2 * Copyright © 2012 Intel Corporation
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
21 * DEALINGS IN THE SOFTWARE.
25 * \file link_varyings.cpp
27 * Linker functions related specifically to linking varyings between shader
32 #include "main/mtypes.h"
33 #include "glsl_symbol_table.h"
34 #include "glsl_parser_extras.h"
35 #include "ir_optimization.h"
37 #include "link_varyings.h"
38 #include "main/macros.h"
39 #include "util/hash_table.h"
44 * Get the varying type stripped of the outermost array if we're processing
45 * a stage whose varyings are arrays indexed by a vertex number (such as
46 * geometry shader inputs).
48 static const glsl_type
*
49 get_varying_type(const ir_variable
*var
, gl_shader_stage stage
)
51 const glsl_type
*type
= var
->type
;
53 if (!var
->data
.patch
&&
54 ((var
->data
.mode
== ir_var_shader_out
&&
55 stage
== MESA_SHADER_TESS_CTRL
) ||
56 (var
->data
.mode
== ir_var_shader_in
&&
57 (stage
== MESA_SHADER_TESS_CTRL
|| stage
== MESA_SHADER_TESS_EVAL
||
58 stage
== MESA_SHADER_GEOMETRY
)))) {
59 assert(type
->is_array());
60 type
= type
->fields
.array
;
67 create_xfb_varying_names(void *mem_ctx
, const glsl_type
*t
, char **name
,
68 size_t name_length
, unsigned *count
,
69 const char *ifc_member_name
,
70 const glsl_type
*ifc_member_t
, char ***varying_names
)
72 if (t
->is_interface()) {
73 size_t new_length
= name_length
;
75 assert(ifc_member_name
&& ifc_member_t
);
76 ralloc_asprintf_rewrite_tail(name
, &new_length
, ".%s", ifc_member_name
);
78 create_xfb_varying_names(mem_ctx
, ifc_member_t
, name
, new_length
, count
,
79 NULL
, NULL
, varying_names
);
80 } else if (t
->is_record()) {
81 for (unsigned i
= 0; i
< t
->length
; i
++) {
82 const char *field
= t
->fields
.structure
[i
].name
;
83 size_t new_length
= name_length
;
85 ralloc_asprintf_rewrite_tail(name
, &new_length
, ".%s", field
);
87 create_xfb_varying_names(mem_ctx
, t
->fields
.structure
[i
].type
, name
,
88 new_length
, count
, NULL
, NULL
,
91 } else if (t
->without_array()->is_record() ||
92 t
->without_array()->is_interface() ||
93 (t
->is_array() && t
->fields
.array
->is_array())) {
94 for (unsigned i
= 0; i
< t
->length
; i
++) {
95 size_t new_length
= name_length
;
97 /* Append the subscript to the current variable name */
98 ralloc_asprintf_rewrite_tail(name
, &new_length
, "[%u]", i
);
100 create_xfb_varying_names(mem_ctx
, t
->fields
.array
, name
, new_length
,
101 count
, ifc_member_name
, ifc_member_t
,
105 (*varying_names
)[(*count
)++] = ralloc_strdup(mem_ctx
, *name
);
110 process_xfb_layout_qualifiers(void *mem_ctx
, const gl_linked_shader
*sh
,
111 struct gl_shader_program
*prog
,
112 unsigned *num_tfeedback_decls
,
113 char ***varying_names
)
115 bool has_xfb_qualifiers
= false;
117 /* We still need to enable transform feedback mode even if xfb_stride is
118 * only applied to a global out. Also we don't bother to propagate
119 * xfb_stride to interface block members so this will catch that case also.
121 for (unsigned j
= 0; j
< MAX_FEEDBACK_BUFFERS
; j
++) {
122 if (prog
->TransformFeedback
.BufferStride
[j
]) {
123 has_xfb_qualifiers
= true;
128 foreach_in_list(ir_instruction
, node
, sh
->ir
) {
129 ir_variable
*var
= node
->as_variable();
130 if (!var
|| var
->data
.mode
!= ir_var_shader_out
)
133 /* From the ARB_enhanced_layouts spec:
135 * "Any shader making any static use (after preprocessing) of any of
136 * these *xfb_* qualifiers will cause the shader to be in a
137 * transform feedback capturing mode and hence responsible for
138 * describing the transform feedback setup. This mode will capture
139 * any output selected by *xfb_offset*, directly or indirectly, to
140 * a transform feedback buffer."
