 * be what we want. This is also whats happening in
1947 * varying_matches::match_comparator().
1954 * Is the given variable a varying variable to be counted against the
1955 * limit in ctx->Const.MaxVarying?
1956 * This includes variables such as texcoords, colors and generic
1957 * varyings, but excludes variables such as gl_FrontFacing and gl_FragCoord.
1960 var_counts_against_varying_limit(gl_shader_stage stage
, const ir_variable
*var
)
1962 /* Only fragment shaders will take a varying variable as an input */
1963 if (stage
== MESA_SHADER_FRAGMENT
&&
1964 var
->data
.mode
== ir_var_shader_in
) {
1965 switch (var
->data
.location
) {
1966 case VARYING_SLOT_POS
:
1967 case VARYING_SLOT_FACE
:
1968 case VARYING_SLOT_PNTC
:
1979 * Visitor class that generates tfeedback_candidate structs describing all
1980 * possible targets of transform feedback.
1982 * tfeedback_candidate structs are stored in the hash table
1983 * tfeedback_candidates, which is passed to the constructor. This hash table
1984 * maps varying names to instances of the tfeedback_candidate struct.
1986 class tfeedback_candidate_generator
: public program_resource_visitor
1989 tfeedback_candidate_generator(void *mem_ctx
,
1990 hash_table
*tfeedback_candidates
)
1992 tfeedback_candidates(tfeedback_candidates
),
1998 void process(ir_variable
*var
)
2000 /* All named varying interface blocks should be flattened by now */
2001 assert(!var
->is_interface_instance());
2003 this->toplevel_var
= var
;
2004 this->varying_floats
= 0;
2005 program_resource_visitor::process(var
, false);
2009 virtual void visit_field(const glsl_type
*type
, const char *name
,
2010 bool /* row_major */,
2011 const glsl_type
* /* record_type */,
2012 const enum glsl_interface_packing
,
2013 bool /* last_field */)
2015 assert(!type
->without_array()->is_record());
2016 assert(!type
->without_array()->is_interface());
2018 tfeedback_candidate
*candidate
2019 = rzalloc(this->mem_ctx
, tfeedback_candidate
);
2020 candidate
->toplevel_var
= this->toplevel_var
;
2021 candidate
->type
= type
;
2022 candidate
->offset
= this->varying_floats
;
2023 _mesa_hash_table_insert(this->tfeedback_candidates
,
2024 ralloc_strdup(this->mem_ctx
, name
),
2026 this->varying_floats
+= type
->component_slots();
2030 * Memory context used to allocate hash table keys and values.
2032 void * const mem_ctx
;
2035 * Hash table in which tfeedback_candidate objects should be stored.
2037 hash_table
* const tfeedback_candidates
;
2040 * Pointer to the toplevel variable that is being traversed.
2042 ir_variable
*toplevel_var
;
2045 * Total number of varying floats that have been visited so far. This is
2046 * used to determine the offset to each varying within the toplevel
2049 unsigned varying_floats
;
2056 populate_consumer_input_sets(void *mem_ctx
, exec_list
*ir
,
2057 hash_table
*consumer_inputs
,
2058 hash_table
*consumer_interface_inputs
,
2059 ir_variable
*consumer_inputs_with_locations
[VARYING_SLOT_TESS_MAX
])
2061 memset(consumer_inputs_with_locations
,
2063 sizeof(consumer_inputs_with_locations
[0]) * VARYING_SLOT_TESS_MAX
);
2065 foreach_in_list(ir_instruction
, node
, ir
) {
2066 ir_variable
*const input_var
= node
->as_variable();
2068 if (input_var
!= NULL
&& input_var
->data
.mode
== ir_var_shader_in
) {
2069 /* All interface blocks should have been lowered by this point */
2070 assert(!input_var
->type
->is_interface());
2072 if (input_var
->data
.explicit_location
) {
2073 /* assign_varying_locations only cares about finding the
2074 * ir_variable at the start of a contiguous location block.
2076 * - For !producer, consumer_inputs_with_locations isn't used.
2078 * - For !consumer, consumer_inputs_with_locations is empty.
2080 * For consumer && producer, if you were trying to set some
2081 * ir_variable to the middle of a location block on the other side
2082 * of producer/consumer, cross_validate_outputs_to_inputs() should
2083 * be link-erroring due to either type mismatch or location
2084 * overlaps. If the variables do match up, then they've got a
2085 * matching data.location and you only looked at
2086 * consumer_inputs_with_locations[var->data.location], not any
2087 * following entries for the array/structure.
