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 * Validate that outputs from one stage match inputs of another
384 cross_validate_outputs_to_inputs(struct gl_shader_program
*prog
,
385 gl_linked_shader
*producer
,
386 gl_linked_shader
*consumer
)
388 glsl_symbol_table parameters
;
389 ir_variable
*explicit_locations
[MAX_VARYINGS_INCL_PATCH
][4] =
392 /* Find all shader outputs in the "producer" stage.
394 foreach_in_list(ir_instruction
, node
, producer
->ir
) {
395 ir_variable
*const var
= node
->as_variable();
397 if (var
== NULL
|| var
->data
.mode
!= ir_var_shader_out
)
400 if (!var
->data
.explicit_location
401 || var
->data
.location
< VARYING_SLOT_VAR0
)
402 parameters
.add_variable(var
);
404 /* User-defined varyings with explicit locations are handled
405 * differently because they do not need to have matching names.
407 const glsl_type
*type
= get_varying_type(var
, producer
->Stage
);
408 unsigned num_elements
= type
->count_attribute_slots(false);
409 unsigned idx
= var
->data
.location
- VARYING_SLOT_VAR0
;
410 unsigned slot_limit
= idx
+ num_elements
;
413 if (type
->without_array()->is_record()) {
414 /* The component qualifier can't be used on structs so just treat
415 * all component slots as used.
419 unsigned dmul
= type
->without_array()->is_64bit() ? 2 : 1;
420 last_comp
= var
->data
.location_frac
+
421 type
->without_array()->vector_elements
* dmul
;
424 while (idx
< slot_limit
) {
425 unsigned i
= var
->data
.location_frac
;
426 while (i
< last_comp
) {
427 if (explicit_locations
[idx
][i
] != NULL
) {
429 "%s shader has multiple outputs explicitly "
430 "assigned to location %d and component %d\n",
431 _mesa_shader_stage_to_string(producer
->Stage
),
432 idx
, var
->data
.location_frac
);
436 /* Make sure all component at this location have the same type.
438 for (unsigned j
= 0; j
< 4; j
++) {
439 if (explicit_locations
[idx
][j
] &&
440 (explicit_locations
[idx
][j
]->type
->without_array()
441 ->base_type
!= type
->without_array()->base_type
)) {
443 "Varyings sharing the same location must "
444 "have the same underlying numerical type. "
445 "Location %u component %u\n", idx
,
446 var
->data
.location_frac
);
451 explicit_locations
[idx
][i
] = var
;
454 /* We need to do some special handling for doubles as dvec3 and
455 * dvec4 consume two consecutive locations. We don't need to
456 * worry about components beginning at anything other than 0 as
457 * the spec does not allow this for dvec3 and dvec4.
459 if (i
== 4 && last_comp
> 4) {
460 last_comp
= last_comp
- 4;
461 /* Bump location index and reset the component index */
472 /* Find all shader inputs in the "consumer" stage. Any variables that have
473 * matching outputs already in the symbol table must have the same type and
476 * Exception: if the consumer is the geometry shader, then the inputs
477 * should be arrays and the type of the array element should match the type
478 * of the corresponding producer output.
480 foreach_in_list(ir_instruction
, node
, consumer
->ir
) {
481 ir_variable
*const input
= node
->as_variable();
483 if (input
== NULL
|| input
->data
.mode
!= ir_var_shader_in
)
486 if (strcmp(input
->name
, "gl_Color") == 0 && input
->data
.used
) {
487 const ir_variable
*const front_color
=
488 parameters
.get_variable("gl_FrontColor");
490 const ir_variable
*const back_color
=
491 parameters
.get_variable("gl_BackColor");
493 cross_validate_front_and_back_color(prog
, input
,
494 front_color
, back_color
,
495 consumer
->Stage
, producer
->Stage
);
496 } else if (strcmp(input
->name
, "gl_SecondaryColor") == 0 && input
->data
.used
) {
497 const ir_variable
*const front_color
=
498 parameters
.get_variable("gl_FrontSecondaryColor");
500 const ir_variable
*const back_color
=
501 parameters
.get_variable("gl_BackSecondaryColor");
503 cross_validate_front_and_back_color(prog
, input
,
504 front_color
, back_color
,
505 consumer
->Stage
, producer
->Stage
);
507 /* The rules for connecting inputs and outputs change in the presence
508 * of explicit locations. In this case, we no longer care about the
509 * names of the variables. Instead, we care only about the
510 * explicitly assigned location.
512 ir_variable
*output
= NULL
;
513 if (input
->data
.explicit_location
514 && input
->data
.location
>= VARYING_SLOT_VAR0
) {
516 const glsl_type
*type
= get_varying_type(input
, consumer
->Stage
);
517 unsigned num_elements
= type
->count_attribute_slots(false);
518 unsigned idx
= input
->data
.location
- VARYING_SLOT_VAR0
;
519 unsigned slot_limit
= idx
+ num_elements
;
521 while (idx
< slot_limit
) {
522 output
= explicit_locations
[idx
][input
->data
.location_frac
];
524 if (output
== NULL
||
525 input
->data
.location
!= output
->data
.location
) {
527 "%s shader input `%s' with explicit location "
528 "has no matching output\n",
529 _mesa_shader_stage_to_string(consumer
->Stage
),
536 output
= parameters
.get_variable(input
->name
);
539 if (output
!= NULL
) {
540 /* Interface blocks have their own validation elsewhere so don't
541 * try validating them here.
543 if (!(input
->get_interface_type() &&
544 output
->get_interface_type()))
545 cross_validate_types_and_qualifiers(prog
, input
, output
,
549 /* Check for input vars with unmatched output vars in prev stage
550 * taking into account that interface blocks could have a matching
551 * output but with different name, so we ignore them.
553 assert(!input
->data
.assigned
);
554 if (input
->data
.used
&& !input
->get_interface_type() &&
555 !input
->data
.explicit_location
&& !prog
->SeparateShader
)
557 "%s shader input `%s' "
558 "has no matching output in the previous stage\n",
559 _mesa_shader_stage_to_string(consumer
->Stage
),
567 * Demote shader inputs and outputs that are not used in other stages, and
568 * remove them via dead code elimination.
571 remove_unused_shader_inputs_and_outputs(bool is_separate_shader_object
,
572 gl_linked_shader
*sh
,
573 enum ir_variable_mode mode
)
575 if (is_separate_shader_object
)
578 foreach_in_list(ir_instruction
, node
, sh
->ir
) {
579 ir_variable
*const var
= node
->as_variable();
581 if (var
== NULL
|| var
->data
.mode
!= int(mode
))
584 /* A shader 'in' or 'out' variable is only really an input or output if
585 * its value is used by other shader stages. This will cause the
586 * variable to have a location assigned.
588 if (var
->data
.is_unmatched_generic_inout
&& !var
->data
.is_xfb_only
) {
589 assert(var
->data
.mode
!= ir_var_temporary
);
591 /* Assign zeros to demoted inputs to allow more optimizations. */
592 if (var
->data
.mode
== ir_var_shader_in
&& !var
->constant_value
)
593 var
->constant_value
= ir_constant::zero(var
, var
->type
);
595 var
->data
.mode
= ir_var_auto
;
599 /* Eliminate code that is now dead due to unused inputs/outputs being
602 while (do_dead_code(sh
->ir
, false))
608 * Initialize this object based on a string that was passed to
609 * glTransformFeedbackVaryings.
611 * If the input is mal-formed, this call still succeeds, but it sets
612 * this->var_name to a mal-formed input, so tfeedback_decl::find_output_var()
613 * will fail to find any matching variable.
616 tfeedback_decl::init(struct gl_context
*ctx
, const void *mem_ctx
,
619 /* We don't have to be pedantic about what is a valid GLSL variable name,
620 * because any variable with an invalid name can't exist in the IR anyway.
624 this->orig_name
= input
;
625 this->lowered_builtin_array_variable
= none
;
626 this->skip_components
= 0;
627 this->next_buffer_separator
= false;
628 this->matched_candidate
= NULL
;
633 if (ctx
->Extensions
.ARB_transform_feedback3
) {
634 /* Parse gl_NextBuffer. */
635 if (strcmp(input
, "gl_NextBuffer") == 0) {
636 this->next_buffer_separator
= true;
640 /* Parse gl_SkipComponents. */
641 if (strcmp(input
, "gl_SkipComponents1") == 0)
642 this->skip_components
= 1;
643 else if (strcmp(input
, "gl_SkipComponents2") == 0)
644 this->skip_components
= 2;
645 else if (strcmp(input
, "gl_SkipComponents3") == 0)
646 this->skip_components
= 3;
647 else if (strcmp(input
, "gl_SkipComponents4") == 0)
648 this->skip_components
= 4;
650 if (this->skip_components
)
654 /* Parse a declaration. */
655 const char *base_name_end
;
656 long subscript
= parse_program_resource_name(input
, &base_name_end
);
657 this->var_name
= ralloc_strndup(mem_ctx
, input
, base_name_end
- input
);
658 if (this->var_name
== NULL
) {
659 _mesa_error_no_memory(__func__
);
663 if (subscript
>= 0) {
664 this->array_subscript
= subscript
;
665 this->is_subscripted
= true;
667 this->is_subscripted
= false;
670 /* For drivers that lower gl_ClipDistance to gl_ClipDistanceMESA, this
671 * class must behave specially to account for the fact that gl_ClipDistance
672 * is converted from a float[8] to a vec4[2].
