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.
329 if (input
->data
.interpolation
!= output
->data
.interpolation
&&
330 prog
->data
->Version
< 440) {
332 "%s shader output `%s' specifies %s "
333 "interpolation qualifier, "
334 "but %s shader input specifies %s "
335 "interpolation qualifier\n",
336 _mesa_shader_stage_to_string(producer_stage
),
338 interpolation_string(output
->data
.interpolation
),
339 _mesa_shader_stage_to_string(consumer_stage
),
340 interpolation_string(input
->data
.interpolation
));
346 * Validate front and back color outputs against single color input
349 cross_validate_front_and_back_color(struct gl_shader_program
*prog
,
350 const ir_variable
*input
,
351 const ir_variable
*front_color
,
352 const ir_variable
*back_color
,
353 gl_shader_stage consumer_stage
,
354 gl_shader_stage producer_stage
)
356 if (front_color
!= NULL
&& front_color
->data
.assigned
)
357 cross_validate_types_and_qualifiers(prog
, input
, front_color
,
358 consumer_stage
, producer_stage
);
360 if (back_color
!= NULL
&& back_color
->data
.assigned
)
361 cross_validate_types_and_qualifiers(prog
, input
, back_color
,
362 consumer_stage
, producer_stage
);
366 * Validate that outputs from one stage match inputs of another
369 cross_validate_outputs_to_inputs(struct gl_shader_program
*prog
,
370 gl_linked_shader
*producer
,
371 gl_linked_shader
*consumer
)
373 glsl_symbol_table parameters
;
374 ir_variable
*explicit_locations
[MAX_VARYINGS_INCL_PATCH
][4] =
377 /* Find all shader outputs in the "producer" stage.
379 foreach_in_list(ir_instruction
, node
, producer
->ir
) {
380 ir_variable
*const var
= node
->as_variable();
382 if (var
== NULL
|| var
->data
.mode
!= ir_var_shader_out
)
385 if (!var
->data
.explicit_location
386 || var
->data
.location
< VARYING_SLOT_VAR0
)
387 parameters
.add_variable(var
);
389 /* User-defined varyings with explicit locations are handled
390 * differently because they do not need to have matching names.
392 const glsl_type
*type
= get_varying_type(var
, producer
->Stage
);
393 unsigned num_elements
= type
->count_attribute_slots(false);
394 unsigned idx
= var
->data
.location
- VARYING_SLOT_VAR0
;
395 unsigned slot_limit
= idx
+ num_elements
;
398 if (type
->without_array()->is_record()) {
399 /* The component qualifier can't be used on structs so just treat
400 * all component slots as used.
404 unsigned dmul
= type
->without_array()->is_64bit() ? 2 : 1;
405 last_comp
= var
->data
.location_frac
+
406 type
->without_array()->vector_elements
* dmul
;
409 while (idx
< slot_limit
) {
410 unsigned i
= var
->data
.location_frac
;
411 while (i
< last_comp
) {
412 if (explicit_locations
[idx
][i
] != NULL
) {
414 "%s shader has multiple outputs explicitly "
415 "assigned to location %d and component %d\n",
416 _mesa_shader_stage_to_string(producer
->Stage
),
417 idx
, var
->data
.location_frac
);
421 /* Make sure all component at this location have the same type.
423 for (unsigned j
= 0; j
< 4; j
++) {
424 if (explicit_locations
[idx
][j
] &&
425 (explicit_locations
[idx
][j
]->type
->without_array()
426 ->base_type
!= type
->without_array()->base_type
)) {
428 "Varyings sharing the same location must "
429 "have the same underlying numerical type. "
430 "Location %u component %u\n", idx
,
431 var
->data
.location_frac
);
436 explicit_locations
[idx
][i
] = var
;
439 /* We need to do some special handling for doubles as dvec3 and
440 * dvec4 consume two consecutive locations. We don't need to
441 * worry about components beginning at anything other than 0 as
442 * the spec does not allow this for dvec3 and dvec4.
444 if (i
== 4 && last_comp
> 4) {
445 last_comp
= last_comp
- 4;
446 /* Bump location index and reset the component index */
457 /* Find all shader inputs in the "consumer" stage. Any variables that have
458 * matching outputs already in the symbol table must have the same type and
461 * Exception: if the consumer is the geometry shader, then the inputs
462 * should be arrays and the type of the array element should match the type
463 * of the corresponding producer output.
465 foreach_in_list(ir_instruction
, node
, consumer
->ir
) {
466 ir_variable
*const input
= node
->as_variable();
468 if (input
== NULL
|| input
->data
.mode
!= ir_var_shader_in
)
471 if (strcmp(input
->name
, "gl_Color") == 0 && input
->data
.used
) {
472 const ir_variable
*const front_color
=
473 parameters
.get_variable("gl_FrontColor");
475 const ir_variable
*const back_color
=
476 parameters
.get_variable("gl_BackColor");
478 cross_validate_front_and_back_color(prog
, input
,
479 front_color
, back_color
,
480 consumer
->Stage
, producer
->Stage
);
481 } else if (strcmp(input
->name
, "gl_SecondaryColor") == 0 && input
->data
.used
) {
482 const ir_variable
*const front_color
=
483 parameters
.get_variable("gl_FrontSecondaryColor");
485 const ir_variable
*const back_color
=
486 parameters
.get_variable("gl_BackSecondaryColor");
488 cross_validate_front_and_back_color(prog
, input
,
489 front_color
, back_color
,
490 consumer
->Stage
, producer
->Stage
);
492 /* The rules for connecting inputs and outputs change in the presence
493 * of explicit locations. In this case, we no longer care about the
494 * names of the variables. Instead, we care only about the
495 * explicitly assigned location.
497 ir_variable
*output
= NULL
;
498 if (input
->data
.explicit_location
499 && input
->data
.location
>= VARYING_SLOT_VAR0
) {
501 const glsl_type
*type
= get_varying_type(input
, consumer
->Stage
);
502 unsigned num_elements
= type
->count_attribute_slots(false);
503 unsigned idx
= input
->data
.location
- VARYING_SLOT_VAR0
;
504 unsigned slot_limit
= idx
+ num_elements
;
506 while (idx
< slot_limit
) {
507 output
= explicit_locations
[idx
][input
->data
.location_frac
];
509 if (output
== NULL
||
510 input
->data
.location
!= output
->data
.location
) {
512 "%s shader input `%s' with explicit location "
513 "has no matching output\n",
514 _mesa_shader_stage_to_string(consumer
->Stage
),
521 output
= parameters
.get_variable(input
->name
);
524 if (output
!= NULL
) {
525 /* Interface blocks have their own validation elsewhere so don't
526 * try validating them here.
528 if (!(input
->get_interface_type() &&
529 output
->get_interface_type()))
530 cross_validate_types_and_qualifiers(prog
, input
, output
,
534 /* Check for input vars with unmatched output vars in prev stage
535 * taking into account that interface blocks could have a matching
536 * output but with different name, so we ignore them.
538 assert(!input
->data
.assigned
);
539 if (input
->data
.used
&& !input
->get_interface_type() &&
540 !input
->data
.explicit_location
&& !prog
->SeparateShader
)
542 "%s shader input `%s' "
543 "has no matching output in the previous stage\n",
544 _mesa_shader_stage_to_string(consumer
->Stage
),
552 * Demote shader inputs and outputs that are not used in other stages, and
553 * remove them via dead code elimination.
556 remove_unused_shader_inputs_and_outputs(bool is_separate_shader_object
,
557 gl_linked_shader
*sh
,
558 enum ir_variable_mode mode
)
560 if (is_separate_shader_object
)
563 foreach_in_list(ir_instruction
, node
, sh
->ir
) {
564 ir_variable
*const var
= node
->as_variable();
566 if (var
== NULL
|| var
->data
.mode
!= int(mode
))
569 /* A shader 'in' or 'out' variable is only really an input or output if
570 * its value is used by other shader stages. This will cause the
571 * variable to have a location assigned.
573 if (var
->data
.is_unmatched_generic_inout
&& !var
->data
.is_xfb_only
) {
574 assert(var
->data
.mode
!= ir_var_temporary
);
576 /* Assign zeros to demoted inputs to allow more optimizations. */
577 if (var
->data
.mode
== ir_var_shader_in
&& !var
->constant_value
)
578 var
->constant_value
= ir_constant::zero(var
, var
->type
);
580 var
->data
.mode
= ir_var_auto
;
584 /* Eliminate code that is now dead due to unused inputs/outputs being
587 while (do_dead_code(sh
->ir
, false))
593 * Initialize this object based on a string that was passed to
594 * glTransformFeedbackVaryings.
596 * If the input is mal-formed, this call still succeeds, but it sets
597 * this->var_name to a mal-formed input, so tfeedback_decl::find_output_var()
598 * will fail to find any matching variable.
601 tfeedback_decl::init(struct gl_context
*ctx
, const void *mem_ctx
,
604 /* We don't have to be pedantic about what is a valid GLSL variable name,
605 * because any variable with an invalid name can't exist in the IR anyway.
609 this->orig_name
= input
;
610 this->lowered_builtin_array_variable
= none
;
611 this->skip_components
= 0;
612 this->next_buffer_separator
= false;
613 this->matched_candidate
= NULL
;
618 if (ctx
->Extensions
.ARB_transform_feedback3
) {
619 /* Parse gl_NextBuffer. */
620 if (strcmp(input
, "gl_NextBuffer") == 0) {
621 this->next_buffer_separator
= true;
625 /* Parse gl_SkipComponents. */
626 if (strcmp(input
, "gl_SkipComponents1") == 0)
627 this->skip_components
= 1;
628 else if (strcmp(input
, "gl_SkipComponents2") == 0)
629 this->skip_components
= 2;
630 else if (strcmp(input
, "gl_SkipComponents3") == 0)
631 this->skip_components
= 3;
632 else if (strcmp(input
, "gl_SkipComponents4") == 0)
633 this->skip_components
= 4;
635 if (this->skip_components
)
639 /* Parse a declaration. */
640 const char *base_name_end
;
641 long subscript
= parse_program_resource_name(input
, &base_name_end
);
642 this->var_name
= ralloc_strndup(mem_ctx
, input
, base_name_end
- input
);
643 if (this->var_name
== NULL
) {
644 _mesa_error_no_memory(__func__
);
648 if (subscript
>= 0) {
649 this->array_subscript
= subscript
;
650 this->is_subscripted
= true;
652 this->is_subscripted
= false;
655 /* For drivers that lower gl_ClipDistance to gl_ClipDistanceMESA, this
656 * class must behave specially to account for the fact that gl_ClipDistance
657 * is converted from a float[8] to a vec4[2].