142 if (var
->data
.explicit_xfb_buffer
|| var
->data
.explicit_xfb_stride
) {
143 has_xfb_qualifiers
= true;
146 if (var
->data
.explicit_xfb_offset
) {
147 *num_tfeedback_decls
+= var
->type
->varying_count();
148 has_xfb_qualifiers
= true;
152 if (*num_tfeedback_decls
== 0)
153 return has_xfb_qualifiers
;
156 *varying_names
= ralloc_array(mem_ctx
, char *, *num_tfeedback_decls
);
157 foreach_in_list(ir_instruction
, node
, sh
->ir
) {
158 ir_variable
*var
= node
->as_variable();
159 if (!var
|| var
->data
.mode
!= ir_var_shader_out
)
162 if (var
->data
.explicit_xfb_offset
) {
164 const glsl_type
*type
, *member_type
;
166 if (var
->data
.from_named_ifc_block
) {
167 type
= var
->get_interface_type();
168 /* Find the member type before it was altered by lowering */
170 type
->fields
.structure
[type
->field_index(var
->name
)].type
;
171 name
= ralloc_strdup(NULL
, type
->without_array()->name
);
175 name
= ralloc_strdup(NULL
, var
->name
);
177 create_xfb_varying_names(mem_ctx
, type
, &name
, strlen(name
), &i
,
178 var
->name
, member_type
, varying_names
);
183 assert(i
== *num_tfeedback_decls
);
184 return has_xfb_qualifiers
;
188 * Validate the types and qualifiers of an output from one stage against the
189 * matching input to another stage.
192 cross_validate_types_and_qualifiers(struct gl_shader_program
*prog
,
193 const ir_variable
*input
,
194 const ir_variable
*output
,
195 gl_shader_stage consumer_stage
,
196 gl_shader_stage producer_stage
)
198 /* Check that the types match between stages.
200 const glsl_type
*type_to_match
= input
->type
;
202 /* VS -> GS, VS -> TCS, VS -> TES, TES -> GS */
203 const bool extra_array_level
= (producer_stage
== MESA_SHADER_VERTEX
&&
204 consumer_stage
!= MESA_SHADER_FRAGMENT
) ||
205 consumer_stage
== MESA_SHADER_GEOMETRY
;
206 if (extra_array_level
) {
207 assert(type_to_match
->is_array());
208 type_to_match
= type_to_match
->fields
.array
;
211 if (type_to_match
!= output
->type
) {
212 /* There is a bit of a special case for gl_TexCoord. This
213 * built-in is unsized by default. Applications that variable
214 * access it must redeclare it with a size. There is some
215 * language in the GLSL spec that implies the fragment shader
216 * and vertex shader do not have to agree on this size. Other
217 * driver behave this way, and one or two applications seem to
220 * Neither declaration needs to be modified here because the array
221 * sizes are fixed later when update_array_sizes is called.
223 * From page 48 (page 54 of the PDF) of the GLSL 1.10 spec:
225 * "Unlike user-defined varying variables, the built-in
226 * varying variables don't have a strict one-to-one
227 * correspondence between the vertex language and the
228 * fragment language."
230 if (!output
->type
->is_array() || !is_gl_identifier(output
->name
)) {
232 "%s shader output `%s' declared as type `%s', "
233 "but %s shader input declared as type `%s'\n",
234 _mesa_shader_stage_to_string(producer_stage
),
237 _mesa_shader_stage_to_string(consumer_stage
),
243 /* Check that all of the qualifiers match between stages.
246 /* According to the OpenGL and OpenGLES GLSL specs, the centroid qualifier
247 * should match until OpenGL 4.3 and OpenGLES 3.1. The OpenGLES 3.0
248 * conformance test suite does not verify that the qualifiers must match.
249 * The deqp test suite expects the opposite (OpenGLES 3.1) behavior for
250 * OpenGLES 3.0 drivers, so we relax the checking in all cases.