674 if (ctx
->Const
.ShaderCompilerOptions
[MESA_SHADER_VERTEX
].LowerCombinedClipCullDistance
&&
675 strcmp(this->var_name
, "gl_ClipDistance") == 0) {
676 this->lowered_builtin_array_variable
= clip_distance
;
678 if (ctx
->Const
.ShaderCompilerOptions
[MESA_SHADER_VERTEX
].LowerCombinedClipCullDistance
&&
679 strcmp(this->var_name
, "gl_CullDistance") == 0) {
680 this->lowered_builtin_array_variable
= cull_distance
;
683 if (ctx
->Const
.LowerTessLevel
&&
684 (strcmp(this->var_name
, "gl_TessLevelOuter") == 0))
685 this->lowered_builtin_array_variable
= tess_level_outer
;
686 if (ctx
->Const
.LowerTessLevel
&&
687 (strcmp(this->var_name
, "gl_TessLevelInner") == 0))
688 this->lowered_builtin_array_variable
= tess_level_inner
;
693 * Determine whether two tfeedback_decl objects refer to the same variable and
694 * array index (if applicable).
697 tfeedback_decl::is_same(const tfeedback_decl
&x
, const tfeedback_decl
&y
)
699 assert(x
.is_varying() && y
.is_varying());
701 if (strcmp(x
.var_name
, y
.var_name
) != 0)
703 if (x
.is_subscripted
!= y
.is_subscripted
)
705 if (x
.is_subscripted
&& x
.array_subscript
!= y
.array_subscript
)
712 * Assign a location and stream ID for this tfeedback_decl object based on the
713 * transform feedback candidate found by find_candidate.
715 * If an error occurs, the error is reported through linker_error() and false
719 tfeedback_decl::assign_location(struct gl_context
*ctx
,
720 struct gl_shader_program
*prog
)
722 assert(this->is_varying());
724 unsigned fine_location
725 = this->matched_candidate
->toplevel_var
->data
.location
* 4
726 + this->matched_candidate
->toplevel_var
->data
.location_frac
727 + this->matched_candidate
->offset
;
728 const unsigned dmul
=
729 this->matched_candidate
->type
->without_array()->is_64bit() ? 2 : 1;
731 if (this->matched_candidate
->type
->is_array()) {
733 const unsigned matrix_cols
=
734 this->matched_candidate
->type
->fields
.array
->matrix_columns
;
735 const unsigned vector_elements
=
736 this->matched_candidate
->type
->fields
.array
->vector_elements
;
737 unsigned actual_array_size
;
738 switch (this->lowered_builtin_array_variable
) {
740 actual_array_size
= prog
->last_vert_prog
?
741 prog
->last_vert_prog
->info
.clip_distance_array_size
: 0;
744 actual_array_size
= prog
->last_vert_prog
?
745 prog
->last_vert_prog
->info
.cull_distance_array_size
: 0;
747 case tess_level_outer
:
748 actual_array_size
= 4;
750 case tess_level_inner
:
751 actual_array_size
= 2;
755 actual_array_size
= this->matched_candidate
->type
->array_size();
759 if (this->is_subscripted
) {
760 /* Check array bounds. */
761 if (this->array_subscript
>= actual_array_size
) {
762 linker_error(prog
, "Transform feedback varying %s has index "
763 "%i, but the array size is %u.",
764 this->orig_name
, this->array_subscript
,
768 unsigned array_elem_size
= this->lowered_builtin_array_variable
?
769 1 : vector_elements
* matrix_cols
* dmul
;
770 fine_location
+= array_elem_size
* this->array_subscript
;
773 this->size
= actual_array_size
;
775 this->vector_elements
= vector_elements
;
776 this->matrix_columns
= matrix_cols
;
777 if (this->lowered_builtin_array_variable
)
778 this->type
= GL_FLOAT
;
780 this->type
= this->matched_candidate
->type
->fields
.array
->gl_type
;
782 /* Regular variable (scalar, vector, or matrix) */
783 if (this->is_subscripted
) {
784 linker_error(prog
, "Transform feedback varying %s requested, "
785 "but %s is not an array.",
786 this->orig_name
, this->var_name
);
790 this->vector_elements
= this->matched_candidate
->type
->vector_elements
;
791 this->matrix_columns
= this->matched_candidate
->type
->matrix_columns
;
792 this->type
= this->matched_candidate
->type
->gl_type
;
794 this->location
= fine_location
/ 4;
795 this->location_frac
= fine_location
% 4;
797 /* From GL_EXT_transform_feedback:
798 * A program will fail to link if:
800 * * the total number of components to capture in any varying
801 * variable in <varyings> is greater than the constant
802 * MAX_TRANSFORM_FEEDBACK_SEPARATE_COMPONENTS_EXT and the
803 * buffer mode is SEPARATE_ATTRIBS_EXT;
805 if (prog
->TransformFeedback
.BufferMode
== GL_SEPARATE_ATTRIBS
&&
806 this->num_components() >
807 ctx
->Const
.MaxTransformFeedbackSeparateComponents
) {
808 linker_error(prog
, "Transform feedback varying %s exceeds "
809 "MAX_TRANSFORM_FEEDBACK_SEPARATE_COMPONENTS.",
814 /* Only transform feedback varyings can be assigned to non-zero streams,
815 * so assign the stream id here.
817 this->stream_id
= this->matched_candidate
->toplevel_var
->data
.stream
;
819 unsigned array_offset
= this->array_subscript
* 4 * dmul
;
820 unsigned struct_offset
= this->matched_candidate
->offset
* 4 * dmul
;
821 this->buffer
= this->matched_candidate
->toplevel_var
->data
.xfb_buffer
;
822 this->offset
= this->matched_candidate
->toplevel_var
->data
.offset
+
823 array_offset
+ struct_offset
;
830 tfeedback_decl::get_num_outputs() const
832 if (!this->is_varying()) {
835 return (this->num_components() + this->location_frac
+ 3)/4;
840 * Update gl_transform_feedback_info to reflect this tfeedback_decl.
842 * If an error occurs, the error is reported through linker_error() and false
846 tfeedback_decl::store(struct gl_context
*ctx
, struct gl_shader_program
*prog
,
847 struct gl_transform_feedback_info
*info
,
848 unsigned buffer
, unsigned buffer_index
,
849 const unsigned max_outputs
, bool *explicit_stride
,
850 bool has_xfb_qualifiers
) const
852 unsigned xfb_offset
= 0;
853 unsigned size
= this->size
;
854 /* Handle gl_SkipComponents. */
855 if (this->skip_components
) {
856 info
->Buffers
[buffer
].Stride
+= this->skip_components
;
857 size
= this->skip_components
;
861 if (this->next_buffer_separator
) {
866 if (has_xfb_qualifiers
) {
867 xfb_offset
= this->offset
/ 4;
869 xfb_offset
= info
->Buffers
[buffer
].Stride
;
871 info
->Varyings
[info
->NumVarying
].Offset
= xfb_offset
* 4;
874 unsigned location
= this->location
;
875 unsigned location_frac
= this->location_frac
;
876 unsigned num_components
= this->num_components();
877 while (num_components
> 0) {
878 unsigned output_size
= MIN2(num_components
, 4 - location_frac
);
879 assert((info
->NumOutputs
== 0 && max_outputs
== 0) ||
880 info
->NumOutputs
< max_outputs
);
882 /* From the ARB_enhanced_layouts spec:
884 * "If such a block member or variable is not written during a shader
885 * invocation, the buffer contents at the assigned offset will be
886 * undefined. Even if there are no static writes to a variable or
887 * member that is assigned a transform feedback offset, the space is
888 * still allocated in the buffer and still affects the stride."
890 if (this->is_varying_written()) {
891 info
->Outputs
[info
->NumOutputs
].ComponentOffset
= location_frac
;
892 info
->Outputs
[info
->NumOutputs
].OutputRegister
= location
;
893 info
->Outputs
[info
->NumOutputs
].NumComponents
= output_size
;
894 info
->Outputs
[info
->NumOutputs
].StreamId
= stream_id
;
895 info
->Outputs
[info
->NumOutputs
].OutputBuffer
= buffer
;
896 info
->Outputs
[info
->NumOutputs
].DstOffset
= xfb_offset
;
899 info
->Buffers
[buffer
].Stream
= this->stream_id
;
900 xfb_offset
+= output_size
;
902 num_components
-= output_size
;
908 if (explicit_stride
&& explicit_stride
[buffer
]) {
909 if (this->is_64bit() && info
->Buffers
[buffer
].Stride
% 2) {
910 linker_error(prog
, "invalid qualifier xfb_stride=%d must be a "
911 "multiple of 8 as its applied to a type that is or "
912 "contains a double.",
913 info
->Buffers
[buffer
].Stride
* 4);
917 if ((this->offset
/ 4) / info
->Buffers
[buffer
].Stride
!=
918 (xfb_offset
- 1) / info
->Buffers
[buffer
].Stride
) {
919 linker_error(prog
, "xfb_offset (%d) overflows xfb_stride (%d) for "
920 "buffer (%d)", xfb_offset
* 4,
921 info
->Buffers
[buffer
].Stride
* 4, buffer
);
925 info
->Buffers
[buffer
].Stride
= xfb_offset
;
928 /* From GL_EXT_transform_feedback:
929 * A program will fail to link if:
931 * * the total number of components to capture is greater than
932 * the constant MAX_TRANSFORM_FEEDBACK_INTERLEAVED_COMPONENTS_EXT
933 * and the buffer mode is INTERLEAVED_ATTRIBS_EXT.
935 * From GL_ARB_enhanced_layouts:
937 * "The resulting stride (implicit or explicit) must be less than or
938 * equal to the implementation-dependent constant
939 * gl_MaxTransformFeedbackInterleavedComponents."