2089 consumer_inputs_with_locations
[input_var
->data
.location
] =
2091 } else if (input_var
->get_interface_type() != NULL
) {
2092 char *const iface_field_name
=
2093 ralloc_asprintf(mem_ctx
, "%s.%s",
2094 input_var
->get_interface_type()->without_array()->name
,
2096 _mesa_hash_table_insert(consumer_interface_inputs
,
2097 iface_field_name
, input_var
);
2099 _mesa_hash_table_insert(consumer_inputs
,
2100 ralloc_strdup(mem_ctx
, input_var
->name
),
2108 * Find a variable from the consumer that "matches" the specified variable
2110 * This function only finds inputs with names that match. There is no
2111 * validation (here) that the types, etc. are compatible.
2114 get_matching_input(void *mem_ctx
,
2115 const ir_variable
*output_var
,
2116 hash_table
*consumer_inputs
,
2117 hash_table
*consumer_interface_inputs
,
2118 ir_variable
*consumer_inputs_with_locations
[VARYING_SLOT_TESS_MAX
])
2120 ir_variable
*input_var
;
2122 if (output_var
->data
.explicit_location
) {
2123 input_var
= consumer_inputs_with_locations
[output_var
->data
.location
];
2124 } else if (output_var
->get_interface_type() != NULL
) {
2125 char *const iface_field_name
=
2126 ralloc_asprintf(mem_ctx
, "%s.%s",
2127 output_var
->get_interface_type()->without_array()->name
,
2129 hash_entry
*entry
= _mesa_hash_table_search(consumer_interface_inputs
, iface_field_name
);
2130 input_var
= entry
? (ir_variable
*) entry
->data
: NULL
;
2132 hash_entry
*entry
= _mesa_hash_table_search(consumer_inputs
, output_var
->name
);
2133 input_var
= entry
? (ir_variable
*) entry
->data
: NULL
;
2136 return (input_var
== NULL
|| input_var
->data
.mode
!= ir_var_shader_in
)
2143 io_variable_cmp(const void *_a
, const void *_b
)
2145 const ir_variable
*const a
= *(const ir_variable
**) _a
;
2146 const ir_variable
*const b
= *(const ir_variable
**) _b
;
2148 if (a
->data
.explicit_location
&& b
->data
.explicit_location
)
2149 return b
->data
.location
- a
->data
.location
;
2151 if (a
->data
.explicit_location
&& !b
->data
.explicit_location
)
2154 if (!a
->data
.explicit_location
&& b
->data
.explicit_location
)
2157 return -strcmp(a
->name
, b
->name
);
2161 * Sort the shader IO variables into canonical order
2164 canonicalize_shader_io(exec_list
*ir
, enum ir_variable_mode io_mode
)
2166 ir_variable
*var_table
[MAX_PROGRAM_OUTPUTS
* 4];
2167 unsigned num_variables
= 0;
2169 foreach_in_list(ir_instruction
, node
, ir
) {
2170 ir_variable
*const var
= node
->as_variable();
2172 if (var
== NULL
|| var
->data
.mode
!= io_mode
)
2175 /* If we have already encountered more I/O variables that could
2176 * successfully link, bail.
2178 if (num_variables
== ARRAY_SIZE(var_table
))
2181 var_table
[num_variables
++] = var
;
2184 if (num_variables
== 0)
2187 /* Sort the list in reverse order (io_variable_cmp handles this). Later
2188 * we're going to push the variables on to the IR list as a stack, so we
2189 * want the last variable (in canonical order) to be first in the list.
2191 qsort(var_table
, num_variables
, sizeof(var_table
[0]), io_variable_cmp
);
2193 /* Remove the variable from it's current location in the IR, and put it at
2196 for (unsigned i
= 0; i
< num_variables
; i
++) {
2197 var_table
[i
]->remove();
2198 ir
->push_head(var_table
[i
]);
2203 * Generate a bitfield map of the explicit locations for shader varyings.
2205 * Note: For Tessellation shaders we are sitting right on the limits of the
2206 * 64 bit map. Per-vertex and per-patch both have separate location domains
2207 * with a max of MAX_VARYING.