659 if (ctx
->Const
.ShaderCompilerOptions
[MESA_SHADER_VERTEX
].LowerCombinedClipCullDistance
&&
660 strcmp(this->var_name
, "gl_ClipDistance") == 0) {
661 this->lowered_builtin_array_variable
= clip_distance
;
663 if (ctx
->Const
.ShaderCompilerOptions
[MESA_SHADER_VERTEX
].LowerCombinedClipCullDistance
&&
664 strcmp(this->var_name
, "gl_CullDistance") == 0) {
665 this->lowered_builtin_array_variable
= cull_distance
;
668 if (ctx
->Const
.LowerTessLevel
&&
669 (strcmp(this->var_name
, "gl_TessLevelOuter") == 0))
670 this->lowered_builtin_array_variable
= tess_level_outer
;
671 if (ctx
->Const
.LowerTessLevel
&&
672 (strcmp(this->var_name
, "gl_TessLevelInner") == 0))
673 this->lowered_builtin_array_variable
= tess_level_inner
;
678 * Determine whether two tfeedback_decl objects refer to the same variable and
679 * array index (if applicable).
682 tfeedback_decl::is_same(const tfeedback_decl
&x
, const tfeedback_decl
&y
)
684 assert(x
.is_varying() && y
.is_varying());
686 if (strcmp(x
.var_name
, y
.var_name
) != 0)
688 if (x
.is_subscripted
!= y
.is_subscripted
)
690 if (x
.is_subscripted
&& x
.array_subscript
!= y
.array_subscript
)
697 * Assign a location and stream ID for this tfeedback_decl object based on the
698 * transform feedback candidate found by find_candidate.
700 * If an error occurs, the error is reported through linker_error() and false
704 tfeedback_decl::assign_location(struct gl_context
*ctx
,
705 struct gl_shader_program
*prog
)
707 assert(this->is_varying());
709 unsigned fine_location
710 = this->matched_candidate
->toplevel_var
->data
.location
* 4
711 + this->matched_candidate
->toplevel_var
->data
.location_frac
712 + this->matched_candidate
->offset
;
713 const unsigned dmul
=
714 this->matched_candidate
->type
->without_array()->is_64bit() ? 2 : 1;
716 if (this->matched_candidate
->type
->is_array()) {
718 const unsigned matrix_cols
=
719 this->matched_candidate
->type
->fields
.array
->matrix_columns
;
720 const unsigned vector_elements
=
721 this->matched_candidate
->type
->fields
.array
->vector_elements
;
722 unsigned actual_array_size
;
723 switch (this->lowered_builtin_array_variable
) {
725 actual_array_size
= prog
->last_vert_prog
?
726 prog
->last_vert_prog
->info
.clip_distance_array_size
: 0;
729 actual_array_size
= prog
->last_vert_prog
?
730 prog
->last_vert_prog
->info
.cull_distance_array_size
: 0;
732 case tess_level_outer
:
733 actual_array_size
= 4;
735 case tess_level_inner
:
736 actual_array_size
= 2;
740 actual_array_size
= this->matched_candidate
->type
->array_size();
744 if (this->is_subscripted
) {
745 /* Check array bounds. */
746 if (this->array_subscript
>= actual_array_size
) {
747 linker_error(prog
, "Transform feedback varying %s has index "
748 "%i, but the array size is %u.",
749 this->orig_name
, this->array_subscript
,
753 unsigned array_elem_size
= this->lowered_builtin_array_variable
?
754 1 : vector_elements
* matrix_cols
* dmul
;
755 fine_location
+= array_elem_size
* this->array_subscript
;
758 this->size
= actual_array_size
;
760 this->vector_elements
= vector_elements
;
761 this->matrix_columns
= matrix_cols
;
762 if (this->lowered_builtin_array_variable
)
763 this->type
= GL_FLOAT
;
765 this->type
= this->matched_candidate
->type
->fields
.array
->gl_type
;
767 /* Regular variable (scalar, vector, or matrix) */
768 if (this->is_subscripted
) {
769 linker_error(prog
, "Transform feedback varying %s requested, "
770 "but %s is not an array.",
771 this->orig_name
, this->var_name
);
775 this->vector_elements
= this->matched_candidate
->type
->vector_elements
;
776 this->matrix_columns
= this->matched_candidate
->type
->matrix_columns
;
777 this->type
= this->matched_candidate
->type
->gl_type
;
779 this->location
= fine_location
/ 4;
780 this->location_frac
= fine_location
% 4;
782 /* From GL_EXT_transform_feedback:
783 * A program will fail to link if:
785 * * the total number of components to capture in any varying
786 * variable in <varyings> is greater than the constant
787 * MAX_TRANSFORM_FEEDBACK_SEPARATE_COMPONENTS_EXT and the
788 * buffer mode is SEPARATE_ATTRIBS_EXT;
790 if (prog
->TransformFeedback
.BufferMode
== GL_SEPARATE_ATTRIBS
&&
791 this->num_components() >
792 ctx
->Const
.MaxTransformFeedbackSeparateComponents
) {
793 linker_error(prog
, "Transform feedback varying %s exceeds "
794 "MAX_TRANSFORM_FEEDBACK_SEPARATE_COMPONENTS.",
799 /* Only transform feedback varyings can be assigned to non-zero streams,
800 * so assign the stream id here.
802 this->stream_id
= this->matched_candidate
->toplevel_var
->data
.stream
;
804 unsigned array_offset
= this->array_subscript
* 4 * dmul
;
805 unsigned struct_offset
= this->matched_candidate
->offset
* 4 * dmul
;
806 this->buffer
= this->matched_candidate
->toplevel_var
->data
.xfb_buffer
;
807 this->offset
= this->matched_candidate
->toplevel_var
->data
.offset
+
808 array_offset
+ struct_offset
;
815 tfeedback_decl::get_num_outputs() const
817 if (!this->is_varying()) {
820 return (this->num_components() + this->location_frac
+ 3)/4;
825 * Update gl_transform_feedback_info to reflect this tfeedback_decl.
827 * If an error occurs, the error is reported through linker_error() and false
831 tfeedback_decl::store(struct gl_context
*ctx
, struct gl_shader_program
*prog
,
832 struct gl_transform_feedback_info
*info
,
833 unsigned buffer
, unsigned buffer_index
,
834 const unsigned max_outputs
, bool *explicit_stride
,
835 bool has_xfb_qualifiers
) const
837 unsigned xfb_offset
= 0;
838 unsigned size
= this->size
;
839 /* Handle gl_SkipComponents. */
840 if (this->skip_components
) {
841 info
->Buffers
[buffer
].Stride
+= this->skip_components
;
842 size
= this->skip_components
;
846 if (this->next_buffer_separator
) {
851 if (has_xfb_qualifiers
) {
852 xfb_offset
= this->offset
/ 4;
854 xfb_offset
= info
->Buffers
[buffer
].Stride
;
856 info
->Varyings
[info
->NumVarying
].Offset
= xfb_offset
* 4;
859 unsigned location
= this->location
;
860 unsigned location_frac
= this->location_frac
;
861 unsigned num_components
= this->num_components();
862 while (num_components
> 0) {
863 unsigned output_size
= MIN2(num_components
, 4 - location_frac
);
864 assert((info
->NumOutputs
== 0 && max_outputs
== 0) ||
865 info
->NumOutputs
< max_outputs
);
867 /* From the ARB_enhanced_layouts spec:
869 * "If such a block member or variable is not written during a shader
870 * invocation, the buffer contents at the assigned offset will be
871 * undefined. Even if there are no static writes to a variable or
872 * member that is assigned a transform feedback offset, the space is
873 * still allocated in the buffer and still affects the stride."
875 if (this->is_varying_written()) {
876 info
->Outputs
[info
->NumOutputs
].ComponentOffset
= location_frac
;
877 info
->Outputs
[info
->NumOutputs
].OutputRegister
= location
;
878 info
->Outputs
[info
->NumOutputs
].NumComponents
= output_size
;
879 info
->Outputs
[info
->NumOutputs
].StreamId
= stream_id
;
880 info
->Outputs
[info
->NumOutputs
].OutputBuffer
= buffer
;
881 info
->Outputs
[info
->NumOutputs
].DstOffset
= xfb_offset
;
884 info
->Buffers
[buffer
].Stream
= this->stream_id
;
885 xfb_offset
+= output_size
;
887 num_components
-= output_size
;
893 if (explicit_stride
&& explicit_stride
[buffer
]) {
894 if (this->is_64bit() && info
->Buffers
[buffer
].Stride
% 2) {
895 linker_error(prog
, "invalid qualifier xfb_stride=%d must be a "
896 "multiple of 8 as its applied to a type that is or "
897 "contains a double.",
898 info
->Buffers
[buffer
].Stride
* 4);
902 if ((this->offset
/ 4) / info
->Buffers
[buffer
].Stride
!=
903 (xfb_offset
- 1) / info
->Buffers
[buffer
].Stride
) {
904 linker_error(prog
, "xfb_offset (%d) overflows xfb_stride (%d) for "
905 "buffer (%d)", xfb_offset
* 4,
906 info
->Buffers
[buffer
].Stride
* 4, buffer
);
910 info
->Buffers
[buffer
].Stride
= xfb_offset
;
913 /* From GL_EXT_transform_feedback:
914 * A program will fail to link if:
916 * * the total number of components to capture is greater than
917 * the constant MAX_TRANSFORM_FEEDBACK_INTERLEAVED_COMPONENTS_EXT
918 * and the buffer mode is INTERLEAVED_ATTRIBS_EXT.