941 if ((prog
->TransformFeedback
.BufferMode
== GL_INTERLEAVED_ATTRIBS
||
942 has_xfb_qualifiers
) &&
943 info
->Buffers
[buffer
].Stride
>
944 ctx
->Const
.MaxTransformFeedbackInterleavedComponents
) {
945 linker_error(prog
, "The MAX_TRANSFORM_FEEDBACK_INTERLEAVED_COMPONENTS "
946 "limit has been exceeded.");
951 info
->Varyings
[info
->NumVarying
].Name
= ralloc_strdup(prog
,
953 info
->Varyings
[info
->NumVarying
].Type
= this->type
;
954 info
->Varyings
[info
->NumVarying
].Size
= size
;
955 info
->Varyings
[info
->NumVarying
].BufferIndex
= buffer_index
;
957 info
->Buffers
[buffer
].NumVaryings
++;
963 const tfeedback_candidate
*
964 tfeedback_decl::find_candidate(gl_shader_program
*prog
,
965 hash_table
*tfeedback_candidates
)
967 const char *name
= this->var_name
;
968 switch (this->lowered_builtin_array_variable
) {
970 name
= this->var_name
;
973 name
= "gl_ClipDistanceMESA";
976 name
= "gl_CullDistanceMESA";
978 case tess_level_outer
:
979 name
= "gl_TessLevelOuterMESA";
981 case tess_level_inner
:
982 name
= "gl_TessLevelInnerMESA";
985 hash_entry
*entry
= _mesa_hash_table_search(tfeedback_candidates
, name
);
987 this->matched_candidate
= entry
?
988 (const tfeedback_candidate
*) entry
->data
: NULL
;
990 if (!this->matched_candidate
) {
991 /* From GL_EXT_transform_feedback:
992 * A program will fail to link if:
994 * * any variable name specified in the <varyings> array is not
995 * declared as an output in the geometry shader (if present) or
996 * the vertex shader (if no geometry shader is present);
998 linker_error(prog
, "Transform feedback varying %s undeclared.",
1002 return this->matched_candidate
;
1007 * Parse all the transform feedback declarations that were passed to
1008 * glTransformFeedbackVaryings() and store them in tfeedback_decl objects.
1010 * If an error occurs, the error is reported through linker_error() and false
1014 parse_tfeedback_decls(struct gl_context
*ctx
, struct gl_shader_program
*prog
,
1015 const void *mem_ctx
, unsigned num_names
,
1016 char **varying_names
, tfeedback_decl
*decls
)
1018 for (unsigned i
= 0; i
< num_names
; ++i
) {
1019 decls
[i
].init(ctx
, mem_ctx
, varying_names
[i
]);
1021 if (!decls
[i
].is_varying())
1024 /* From GL_EXT_transform_feedback:
1025 * A program will fail to link if:
1027 * * any two entries in the <varyings> array specify the same varying
1030 * We interpret this to mean "any two entries in the <varyings> array
1031 * specify the same varying variable and array index", since transform
1032 * feedback of arrays would be useless otherwise.
1034 for (unsigned j
= 0; j
< i
; ++j
) {
1035 if (!decls
[j
].is_varying())
1038 if (tfeedback_decl::is_same(decls
[i
], decls
[j
])) {
1039 linker_error(prog
, "Transform feedback varying %s specified "
1040 "more than once.", varying_names
[i
]);
1050 cmp_xfb_offset(const void * x_generic
, const void * y_generic
)
1052 tfeedback_decl
*x
= (tfeedback_decl
*) x_generic
;
1053 tfeedback_decl
*y
= (tfeedback_decl
*) y_generic
;
1055 if (x
->get_buffer() != y
->get_buffer())
1056 return x
->get_buffer() - y
->get_buffer();
1057 return x
->get_offset() - y
->get_offset();
1061 * Store transform feedback location assignments into
1062 * prog->sh.LinkedTransformFeedback based on the data stored in
1065 * If an error occurs, the error is reported through linker_error() and false
1069 store_tfeedback_info(struct gl_context
*ctx
, struct gl_shader_program
*prog
,
1070 unsigned num_tfeedback_decls
,
1071 tfeedback_decl
*tfeedback_decls
, bool has_xfb_qualifiers
)
1073 if (!prog
->last_vert_prog
)
1076 /* Make sure MaxTransformFeedbackBuffers is less than 32 so the bitmask for
1077 * tracking the number of buffers doesn't overflow.
1079 assert(ctx
->Const
.MaxTransformFeedbackBuffers
< 32);
1081 bool separate_attribs_mode
=
1082 prog
->TransformFeedback
.BufferMode
== GL_SEPARATE_ATTRIBS
;
1084 struct gl_program
*xfb_prog
= prog
->last_vert_prog
;
1085 xfb_prog
->sh
.LinkedTransformFeedback
=
1086 rzalloc(xfb_prog
, struct gl_transform_feedback_info
);
1088 /* The xfb_offset qualifier does not have to be used in increasing order
1089 * however some drivers expect to receive the list of transform feedback
1090 * declarations in order so sort it now for convenience.
1092 if (has_xfb_qualifiers
)
1093 qsort(tfeedback_decls
, num_tfeedback_decls
, sizeof(*tfeedback_decls
),
1096 xfb_prog
->sh
.LinkedTransformFeedback
->Varyings
=
1097 rzalloc_array(xfb_prog
, struct gl_transform_feedback_varying_info
,
1098 num_tfeedback_decls
);
1100 unsigned num_outputs
= 0;
1101 for (unsigned i
= 0; i
< num_tfeedback_decls
; ++i
) {
1102 if (tfeedback_decls
[i
].is_varying_written())
1103 num_outputs
+= tfeedback_decls
[i
].get_num_outputs();
1106 xfb_prog
->sh
.LinkedTransformFeedback
->Outputs
=
1107 rzalloc_array(xfb_prog
, struct gl_transform_feedback_output
,
1110 unsigned num_buffers
= 0;
1111 unsigned buffers
= 0;
1113 if (!has_xfb_qualifiers
&& separate_attribs_mode
) {
1114 /* GL_SEPARATE_ATTRIBS */
1115 for (unsigned i
= 0; i
< num_tfeedback_decls
; ++i
) {
1116 if (!tfeedback_decls
[i
].store(ctx
, prog
,
1117 xfb_prog
->sh
.LinkedTransformFeedback
,
1118 num_buffers
, num_buffers
, num_outputs
,
1119 NULL
, has_xfb_qualifiers
))
1122 buffers
|= 1 << num_buffers
;
1127 /* GL_INVERLEAVED_ATTRIBS */
1128 int buffer_stream_id
= -1;
1130 num_tfeedback_decls
? tfeedback_decls
[0].get_buffer() : 0;
1131 bool explicit_stride
[MAX_FEEDBACK_BUFFERS
] = { false };
1133 /* Apply any xfb_stride global qualifiers */
1134 if (has_xfb_qualifiers
) {
1135 for (unsigned j
= 0; j
< MAX_FEEDBACK_BUFFERS
; j
++) {
1136 if (prog
->TransformFeedback
.BufferStride
[j
]) {
1138 explicit_stride
[j
] = true;
1139 xfb_prog
->sh
.LinkedTransformFeedback
->Buffers
[j
].Stride
=
1140 prog
->TransformFeedback
.BufferStride
[j
] / 4;
1145 for (unsigned i
= 0; i
< num_tfeedback_decls
; ++i
) {
1146 if (has_xfb_qualifiers
&&
1147 buffer
!= tfeedback_decls
[i
].get_buffer()) {
1148 /* we have moved to the next buffer so reset stream id */
1149 buffer_stream_id
= -1;
1153 if (tfeedback_decls
[i
].is_next_buffer_separator()) {
1154 if (!tfeedback_decls
[i
].store(ctx
, prog
,
1155 xfb_prog
->sh
.LinkedTransformFeedback
,
1156 buffer
, num_buffers
, num_outputs
,
1157 explicit_stride
, has_xfb_qualifiers
))
1160 buffer_stream_id
= -1;
1162 } else if (tfeedback_decls
[i
].is_varying()) {
1163 if (buffer_stream_id
== -1) {
1164 /* First varying writing to this buffer: remember its stream */
1165 buffer_stream_id
= (int) tfeedback_decls
[i
].get_stream_id();
1166 } else if (buffer_stream_id
!=
1167 (int) tfeedback_decls
[i
].get_stream_id()) {
1168 /* Varying writes to the same buffer from a different stream */
1170 "Transform feedback can't capture varyings belonging "
1171 "to different vertex streams in a single buffer. "
1172 "Varying %s writes to buffer from stream %u, other "
1173 "varyings in the same buffer write from stream %u.",
1174 tfeedback_decls
[i
].name(),
1175 tfeedback_decls
[i
].get_stream_id(),
1181 if (has_xfb_qualifiers
) {
1182 buffer
= tfeedback_decls
[i
].get_buffer();
1184 buffer
= num_buffers
;
1186 buffers
|= 1 << buffer
;
1188 if (!tfeedback_decls
[i
].store(ctx
, prog
,
1189 xfb_prog
->sh
.LinkedTransformFeedback
,
1190 buffer
, num_buffers
, num_outputs
,
1191 explicit_stride
, has_xfb_qualifiers
))
1196 assert(xfb_prog
->sh
.LinkedTransformFeedback
->NumOutputs
== num_outputs
);
1198 xfb_prog
->sh
.LinkedTransformFeedback
->ActiveBuffers
= buffers
;
1205 * Data structure recording the relationship between outputs of one shader
1206 * stage (the "producer") and inputs of another (the "consumer").
1208 class varying_matches
1211 varying_matches(bool disable_varying_packing
, bool xfb_enabled
,
1212 bool enhanced_layouts_enabled
,
1213 gl_shader_stage producer_stage
,
1214 gl_shader_stage consumer_stage
);
1216 void record(ir_variable
*producer_var
, ir_variable
*consumer_var
);
1217 unsigned assign_locations(struct gl_shader_program
*prog
,
1218 uint8_t *components
,
1219 uint64_t reserved_slots
);
1220 void store_locations() const;
1223 bool is_varying_packing_safe(const glsl_type
*type
,
1224 const ir_variable
*var
);
1227 * If true, this driver disables varying packing, so all varyings need to
1228 * be aligned on slot boundaries, and take up a number of slots equal to
1229 * their number of matrix columns times their array size.