2210 reserved_varying_slot(struct gl_linked_shader
*stage
,
2211 ir_variable_mode io_mode
)
2213 assert(io_mode
== ir_var_shader_in
|| io_mode
== ir_var_shader_out
);
2214 /* Avoid an overflow of the returned value */
2215 assert(MAX_VARYINGS_INCL_PATCH
<= 64);
2223 foreach_in_list(ir_instruction
, node
, stage
->ir
) {
2224 ir_variable
*const var
= node
->as_variable();
2226 if (var
== NULL
|| var
->data
.mode
!= io_mode
||
2227 !var
->data
.explicit_location
||
2228 var
->data
.location
< VARYING_SLOT_VAR0
)
2231 var_slot
= var
->data
.location
- VARYING_SLOT_VAR0
;
2233 unsigned num_elements
= get_varying_type(var
, stage
->Stage
)
2234 ->count_attribute_slots(io_mode
== ir_var_shader_in
&&
2235 stage
->Stage
== MESA_SHADER_VERTEX
);
2236 for (unsigned i
= 0; i
< num_elements
; i
++) {
2237 if (var_slot
>= 0 && var_slot
< MAX_VARYINGS_INCL_PATCH
)
2238 slots
|= UINT64_C(1) << var_slot
;
2248 * Assign locations for all variables that are produced in one pipeline stage
2249 * (the "producer") and consumed in the next stage (the "consumer").
2251 * Variables produced by the producer may also be consumed by transform
2254 * \param num_tfeedback_decls is the number of declarations indicating
2255 * variables that may be consumed by transform feedback.
2257 * \param tfeedback_decls is a pointer to an array of tfeedback_decl objects
2258 * representing the result of parsing the strings passed to
2259 * glTransformFeedbackVaryings(). assign_location() will be called for
2260 * each of these objects that matches one of the outputs of the
2263 * When num_tfeedback_decls is nonzero, it is permissible for the consumer to
2264 * be NULL. In this case, varying locations are assigned solely based on the
2265 * requirements of transform feedback.
2268 assign_varying_locations(struct gl_context
*ctx
,
2270 struct gl_shader_program
*prog
,
2271 gl_linked_shader
*producer
,
2272 gl_linked_shader
*consumer
,
2273 unsigned num_tfeedback_decls
,
2274 tfeedback_decl
*tfeedback_decls
,
2275 const uint64_t reserved_slots
)
2277 /* Tessellation shaders treat inputs and outputs as shared memory and can
2278 * access inputs and outputs of other invocations.
2279 * Therefore, they can't be lowered to temps easily (and definitely not
2282 bool unpackable_tess
=
2283 (consumer
&& consumer
->Stage
== MESA_SHADER_TESS_EVAL
) ||
2284 (consumer
&& consumer
->Stage
== MESA_SHADER_TESS_CTRL
) ||
2285 (producer
&& producer
->Stage
== MESA_SHADER_TESS_CTRL
);
2287 /* Transform feedback code assumes varying arrays are packed, so if the
2288 * driver has disabled varying packing, make sure to at least enable
2289 * packing required by transform feedback.
2292 ctx
->Extensions
.EXT_transform_feedback
&& !unpackable_tess
;
2294 /* Disable packing on outward facing interfaces for SSO because in ES we
2295 * need to retain the unpacked varying information for draw time
2298 * Packing is still enabled on individual arrays, structs, and matrices as
2299 * these are required by the transform feedback code and it is still safe
2300 * to do so. We also enable packing when a varying is only used for
2301 * transform feedback and its not a SSO.
2303 bool disable_varying_packing
=
2304 ctx
->Const
.DisableVaryingPacking
|| unpackable_tess
;
2305 if (prog
->SeparateShader
&& (producer
== NULL
|| consumer
== NULL
))
2306 disable_varying_packing
= true;
2308 varying_matches
matches(disable_varying_packing
, xfb_enabled
,
2309 ctx
->Extensions
.ARB_enhanced_layouts
,
2310 producer
? producer
->Stage
: MESA_SHADER_NONE
,
2311 consumer
? consumer
->Stage
: MESA_SHADER_NONE
);
2312 hash_table
*tfeedback_candidates
=
2313 _mesa_hash_table_create(NULL
, _mesa_key_hash_string
,
2314 _mesa_key_string_equal
);
2315 hash_table
*consumer_inputs
=
2316 _mesa_hash_table_create(NULL
, _mesa_key_hash_string
,
2317 _mesa_key_string_equal
);
2318 hash_table
*consumer_interface_inputs
=
2319 _mesa_hash_table_create(NULL
, _mesa_key_hash_string
,
2320 _mesa_key_string_equal
);
2321 ir_variable
*consumer_inputs_with_locations
[VARYING_SLOT_TESS_MAX
] = {
2325 unsigned consumer_vertices
= 0;
2326 if (consumer
&& consumer
->Stage
== MESA_SHADER_GEOMETRY
)
2327 consumer_vertices
= prog
->Geom
.VerticesIn
;
2329 /* Operate in a total of four passes.