920 * From GL_ARB_enhanced_layouts:
922 * "The resulting stride (implicit or explicit) must be less than or
923 * equal to the implementation-dependent constant
924 * gl_MaxTransformFeedbackInterleavedComponents."
926 if ((prog
->TransformFeedback
.BufferMode
== GL_INTERLEAVED_ATTRIBS
||
927 has_xfb_qualifiers
) &&
928 info
->Buffers
[buffer
].Stride
>
929 ctx
->Const
.MaxTransformFeedbackInterleavedComponents
) {
930 linker_error(prog
, "The MAX_TRANSFORM_FEEDBACK_INTERLEAVED_COMPONENTS "
931 "limit has been exceeded.");
936 info
->Varyings
[info
->NumVarying
].Name
= ralloc_strdup(prog
,
938 info
->Varyings
[info
->NumVarying
].Type
= this->type
;
939 info
->Varyings
[info
->NumVarying
].Size
= size
;
940 info
->Varyings
[info
->NumVarying
].BufferIndex
= buffer_index
;
942 info
->Buffers
[buffer
].NumVaryings
++;
948 const tfeedback_candidate
*
949 tfeedback_decl::find_candidate(gl_shader_program
*prog
,
950 hash_table
*tfeedback_candidates
)
952 const char *name
= this->var_name
;
953 switch (this->lowered_builtin_array_variable
) {
955 name
= this->var_name
;
958 name
= "gl_ClipDistanceMESA";
961 name
= "gl_CullDistanceMESA";
963 case tess_level_outer
:
964 name
= "gl_TessLevelOuterMESA";
966 case tess_level_inner
:
967 name
= "gl_TessLevelInnerMESA";
970 hash_entry
*entry
= _mesa_hash_table_search(tfeedback_candidates
, name
);
972 this->matched_candidate
= entry
?
973 (const tfeedback_candidate
*) entry
->data
: NULL
;
975 if (!this->matched_candidate
) {
976 /* From GL_EXT_transform_feedback:
977 * A program will fail to link if:
979 * * any variable name specified in the <varyings> array is not
980 * declared as an output in the geometry shader (if present) or
981 * the vertex shader (if no geometry shader is present);
983 linker_error(prog
, "Transform feedback varying %s undeclared.",
987 return this->matched_candidate
;
992 * Parse all the transform feedback declarations that were passed to
993 * glTransformFeedbackVaryings() and store them in tfeedback_decl objects.
995 * If an error occurs, the error is reported through linker_error() and false
999 parse_tfeedback_decls(struct gl_context
*ctx
, struct gl_shader_program
*prog
,
1000 const void *mem_ctx
, unsigned num_names
,
1001 char **varying_names
, tfeedback_decl
*decls
)
1003 for (unsigned i
= 0; i
< num_names
; ++i
) {
1004 decls
[i
].init(ctx
, mem_ctx
, varying_names
[i
]);
1006 if (!decls
[i
].is_varying())
1009 /* From GL_EXT_transform_feedback:
1010 * A program will fail to link if:
1012 * * any two entries in the <varyings> array specify the same varying
1015 * We interpret this to mean "any two entries in the <varyings> array
1016 * specify the same varying variable and array index", since transform
1017 * feedback of arrays would be useless otherwise.
1019 for (unsigned j
= 0; j
< i
; ++j
) {
1020 if (!decls
[j
].is_varying())
1023 if (tfeedback_decl::is_same(decls
[i
], decls
[j
])) {
1024 linker_error(prog
, "Transform feedback varying %s specified "
1025 "more than once.", varying_names
[i
]);
1035 cmp_xfb_offset(const void * x_generic
, const void * y_generic
)
1037 tfeedback_decl
*x
= (tfeedback_decl
*) x_generic
;
1038 tfeedback_decl
*y
= (tfeedback_decl
*) y_generic
;
1040 if (x
->get_buffer() != y
->get_buffer())
1041 return x
->get_buffer() - y
->get_buffer();
1042 return x
->get_offset() - y
->get_offset();
1046 * Store transform feedback location assignments into
1047 * prog->sh.LinkedTransformFeedback based on the data stored in
1050 * If an error occurs, the error is reported through linker_error() and false
1054 store_tfeedback_info(struct gl_context
*ctx
, struct gl_shader_program
*prog
,
1055 unsigned num_tfeedback_decls
,
1056 tfeedback_decl
*tfeedback_decls
, bool has_xfb_qualifiers
)
1058 if (!prog
->last_vert_prog
)
1061 /* Make sure MaxTransformFeedbackBuffers is less than 32 so the bitmask for
1062 * tracking the number of buffers doesn't overflow.
1064 assert(ctx
->Const
.MaxTransformFeedbackBuffers
< 32);
1066 bool separate_attribs_mode
=
1067 prog
->TransformFeedback
.BufferMode
== GL_SEPARATE_ATTRIBS
;
1069 struct gl_program
*xfb_prog
= prog
->last_vert_prog
;
1070 xfb_prog
->sh
.LinkedTransformFeedback
=
1071 rzalloc(xfb_prog
, struct gl_transform_feedback_info
);
1073 /* The xfb_offset qualifier does not have to be used in increasing order
1074 * however some drivers expect to receive the list of transform feedback
1075 * declarations in order so sort it now for convenience.
1077 if (has_xfb_qualifiers
)
1078 qsort(tfeedback_decls
, num_tfeedback_decls
, sizeof(*tfeedback_decls
),
1081 xfb_prog
->sh
.LinkedTransformFeedback
->Varyings
=
1082 rzalloc_array(xfb_prog
, struct gl_transform_feedback_varying_info
,
1083 num_tfeedback_decls
);
1085 unsigned num_outputs
= 0;
1086 for (unsigned i
= 0; i
< num_tfeedback_decls
; ++i
) {
1087 if (tfeedback_decls
[i
].is_varying_written())
1088 num_outputs
+= tfeedback_decls
[i
].get_num_outputs();
1091 xfb_prog
->sh
.LinkedTransformFeedback
->Outputs
=
1092 rzalloc_array(xfb_prog
, struct gl_transform_feedback_output
,
1095 unsigned num_buffers
= 0;
1096 unsigned buffers
= 0;
1098 if (!has_xfb_qualifiers
&& separate_attribs_mode
) {
1099 /* GL_SEPARATE_ATTRIBS */
1100 for (unsigned i
= 0; i
< num_tfeedback_decls
; ++i
) {
1101 if (!tfeedback_decls
[i
].store(ctx
, prog
,
1102 xfb_prog
->sh
.LinkedTransformFeedback
,
1103 num_buffers
, num_buffers
, num_outputs
,
1104 NULL
, has_xfb_qualifiers
))
1107 buffers
|= 1 << num_buffers
;
1112 /* GL_INVERLEAVED_ATTRIBS */
1113 int buffer_stream_id
= -1;
1115 num_tfeedback_decls
? tfeedback_decls
[0].get_buffer() : 0;
1116 bool explicit_stride
[MAX_FEEDBACK_BUFFERS
] = { false };
1118 /* Apply any xfb_stride global qualifiers */
1119 if (has_xfb_qualifiers
) {
1120 for (unsigned j
= 0; j
< MAX_FEEDBACK_BUFFERS
; j
++) {
1121 if (prog
->TransformFeedback
.BufferStride
[j
]) {
1123 explicit_stride
[j
] = true;
1124 xfb_prog
->sh
.LinkedTransformFeedback
->Buffers
[j
].Stride
=
1125 prog
->TransformFeedback
.BufferStride
[j
] / 4;
1130 for (unsigned i
= 0; i
< num_tfeedback_decls
; ++i
) {
1131 if (has_xfb_qualifiers
&&
1132 buffer
!= tfeedback_decls
[i
].get_buffer()) {
1133 /* we have moved to the next buffer so reset stream id */
1134 buffer_stream_id
= -1;
1138 if (tfeedback_decls
[i
].is_next_buffer_separator()) {
1139 if (!tfeedback_decls
[i
].store(ctx
, prog
,
1140 xfb_prog
->sh
.LinkedTransformFeedback
,
1141 buffer
, num_buffers
, num_outputs
,
1142 explicit_stride
, has_xfb_qualifiers
))
1145 buffer_stream_id
= -1;
1147 } else if (tfeedback_decls
[i
].is_varying()) {
1148 if (buffer_stream_id
== -1) {
1149 /* First varying writing to this buffer: remember its stream */
1150 buffer_stream_id
= (int) tfeedback_decls
[i
].get_stream_id();
1151 } else if (buffer_stream_id
!=
1152 (int) tfeedback_decls
[i
].get_stream_id()) {
1153 /* Varying writes to the same buffer from a different stream */
1155 "Transform feedback can't capture varyings belonging "
1156 "to different vertex streams in a single buffer. "
1157 "Varying %s writes to buffer from stream %u, other "
1158 "varyings in the same buffer write from stream %u.",
1159 tfeedback_decls
[i
].name(),
1160 tfeedback_decls
[i
].get_stream_id(),
1166 if (has_xfb_qualifiers
) {
1167 buffer
= tfeedback_decls
[i
].get_buffer();
1169 buffer
= num_buffers
;
1171 buffers
|= 1 << buffer
;
1173 if (!tfeedback_decls
[i
].store(ctx
, prog
,
1174 xfb_prog
->sh
.LinkedTransformFeedback
,
1175 buffer
, num_buffers
, num_outputs
,
1176 explicit_stride
, has_xfb_qualifiers
))
1181 assert(xfb_prog
->sh
.LinkedTransformFeedback
->NumOutputs
== num_outputs
);
1183 xfb_prog
->sh
.LinkedTransformFeedback
->ActiveBuffers
= buffers
;
1190 * Data structure recording the relationship between outputs of one shader
1191 * stage (the "producer") and inputs of another (the "consumer").