1231 * Packing may also be disabled because our current packing method is not
1232 * safe in SSO or versions of OpenGL where interpolation qualifiers are not
1233 * guaranteed to match across stages.
1235 const bool disable_varying_packing
;
1238 * If true, this driver has transform feedback enabled. The transform
1239 * feedback code requires at least some packing be done even when varying
1240 * packing is disabled, fortunately where transform feedback requires
1241 * packing it's safe to override the disabled setting. See
1242 * is_varying_packing_safe().
1244 const bool xfb_enabled
;
1246 const bool enhanced_layouts_enabled
;
1249 * Enum representing the order in which varyings are packed within a
1252 * Currently we pack vec4's first, then vec2's, then scalar values, then
1253 * vec3's. This order ensures that the only vectors that are at risk of
1254 * having to be "double parked" (split between two adjacent varying slots)
1257 enum packing_order_enum
{
1260 PACKING_ORDER_SCALAR
,
1264 static unsigned compute_packing_class(const ir_variable
*var
);
1265 static packing_order_enum
compute_packing_order(const ir_variable
*var
);
1266 static int match_comparator(const void *x_generic
, const void *y_generic
);
1267 static int xfb_comparator(const void *x_generic
, const void *y_generic
);
1270 * Structure recording the relationship between a single producer output
1271 * and a single consumer input.
1275 * Packing class for this varying, computed by compute_packing_class().
1277 unsigned packing_class
;
1280 * Packing order for this varying, computed by compute_packing_order().
1282 packing_order_enum packing_order
;
1283 unsigned num_components
;
1286 * The output variable in the producer stage.
1288 ir_variable
*producer_var
;
1291 * The input variable in the consumer stage.
1293 ir_variable
*consumer_var
;
1296 * The location which has been assigned for this varying. This is
1297 * expressed in multiples of a float, with the first generic varying
1298 * (i.e. the one referred to by VARYING_SLOT_VAR0) represented by the
1301 unsigned generic_location
;
1305 * The number of elements in the \c matches array that are currently in
1308 unsigned num_matches
;
1311 * The number of elements that were set aside for the \c matches array when
1314 unsigned matches_capacity
;
1316 gl_shader_stage producer_stage
;
1317 gl_shader_stage consumer_stage
;
1320 } /* anonymous namespace */
1322 varying_matches::varying_matches(bool disable_varying_packing
,
1324 bool enhanced_layouts_enabled
,
1325 gl_shader_stage producer_stage
,
1326 gl_shader_stage consumer_stage
)
1327 : disable_varying_packing(disable_varying_packing
),
1328 xfb_enabled(xfb_enabled
),
1329 enhanced_layouts_enabled(enhanced_layouts_enabled
),
1330 producer_stage(producer_stage
),
1331 consumer_stage(consumer_stage
)
1333 /* Note: this initial capacity is rather arbitrarily chosen to be large
1334 * enough for many cases without wasting an unreasonable amount of space.
1335 * varying_matches::record() will resize the array if there are more than
1336 * this number of varyings.
1338 this->matches_capacity
= 8;
1339 this->matches
= (match
*)
1340 malloc(sizeof(*this->matches
) * this->matches_capacity
);
1341 this->num_matches
= 0;
1345 varying_matches::~varying_matches()
1347 free(this->matches
);
1352 * Packing is always safe on individual arrays, structures, and matrices. It
1353 * is also safe if the varying is only used for transform feedback.
1356 varying_matches::is_varying_packing_safe(const glsl_type
*type
,
1357 const ir_variable
*var
)
1359 if (consumer_stage
== MESA_SHADER_TESS_EVAL
||
1360 consumer_stage
== MESA_SHADER_TESS_CTRL
||
1361 producer_stage
== MESA_SHADER_TESS_CTRL
)
1364 return xfb_enabled
&& (type
->is_array() || type
->is_record() ||
1365 type
->is_matrix() || var
->data
.is_xfb_only
);
1370 * Record the given producer/consumer variable pair in the list of variables
1371 * that should later be assigned locations.
1373 * It is permissible for \c consumer_var to be NULL (this happens if a
1374 * variable is output by the producer and consumed by transform feedback, but
1375 * not consumed by the consumer).
1377 * If \c producer_var has already been paired up with a consumer_var, or
1378 * producer_var is part of fixed pipeline functionality (and hence already has
1379 * a location assigned), this function has no effect.
1381 * Note: as a side effect this function may change the interpolation type of
1382 * \c producer_var, but only when the change couldn't possibly affect
1386 varying_matches::record(ir_variable
*producer_var
, ir_variable
*consumer_var
)
1388 assert(producer_var
!= NULL
|| consumer_var
!= NULL
);
1390 if ((producer_var
&& (!producer_var
->data
.is_unmatched_generic_inout
||
1391 producer_var
->data
.explicit_location
)) ||
1392 (consumer_var
&& (!consumer_var
->data
.is_unmatched_generic_inout
||
1393 consumer_var
->data
.explicit_location
))) {
1394 /* Either a location already exists for this variable (since it is part
1395 * of fixed functionality), or it has already been recorded as part of a
1401 bool needs_flat_qualifier
= consumer_var
== NULL
&&
1402 (producer_var
->type
->contains_integer() ||
1403 producer_var
->type
->contains_double());
1405 if (!disable_varying_packing
&&
1406 (needs_flat_qualifier
||
1407 (consumer_stage
!= MESA_SHADER_NONE
&& consumer_stage
!= MESA_SHADER_FRAGMENT
))) {
1408 /* Since this varying is not being consumed by the fragment shader, its
1409 * interpolation type varying cannot possibly affect rendering.
1410 * Also, this variable is non-flat and is (or contains) an integer
1412 * If the consumer stage is unknown, don't modify the interpolation
1413 * type as it could affect rendering later with separate shaders.
1415 * lower_packed_varyings requires all integer varyings to flat,
1416 * regardless of where they appear. We can trivially satisfy that
1417 * requirement by changing the interpolation type to flat here.
1420 producer_var
->data
.centroid
= false;
1421 producer_var
->data
.sample
= false;
1422 producer_var
->data
.interpolation
= INTERP_MODE_FLAT
;
1426 consumer_var
->data
.centroid
= false;
1427 consumer_var
->data
.sample
= false;
1428 consumer_var
->data
.interpolation
= INTERP_MODE_FLAT
;
1432 if (this->num_matches
== this->matches_capacity
) {
1433 this->matches_capacity
*= 2;
1434 this->matches
= (match
*)
1435 realloc(this->matches
,
1436 sizeof(*this->matches
) * this->matches_capacity
);
1439 /* We must use the consumer to compute the packing class because in GL4.4+
1440 * there is no guarantee interpolation qualifiers will match across stages.
1442 * From Section 4.5 (Interpolation Qualifiers) of the GLSL 4.30 spec:
1444 * "The type and presence of interpolation qualifiers of variables with
1445 * the same name declared in all linked shaders for the same cross-stage
1446 * interface must match, otherwise the link command will fail.
1448 * When comparing an output from one stage to an input of a subsequent
1449 * stage, the input and output don't match if their interpolation
1450 * qualifiers (or lack thereof) are not the same."
1452 * This text was also in at least revison 7 of the 4.40 spec but is no
1453 * longer in revision 9 and not in the 4.50 spec.
1455 const ir_variable
*const var
= (consumer_var
!= NULL
)
1456 ? consumer_var
: producer_var
;
1457 const gl_shader_stage stage
= (consumer_var
!= NULL
)
1458 ? consumer_stage
: producer_stage
;
1459 const glsl_type
*type
= get_varying_type(var
, stage
);
1461 if (producer_var
&& consumer_var
&&
1462 consumer_var
->data
.must_be_shader_input
) {
1463 producer_var
->data
.must_be_shader_input
= 1;
1466 this->matches
[this->num_matches
].packing_class
1467 = this->compute_packing_class(var
);
1468 this->matches
[this->num_matches
].packing_order
1469 = this->compute_packing_order(var
);
1470 if ((this->disable_varying_packing
&& !is_varying_packing_safe(type
, var
)) ||
1471 var
->data
.must_be_shader_input
) {
1472 unsigned slots
= type
->count_attribute_slots(false);
1473 this->matches
[this->num_matches
].num_components
= slots
* 4;
1475 this->matches
[this->num_matches
].num_components
1476 = type
->component_slots();
1479 this->matches
[this->num_matches
].producer_var
= producer_var
;
1480 this->matches
[this->num_matches
].consumer_var
= consumer_var
;
1481 this->num_matches
++;
1483 producer_var
->data
.is_unmatched_generic_inout
= 0;
1485 consumer_var
->data
.is_unmatched_generic_inout
= 0;
1490 * Choose locations for all of the variable matches that were previously
1491 * passed to varying_matches::record().
1494 varying_matches::assign_locations(struct gl_shader_program
*prog
,
1495 uint8_t *components
,
1496 uint64_t reserved_slots
)
1498 /* If packing has been disabled then we cannot safely sort the varyings by
1499 * class as it may mean we are using a version of OpenGL where
1500 * interpolation qualifiers are not guaranteed to be matching across
1501 * shaders, sorting in this case could result in mismatching shader
1503 * When packing is disabled the sort orders varyings used by transform
1504 * feedback first, but also depends on *undefined behaviour* of qsort to
1505 * reverse the order of the varyings. See: xfb_comparator().