252 if (false /* always skip the centroid check */ &&
253 prog
->data
->Version
< (prog
->IsES
? 310 : 430) &&
254 input
->data
.centroid
!= output
->data
.centroid
) {
256 "%s shader output `%s' %s centroid qualifier, "
257 "but %s shader input %s centroid qualifier\n",
258 _mesa_shader_stage_to_string(producer_stage
),
260 (output
->data
.centroid
) ? "has" : "lacks",
261 _mesa_shader_stage_to_string(consumer_stage
),
262 (input
->data
.centroid
) ? "has" : "lacks");
266 if (input
->data
.sample
!= output
->data
.sample
) {
268 "%s shader output `%s' %s sample qualifier, "
269 "but %s shader input %s sample qualifier\n",
270 _mesa_shader_stage_to_string(producer_stage
),
272 (output
->data
.sample
) ? "has" : "lacks",
273 _mesa_shader_stage_to_string(consumer_stage
),
274 (input
->data
.sample
) ? "has" : "lacks");
278 if (input
->data
.patch
!= output
->data
.patch
) {
280 "%s shader output `%s' %s patch qualifier, "
281 "but %s shader input %s patch qualifier\n",
282 _mesa_shader_stage_to_string(producer_stage
),
284 (output
->data
.patch
) ? "has" : "lacks",
285 _mesa_shader_stage_to_string(consumer_stage
),
286 (input
->data
.patch
) ? "has" : "lacks");
290 /* The GLSL 4.30 and GLSL ES 3.00 specifications say:
292 * "As only outputs need be declared with invariant, an output from
293 * one shader stage will still match an input of a subsequent stage
294 * without the input being declared as invariant."
296 * while GLSL 4.20 says:
298 * "For variables leaving one shader and coming into another shader,
299 * the invariant keyword has to be used in both shaders, or a link
300 * error will result."
302 * and GLSL ES 1.00 section 4.6.4 "Invariance and Linking" says:
304 * "The invariance of varyings that are declared in both the vertex
305 * and fragment shaders must match."
307 if (input
->data
.invariant
!= output
->data
.invariant
&&
308 prog
->data
->Version
< (prog
->IsES
? 300 : 430)) {
310 "%s shader output `%s' %s invariant qualifier, "
311 "but %s shader input %s invariant qualifier\n",
312 _mesa_shader_stage_to_string(producer_stage
),
314 (output
->data
.invariant
) ? "has" : "lacks",
315 _mesa_shader_stage_to_string(consumer_stage
),
316 (input
->data
.invariant
) ? "has" : "lacks");
320 /* GLSL >= 4.40 removes text requiring interpolation qualifiers
321 * to match cross stage, they must only match within the same stage.
323 * From page 84 (page 90 of the PDF) of the GLSL 4.40 spec:
325 * "It is a link-time error if, within the same stage, the interpolation
326 * qualifiers of variables of the same name do not match.
328 * Section 4.3.9 (Interpolation) of the GLSL ES 3.00 spec says:
330 * "When no interpolation qualifier is present, smooth interpolation
333 * So we match variables where one is smooth and the other has no explicit
336 unsigned input_interpolation
= input
->data
.interpolation
;
337 unsigned output_interpolation
= output
->data
.interpolation
;
339 if (input_interpolation
== INTERP_MODE_NONE
)
340 input_interpolation
= INTERP_MODE_SMOOTH
;
341 if (output_interpolation
== INTERP_MODE_NONE
)
342 output_interpolation
= INTERP_MODE_SMOOTH
;
344 if (input_interpolation
!= output_interpolation
&&
345 prog
->data
->Version
< 440) {
347 "%s shader output `%s' specifies %s "
348 "interpolation qualifier, "
349 "but %s shader input specifies %s "
350 "interpolation qualifier\n",
351 _mesa_shader_stage_to_string(producer_stage
),
353 interpolation_string(output
->data
.interpolation
),
354 _mesa_shader_stage_to_string(consumer_stage
),
355 interpolation_string(input
->data
.interpolation
));
361 * Validate front and back color outputs against single color input
364 cross_validate_front_and_back_color(struct gl_shader_program
*prog
,
365 const ir_variable
*input
,
366 const ir_variable
*front_color
,
367 const ir_variable
*back_color
,
368 gl_shader_stage consumer_stage
,
369 gl_shader_stage producer_stage
)
371 if (front_color
!= NULL
&& front_color
->data
.assigned
)
372 cross_validate_types_and_qualifiers(prog
, input
, front_color
,
373 consumer_stage
, producer_stage
);
375 if (back_color
!= NULL
&& back_color
->data
.assigned
)
376 cross_validate_types_and_qualifiers(prog
, input
, back_color
,
377 consumer_stage
, producer_stage
);
381 compute_variable_location_slot(ir_variable
*var
, gl_shader_stage stage
)
383 unsigned location_start
= VARYING_SLOT_VAR0
;
386 case MESA_SHADER_VERTEX
:
387 if (var
->data
.mode
== ir_var_shader_in
)
388 location_start
= VERT_ATTRIB_GENERIC0
;
390 case MESA_SHADER_TESS_CTRL
:
391 case MESA_SHADER_TESS_EVAL
:
393 location_start
= VARYING_SLOT_PATCH0
;
395 case MESA_SHADER_FRAGMENT
:
396 if (var
->data
.mode
== ir_var_shader_out
)
397 location_start
= FRAG_RESULT_DATA0
;
403 return var
->data
.location
- location_start
;
406 struct explicit_location_info
{
409 unsigned interpolation
;
416 check_location_aliasing(struct explicit_location_info explicit_locations
[][4],
420 unsigned location_limit
,
421 const glsl_type
*type
,
422 unsigned interpolation
,
426 gl_shader_program
*prog
,
427 gl_shader_stage stage
)
430 if (type
->without_array()->is_record()) {
431 /* The component qualifier can't be used on structs so just treat
432 * all component slots as used.