1193 class varying_matches
1196 varying_matches(bool disable_varying_packing
, bool xfb_enabled
,
1197 bool enhanced_layouts_enabled
,
1198 gl_shader_stage producer_stage
,
1199 gl_shader_stage consumer_stage
);
1201 void record(ir_variable
*producer_var
, ir_variable
*consumer_var
);
1202 unsigned assign_locations(struct gl_shader_program
*prog
,
1203 uint8_t *components
,
1204 uint64_t reserved_slots
);
1205 void store_locations() const;
1208 bool is_varying_packing_safe(const glsl_type
*type
,
1209 const ir_variable
*var
);
1212 * If true, this driver disables varying packing, so all varyings need to
1213 * be aligned on slot boundaries, and take up a number of slots equal to
1214 * their number of matrix columns times their array size.
1216 * Packing may also be disabled because our current packing method is not
1217 * safe in SSO or versions of OpenGL where interpolation qualifiers are not
1218 * guaranteed to match across stages.
1220 const bool disable_varying_packing
;
1223 * If true, this driver has transform feedback enabled. The transform
1224 * feedback code requires at least some packing be done even when varying
1225 * packing is disabled, fortunately where transform feedback requires
1226 * packing it's safe to override the disabled setting. See
1227 * is_varying_packing_safe().
1229 const bool xfb_enabled
;
1231 const bool enhanced_layouts_enabled
;
1234 * Enum representing the order in which varyings are packed within a
1237 * Currently we pack vec4's first, then vec2's, then scalar values, then
1238 * vec3's. This order ensures that the only vectors that are at risk of
1239 * having to be "double parked" (split between two adjacent varying slots)
1242 enum packing_order_enum
{
1245 PACKING_ORDER_SCALAR
,
1249 static unsigned compute_packing_class(const ir_variable
*var
);
1250 static packing_order_enum
compute_packing_order(const ir_variable
*var
);
1251 static int match_comparator(const void *x_generic
, const void *y_generic
);
1252 static int xfb_comparator(const void *x_generic
, const void *y_generic
);
1255 * Structure recording the relationship between a single producer output
1256 * and a single consumer input.
1260 * Packing class for this varying, computed by compute_packing_class().
1262 unsigned packing_class
;
1265 * Packing order for this varying, computed by compute_packing_order().
1267 packing_order_enum packing_order
;
1268 unsigned num_components
;
1271 * The output variable in the producer stage.
1273 ir_variable
*producer_var
;
1276 * The input variable in the consumer stage.
1278 ir_variable
*consumer_var
;
1281 * The location which has been assigned for this varying. This is
1282 * expressed in multiples of a float, with the first generic varying
1283 * (i.e. the one referred to by VARYING_SLOT_VAR0) represented by the
1286 unsigned generic_location
;
1290 * The number of elements in the \c matches array that are currently in
1293 unsigned num_matches
;
1296 * The number of elements that were set aside for the \c matches array when
1299 unsigned matches_capacity
;
1301 gl_shader_stage producer_stage
;
1302 gl_shader_stage consumer_stage
;
1305 } /* anonymous namespace */
1307 varying_matches::varying_matches(bool disable_varying_packing
,
1309 bool enhanced_layouts_enabled
,
1310 gl_shader_stage producer_stage
,
1311 gl_shader_stage consumer_stage
)
1312 : disable_varying_packing(disable_varying_packing
),
1313 xfb_enabled(xfb_enabled
),
1314 enhanced_layouts_enabled(enhanced_layouts_enabled
),
1315 producer_stage(producer_stage
),
1316 consumer_stage(consumer_stage
)
1318 /* Note: this initial capacity is rather arbitrarily chosen to be large
1319 * enough for many cases without wasting an unreasonable amount of space.
1320 * varying_matches::record() will resize the array if there are more than
1321 * this number of varyings.
1323 this->matches_capacity
= 8;
1324 this->matches
= (match
*)
1325 malloc(sizeof(*this->matches
) * this->matches_capacity
);
1326 this->num_matches
= 0;
1330 varying_matches::~varying_matches()
1332 free(this->matches
);
1337 * Packing is always safe on individual arrays, structures, and matrices. It
1338 * is also safe if the varying is only used for transform feedback.
1341 varying_matches::is_varying_packing_safe(const glsl_type
*type
,
1342 const ir_variable
*var
)
1344 if (consumer_stage
== MESA_SHADER_TESS_EVAL
||
1345 consumer_stage
== MESA_SHADER_TESS_CTRL
||
1346 producer_stage
== MESA_SHADER_TESS_CTRL
)
1349 return xfb_enabled
&& (type
->is_array() || type
->is_record() ||
1350 type
->is_matrix() || var
->data
.is_xfb_only
);
1355 * Record the given producer/consumer variable pair in the list of variables
1356 * that should later be assigned locations.
1358 * It is permissible for \c consumer_var to be NULL (this happens if a
1359 * variable is output by the producer and consumed by transform feedback, but
1360 * not consumed by the consumer).
1362 * If \c producer_var has already been paired up with a consumer_var, or
1363 * producer_var is part of fixed pipeline functionality (and hence already has
1364 * a location assigned), this function has no effect.
1366 * Note: as a side effect this function may change the interpolation type of
1367 * \c producer_var, but only when the change couldn't possibly affect
1371 varying_matches::record(ir_variable
*producer_var
, ir_variable
*consumer_var
)
1373 assert(producer_var
!= NULL
|| consumer_var
!= NULL
);
1375 if ((producer_var
&& (!producer_var
->data
.is_unmatched_generic_inout
||
1376 producer_var
->data
.explicit_location
)) ||
1377 (consumer_var
&& (!consumer_var
->data
.is_unmatched_generic_inout
||
1378 consumer_var
->data
.explicit_location
))) {
1379 /* Either a location already exists for this variable (since it is part
1380 * of fixed functionality), or it has already been recorded as part of a
1386 bool needs_flat_qualifier
= consumer_var
== NULL
&&
1387 (producer_var
->type
->contains_integer() ||
1388 producer_var
->type
->contains_double());
1390 if (!disable_varying_packing
&&
1391 (needs_flat_qualifier
||
1392 (consumer_stage
!= MESA_SHADER_NONE
&& consumer_stage
!= MESA_SHADER_FRAGMENT
))) {
1393 /* Since this varying is not being consumed by the fragment shader, its
1394 * interpolation type varying cannot possibly affect rendering.
1395 * Also, this variable is non-flat and is (or contains) an integer
1397 * If the consumer stage is unknown, don't modify the interpolation
1398 * type as it could affect rendering later with separate shaders.
1400 * lower_packed_varyings requires all integer varyings to flat,
1401 * regardless of where they appear. We can trivially satisfy that
1402 * requirement by changing the interpolation type to flat here.
1405 producer_var
->data
.centroid
= false;
1406 producer_var
->data
.sample
= false;
1407 producer_var
->data
.interpolation
= INTERP_MODE_FLAT
;
1411 consumer_var
->data
.centroid
= false;
1412 consumer_var
->data
.sample
= false;
1413 consumer_var
->data
.interpolation
= INTERP_MODE_FLAT
;
1417 if (this->num_matches
== this->matches_capacity
) {
1418 this->matches_capacity
*= 2;
1419 this->matches
= (match
*)
1420 realloc(this->matches
,
1421 sizeof(*this->matches
) * this->matches_capacity
);
1424 /* We must use the consumer to compute the packing class because in GL4.4+
1425 * there is no guarantee interpolation qualifiers will match across stages.
1427 * From Section 4.5 (Interpolation Qualifiers) of the GLSL 4.30 spec:
1429 * "The type and presence of interpolation qualifiers of variables with
1430 * the same name declared in all linked shaders for the same cross-stage
1431 * interface must match, otherwise the link command will fail.
1433 * When comparing an output from one stage to an input of a subsequent
1434 * stage, the input and output don't match if their interpolation
1435 * qualifiers (or lack thereof) are not the same."
1437 * This text was also in at least revison 7 of the 4.40 spec but is no
1438 * longer in revision 9 and not in the 4.50 spec.
1440 const ir_variable
*const var
= (consumer_var
!= NULL
)
1441 ? consumer_var
: producer_var
;
1442 const gl_shader_stage stage
= (consumer_var
!= NULL
)
1443 ? consumer_stage
: producer_stage
;
1444 const glsl_type
*type
= get_varying_type(var
, stage
);
1446 if (producer_var
&& consumer_var
&&
1447 consumer_var
->data
.must_be_shader_input
) {
1448 producer_var
->data
.must_be_shader_input
= 1;
1451 this->matches
[this->num_matches
].packing_class
1452 = this->compute_packing_class(var
);
1453 this->matches
[this->num_matches
].packing_order
1454 = this->compute_packing_order(var
);
1455 if ((this->disable_varying_packing
&& !is_varying_packing_safe(type
, var
)) ||
1456 var
->data
.must_be_shader_input
) {
1457 unsigned slots
= type
->count_attribute_slots(false);
1458 this->matches
[this->num_matches
].num_components
= slots
* 4;
1460 this->matches
[this->num_matches
].num_components
1461 = type
->component_slots();
1464 this->matches
[this->num_matches
].producer_var
= producer_var
;
1465 this->matches
[this->num_matches
].consumer_var
= consumer_var
;
1466 this->num_matches
++;
1468 producer_var
->data
.is_unmatched_generic_inout
= 0;
1470 consumer_var
->data
.is_unmatched_generic_inout
= 0;
1475 * Choose locations for all of the variable matches that were previously
1476 * passed to varying_matches::record().
1479 varying_matches::assign_locations(struct gl_shader_program
*prog
,
1480 uint8_t *components
,
1481 uint64_t reserved_slots
)
1483 /* If packing has been disabled then we cannot safely sort the varyings by
1484 * class as it may mean we are using a version of OpenGL where
1485 * interpolation qualifiers are not guaranteed to be matching across
1486 * shaders, sorting in this case could result in mismatching shader
1488 * When packing is disabled the sort orders varyings used by transform
1489 * feedback first, but also depends on *undefined behaviour* of qsort to
1490 * reverse the order of the varyings. See: xfb_comparator().