2331 * 1. Sort inputs / outputs into a canonical order. This is necessary so
2332 * that inputs / outputs of separable shaders will be assigned
2333 * predictable locations regardless of the order in which declarations
2334 * appeared in the shader source.
2336 * 2. Assign locations for any matching inputs and outputs.
2338 * 3. Mark output variables in the producer that do not have locations as
2339 * not being outputs. This lets the optimizer eliminate them.
2341 * 4. Mark input variables in the consumer that do not have locations as
2342 * not being inputs. This lets the optimizer eliminate them.
2345 canonicalize_shader_io(consumer
->ir
, ir_var_shader_in
);
2348 canonicalize_shader_io(producer
->ir
, ir_var_shader_out
);
2351 linker::populate_consumer_input_sets(mem_ctx
, consumer
->ir
,
2353 consumer_interface_inputs
,
2354 consumer_inputs_with_locations
);
2357 foreach_in_list(ir_instruction
, node
, producer
->ir
) {
2358 ir_variable
*const output_var
= node
->as_variable();
2360 if (output_var
== NULL
|| output_var
->data
.mode
!= ir_var_shader_out
)
2363 /* Only geometry shaders can use non-zero streams */
2364 assert(output_var
->data
.stream
== 0 ||
2365 (output_var
->data
.stream
< MAX_VERTEX_STREAMS
&&
2366 producer
->Stage
== MESA_SHADER_GEOMETRY
));
2368 if (num_tfeedback_decls
> 0) {
2369 tfeedback_candidate_generator
g(mem_ctx
, tfeedback_candidates
);
2370 g
.process(output_var
);
2373 ir_variable
*const input_var
=
2374 linker::get_matching_input(mem_ctx
, output_var
, consumer_inputs
,
2375 consumer_interface_inputs
,
2376 consumer_inputs_with_locations
);
2378 /* If a matching input variable was found, add this output (and the
2379 * input) to the set. If this is a separable program and there is no
2380 * consumer stage, add the output.
2382 * Always add TCS outputs. They are shared by all invocations
2383 * within a patch and can be used as shared memory.
2385 if (input_var
|| (prog
->SeparateShader
&& consumer
== NULL
) ||
2386 producer
->Stage
== MESA_SHADER_TESS_CTRL
) {
2387 matches
.record(output_var
, input_var
);
2390 /* Only stream 0 outputs can be consumed in the next stage */
2391 if (input_var
&& output_var
->data
.stream
!= 0) {
2392 linker_error(prog
, "output %s is assigned to stream=%d but "
2393 "is linked to an input, which requires stream=0",
2394 output_var
->name
, output_var
->data
.stream
);
2399 /* If there's no producer stage, then this must be a separable program.
2400 * For example, we may have a program that has just a fragment shader.
2401 * Later this program will be used with some arbitrary vertex (or
2402 * geometry) shader program. This means that locations must be assigned
2403 * for all the inputs.
2405 foreach_in_list(ir_instruction
, node
, consumer
->ir
) {
2406 ir_variable
*const input_var
= node
->as_variable();
2408 if (input_var
== NULL
|| input_var
->data
.mode
!= ir_var_shader_in
)
2411 matches
.record(NULL
, input_var
);
2415 for (unsigned i
= 0; i
< num_tfeedback_decls
; ++i
) {
2416 if (!tfeedback_decls
[i
].is_varying())
2419 const tfeedback_candidate
*matched_candidate
2420 = tfeedback_decls
[i
].find_candidate(prog
, tfeedback_candidates
);
2422 if (matched_candidate
== NULL
) {
2423 _mesa_hash_table_destroy(tfeedback_candidates
, NULL
);
2427 /* Mark xfb varyings as always active */
2428 matched_candidate
->toplevel_var
->data
.always_active_io
= 1;
2430 /* Mark any corresponding inputs as always active also. We must do this
2431 * because we have a NIR pass that lowers vectors to scalars and another
2432 * that removes unused varyings.