1507 if (!this->disable_varying_packing
) {
1508 /* Sort varying matches into an order that makes them easy to pack. */
1509 qsort(this->matches
, this->num_matches
, sizeof(*this->matches
),
1510 &varying_matches::match_comparator
);
1512 /* Only sort varyings that are only used by transform feedback. */
1513 qsort(this->matches
, this->num_matches
, sizeof(*this->matches
),
1514 &varying_matches::xfb_comparator
);
1517 unsigned generic_location
= 0;
1518 unsigned generic_patch_location
= MAX_VARYING
*4;
1519 bool previous_var_xfb_only
= false;
1521 for (unsigned i
= 0; i
< this->num_matches
; i
++) {
1522 unsigned *location
= &generic_location
;
1524 const ir_variable
*var
;
1525 const glsl_type
*type
;
1526 bool is_vertex_input
= false;
1527 if (matches
[i
].consumer_var
) {
1528 var
= matches
[i
].consumer_var
;
1529 type
= get_varying_type(var
, consumer_stage
);
1530 if (consumer_stage
== MESA_SHADER_VERTEX
)
1531 is_vertex_input
= true;
1533 var
= matches
[i
].producer_var
;
1534 type
= get_varying_type(var
, producer_stage
);
1537 if (var
->data
.patch
)
1538 location
= &generic_patch_location
;
1540 /* Advance to the next slot if this varying has a different packing
1541 * class than the previous one, and we're not already on a slot
1544 * Also advance to the next slot if packing is disabled. This makes sure
1545 * we don't assign varyings the same locations which is possible
1546 * because we still pack individual arrays, records and matrices even
1547 * when packing is disabled. Note we don't advance to the next slot if
1548 * we can pack varyings together that are only used for transform
1551 if (var
->data
.must_be_shader_input
||
1552 (this->disable_varying_packing
&&
1553 !(previous_var_xfb_only
&& var
->data
.is_xfb_only
)) ||
1554 (i
> 0 && this->matches
[i
- 1].packing_class
1555 != this->matches
[i
].packing_class
)) {
1556 *location
= ALIGN(*location
, 4);
1559 previous_var_xfb_only
= var
->data
.is_xfb_only
;
1561 /* The number of components taken up by this variable. For vertex shader
1562 * inputs, we use the number of slots * 4, as they have different
1565 unsigned num_components
= is_vertex_input
?
1566 type
->count_attribute_slots(is_vertex_input
) * 4 :
1567 this->matches
[i
].num_components
;
1569 /* The last slot for this variable, inclusive. */
1570 unsigned slot_end
= *location
+ num_components
- 1;
1572 /* FIXME: We could be smarter in the below code and loop back over
1573 * trying to fill any locations that we skipped because we couldn't pack
1574 * the varying between an explicit location. For now just let the user
1575 * hit the linking error if we run out of room and suggest they use
1576 * explicit locations.
1578 while (slot_end
< MAX_VARYING
* 4u) {
1579 const unsigned slots
= (slot_end
/ 4u) - (*location
/ 4u) + 1;
1580 const uint64_t slot_mask
= ((1ull << slots
) - 1) << (*location
/ 4u);
1583 if (reserved_slots
& slot_mask
) {
1584 *location
= ALIGN(*location
+ 1, 4);
1585 slot_end
= *location
+ num_components
- 1;
1592 if (!var
->data
.patch
&& slot_end
>= MAX_VARYING
* 4u) {
1593 linker_error(prog
, "insufficient contiguous locations available for "
1594 "%s it is possible an array or struct could not be "
1595 "packed between varyings with explicit locations. Try "
1596 "using an explicit location for arrays and structs.",
1600 if (slot_end
< MAX_VARYINGS_INCL_PATCH
* 4u) {
1601 for (unsigned j
= *location
/ 4u; j
< slot_end
/ 4u; j
++)
1603 components
[slot_end
/ 4u] = (slot_end
& 3) + 1;
1606 this->matches
[i
].generic_location
= *location
;
1608 *location
= slot_end
+ 1;
1611 return (generic_location
+ 3) / 4;
1616 * Update the producer and consumer shaders to reflect the locations
1617 * assignments that were made by varying_matches::assign_locations().
1620 varying_matches::store_locations() const
1622 /* Check is location needs to be packed with lower_packed_varyings() or if
1623 * we can just use ARB_enhanced_layouts packing.
1625 bool pack_loc
[MAX_VARYINGS_INCL_PATCH
] = { 0 };
1626 const glsl_type
*loc_type
[MAX_VARYINGS_INCL_PATCH
][4] = { {NULL
, NULL
} };
1628 for (unsigned i
= 0; i
< this->num_matches
; i
++) {
1629 ir_variable
*producer_var
= this->matches
[i
].producer_var
;
1630 ir_variable
*consumer_var
= this->matches
[i
].consumer_var
;
1631 unsigned generic_location
= this->matches
[i
].generic_location
;
1632 unsigned slot
= generic_location
/ 4;
1633 unsigned offset
= generic_location
% 4;
1636 producer_var
->data
.location
= VARYING_SLOT_VAR0
+ slot
;
1637 producer_var
->data
.location_frac
= offset
;
1641 assert(consumer_var
->data
.location
== -1);
1642 consumer_var
->data
.location
= VARYING_SLOT_VAR0
+ slot
;
1643 consumer_var
->data
.location_frac
= offset
;
1646 /* Find locations suitable for native packing via
1647 * ARB_enhanced_layouts.
1649 if (producer_var
&& consumer_var
) {
1650 if (enhanced_layouts_enabled
) {
1651 const glsl_type
*type
=
1652 get_varying_type(producer_var
, producer_stage
);
1653 if (type
->is_array() || type
->is_matrix() || type
->is_record() ||
1654 type
->is_double()) {
1655 unsigned comp_slots
= type
->component_slots() + offset
;
1656 unsigned slots
= comp_slots
/ 4;
1660 for (unsigned j
= 0; j
< slots
; j
++) {
1661 pack_loc
[slot
+ j
] = true;
1663 } else if (offset
+ type
->vector_elements
> 4) {
1664 pack_loc
[slot
] = true;
1665 pack_loc
[slot
+ 1] = true;
1667 loc_type
[slot
][offset
] = type
;
1673 /* Attempt to use ARB_enhanced_layouts for more efficient packing if
1676 if (enhanced_layouts_enabled
) {
1677 for (unsigned i
= 0; i
< this->num_matches
; i
++) {
1678 ir_variable
*producer_var
= this->matches
[i
].producer_var
;
1679 ir_variable
*consumer_var
= this->matches
[i
].consumer_var
;
1680 unsigned generic_location
= this->matches
[i
].generic_location
;
1681 unsigned slot
= generic_location
/ 4;
1683 if (pack_loc
[slot
] || !producer_var
|| !consumer_var
)
1686 const glsl_type
*type
=
1687 get_varying_type(producer_var
, producer_stage
);
1688 bool type_match
= true;
1689 for (unsigned j
= 0; j
< 4; j
++) {
1690 if (loc_type
[slot
][j
]) {
1691 if (type
->base_type
!= loc_type
[slot
][j
]->base_type
)
1697 producer_var
->data
.explicit_location
= 1;
1698 consumer_var
->data
.explicit_location
= 1;
1699 producer_var
->data
.explicit_component
= 1;
1700 consumer_var
->data
.explicit_component
= 1;
1708 * Compute the "packing class" of the given varying. This is an unsigned
1709 * integer with the property that two variables in the same packing class can
1710 * be safely backed into the same vec4.
1713 varying_matches::compute_packing_class(const ir_variable
*var
)
1715 /* Without help from the back-end, there is no way to pack together
1716 * variables with different interpolation types, because
1717 * lower_packed_varyings must choose exactly one interpolation type for
1718 * each packed varying it creates.
1720 * However, we can safely pack together floats, ints, and uints, because:
1722 * - varyings of base type "int" and "uint" must use the "flat"
1723 * interpolation type, which can only occur in GLSL 1.30 and above.
1725 * - On platforms that support GLSL 1.30 and above, lower_packed_varyings
1726 * can store flat floats as ints without losing any information (using
1727 * the ir_unop_bitcast_* opcodes).
1729 * Therefore, the packing class depends only on the interpolation type.
1731 unsigned packing_class
= var
->data
.centroid
| (var
->data
.sample
<< 1) |
1732 (var
->data
.patch
<< 2) |
1733 (var
->data
.must_be_shader_input
<< 3);
1735 packing_class
+= var
->is_interpolation_flat()
1736 ? unsigned(INTERP_MODE_FLAT
) : var
->data
.interpolation
;
1737 return packing_class
;
1742 * Compute the "packing order" of the given varying. This is a sort key we
1743 * use to determine when to attempt to pack the given varying relative to
1744 * other varyings in the same packing class.
1746 varying_matches::packing_order_enum
1747 varying_matches::compute_packing_order(const ir_variable
*var
)
1749 const glsl_type
*element_type
= var
->type
;
1751 while (element_type
->is_array()) {
1752 element_type
= element_type
->fields
.array
;
1755 switch (element_type
->component_slots() % 4) {
1756 case 1: return PACKING_ORDER_SCALAR
;
1757 case 2: return PACKING_ORDER_VEC2
;
1758 case 3: return PACKING_ORDER_VEC3
;
1759 case 0: return PACKING_ORDER_VEC4
;
1761 assert(!"Unexpected value of vector_elements");
1762 return PACKING_ORDER_VEC4
;
1768 * Comparison function passed to qsort() to sort varyings by packing_class and
1769 * then by packing_order.
1772 varying_matches::match_comparator(const void *x_generic
, const void *y_generic
)
1774 const match
*x
= (const match
*) x_generic
;
1775 const match
*y
= (const match
*) y_generic
;
1777 if (x
->packing_class
!= y
->packing_class
)
1778 return x
->packing_class
- y
->packing_class
;
1779 return x
->packing_order
- y
->packing_order
;
1784 * Comparison function passed to qsort() to sort varyings used only by
1785 * transform feedback when packing of other varyings is disabled.