1492 if (!this->disable_varying_packing
) {
1493 /* Sort varying matches into an order that makes them easy to pack. */
1494 qsort(this->matches
, this->num_matches
, sizeof(*this->matches
),
1495 &varying_matches::match_comparator
);
1497 /* Only sort varyings that are only used by transform feedback. */
1498 qsort(this->matches
, this->num_matches
, sizeof(*this->matches
),
1499 &varying_matches::xfb_comparator
);
1502 unsigned generic_location
= 0;
1503 unsigned generic_patch_location
= MAX_VARYING
*4;
1504 bool previous_var_xfb_only
= false;
1506 for (unsigned i
= 0; i
< this->num_matches
; i
++) {
1507 unsigned *location
= &generic_location
;
1509 const ir_variable
*var
;
1510 const glsl_type
*type
;
1511 bool is_vertex_input
= false;
1512 if (matches
[i
].consumer_var
) {
1513 var
= matches
[i
].consumer_var
;
1514 type
= get_varying_type(var
, consumer_stage
);
1515 if (consumer_stage
== MESA_SHADER_VERTEX
)
1516 is_vertex_input
= true;
1518 var
= matches
[i
].producer_var
;
1519 type
= get_varying_type(var
, producer_stage
);
1522 if (var
->data
.patch
)
1523 location
= &generic_patch_location
;
1525 /* Advance to the next slot if this varying has a different packing
1526 * class than the previous one, and we're not already on a slot
1529 * Also advance to the next slot if packing is disabled. This makes sure
1530 * we don't assign varyings the same locations which is possible
1531 * because we still pack individual arrays, records and matrices even
1532 * when packing is disabled. Note we don't advance to the next slot if
1533 * we can pack varyings together that are only used for transform
1536 if (var
->data
.must_be_shader_input
||
1537 (this->disable_varying_packing
&&
1538 !(previous_var_xfb_only
&& var
->data
.is_xfb_only
)) ||
1539 (i
> 0 && this->matches
[i
- 1].packing_class
1540 != this->matches
[i
].packing_class
)) {
1541 *location
= ALIGN(*location
, 4);
1544 previous_var_xfb_only
= var
->data
.is_xfb_only
;
1546 /* The number of components taken up by this variable. For vertex shader
1547 * inputs, we use the number of slots * 4, as they have different
1550 unsigned num_components
= is_vertex_input
?
1551 type
->count_attribute_slots(is_vertex_input
) * 4 :
1552 this->matches
[i
].num_components
;
1554 /* The last slot for this variable, inclusive. */
1555 unsigned slot_end
= *location
+ num_components
- 1;
1557 /* FIXME: We could be smarter in the below code and loop back over
1558 * trying to fill any locations that we skipped because we couldn't pack
1559 * the varying between an explicit location. For now just let the user
1560 * hit the linking error if we run out of room and suggest they use
1561 * explicit locations.
1563 while (slot_end
< MAX_VARYING
* 4u) {
1564 const unsigned slots
= (slot_end
/ 4u) - (*location
/ 4u) + 1;
1565 const uint64_t slot_mask
= ((1ull << slots
) - 1) << (*location
/ 4u);
1568 if (reserved_slots
& slot_mask
) {
1569 *location
= ALIGN(*location
+ 1, 4);
1570 slot_end
= *location
+ num_components
- 1;
1577 if (!var
->data
.patch
&& slot_end
>= MAX_VARYING
* 4u) {
1578 linker_error(prog
, "insufficient contiguous locations available for "
1579 "%s it is possible an array or struct could not be "
1580 "packed between varyings with explicit locations. Try "
1581 "using an explicit location for arrays and structs.",
1585 if (slot_end
< MAX_VARYINGS_INCL_PATCH
* 4u) {
1586 for (unsigned j
= *location
/ 4u; j
< slot_end
/ 4u; j
++)
1588 components
[slot_end
/ 4u] = (slot_end
& 3) + 1;
1591 this->matches
[i
].generic_location
= *location
;
1593 *location
= slot_end
+ 1;
1596 return (generic_location
+ 3) / 4;
1601 * Update the producer and consumer shaders to reflect the locations
1602 * assignments that were made by varying_matches::assign_locations().
1605 varying_matches::store_locations() const
1607 /* Check is location needs to be packed with lower_packed_varyings() or if
1608 * we can just use ARB_enhanced_layouts packing.
1610 bool pack_loc
[MAX_VARYINGS_INCL_PATCH
] = { 0 };
1611 const glsl_type
*loc_type
[MAX_VARYINGS_INCL_PATCH
][4] = { {NULL
, NULL
} };
1613 for (unsigned i
= 0; i
< this->num_matches
; i
++) {
1614 ir_variable
*producer_var
= this->matches
[i
].producer_var
;
1615 ir_variable
*consumer_var
= this->matches
[i
].consumer_var
;
1616 unsigned generic_location
= this->matches
[i
].generic_location
;
1617 unsigned slot
= generic_location
/ 4;
1618 unsigned offset
= generic_location
% 4;
1621 producer_var
->data
.location
= VARYING_SLOT_VAR0
+ slot
;
1622 producer_var
->data
.location_frac
= offset
;
1626 assert(consumer_var
->data
.location
== -1);
1627 consumer_var
->data
.location
= VARYING_SLOT_VAR0
+ slot
;
1628 consumer_var
->data
.location_frac
= offset
;
1631 /* Find locations suitable for native packing via
1632 * ARB_enhanced_layouts.
1634 if (producer_var
&& consumer_var
) {
1635 if (enhanced_layouts_enabled
) {
1636 const glsl_type
*type
=
1637 get_varying_type(producer_var
, producer_stage
);
1638 if (type
->is_array() || type
->is_matrix() || type
->is_record() ||
1639 type
->is_double()) {
1640 unsigned comp_slots
= type
->component_slots() + offset
;
1641 unsigned slots
= comp_slots
/ 4;
1645 for (unsigned j
= 0; j
< slots
; j
++) {
1646 pack_loc
[slot
+ j
] = true;
1648 } else if (offset
+ type
->vector_elements
> 4) {
1649 pack_loc
[slot
] = true;
1650 pack_loc
[slot
+ 1] = true;
1652 loc_type
[slot
][offset
] = type
;
1658 /* Attempt to use ARB_enhanced_layouts for more efficient packing if
1661 if (enhanced_layouts_enabled
) {
1662 for (unsigned i
= 0; i
< this->num_matches
; i
++) {
1663 ir_variable
*producer_var
= this->matches
[i
].producer_var
;
1664 ir_variable
*consumer_var
= this->matches
[i
].consumer_var
;
1665 unsigned generic_location
= this->matches
[i
].generic_location
;
1666 unsigned slot
= generic_location
/ 4;
1668 if (pack_loc
[slot
] || !producer_var
|| !consumer_var
)
1671 const glsl_type
*type
=
1672 get_varying_type(producer_var
, producer_stage
);
1673 bool type_match
= true;
1674 for (unsigned j
= 0; j
< 4; j
++) {
1675 if (loc_type
[slot
][j
]) {
1676 if (type
->base_type
!= loc_type
[slot
][j
]->base_type
)
1682 producer_var
->data
.explicit_location
= 1;
1683 consumer_var
->data
.explicit_location
= 1;
1684 producer_var
->data
.explicit_component
= 1;
1685 consumer_var
->data
.explicit_component
= 1;
1693 * Compute the "packing class" of the given varying. This is an unsigned
1694 * integer with the property that two variables in the same packing class can
1695 * be safely backed into the same vec4.
1698 varying_matches::compute_packing_class(const ir_variable
*var
)
1700 /* Without help from the back-end, there is no way to pack together
1701 * variables with different interpolation types, because
1702 * lower_packed_varyings must choose exactly one interpolation type for
1703 * each packed varying it creates.
1705 * However, we can safely pack together floats, ints, and uints, because:
1707 * - varyings of base type "int" and "uint" must use the "flat"
1708 * interpolation type, which can only occur in GLSL 1.30 and above.
1710 * - On platforms that support GLSL 1.30 and above, lower_packed_varyings
1711 * can store flat floats as ints without losing any information (using
1712 * the ir_unop_bitcast_* opcodes).
1714 * Therefore, the packing class depends only on the interpolation type.
1716 unsigned packing_class
= var
->data
.centroid
| (var
->data
.sample
<< 1) |
1717 (var
->data
.patch
<< 2) |
1718 (var
->data
.must_be_shader_input
<< 3);
1720 packing_class
+= var
->is_interpolation_flat()
1721 ? unsigned(INTERP_MODE_FLAT
) : var
->data
.interpolation
;
1722 return packing_class
;
1727 * Compute the "packing order" of the given varying. This is a sort key we
1728 * use to determine when to attempt to pack the given varying relative to
1729 * other varyings in the same packing class.
1731 varying_matches::packing_order_enum
1732 varying_matches::compute_packing_order(const ir_variable
*var
)
1734 const glsl_type
*element_type
= var
->type
;
1736 while (element_type
->is_array()) {
1737 element_type
= element_type
->fields
.array
;
1740 switch (element_type
->component_slots() % 4) {
1741 case 1: return PACKING_ORDER_SCALAR
;
1742 case 2: return PACKING_ORDER_VEC2
;
1743 case 3: return PACKING_ORDER_VEC3
;
1744 case 0: return PACKING_ORDER_VEC4
;
1746 assert(!"Unexpected value of vector_elements");
1747 return PACKING_ORDER_VEC4
;
1753 * Comparison function passed to qsort() to sort varyings by packing_class and
1754 * then by packing_order.
1757 varying_matches::match_comparator(const void *x_generic
, const void *y_generic
)
1759 const match
*x
= (const match
*) x_generic
;
1760 const match
*y
= (const match
*) y_generic
;
1762 if (x
->packing_class
!= y
->packing_class
)
1763 return x
->packing_class
- y
->packing_class
;
1764 return x
->packing_order
- y
->packing_order
;
1769 * Comparison function passed to qsort() to sort varyings used only by
1770 * transform feedback when packing of other varyings is disabled.