2433 * We don't split varyings marked as always active because there is no
2434 * point in doing so. This means we need to mark both sides of the
2435 * interface as always active otherwise we will have a mismatch and
2436 * start removing things we shouldn't.
2438 ir_variable
*const input_var
=
2439 linker::get_matching_input(mem_ctx
, matched_candidate
->toplevel_var
,
2441 consumer_interface_inputs
,
2442 consumer_inputs_with_locations
);
2444 input_var
->data
.always_active_io
= 1;
2446 if (matched_candidate
->toplevel_var
->data
.is_unmatched_generic_inout
) {
2447 matched_candidate
->toplevel_var
->data
.is_xfb_only
= 1;
2448 matches
.record(matched_candidate
->toplevel_var
, NULL
);
2452 _mesa_hash_table_destroy(consumer_inputs
, NULL
);
2453 _mesa_hash_table_destroy(consumer_interface_inputs
, NULL
);
2455 uint8_t components
[MAX_VARYINGS_INCL_PATCH
] = {0};
2456 const unsigned slots_used
= matches
.assign_locations(
2457 prog
, components
, reserved_slots
);
2458 matches
.store_locations();
2460 for (unsigned i
= 0; i
< num_tfeedback_decls
; ++i
) {
2461 if (!tfeedback_decls
[i
].is_varying())
2464 if (!tfeedback_decls
[i
].assign_location(ctx
, prog
)) {
2465 _mesa_hash_table_destroy(tfeedback_candidates
, NULL
);
2469 _mesa_hash_table_destroy(tfeedback_candidates
, NULL
);
2471 if (consumer
&& producer
) {
2472 foreach_in_list(ir_instruction
, node
, consumer
->ir
) {
2473 ir_variable
*const var
= node
->as_variable();
2475 if (var
&& var
->data
.mode
== ir_var_shader_in
&&
2476 var
->data
.is_unmatched_generic_inout
) {
2477 if (!prog
->IsES
&& prog
->data
->Version
<= 120) {
2478 /* On page 25 (page 31 of the PDF) of the GLSL 1.20 spec:
2480 * Only those varying variables used (i.e. read) in
2481 * the fragment shader executable must be written to
2482 * by the vertex shader executable; declaring
2483 * superfluous varying variables in a vertex shader is
2486 * We interpret this text as meaning that the VS must
2487 * write the variable for the FS to read it. See
2488 * "glsl1-varying read but not written" in piglit.
2490 linker_error(prog
, "%s shader varying %s not written "
2492 _mesa_shader_stage_to_string(consumer
->Stage
),
2494 _mesa_shader_stage_to_string(producer
->Stage
));
2496 linker_warning(prog
, "%s shader varying %s not written "
2498 _mesa_shader_stage_to_string(consumer
->Stage
),
2500 _mesa_shader_stage_to_string(producer
->Stage
));
2505 /* Now that validation is done its safe to remove unused varyings. As
2506 * we have both a producer and consumer its safe to remove unused
2507 * varyings even if the program is a SSO because the stages are being
2508 * linked together i.e. we have a multi-stage SSO.