1788 varying_matches::xfb_comparator(const void *x_generic
, const void *y_generic
)
1790 const match
*x
= (const match
*) x_generic
;
1792 if (x
->producer_var
!= NULL
&& x
->producer_var
->data
.is_xfb_only
)
1793 return match_comparator(x_generic
, y_generic
);
1795 /* FIXME: When the comparator returns 0 it means the elements being
1796 * compared are equivalent. However the qsort documentation says:
1798 * "The order of equivalent elements is undefined."
1800 * In practice the sort ends up reversing the order of the varyings which
1801 * means locations are also assigned in this reversed order and happens to
1802 * be what we want. This is also whats happening in
1803 * varying_matches::match_comparator().
1810 * Is the given variable a varying variable to be counted against the
1811 * limit in ctx->Const.MaxVarying?
1812 * This includes variables such as texcoords, colors and generic
1813 * varyings, but excludes variables such as gl_FrontFacing and gl_FragCoord.
1816 var_counts_against_varying_limit(gl_shader_stage stage
, const ir_variable
*var
)
1818 /* Only fragment shaders will take a varying variable as an input */
1819 if (stage
== MESA_SHADER_FRAGMENT
&&
1820 var
->data
.mode
== ir_var_shader_in
) {
1821 switch (var
->data
.location
) {
1822 case VARYING_SLOT_POS
:
1823 case VARYING_SLOT_FACE
:
1824 case VARYING_SLOT_PNTC
:
1835 * Visitor class that generates tfeedback_candidate structs describing all
1836 * possible targets of transform feedback.
1838 * tfeedback_candidate structs are stored in the hash table
1839 * tfeedback_candidates, which is passed to the constructor. This hash table
1840 * maps varying names to instances of the tfeedback_candidate struct.
1842 class tfeedback_candidate_generator
: public program_resource_visitor
1845 tfeedback_candidate_generator(void *mem_ctx
,
1846 hash_table
*tfeedback_candidates
)
1848 tfeedback_candidates(tfeedback_candidates
),
1854 void process(ir_variable
*var
)
1856 /* All named varying interface blocks should be flattened by now */
1857 assert(!var
->is_interface_instance());
1859 this->toplevel_var
= var
;
1860 this->varying_floats
= 0;
1861 program_resource_visitor::process(var
, false);
1865 virtual void visit_field(const glsl_type
*type
, const char *name
,
1866 bool /* row_major */,
1867 const glsl_type
* /* record_type */,
1868 const enum glsl_interface_packing
,
1869 bool /* last_field */)
1871 assert(!type
->without_array()->is_record());
1872 assert(!type
->without_array()->is_interface());
1874 tfeedback_candidate
*candidate
1875 = rzalloc(this->mem_ctx
, tfeedback_candidate
);
1876 candidate
->toplevel_var
= this->toplevel_var
;
1877 candidate
->type
= type
;
1878 candidate
->offset
= this->varying_floats
;
1879 _mesa_hash_table_insert(this->tfeedback_candidates
,
1880 ralloc_strdup(this->mem_ctx
, name
),
1882 this->varying_floats
+= type
->component_slots();
1886 * Memory context used to allocate hash table keys and values.
1888 void * const mem_ctx
;
1891 * Hash table in which tfeedback_candidate objects should be stored.
1893 hash_table
* const tfeedback_candidates
;
1896 * Pointer to the toplevel variable that is being traversed.
1898 ir_variable
*toplevel_var
;
1901 * Total number of varying floats that have been visited so far. This is
1902 * used to determine the offset to each varying within the toplevel
1905 unsigned varying_floats
;
1912 populate_consumer_input_sets(void *mem_ctx
, exec_list
*ir
,
1913 hash_table
*consumer_inputs
,
1914 hash_table
*consumer_interface_inputs
,
1915 ir_variable
*consumer_inputs_with_locations
[VARYING_SLOT_TESS_MAX
])
1917 memset(consumer_inputs_with_locations
,
1919 sizeof(consumer_inputs_with_locations
[0]) * VARYING_SLOT_TESS_MAX
);
1921 foreach_in_list(ir_instruction
, node
, ir
) {
1922 ir_variable
*const input_var
= node
->as_variable();
1924 if (input_var
!= NULL
&& input_var
->data
.mode
== ir_var_shader_in
) {
1925 /* All interface blocks should have been lowered by this point */
1926 assert(!input_var
->type
->is_interface());
1928 if (input_var
->data
.explicit_location
) {
1929 /* assign_varying_locations only cares about finding the
1930 * ir_variable at the start of a contiguous location block.
1932 * - For !producer, consumer_inputs_with_locations isn't used.
1934 * - For !consumer, consumer_inputs_with_locations is empty.
1936 * For consumer && producer, if you were trying to set some
1937 * ir_variable to the middle of a location block on the other side
1938 * of producer/consumer, cross_validate_outputs_to_inputs() should
1939 * be link-erroring due to either type mismatch or location
1940 * overlaps. If the variables do match up, then they've got a
1941 * matching data.location and you only looked at
1942 * consumer_inputs_with_locations[var->data.location], not any
1943 * following entries for the array/structure.
1945 consumer_inputs_with_locations
[input_var
->data
.location
] =
1947 } else if (input_var
->get_interface_type() != NULL
) {
1948 char *const iface_field_name
=
1949 ralloc_asprintf(mem_ctx
, "%s.%s",
1950 input_var
->get_interface_type()->without_array()->name
,
1952 _mesa_hash_table_insert(consumer_interface_inputs
,
1953 iface_field_name
, input_var
);
1955 _mesa_hash_table_insert(consumer_inputs
,
1956 ralloc_strdup(mem_ctx
, input_var
->name
),
1964 * Find a variable from the consumer that "matches" the specified variable
1966 * This function only finds inputs with names that match. There is no
1967 * validation (here) that the types, etc. are compatible.
1970 get_matching_input(void *mem_ctx
,
1971 const ir_variable
*output_var
,
1972 hash_table
*consumer_inputs
,
1973 hash_table
*consumer_interface_inputs
,
1974 ir_variable
*consumer_inputs_with_locations
[VARYING_SLOT_TESS_MAX
])
1976 ir_variable
*input_var
;
1978 if (output_var
->data
.explicit_location
) {
1979 input_var
= consumer_inputs_with_locations
[output_var
->data
.location
];
1980 } else if (output_var
->get_interface_type() != NULL
) {
1981 char *const iface_field_name
=
1982 ralloc_asprintf(mem_ctx
, "%s.%s",
1983 output_var
->get_interface_type()->without_array()->name
,
1985 hash_entry
*entry
= _mesa_hash_table_search(consumer_interface_inputs
, iface_field_name
);
1986 input_var
= entry
? (ir_variable
*) entry
->data
: NULL
;
1988 hash_entry
*entry
= _mesa_hash_table_search(consumer_inputs
, output_var
->name
);
1989 input_var
= entry
? (ir_variable
*) entry
->data
: NULL
;
1992 return (input_var
== NULL
|| input_var
->data
.mode
!= ir_var_shader_in
)
1999 io_variable_cmp(const void *_a
, const void *_b
)
2001 const ir_variable
*const a
= *(const ir_variable
**) _a
;
2002 const ir_variable
*const b
= *(const ir_variable
**) _b
;
2004 if (a
->data
.explicit_location
&& b
->data
.explicit_location
)
2005 return b
->data
.location
- a
->data
.location
;
2007 if (a
->data
.explicit_location
&& !b
->data
.explicit_location
)
2010 if (!a
->data
.explicit_location
&& b
->data
.explicit_location
)
2013 return -strcmp(a
->name
, b
->name
);
2017 * Sort the shader IO variables into canonical order
2020 canonicalize_shader_io(exec_list
*ir
, enum ir_variable_mode io_mode
)
2022 ir_variable
*var_table
[MAX_PROGRAM_OUTPUTS
* 4];
2023 unsigned num_variables
= 0;
2025 foreach_in_list(ir_instruction
, node
, ir
) {
2026 ir_variable
*const var
= node
->as_variable();
2028 if (var
== NULL
|| var
->data
.mode
!= io_mode
)
2031 /* If we have already encountered more I/O variables that could
2032 * successfully link, bail.
2034 if (num_variables
== ARRAY_SIZE(var_table
))
2037 var_table
[num_variables
++] = var
;
2040 if (num_variables
== 0)
2043 /* Sort the list in reverse order (io_variable_cmp handles this). Later
2044 * we're going to push the variables on to the IR list as a stack, so we
2045 * want the last variable (in canonical order) to be first in the list.
2047 qsort(var_table
, num_variables
, sizeof(var_table
[0]), io_variable_cmp
);
2049 /* Remove the variable from it's current location in the IR, and put it at
2052 for (unsigned i
= 0; i
< num_variables
; i
++) {
2053 var_table
[i
]->remove();
2054 ir
->push_head(var_table
[i
]);
2059 * Generate a bitfield map of the explicit locations for shader varyings.
2061 * Note: For Tessellation shaders we are sitting right on the limits of the
2062 * 64 bit map. Per-vertex and per-patch both have separate location domains
2063 * with a max of MAX_VARYING.
2066 reserved_varying_slot(struct gl_linked_shader
*stage
,
2067 ir_variable_mode io_mode
)
2069 assert(io_mode
== ir_var_shader_in
|| io_mode
== ir_var_shader_out
);
2070 /* Avoid an overflow of the returned value */
2071 assert(MAX_VARYINGS_INCL_PATCH
<= 64);
2079 foreach_in_list(ir_instruction
, node
, stage
->ir
) {
2080 ir_variable
*const var
= node
->as_variable();
2082 if (var
== NULL
|| var
->data
.mode
!= io_mode
||
2083 !var
->data
.explicit_location
||
2084 var
->data
.location
< VARYING_SLOT_VAR0
)
2087 var_slot
= var
->data
.location
- VARYING_SLOT_VAR0
;
2089 unsigned num_elements
= get_varying_type(var
, stage
->Stage
)
2090 ->count_attribute_slots(io_mode
== ir_var_shader_in
&&
2091 stage
->Stage
== MESA_SHADER_VERTEX
);
2092 for (unsigned i
= 0; i
< num_elements
; i
++) {
2093 if (var_slot
>= 0 && var_slot
< MAX_VARYINGS_INCL_PATCH
)
2094 slots
|= UINT64_C(1) << var_slot
;
2104 * Assign locations for all variables that are produced in one pipeline stage
2105 * (the "producer") and consumed in the next stage (the "consumer").