1773 varying_matches::xfb_comparator(const void *x_generic
, const void *y_generic
)
1775 const match
*x
= (const match
*) x_generic
;
1777 if (x
->producer_var
!= NULL
&& x
->producer_var
->data
.is_xfb_only
)
1778 return match_comparator(x_generic
, y_generic
);
1780 /* FIXME: When the comparator returns 0 it means the elements being
1781 * compared are equivalent. However the qsort documentation says:
1783 * "The order of equivalent elements is undefined."
1785 * In practice the sort ends up reversing the order of the varyings which
1786 * means locations are also assigned in this reversed order and happens to
1787 * be what we want. This is also whats happening in
1788 * varying_matches::match_comparator().
1795 * Is the given variable a varying variable to be counted against the
1796 * limit in ctx->Const.MaxVarying?
1797 * This includes variables such as texcoords, colors and generic
1798 * varyings, but excludes variables such as gl_FrontFacing and gl_FragCoord.
1801 var_counts_against_varying_limit(gl_shader_stage stage
, const ir_variable
*var
)
1803 /* Only fragment shaders will take a varying variable as an input */
1804 if (stage
== MESA_SHADER_FRAGMENT
&&
1805 var
->data
.mode
== ir_var_shader_in
) {
1806 switch (var
->data
.location
) {
1807 case VARYING_SLOT_POS
:
1808 case VARYING_SLOT_FACE
:
1809 case VARYING_SLOT_PNTC
:
1820 * Visitor class that generates tfeedback_candidate structs describing all
1821 * possible targets of transform feedback.
1823 * tfeedback_candidate structs are stored in the hash table
1824 * tfeedback_candidates, which is passed to the constructor. This hash table
1825 * maps varying names to instances of the tfeedback_candidate struct.
1827 class tfeedback_candidate_generator
: public program_resource_visitor
1830 tfeedback_candidate_generator(void *mem_ctx
,
1831 hash_table
*tfeedback_candidates
)
1833 tfeedback_candidates(tfeedback_candidates
),
1839 void process(ir_variable
*var
)
1841 /* All named varying interface blocks should be flattened by now */
1842 assert(!var
->is_interface_instance());
1844 this->toplevel_var
= var
;
1845 this->varying_floats
= 0;
1846 program_resource_visitor::process(var
, false);
1850 virtual void visit_field(const glsl_type
*type
, const char *name
,
1851 bool /* row_major */,
1852 const glsl_type
* /* record_type */,
1853 const enum glsl_interface_packing
,
1854 bool /* last_field */)
1856 assert(!type
->without_array()->is_record());
1857 assert(!type
->without_array()->is_interface());
1859 tfeedback_candidate
*candidate
1860 = rzalloc(this->mem_ctx
, tfeedback_candidate
);
1861 candidate
->toplevel_var
= this->toplevel_var
;
1862 candidate
->type
= type
;
1863 candidate
->offset
= this->varying_floats
;
1864 _mesa_hash_table_insert(this->tfeedback_candidates
,
1865 ralloc_strdup(this->mem_ctx
, name
),
1867 this->varying_floats
+= type
->component_slots();
1871 * Memory context used to allocate hash table keys and values.
1873 void * const mem_ctx
;
1876 * Hash table in which tfeedback_candidate objects should be stored.
1878 hash_table
* const tfeedback_candidates
;
1881 * Pointer to the toplevel variable that is being traversed.
1883 ir_variable
*toplevel_var
;
1886 * Total number of varying floats that have been visited so far. This is
1887 * used to determine the offset to each varying within the toplevel
1890 unsigned varying_floats
;
1897 populate_consumer_input_sets(void *mem_ctx
, exec_list
*ir
,
1898 hash_table
*consumer_inputs
,
1899 hash_table
*consumer_interface_inputs
,
1900 ir_variable
*consumer_inputs_with_locations
[VARYING_SLOT_TESS_MAX
])
1902 memset(consumer_inputs_with_locations
,
1904 sizeof(consumer_inputs_with_locations
[0]) * VARYING_SLOT_TESS_MAX
);
1906 foreach_in_list(ir_instruction
, node
, ir
) {
1907 ir_variable
*const input_var
= node
->as_variable();
1909 if (input_var
!= NULL
&& input_var
->data
.mode
== ir_var_shader_in
) {
1910 /* All interface blocks should have been lowered by this point */
1911 assert(!input_var
->type
->is_interface());
1913 if (input_var
->data
.explicit_location
) {
1914 /* assign_varying_locations only cares about finding the
1915 * ir_variable at the start of a contiguous location block.
1917 * - For !producer, consumer_inputs_with_locations isn't used.
1919 * - For !consumer, consumer_inputs_with_locations is empty.
1921 * For consumer && producer, if you were trying to set some
1922 * ir_variable to the middle of a location block on the other side
1923 * of producer/consumer, cross_validate_outputs_to_inputs() should
1924 * be link-erroring due to either type mismatch or location
1925 * overlaps. If the variables do match up, then they've got a
1926 * matching data.location and you only looked at
1927 * consumer_inputs_with_locations[var->data.location], not any
1928 * following entries for the array/structure.
1930 consumer_inputs_with_locations
[input_var
->data
.location
] =
1932 } else if (input_var
->get_interface_type() != NULL
) {
1933 char *const iface_field_name
=
1934 ralloc_asprintf(mem_ctx
, "%s.%s",
1935 input_var
->get_interface_type()->without_array()->name
,
1937 _mesa_hash_table_insert(consumer_interface_inputs
,
1938 iface_field_name
, input_var
);
1940 _mesa_hash_table_insert(consumer_inputs
,
1941 ralloc_strdup(mem_ctx
, input_var
->name
),
1949 * Find a variable from the consumer that "matches" the specified variable
1951 * This function only finds inputs with names that match. There is no
1952 * validation (here) that the types, etc. are compatible.
1955 get_matching_input(void *mem_ctx
,
1956 const ir_variable
*output_var
,
1957 hash_table
*consumer_inputs
,
1958 hash_table
*consumer_interface_inputs
,
1959 ir_variable
*consumer_inputs_with_locations
[VARYING_SLOT_TESS_MAX
])
1961 ir_variable
*input_var
;
1963 if (output_var
->data
.explicit_location
) {
1964 input_var
= consumer_inputs_with_locations
[output_var
->data
.location
];
1965 } else if (output_var
->get_interface_type() != NULL
) {
1966 char *const iface_field_name
=
1967 ralloc_asprintf(mem_ctx
, "%s.%s",
1968 output_var
->get_interface_type()->without_array()->name
,
1970 hash_entry
*entry
= _mesa_hash_table_search(consumer_interface_inputs
, iface_field_name
);
1971 input_var
= entry
? (ir_variable
*) entry
->data
: NULL
;
1973 hash_entry
*entry
= _mesa_hash_table_search(consumer_inputs
, output_var
->name
);
1974 input_var
= entry
? (ir_variable
*) entry
->data
: NULL
;
1977 return (input_var
== NULL
|| input_var
->data
.mode
!= ir_var_shader_in
)
1984 io_variable_cmp(const void *_a
, const void *_b
)
1986 const ir_variable
*const a
= *(const ir_variable
**) _a
;
1987 const ir_variable
*const b
= *(const ir_variable
**) _b
;
1989 if (a
->data
.explicit_location
&& b
->data
.explicit_location
)
1990 return b
->data
.location
- a
->data
.location
;
1992 if (a
->data
.explicit_location
&& !b
->data
.explicit_location
)
1995 if (!a
->data
.explicit_location
&& b
->data
.explicit_location
)
1998 return -strcmp(a
->name
, b
->name
);
2002 * Sort the shader IO variables into canonical order
2005 canonicalize_shader_io(exec_list
*ir
, enum ir_variable_mode io_mode
)
2007 ir_variable
*var_table
[MAX_PROGRAM_OUTPUTS
* 4];
2008 unsigned num_variables
= 0;
2010 foreach_in_list(ir_instruction
, node
, ir
) {
2011 ir_variable
*const var
= node
->as_variable();
2013 if (var
== NULL
|| var
->data
.mode
!= io_mode
)
2016 /* If we have already encountered more I/O variables that could
2017 * successfully link, bail.
2019 if (num_variables
== ARRAY_SIZE(var_table
))
2022 var_table
[num_variables
++] = var
;
2025 if (num_variables
== 0)
2028 /* Sort the list in reverse order (io_variable_cmp handles this). Later
2029 * we're going to push the variables on to the IR list as a stack, so we
2030 * want the last variable (in canonical order) to be first in the list.
2032 qsort(var_table
, num_variables
, sizeof(var_table
[0]), io_variable_cmp
);
2034 /* Remove the variable from it's current location in the IR, and put it at
2037 for (unsigned i
= 0; i
< num_variables
; i
++) {
2038 var_table
[i
]->remove();
2039 ir
->push_head(var_table
[i
]);
2044 * Generate a bitfield map of the explicit locations for shader varyings.
2046 * Note: For Tessellation shaders we are sitting right on the limits of the
2047 * 64 bit map. Per-vertex and per-patch both have separate location domains
2048 * with a max of MAX_VARYING.