436 unsigned dmul
= type
->without_array()->is_64bit() ? 2 : 1;
437 last_comp
= component
+ type
->without_array()->vector_elements
* dmul
;
440 while (location
< location_limit
) {
441 unsigned i
= component
;
443 /* If there are other outputs assigned to the same location
444 * they must have the same interpolation
448 /* Skip the components used by this output, we only care about
449 * other outputs in the same location
456 struct explicit_location_info
*info
=
457 &explicit_locations
[location
][comp
];
460 if (info
->interpolation
!= interpolation
) {
462 "%s shader has multiple outputs at explicit "
463 "location %u with different interpolation "
465 _mesa_shader_stage_to_string(stage
), location
);
469 if (info
->centroid
!= centroid
||
470 info
->sample
!= sample
||
471 info
->patch
!= patch
) {
473 "%s shader has multiple outputs at explicit "
474 "location %u with different aux storage\n",
475 _mesa_shader_stage_to_string(stage
), location
);
483 /* Component aliasing is not allowed */
484 while (i
< last_comp
) {
485 if (explicit_locations
[location
][i
].var
!= NULL
) {
487 "%s shader has multiple outputs explicitly "
488 "assigned to location %d and component %d\n",
489 _mesa_shader_stage_to_string(stage
),
490 location
, component
);
494 /* Make sure all component at this location have the same type.
496 for (unsigned j
= 0; j
< 4; j
++) {
497 if (explicit_locations
[location
][j
].var
&&
498 explicit_locations
[location
][j
].base_type
!=
499 type
->without_array()->base_type
) {
501 "Varyings sharing the same location must "
502 "have the same underlying numerical type. "
503 "Location %u component %u\n", location
, component
);
508 explicit_locations
[location
][i
].var
= var
;
509 explicit_locations
[location
][i
].base_type
=
510 type
->without_array()->base_type
;
511 explicit_locations
[location
][i
].interpolation
= interpolation
;
512 explicit_locations
[location
][i
].centroid
= centroid
;
513 explicit_locations
[location
][i
].sample
= sample
;
514 explicit_locations
[location
][i
].patch
= patch
;
517 /* We need to do some special handling for doubles as dvec3 and
518 * dvec4 consume two consecutive locations. We don't need to
519 * worry about components beginning at anything other than 0 as
520 * the spec does not allow this for dvec3 and dvec4.
522 if (i
== 4 && last_comp
> 4) {
523 last_comp
= last_comp
- 4;
524 /* Bump location index and reset the component index */
537 validate_explicit_variable_location(struct gl_context
*ctx
,
538 struct explicit_location_info explicit_locations
[][4],
540 gl_shader_program
*prog
,
541 gl_linked_shader
*sh
)
543 const glsl_type
*type
= get_varying_type(var
, sh
->Stage
);
544 unsigned num_elements
= type
->count_attribute_slots(false);
545 unsigned idx
= compute_variable_location_slot(var
, sh
->Stage
);
546 unsigned slot_limit
= idx
+ num_elements
;
548 /* Vertex shader inputs and fragment shader outputs are validated in
549 * assign_attribute_or_color_locations() so we should not attempt to
550 * validate them again here.