2510 remove_unused_shader_inputs_and_outputs(false, producer
,
2512 remove_unused_shader_inputs_and_outputs(false, consumer
,
2517 lower_packed_varyings(mem_ctx
, slots_used
, components
, ir_var_shader_out
,
2518 0, producer
, disable_varying_packing
,
2523 lower_packed_varyings(mem_ctx
, slots_used
, components
, ir_var_shader_in
,
2524 consumer_vertices
, consumer
,
2525 disable_varying_packing
, xfb_enabled
);
2532 check_against_output_limit(struct gl_context
*ctx
,
2533 struct gl_shader_program
*prog
,
2534 gl_linked_shader
*producer
,
2535 unsigned num_explicit_locations
)
2537 unsigned output_vectors
= num_explicit_locations
;
2539 foreach_in_list(ir_instruction
, node
, producer
->ir
) {
2540 ir_variable
*const var
= node
->as_variable();
2542 if (var
&& !var
->data
.explicit_location
&&
2543 var
->data
.mode
== ir_var_shader_out
&&
2544 var_counts_against_varying_limit(producer
->Stage
, var
)) {
2545 /* outputs for fragment shader can't be doubles */
2546 output_vectors
+= var
->type
->count_attribute_slots(false);
2550 assert(producer
->Stage
!= MESA_SHADER_FRAGMENT
);
2551 unsigned max_output_components
=
2552 ctx
->Const
.Program
[producer
->Stage
].MaxOutputComponents
;
2554 const unsigned output_components
= output_vectors
* 4;
2555 if (output_components
> max_output_components
) {
2556 if (ctx
->API
== API_OPENGLES2
|| prog
->IsES
)
2557 linker_error(prog
, "%s shader uses too many output vectors "
2559 _mesa_shader_stage_to_string(producer
->Stage
),
2561 max_output_components
/ 4);
2563 linker_error(prog
, "%s shader uses too many output components "
2565 _mesa_shader_stage_to_string(producer
->Stage
),
2567 max_output_components
);
2576 check_against_input_limit(struct gl_context
*ctx
,
2577 struct gl_shader_program
*prog
,
2578 gl_linked_shader
*consumer
,
2579 unsigned num_explicit_locations
)
2581 unsigned input_vectors
= num_explicit_locations
;
2583 foreach_in_list(ir_instruction
, node
, consumer
->ir
) {
2584 ir_variable
*const var
= node
->as_variable();
2586 if (var
&& !var
->data
.explicit_location
&&
2587 var
->data
.mode
== ir_var_shader_in
&&
2588 var_counts_against_varying_limit(consumer
->Stage
, var
)) {
2589 /* vertex inputs aren't varying counted */
2590 input_vectors
+= var
->type
->count_attribute_slots(false);
2594 assert(consumer
->Stage
!= MESA_SHADER_VERTEX
);
2595 unsigned max_input_components
=
2596 ctx
->Const
.Program
[consumer
->Stage
].MaxInputComponents
;
2598 const unsigned input_components
= input_vectors
* 4;
2599 if (input_components
> max_input_components
) {
2600 if (ctx
->API
== API_OPENGLES2
|| prog
->IsES
)
2601 linker_error(prog
, "%s shader uses too many input vectors "
2603 _mesa_shader_stage_to_string(consumer
->Stage
),
2605 max_input_components
/ 4);
2607 linker_error(prog
, "%s shader uses too many input components "
2609 _mesa_shader_stage_to_string(consumer
->Stage
),
2611 max_input_components
);
2620 link_varyings(struct gl_shader_program
*prog
, unsigned first
, unsigned last
,
2621 struct gl_context
*ctx
, void *mem_ctx
)
2623 bool has_xfb_qualifiers
= false;
2624 unsigned num_tfeedback_decls
= 0;
2625 char **varying_names
= NULL
;
2626 tfeedback_decl
*tfeedback_decls
= NULL
;
2628 /* From the ARB_enhanced_layouts spec:
2630 * "If the shader used to record output variables for transform feedback
2631 * varyings uses the "xfb_buffer", "xfb_offset", or "xfb_stride" layout
2632 * qualifiers, the values specified by TransformFeedbackVaryings are
2633 * ignored, and the set of variables captured for transform feedback is
2634 * instead derived from the specified layout qualifiers."
2636 for (int i
= MESA_SHADER_FRAGMENT
- 1; i
>= 0; i
--) {
2637 /* Find last stage before fragment shader */
2638 if (prog
->_LinkedShaders
[i
]) {
2639 has_xfb_qualifiers
=
2640 process_xfb_layout_qualifiers(mem_ctx
, prog
->_LinkedShaders
[i
],
2641 prog
, &num_tfeedback_decls
,
2647 if (!has_xfb_qualifiers
) {
2648 num_tfeedback_decls
= prog
->TransformFeedback
.NumVarying
;
2649 varying_names
= prog
->TransformFeedback
.VaryingNames
;
2652 if (num_tfeedback_decls
!= 0) {
2653 /* From GL_EXT_transform_feedback:
2654 * A program will fail to link if:
2656 * * the <count> specified by TransformFeedbackVaryingsEXT is
2657 * non-zero, but the program object has no vertex or geometry
2660 if (first
>= MESA_SHADER_FRAGMENT
) {
2661 linker_error(prog
, "Transform feedback varyings specified, but "
2662 "no vertex, tessellation, or geometry shader is "
2667 tfeedback_decls
= rzalloc_array(mem_ctx
, tfeedback_decl
,
2668 num_tfeedback_decls
);
2669 if (!parse_tfeedback_decls(ctx
, prog
, mem_ctx
, num_tfeedback_decls
,
2670 varying_names
, tfeedback_decls
))
2674 /* If there is no fragment shader we need to set transform feedback.