2051 reserved_varying_slot(struct gl_linked_shader
*stage
,
2052 ir_variable_mode io_mode
)
2054 assert(io_mode
== ir_var_shader_in
|| io_mode
== ir_var_shader_out
);
2055 /* Avoid an overflow of the returned value */
2056 assert(MAX_VARYINGS_INCL_PATCH
<= 64);
2064 foreach_in_list(ir_instruction
, node
, stage
->ir
) {
2065 ir_variable
*const var
= node
->as_variable();
2067 if (var
== NULL
|| var
->data
.mode
!= io_mode
||
2068 !var
->data
.explicit_location
||
2069 var
->data
.location
< VARYING_SLOT_VAR0
)
2072 var_slot
= var
->data
.location
- VARYING_SLOT_VAR0
;
2074 unsigned num_elements
= get_varying_type(var
, stage
->Stage
)
2075 ->count_attribute_slots(io_mode
== ir_var_shader_in
&&
2076 stage
->Stage
== MESA_SHADER_VERTEX
);
2077 for (unsigned i
= 0; i
< num_elements
; i
++) {
2078 if (var_slot
>= 0 && var_slot
< MAX_VARYINGS_INCL_PATCH
)
2079 slots
|= UINT64_C(1) << var_slot
;
2089 * Assign locations for all variables that are produced in one pipeline stage
2090 * (the "producer") and consumed in the next stage (the "consumer").
2092 * Variables produced by the producer may also be consumed by transform
2095 * \param num_tfeedback_decls is the number of declarations indicating
2096 * variables that may be consumed by transform feedback.
2098 * \param tfeedback_decls is a pointer to an array of tfeedback_decl objects
2099 * representing the result of parsing the strings passed to
2100 * glTransformFeedbackVaryings(). assign_location() will be called for
2101 * each of these objects that matches one of the outputs of the
2104 * When num_tfeedback_decls is nonzero, it is permissible for the consumer to
2105 * be NULL. In this case, varying locations are assigned solely based on the
2106 * requirements of transform feedback.
2109 assign_varying_locations(struct gl_context
*ctx
,
2111 struct gl_shader_program
*prog
,
2112 gl_linked_shader
*producer
,
2113 gl_linked_shader
*consumer
,
2114 unsigned num_tfeedback_decls
,
2115 tfeedback_decl
*tfeedback_decls
,
2116 const uint64_t reserved_slots
)
2118 /* Tessellation shaders treat inputs and outputs as shared memory and can
2119 * access inputs and outputs of other invocations.
2120 * Therefore, they can't be lowered to temps easily (and definitely not
2123 bool unpackable_tess
=
2124 (consumer
&& consumer
->Stage
== MESA_SHADER_TESS_EVAL
) ||
2125 (consumer
&& consumer
->Stage
== MESA_SHADER_TESS_CTRL
) ||
2126 (producer
&& producer
->Stage
== MESA_SHADER_TESS_CTRL
);
2128 /* Transform feedback code assumes varying arrays are packed, so if the
2129 * driver has disabled varying packing, make sure to at least enable
2130 * packing required by transform feedback.
2133 ctx
->Extensions
.EXT_transform_feedback
&& !unpackable_tess
;
2135 /* Disable packing on outward facing interfaces for SSO because in ES we
2136 * need to retain the unpacked varying information for draw time
2139 * Packing is still enabled on individual arrays, structs, and matrices as
2140 * these are required by the transform feedback code and it is still safe
2141 * to do so. We also enable packing when a varying is only used for
2142 * transform feedback and its not a SSO.
2144 bool disable_varying_packing
=
2145 ctx
->Const
.DisableVaryingPacking
|| unpackable_tess
;
2146 if (prog
->SeparateShader
&& (producer
== NULL
|| consumer
== NULL
))
2147 disable_varying_packing
= true;
2149 varying_matches
matches(disable_varying_packing
, xfb_enabled
,
2150 ctx
->Extensions
.ARB_enhanced_layouts
,
2151 producer
? producer
->Stage
: MESA_SHADER_NONE
,
2152 consumer
? consumer
->Stage
: MESA_SHADER_NONE
);
2153 hash_table
*tfeedback_candidates
=
2154 _mesa_hash_table_create(NULL
, _mesa_key_hash_string
,
2155 _mesa_key_string_equal
);
2156 hash_table
*consumer_inputs
=
2157 _mesa_hash_table_create(NULL
, _mesa_key_hash_string
,
2158 _mesa_key_string_equal
);
2159 hash_table
*consumer_interface_inputs
=
2160 _mesa_hash_table_create(NULL
, _mesa_key_hash_string
,
2161 _mesa_key_string_equal
);
2162 ir_variable
*consumer_inputs_with_locations
[VARYING_SLOT_TESS_MAX
] = {
2166 unsigned consumer_vertices
= 0;
2167 if (consumer
&& consumer
->Stage
== MESA_SHADER_GEOMETRY
)
2168 consumer_vertices
= prog
->Geom
.VerticesIn
;
2170 /* Operate in a total of four passes.
2172 * 1. Sort inputs / outputs into a canonical order. This is necessary so
2173 * that inputs / outputs of separable shaders will be assigned
2174 * predictable locations regardless of the order in which declarations
2175 * appeared in the shader source.
2177 * 2. Assign locations for any matching inputs and outputs.
2179 * 3. Mark output variables in the producer that do not have locations as
2180 * not being outputs. This lets the optimizer eliminate them.
2182 * 4. Mark input variables in the consumer that do not have locations as
2183 * not being inputs. This lets the optimizer eliminate them.
2186 canonicalize_shader_io(consumer
->ir
, ir_var_shader_in
);
2189 canonicalize_shader_io(producer
->ir
, ir_var_shader_out
);
2192 linker::populate_consumer_input_sets(mem_ctx
, consumer
->ir
,
2194 consumer_interface_inputs
,
2195 consumer_inputs_with_locations
);
2198 foreach_in_list(ir_instruction
, node
, producer
->ir
) {
2199 ir_variable
*const output_var
= node
->as_variable();
2201 if (output_var
== NULL
|| output_var
->data
.mode
!= ir_var_shader_out
)
2204 /* Only geometry shaders can use non-zero streams */
2205 assert(output_var
->data
.stream
== 0 ||
2206 (output_var
->data
.stream
< MAX_VERTEX_STREAMS
&&
2207 producer
->Stage
== MESA_SHADER_GEOMETRY
));
2209 if (num_tfeedback_decls
> 0) {
2210 tfeedback_candidate_generator
g(mem_ctx
, tfeedback_candidates
);
2211 g
.process(output_var
);
2214 ir_variable
*const input_var
=
2215 linker::get_matching_input(mem_ctx
, output_var
, consumer_inputs
,
2216 consumer_interface_inputs
,
2217 consumer_inputs_with_locations
);
2219 /* If a matching input variable was found, add this output (and the
2220 * input) to the set. If this is a separable program and there is no
2221 * consumer stage, add the output.
2223 * Always add TCS outputs. They are shared by all invocations
2224 * within a patch and can be used as shared memory.
2226 if (input_var
|| (prog
->SeparateShader
&& consumer
== NULL
) ||
2227 producer
->Stage
== MESA_SHADER_TESS_CTRL
) {
2228 matches
.record(output_var
, input_var
);
2231 /* Only stream 0 outputs can be consumed in the next stage */
2232 if (input_var
&& output_var
->data
.stream
!= 0) {
2233 linker_error(prog
, "output %s is assigned to stream=%d but "
2234 "is linked to an input, which requires stream=0",
2235 output_var
->name
, output_var
->data
.stream
);
2240 /* If there's no producer stage, then this must be a separable program.
2241 * For example, we may have a program that has just a fragment shader.
2242 * Later this program will be used with some arbitrary vertex (or
2243 * geometry) shader program. This means that locations must be assigned
2244 * for all the inputs.
2246 foreach_in_list(ir_instruction
, node
, consumer
->ir
) {
2247 ir_variable
*const input_var
= node
->as_variable();
2249 if (input_var
== NULL
|| input_var
->data
.mode
!= ir_var_shader_in
)
2252 matches
.record(NULL
, input_var
);
2256 _mesa_hash_table_destroy(consumer_inputs
, NULL
);
2257 _mesa_hash_table_destroy(consumer_interface_inputs
, NULL
);
2259 for (unsigned i
= 0; i
< num_tfeedback_decls
; ++i
) {
2260 if (!tfeedback_decls
[i
].is_varying())
2263 const tfeedback_candidate
*matched_candidate
2264 = tfeedback_decls
[i
].find_candidate(prog
, tfeedback_candidates
);
2266 if (matched_candidate
== NULL
) {
2267 _mesa_hash_table_destroy(tfeedback_candidates
, NULL
);
2271 /* Mark xfb varyings as always active */
2272 matched_candidate
->toplevel_var
->data
.always_active_io
= 1;
2274 if (matched_candidate
->toplevel_var
->data
.is_unmatched_generic_inout
) {
2275 matched_candidate
->toplevel_var
->data
.is_xfb_only
= 1;
2276 matches
.record(matched_candidate
->toplevel_var
, NULL
);
2280 uint8_t components
[MAX_VARYINGS_INCL_PATCH
] = {0};
2281 const unsigned slots_used
= matches
.assign_locations(
2282 prog
, components
, reserved_slots
);
2283 matches
.store_locations();
2285 for (unsigned i
= 0; i
< num_tfeedback_decls
; ++i
) {
2286 if (!tfeedback_decls
[i
].is_varying())
2289 if (!tfeedback_decls
[i
].assign_location(ctx
, prog
)) {
2290 _mesa_hash_table_destroy(tfeedback_candidates
, NULL
);
2294 _mesa_hash_table_destroy(tfeedback_candidates
, NULL
);
2296 if (consumer
&& producer
) {
2297 foreach_in_list(ir_instruction
, node
, consumer
->ir
) {
2298 ir_variable
*const var
= node
->as_variable();
2300 if (var
&& var
->data
.mode
== ir_var_shader_in
&&
2301 var
->data
.is_unmatched_generic_inout
) {
2302 if (!prog
->IsES
&& prog
->data
->Version
<= 120) {
2303 /* On page 25 (page 31 of the PDF) of the GLSL 1.20 spec:
2305 * Only those varying variables used (i.e. read) in
2306 * the fragment shader executable must be written to
2307 * by the vertex shader executable; declaring
2308 * superfluous varying variables in a vertex shader is
2311 * We interpret this text as meaning that the VS must
2312 * write the variable for the FS to read it. See
2313 * "glsl1-varying read but not written" in piglit.