2107 * Variables produced by the producer may also be consumed by transform
2110 * \param num_tfeedback_decls is the number of declarations indicating
2111 * variables that may be consumed by transform feedback.
2113 * \param tfeedback_decls is a pointer to an array of tfeedback_decl objects
2114 * representing the result of parsing the strings passed to
2115 * glTransformFeedbackVaryings(). assign_location() will be called for
2116 * each of these objects that matches one of the outputs of the
2119 * When num_tfeedback_decls is nonzero, it is permissible for the consumer to
2120 * be NULL. In this case, varying locations are assigned solely based on the
2121 * requirements of transform feedback.
2124 assign_varying_locations(struct gl_context
*ctx
,
2126 struct gl_shader_program
*prog
,
2127 gl_linked_shader
*producer
,
2128 gl_linked_shader
*consumer
,
2129 unsigned num_tfeedback_decls
,
2130 tfeedback_decl
*tfeedback_decls
,
2131 const uint64_t reserved_slots
)
2133 /* Tessellation shaders treat inputs and outputs as shared memory and can
2134 * access inputs and outputs of other invocations.
2135 * Therefore, they can't be lowered to temps easily (and definitely not
2138 bool unpackable_tess
=
2139 (consumer
&& consumer
->Stage
== MESA_SHADER_TESS_EVAL
) ||
2140 (consumer
&& consumer
->Stage
== MESA_SHADER_TESS_CTRL
) ||
2141 (producer
&& producer
->Stage
== MESA_SHADER_TESS_CTRL
);
2143 /* Transform feedback code assumes varying arrays are packed, so if the
2144 * driver has disabled varying packing, make sure to at least enable
2145 * packing required by transform feedback.
2148 ctx
->Extensions
.EXT_transform_feedback
&& !unpackable_tess
;
2150 /* Disable packing on outward facing interfaces for SSO because in ES we
2151 * need to retain the unpacked varying information for draw time
2154 * Packing is still enabled on individual arrays, structs, and matrices as
2155 * these are required by the transform feedback code and it is still safe
2156 * to do so. We also enable packing when a varying is only used for
2157 * transform feedback and its not a SSO.
2159 bool disable_varying_packing
=
2160 ctx
->Const
.DisableVaryingPacking
|| unpackable_tess
;
2161 if (prog
->SeparateShader
&& (producer
== NULL
|| consumer
== NULL
))
2162 disable_varying_packing
= true;
2164 varying_matches
matches(disable_varying_packing
, xfb_enabled
,
2165 ctx
->Extensions
.ARB_enhanced_layouts
,
2166 producer
? producer
->Stage
: MESA_SHADER_NONE
,
2167 consumer
? consumer
->Stage
: MESA_SHADER_NONE
);
2168 hash_table
*tfeedback_candidates
=
2169 _mesa_hash_table_create(NULL
, _mesa_key_hash_string
,
2170 _mesa_key_string_equal
);
2171 hash_table
*consumer_inputs
=
2172 _mesa_hash_table_create(NULL
, _mesa_key_hash_string
,
2173 _mesa_key_string_equal
);
2174 hash_table
*consumer_interface_inputs
=
2175 _mesa_hash_table_create(NULL
, _mesa_key_hash_string
,
2176 _mesa_key_string_equal
);
2177 ir_variable
*consumer_inputs_with_locations
[VARYING_SLOT_TESS_MAX
] = {
2181 unsigned consumer_vertices
= 0;
2182 if (consumer
&& consumer
->Stage
== MESA_SHADER_GEOMETRY
)
2183 consumer_vertices
= prog
->Geom
.VerticesIn
;
2185 /* Operate in a total of four passes.
2187 * 1. Sort inputs / outputs into a canonical order. This is necessary so
2188 * that inputs / outputs of separable shaders will be assigned
2189 * predictable locations regardless of the order in which declarations
2190 * appeared in the shader source.
2192 * 2. Assign locations for any matching inputs and outputs.
2194 * 3. Mark output variables in the producer that do not have locations as
2195 * not being outputs. This lets the optimizer eliminate them.
2197 * 4. Mark input variables in the consumer that do not have locations as
2198 * not being inputs. This lets the optimizer eliminate them.
2201 canonicalize_shader_io(consumer
->ir
, ir_var_shader_in
);
2204 canonicalize_shader_io(producer
->ir
, ir_var_shader_out
);
2207 linker::populate_consumer_input_sets(mem_ctx
, consumer
->ir
,
2209 consumer_interface_inputs
,
2210 consumer_inputs_with_locations
);
2213 foreach_in_list(ir_instruction
, node
, producer
->ir
) {
2214 ir_variable
*const output_var
= node
->as_variable();
2216 if (output_var
== NULL
|| output_var
->data
.mode
!= ir_var_shader_out
)
2219 /* Only geometry shaders can use non-zero streams */
2220 assert(output_var
->data
.stream
== 0 ||
2221 (output_var
->data
.stream
< MAX_VERTEX_STREAMS
&&
2222 producer
->Stage
== MESA_SHADER_GEOMETRY
));
2224 if (num_tfeedback_decls
> 0) {
2225 tfeedback_candidate_generator
g(mem_ctx
, tfeedback_candidates
);
2226 g
.process(output_var
);
2229 ir_variable
*const input_var
=
2230 linker::get_matching_input(mem_ctx
, output_var
, consumer_inputs
,
2231 consumer_interface_inputs
,
2232 consumer_inputs_with_locations
);
2234 /* If a matching input variable was found, add this output (and the
2235 * input) to the set. If this is a separable program and there is no
2236 * consumer stage, add the output.
2238 * Always add TCS outputs. They are shared by all invocations
2239 * within a patch and can be used as shared memory.
2241 if (input_var
|| (prog
->SeparateShader
&& consumer
== NULL
) ||
2242 producer
->Stage
== MESA_SHADER_TESS_CTRL
) {
2243 matches
.record(output_var
, input_var
);
2246 /* Only stream 0 outputs can be consumed in the next stage */
2247 if (input_var
&& output_var
->data
.stream
!= 0) {
2248 linker_error(prog
, "output %s is assigned to stream=%d but "
2249 "is linked to an input, which requires stream=0",
2250 output_var
->name
, output_var
->data
.stream
);
2255 /* If there's no producer stage, then this must be a separable program.
2256 * For example, we may have a program that has just a fragment shader.
2257 * Later this program will be used with some arbitrary vertex (or
2258 * geometry) shader program. This means that locations must be assigned
2259 * for all the inputs.
2261 foreach_in_list(ir_instruction
, node
, consumer
->ir
) {
2262 ir_variable
*const input_var
= node
->as_variable();
2264 if (input_var
== NULL
|| input_var
->data
.mode
!= ir_var_shader_in
)
2267 matches
.record(NULL
, input_var
);
2271 _mesa_hash_table_destroy(consumer_inputs
, NULL
);
2272 _mesa_hash_table_destroy(consumer_interface_inputs
, NULL
);
2274 for (unsigned i
= 0; i
< num_tfeedback_decls
; ++i
) {
2275 if (!tfeedback_decls
[i
].is_varying())
2278 const tfeedback_candidate
*matched_candidate
2279 = tfeedback_decls
[i
].find_candidate(prog
, tfeedback_candidates
);
2281 if (matched_candidate
== NULL
) {
2282 _mesa_hash_table_destroy(tfeedback_candidates
, NULL
);
2286 /* Mark xfb varyings as always active */
2287 matched_candidate
->toplevel_var
->data
.always_active_io
= 1;
2289 if (matched_candidate
->toplevel_var
->data
.is_unmatched_generic_inout
) {
2290 matched_candidate
->toplevel_var
->data
.is_xfb_only
= 1;
2291 matches
.record(matched_candidate
->toplevel_var
, NULL
);
2295 uint8_t components
[MAX_VARYINGS_INCL_PATCH
] = {0};
2296 const unsigned slots_used
= matches
.assign_locations(
2297 prog
, components
, reserved_slots
);
2298 matches
.store_locations();
2300 for (unsigned i
= 0; i
< num_tfeedback_decls
; ++i
) {
2301 if (!tfeedback_decls
[i
].is_varying())
2304 if (!tfeedback_decls
[i
].assign_location(ctx
, prog
)) {
2305 _mesa_hash_table_destroy(tfeedback_candidates
, NULL
);
2309 _mesa_hash_table_destroy(tfeedback_candidates
, NULL
);
2311 if (consumer
&& producer
) {
2312 foreach_in_list(ir_instruction
, node
, consumer
->ir
) {
2313 ir_variable
*const var
= node
->as_variable();
2315 if (var
&& var
->data
.mode
== ir_var_shader_in
&&
2316 var
->data
.is_unmatched_generic_inout
) {
2317 if (!prog
->IsES
&& prog
->data
->Version
<= 120) {
2318 /* On page 25 (page 31 of the PDF) of the GLSL 1.20 spec:
2320 * Only those varying variables used (i.e. read) in
2321 * the fragment shader executable must be written to
2322 * by the vertex shader executable; declaring
2323 * superfluous varying variables in a vertex shader is
2326 * We interpret this text as meaning that the VS must
2327 * write the variable for the FS to read it. See
2328 * "glsl1-varying read but not written" in piglit.