2676 * For SSO we also need to assign output locations. We assign them here
2677 * because we need to do it for both single stage programs and multi stage
2680 if (last
< MESA_SHADER_FRAGMENT
&&
2681 (num_tfeedback_decls
!= 0 || prog
->SeparateShader
)) {
2682 const uint64_t reserved_out_slots
=
2683 reserved_varying_slot(prog
->_LinkedShaders
[last
], ir_var_shader_out
);
2684 if (!assign_varying_locations(ctx
, mem_ctx
, prog
,
2685 prog
->_LinkedShaders
[last
], NULL
,
2686 num_tfeedback_decls
, tfeedback_decls
,
2687 reserved_out_slots
))
2691 if (last
<= MESA_SHADER_FRAGMENT
) {
2692 /* Remove unused varyings from the first/last stage unless SSO */
2693 remove_unused_shader_inputs_and_outputs(prog
->SeparateShader
,
2694 prog
->_LinkedShaders
[first
],
2696 remove_unused_shader_inputs_and_outputs(prog
->SeparateShader
,
2697 prog
->_LinkedShaders
[last
],
2700 /* If the program is made up of only a single stage */
2701 if (first
== last
) {
2702 gl_linked_shader
*const sh
= prog
->_LinkedShaders
[last
];
2704 do_dead_builtin_varyings(ctx
, NULL
, sh
, 0, NULL
);
2705 do_dead_builtin_varyings(ctx
, sh
, NULL
, num_tfeedback_decls
,
2708 if (prog
->SeparateShader
) {
2709 const uint64_t reserved_slots
=
2710 reserved_varying_slot(sh
, ir_var_shader_in
);
2712 /* Assign input locations for SSO, output locations are already
2715 if (!assign_varying_locations(ctx
, mem_ctx
, prog
,
2716 NULL
/* producer */,
2718 0 /* num_tfeedback_decls */,
2719 NULL
/* tfeedback_decls */,
2724 /* Linking the stages in the opposite order (from fragment to vertex)
2725 * ensures that inter-shader outputs written to in an earlier stage
2726 * are eliminated if they are (transitively) not used in a later
2730 for (int i
= next
- 1; i
>= 0; i
--) {
2731 if (prog
->_LinkedShaders
[i
] == NULL
&& i
!= 0)
2734 gl_linked_shader
*const sh_i
= prog
->_LinkedShaders
[i
];
2735 gl_linked_shader
*const sh_next
= prog
->_LinkedShaders
[next
];
2737 const uint64_t reserved_out_slots
=
2738 reserved_varying_slot(sh_i
, ir_var_shader_out
);
2739 const uint64_t reserved_in_slots
=
2740 reserved_varying_slot(sh_next
, ir_var_shader_in
);
2742 do_dead_builtin_varyings(ctx
, sh_i
, sh_next
,
2743 next
== MESA_SHADER_FRAGMENT
? num_tfeedback_decls
: 0,
2746 if (!assign_varying_locations(ctx
, mem_ctx
, prog
, sh_i
, sh_next
,
2747 next
== MESA_SHADER_FRAGMENT
? num_tfeedback_decls
: 0,
2749 reserved_out_slots
| reserved_in_slots
))
2752 /* This must be done after all dead varyings are eliminated. */
2754 unsigned slots_used
= _mesa_bitcount_64(reserved_out_slots
);
2755 if (!check_against_output_limit(ctx
, prog
, sh_i
, slots_used
)) {
2760 unsigned slots_used
= _mesa_bitcount_64(reserved_in_slots
);
2761 if (!check_against_input_limit(ctx
, prog
, sh_next
, slots_used
))
2769 if (!store_tfeedback_info(ctx
, prog
, num_tfeedback_decls
, tfeedback_decls
,
2770 has_xfb_qualifiers
))