2315 linker_error(prog
, "%s shader varying %s not written "
2317 _mesa_shader_stage_to_string(consumer
->Stage
),
2319 _mesa_shader_stage_to_string(producer
->Stage
));
2321 linker_warning(prog
, "%s shader varying %s not written "
2323 _mesa_shader_stage_to_string(consumer
->Stage
),
2325 _mesa_shader_stage_to_string(producer
->Stage
));
2330 /* Now that validation is done its safe to remove unused varyings. As
2331 * we have both a producer and consumer its safe to remove unused
2332 * varyings even if the program is a SSO because the stages are being
2333 * linked together i.e. we have a multi-stage SSO.
2335 remove_unused_shader_inputs_and_outputs(false, producer
,
2337 remove_unused_shader_inputs_and_outputs(false, consumer
,
2342 lower_packed_varyings(mem_ctx
, slots_used
, components
, ir_var_shader_out
,
2343 0, producer
, disable_varying_packing
,
2348 lower_packed_varyings(mem_ctx
, slots_used
, components
, ir_var_shader_in
,
2349 consumer_vertices
, consumer
,
2350 disable_varying_packing
, xfb_enabled
);
2357 check_against_output_limit(struct gl_context
*ctx
,
2358 struct gl_shader_program
*prog
,
2359 gl_linked_shader
*producer
,
2360 unsigned num_explicit_locations
)
2362 unsigned output_vectors
= num_explicit_locations
;
2364 foreach_in_list(ir_instruction
, node
, producer
->ir
) {
2365 ir_variable
*const var
= node
->as_variable();
2367 if (var
&& !var
->data
.explicit_location
&&
2368 var
->data
.mode
== ir_var_shader_out
&&
2369 var_counts_against_varying_limit(producer
->Stage
, var
)) {
2370 /* outputs for fragment shader can't be doubles */
2371 output_vectors
+= var
->type
->count_attribute_slots(false);
2375 assert(producer
->Stage
!= MESA_SHADER_FRAGMENT
);
2376 unsigned max_output_components
=
2377 ctx
->Const
.Program
[producer
->Stage
].MaxOutputComponents
;
2379 const unsigned output_components
= output_vectors
* 4;
2380 if (output_components
> max_output_components
) {
2381 if (ctx
->API
== API_OPENGLES2
|| prog
->IsES
)
2382 linker_error(prog
, "%s shader uses too many output vectors "
2384 _mesa_shader_stage_to_string(producer
->Stage
),
2386 max_output_components
/ 4);
2388 linker_error(prog
, "%s shader uses too many output components "
2390 _mesa_shader_stage_to_string(producer
->Stage
),
2392 max_output_components
);
2401 check_against_input_limit(struct gl_context
*ctx
,
2402 struct gl_shader_program
*prog
,
2403 gl_linked_shader
*consumer
,
2404 unsigned num_explicit_locations
)
2406 unsigned input_vectors
= num_explicit_locations
;
2408 foreach_in_list(ir_instruction
, node
, consumer
->ir
) {
2409 ir_variable
*const var
= node
->as_variable();
2411 if (var
&& !var
->data
.explicit_location
&&
2412 var
->data
.mode
== ir_var_shader_in
&&
2413 var_counts_against_varying_limit(consumer
->Stage
, var
)) {
2414 /* vertex inputs aren't varying counted */
2415 input_vectors
+= var
->type
->count_attribute_slots(false);
2419 assert(consumer
->Stage
!= MESA_SHADER_VERTEX
);
2420 unsigned max_input_components
=
2421 ctx
->Const
.Program
[consumer
->Stage
].MaxInputComponents
;
2423 const unsigned input_components
= input_vectors
* 4;
2424 if (input_components
> max_input_components
) {
2425 if (ctx
->API
== API_OPENGLES2
|| prog
->IsES
)
2426 linker_error(prog
, "%s shader uses too many input vectors "
2428 _mesa_shader_stage_to_string(consumer
->Stage
),
2430 max_input_components
/ 4);
2432 linker_error(prog
, "%s shader uses too many input components "
2434 _mesa_shader_stage_to_string(consumer
->Stage
),
2436 max_input_components
);
2445 link_varyings(struct gl_shader_program
*prog
, unsigned first
, unsigned last
,
2446 struct gl_context
*ctx
, void *mem_ctx
)
2448 bool has_xfb_qualifiers
= false;
2449 unsigned num_tfeedback_decls
= 0;
2450 char **varying_names
= NULL
;
2451 tfeedback_decl
*tfeedback_decls
= NULL
;
2453 /* From the ARB_enhanced_layouts spec:
2455 * "If the shader used to record output variables for transform feedback
2456 * varyings uses the "xfb_buffer", "xfb_offset", or "xfb_stride" layout
2457 * qualifiers, the values specified by TransformFeedbackVaryings are
2458 * ignored, and the set of variables captured for transform feedback is
2459 * instead derived from the specified layout qualifiers."
2461 for (int i
= MESA_SHADER_FRAGMENT
- 1; i
>= 0; i
--) {
2462 /* Find last stage before fragment shader */
2463 if (prog
->_LinkedShaders
[i
]) {
2464 has_xfb_qualifiers
=
2465 process_xfb_layout_qualifiers(mem_ctx
, prog
->_LinkedShaders
[i
],
2466 prog
, &num_tfeedback_decls
,
2472 if (!has_xfb_qualifiers
) {
2473 num_tfeedback_decls
= prog
->TransformFeedback
.NumVarying
;
2474 varying_names
= prog
->TransformFeedback
.VaryingNames
;
2477 if (num_tfeedback_decls
!= 0) {
2478 /* From GL_EXT_transform_feedback:
2479 * A program will fail to link if:
2481 * * the <count> specified by TransformFeedbackVaryingsEXT is
2482 * non-zero, but the program object has no vertex or geometry
2485 if (first
>= MESA_SHADER_FRAGMENT
) {
2486 linker_error(prog
, "Transform feedback varyings specified, but "
2487 "no vertex, tessellation, or geometry shader is "
2492 tfeedback_decls
= rzalloc_array(mem_ctx
, tfeedback_decl
,
2493 num_tfeedback_decls
);
2494 if (!parse_tfeedback_decls(ctx
, prog
, mem_ctx
, num_tfeedback_decls
,
2495 varying_names
, tfeedback_decls
))
2499 /* If there is no fragment shader we need to set transform feedback.
2501 * For SSO we also need to assign output locations. We assign them here
2502 * because we need to do it for both single stage programs and multi stage
2505 if (last
< MESA_SHADER_FRAGMENT
&&
2506 (num_tfeedback_decls
!= 0 || prog
->SeparateShader
)) {
2507 const uint64_t reserved_out_slots
=
2508 reserved_varying_slot(prog
->_LinkedShaders
[last
], ir_var_shader_out
);
2509 if (!assign_varying_locations(ctx
, mem_ctx
, prog
,
2510 prog
->_LinkedShaders
[last
], NULL
,
2511 num_tfeedback_decls
, tfeedback_decls
,
2512 reserved_out_slots
))
2516 if (last
<= MESA_SHADER_FRAGMENT
) {
2517 /* Remove unused varyings from the first/last stage unless SSO */
2518 remove_unused_shader_inputs_and_outputs(prog
->SeparateShader
,
2519 prog
->_LinkedShaders
[first
],
2521 remove_unused_shader_inputs_and_outputs(prog
->SeparateShader
,
2522 prog
->_LinkedShaders
[last
],
2525 /* If the program is made up of only a single stage */
2526 if (first
== last
) {
2527 gl_linked_shader
*const sh
= prog
->_LinkedShaders
[last
];
2529 do_dead_builtin_varyings(ctx
, NULL
, sh
, 0, NULL
);
2530 do_dead_builtin_varyings(ctx
, sh
, NULL
, num_tfeedback_decls
,
2533 if (prog
->SeparateShader
) {
2534 const uint64_t reserved_slots
=
2535 reserved_varying_slot(sh
, ir_var_shader_in
);
2537 /* Assign input locations for SSO, output locations are already
2540 if (!assign_varying_locations(ctx
, mem_ctx
, prog
,
2541 NULL
/* producer */,
2543 0 /* num_tfeedback_decls */,
2544 NULL
/* tfeedback_decls */,
2549 /* Linking the stages in the opposite order (from fragment to vertex)
2550 * ensures that inter-shader outputs written to in an earlier stage
2551 * are eliminated if they are (transitively) not used in a later
2555 for (int i
= next
- 1; i
>= 0; i
--) {
2556 if (prog
->_LinkedShaders
[i
] == NULL
&& i
!= 0)
2559 gl_linked_shader
*const sh_i
= prog
->_LinkedShaders
[i
];
2560 gl_linked_shader
*const sh_next
= prog
->_LinkedShaders
[next
];
2562 const uint64_t reserved_out_slots
=
2563 reserved_varying_slot(sh_i
, ir_var_shader_out
);
2564 const uint64_t reserved_in_slots
=
2565 reserved_varying_slot(sh_next
, ir_var_shader_in
);
2567 do_dead_builtin_varyings(ctx
, sh_i
, sh_next
,
2568 next
== MESA_SHADER_FRAGMENT
? num_tfeedback_decls
: 0,
2571 if (!assign_varying_locations(ctx
, mem_ctx
, prog
, sh_i
, sh_next
,
2572 next
== MESA_SHADER_FRAGMENT
? num_tfeedback_decls
: 0,
2574 reserved_out_slots
| reserved_in_slots
))
2577 /* This must be done after all dead varyings are eliminated. */
2579 unsigned slots_used
= _mesa_bitcount_64(reserved_out_slots
);
2580 if (!check_against_output_limit(ctx
, prog
, sh_i
, slots_used
)) {
2585 unsigned slots_used
= _mesa_bitcount_64(reserved_in_slots
);
2586 if (!check_against_input_limit(ctx
, prog
, sh_next
, slots_used
))
2594 if (!store_tfeedback_info(ctx
, prog
, num_tfeedback_decls
, tfeedback_decls
,
2595 has_xfb_qualifiers
))