2330 linker_error(prog
, "%s shader varying %s not written "
2332 _mesa_shader_stage_to_string(consumer
->Stage
),
2334 _mesa_shader_stage_to_string(producer
->Stage
));
2336 linker_warning(prog
, "%s shader varying %s not written "
2338 _mesa_shader_stage_to_string(consumer
->Stage
),
2340 _mesa_shader_stage_to_string(producer
->Stage
));
2345 /* Now that validation is done its safe to remove unused varyings. As
2346 * we have both a producer and consumer its safe to remove unused
2347 * varyings even if the program is a SSO because the stages are being
2348 * linked together i.e. we have a multi-stage SSO.
2350 remove_unused_shader_inputs_and_outputs(false, producer
,
2352 remove_unused_shader_inputs_and_outputs(false, consumer
,
2357 lower_packed_varyings(mem_ctx
, slots_used
, components
, ir_var_shader_out
,
2358 0, producer
, disable_varying_packing
,
2363 lower_packed_varyings(mem_ctx
, slots_used
, components
, ir_var_shader_in
,
2364 consumer_vertices
, consumer
,
2365 disable_varying_packing
, xfb_enabled
);
2372 check_against_output_limit(struct gl_context
*ctx
,
2373 struct gl_shader_program
*prog
,
2374 gl_linked_shader
*producer
,
2375 unsigned num_explicit_locations
)
2377 unsigned output_vectors
= num_explicit_locations
;
2379 foreach_in_list(ir_instruction
, node
, producer
->ir
) {
2380 ir_variable
*const var
= node
->as_variable();
2382 if (var
&& !var
->data
.explicit_location
&&
2383 var
->data
.mode
== ir_var_shader_out
&&
2384 var_counts_against_varying_limit(producer
->Stage
, var
)) {
2385 /* outputs for fragment shader can't be doubles */
2386 output_vectors
+= var
->type
->count_attribute_slots(false);
2390 assert(producer
->Stage
!= MESA_SHADER_FRAGMENT
);
2391 unsigned max_output_components
=
2392 ctx
->Const
.Program
[producer
->Stage
].MaxOutputComponents
;
2394 const unsigned output_components
= output_vectors
* 4;
2395 if (output_components
> max_output_components
) {
2396 if (ctx
->API
== API_OPENGLES2
|| prog
->IsES
)
2397 linker_error(prog
, "%s shader uses too many output vectors "
2399 _mesa_shader_stage_to_string(producer
->Stage
),
2401 max_output_components
/ 4);
2403 linker_error(prog
, "%s shader uses too many output components "
2405 _mesa_shader_stage_to_string(producer
->Stage
),
2407 max_output_components
);
2416 check_against_input_limit(struct gl_context
*ctx
,
2417 struct gl_shader_program
*prog
,
2418 gl_linked_shader
*consumer
,
2419 unsigned num_explicit_locations
)
2421 unsigned input_vectors
= num_explicit_locations
;
2423 foreach_in_list(ir_instruction
, node
, consumer
->ir
) {
2424 ir_variable
*const var
= node
->as_variable();
2426 if (var
&& !var
->data
.explicit_location
&&
2427 var
->data
.mode
== ir_var_shader_in
&&
2428 var_counts_against_varying_limit(consumer
->Stage
, var
)) {
2429 /* vertex inputs aren't varying counted */
2430 input_vectors
+= var
->type
->count_attribute_slots(false);
2434 assert(consumer
->Stage
!= MESA_SHADER_VERTEX
);
2435 unsigned max_input_components
=
2436 ctx
->Const
.Program
[consumer
->Stage
].MaxInputComponents
;
2438 const unsigned input_components
= input_vectors
* 4;
2439 if (input_components
> max_input_components
) {
2440 if (ctx
->API
== API_OPENGLES2
|| prog
->IsES
)
2441 linker_error(prog
, "%s shader uses too many input vectors "
2443 _mesa_shader_stage_to_string(consumer
->Stage
),
2445 max_input_components
/ 4);
2447 linker_error(prog
, "%s shader uses too many input components "
2449 _mesa_shader_stage_to_string(consumer
->Stage
),
2451 max_input_components
);
2460 link_varyings(struct gl_shader_program
*prog
, unsigned first
, unsigned last
,
2461 struct gl_context
*ctx
, void *mem_ctx
)
2463 bool has_xfb_qualifiers
= false;
2464 unsigned num_tfeedback_decls
= 0;
2465 char **varying_names
= NULL
;
2466 tfeedback_decl
*tfeedback_decls
= NULL
;
2468 /* From the ARB_enhanced_layouts spec:
2470 * "If the shader used to record output variables for transform feedback
2471 * varyings uses the "xfb_buffer", "xfb_offset", or "xfb_stride" layout
2472 * qualifiers, the values specified by TransformFeedbackVaryings are
2473 * ignored, and the set of variables captured for transform feedback is
2474 * instead derived from the specified layout qualifiers."
2476 for (int i
= MESA_SHADER_FRAGMENT
- 1; i
>= 0; i
--) {
2477 /* Find last stage before fragment shader */
2478 if (prog
->_LinkedShaders
[i
]) {
2479 has_xfb_qualifiers
=
2480 process_xfb_layout_qualifiers(mem_ctx
, prog
->_LinkedShaders
[i
],
2481 prog
, &num_tfeedback_decls
,
2487 if (!has_xfb_qualifiers
) {
2488 num_tfeedback_decls
= prog
->TransformFeedback
.NumVarying
;
2489 varying_names
= prog
->TransformFeedback
.VaryingNames
;
2492 if (num_tfeedback_decls
!= 0) {
2493 /* From GL_EXT_transform_feedback:
2494 * A program will fail to link if:
2496 * * the <count> specified by TransformFeedbackVaryingsEXT is
2497 * non-zero, but the program object has no vertex or geometry
2500 if (first
>= MESA_SHADER_FRAGMENT
) {
2501 linker_error(prog
, "Transform feedback varyings specified, but "
2502 "no vertex, tessellation, or geometry shader is "
2507 tfeedback_decls
= rzalloc_array(mem_ctx
, tfeedback_decl
,
2508 num_tfeedback_decls
);
2509 if (!parse_tfeedback_decls(ctx
, prog
, mem_ctx
, num_tfeedback_decls
,
2510 varying_names
, tfeedback_decls
))
2514 /* If there is no fragment shader we need to set transform feedback.
2516 * For SSO we also need to assign output locations. We assign them here
2517 * because we need to do it for both single stage programs and multi stage
2520 if (last
< MESA_SHADER_FRAGMENT
&&
2521 (num_tfeedback_decls
!= 0 || prog
->SeparateShader
)) {
2522 const uint64_t reserved_out_slots
=
2523 reserved_varying_slot(prog
->_LinkedShaders
[last
], ir_var_shader_out
);
2524 if (!assign_varying_locations(ctx
, mem_ctx
, prog
,
2525 prog
->_LinkedShaders
[last
], NULL
,
2526 num_tfeedback_decls
, tfeedback_decls
,
2527 reserved_out_slots
))
2531 if (last
<= MESA_SHADER_FRAGMENT
) {
2532 /* Remove unused varyings from the first/last stage unless SSO */
2533 remove_unused_shader_inputs_and_outputs(prog
->SeparateShader
,
2534 prog
->_LinkedShaders
[first
],
2536 remove_unused_shader_inputs_and_outputs(prog
->SeparateShader
,
2537 prog
->_LinkedShaders
[last
],
2540 /* If the program is made up of only a single stage */
2541 if (first
== last
) {
2542 gl_linked_shader
*const sh
= prog
->_LinkedShaders
[last
];
2544 do_dead_builtin_varyings(ctx
, NULL
, sh
, 0, NULL
);
2545 do_dead_builtin_varyings(ctx
, sh
, NULL
, num_tfeedback_decls
,
2548 if (prog
->SeparateShader
) {
2549 const uint64_t reserved_slots
=
2550 reserved_varying_slot(sh
, ir_var_shader_in
);
2552 /* Assign input locations for SSO, output locations are already
2555 if (!assign_varying_locations(ctx
, mem_ctx
, prog
,
2556 NULL
/* producer */,
2558 0 /* num_tfeedback_decls */,
2559 NULL
/* tfeedback_decls */,
2564 /* Linking the stages in the opposite order (from fragment to vertex)
2565 * ensures that inter-shader outputs written to in an earlier stage
2566 * are eliminated if they are (transitively) not used in a later
2570 for (int i
= next
- 1; i
>= 0; i
--) {
2571 if (prog
->_LinkedShaders
[i
] == NULL
&& i
!= 0)
2574 gl_linked_shader
*const sh_i
= prog
->_LinkedShaders
[i
];
2575 gl_linked_shader
*const sh_next
= prog
->_LinkedShaders
[next
];
2577 const uint64_t reserved_out_slots
=
2578 reserved_varying_slot(sh_i
, ir_var_shader_out
);
2579 const uint64_t reserved_in_slots
=
2580 reserved_varying_slot(sh_next
, ir_var_shader_in
);
2582 do_dead_builtin_varyings(ctx
, sh_i
, sh_next
,
2583 next
== MESA_SHADER_FRAGMENT
? num_tfeedback_decls
: 0,
2586 if (!assign_varying_locations(ctx
, mem_ctx
, prog
, sh_i
, sh_next
,
2587 next
== MESA_SHADER_FRAGMENT
? num_tfeedback_decls
: 0,
2589 reserved_out_slots
| reserved_in_slots
))
2592 /* This must be done after all dead varyings are eliminated. */
2594 unsigned slots_used
= _mesa_bitcount_64(reserved_out_slots
);
2595 if (!check_against_output_limit(ctx
, prog
, sh_i
, slots_used
)) {
2600 unsigned slots_used
= _mesa_bitcount_64(reserved_in_slots
);
2601 if (!check_against_input_limit(ctx
, prog
, sh_next
, slots_used
))
2609 if (!store_tfeedback_info(ctx
, prog
, num_tfeedback_decls
, tfeedback_decls
,
2610 has_xfb_qualifiers
))