553 if (var
->data
.mode
== ir_var_shader_out
) {
554 assert(sh
->Stage
!= MESA_SHADER_FRAGMENT
);
556 ctx
->Const
.Program
[sh
->Stage
].MaxOutputComponents
/ 4;
558 assert(var
->data
.mode
== ir_var_shader_in
);
559 assert(sh
->Stage
!= MESA_SHADER_VERTEX
);
561 ctx
->Const
.Program
[sh
->Stage
].MaxInputComponents
/ 4;
564 if (slot_limit
> slot_max
) {
566 "Invalid location %u in %s shader\n",
567 idx
, _mesa_shader_stage_to_string(sh
->Stage
));
571 if (type
->without_array()->is_interface()) {
572 for (unsigned i
= 0; i
< type
->without_array()->length
; i
++) {
573 glsl_struct_field
*field
= &type
->fields
.structure
[i
];
574 unsigned field_location
= field
->location
-
575 (field
->patch
? VARYING_SLOT_PATCH0
: VARYING_SLOT_VAR0
);
576 if (!check_location_aliasing(explicit_locations
, var
,
578 0, field_location
+ 1,
580 field
->interpolation
,
588 } else if (!check_location_aliasing(explicit_locations
, var
,
589 idx
, var
->data
.location_frac
,
591 var
->data
.interpolation
,
603 * Validate that outputs from one stage match inputs of another
606 cross_validate_outputs_to_inputs(struct gl_context
*ctx
,
607 struct gl_shader_program
*prog
,
608 gl_linked_shader
*producer
,
609 gl_linked_shader
*consumer
)
611 glsl_symbol_table parameters
;
612 struct explicit_location_info explicit_locations
[MAX_VARYING
][4] = { 0 };
614 /* Find all shader outputs in the "producer" stage.
616 foreach_in_list(ir_instruction
, node
, producer
->ir
) {
617 ir_variable
*const var
= node
->as_variable();
619 if (var
== NULL
|| var
->data
.mode
!= ir_var_shader_out
)
622 if (!var
->data
.explicit_location
623 || var
->data
.location
< VARYING_SLOT_VAR0
)
624 parameters
.add_variable(var
);
626 /* User-defined varyings with explicit locations are handled
627 * differently because they do not need to have matching names.
629 if (!validate_explicit_variable_location(ctx
,
631 var
, prog
, producer
)) {
638 /* Find all shader inputs in the "consumer" stage. Any variables that have
639 * matching outputs already in the symbol table must have the same type and
642 * Exception: if the consumer is the geometry shader, then the inputs
643 * should be arrays and the type of the array element should match the type
644 * of the corresponding producer output.
646 foreach_in_list(ir_instruction
, node
, consumer
->ir
) {
647 ir_variable
*const input
= node
->as_variable();
649 if (input
== NULL
|| input
->data
.mode
!= ir_var_shader_in
)
652 if (strcmp(input
->name
, "gl_Color") == 0 && input
->data
.used
) {
653 const ir_variable
*const front_color
=
654 parameters
.get_variable("gl_FrontColor");
656 const ir_variable
*const back_color
=
657 parameters
.get_variable("gl_BackColor");
659 cross_validate_front_and_back_color(prog
, input
,
660 front_color
, back_color
,
661 consumer
->Stage
, producer
->Stage
);
662 } else if (strcmp(input
->name
, "gl_SecondaryColor") == 0 && input
->data
.used
) {
663 const ir_variable
*const front_color
=
664 parameters
.get_variable("gl_FrontSecondaryColor");
666 const ir_variable
*const back_color
=
667 parameters
.get_variable("gl_BackSecondaryColor");
669 cross_validate_front_and_back_color(prog
, input
,
670 front_color
, back_color
,
671 consumer
->Stage
, producer
->Stage
);
673 /* The rules for connecting inputs and outputs change in the presence
674 * of explicit locations. In this case, we no longer care about the
675 * names of the variables. Instead, we care only about the
676 * explicitly assigned location.
678 ir_variable
*output
= NULL
;
679 if (input
->data
.explicit_location
680 && input
->data
.location
>= VARYING_SLOT_VAR0
) {
682 const glsl_type
*type
= get_varying_type(input
, consumer
->Stage
);
683 unsigned num_elements
= type
->count_attribute_slots(false);
685 compute_variable_location_slot(input
, consumer
->Stage
);
686 unsigned slot_limit
= idx
+ num_elements
;
688 while (idx
< slot_limit
) {
689 if (idx
>= MAX_VARYING
) {
691 "Invalid location %u in %s shader\n", idx
,
692 _mesa_shader_stage_to_string(consumer
->Stage
));
696 output
= explicit_locations
[idx
][input
->data
.location_frac
].var
;
698 if (output
== NULL
||
699 input
->data
.location
!= output
->data
.location
) {
701 "%s shader input `%s' with explicit location "
702 "has no matching output\n",
703 _mesa_shader_stage_to_string(consumer
->Stage
),
710 output
= parameters
.get_variable(input
->name
);
713 if (output
!= NULL
) {
714 /* Interface blocks have their own validation elsewhere so don't
715 * try validating them here.
717 if (!(input
->get_interface_type() &&
718 output
->get_interface_type()))
719 cross_validate_types_and_qualifiers(prog
, input
, output
,
723 /* Check for input vars with unmatched output vars in prev stage
724 * taking into account that interface blocks could have a matching
725 * output but with different name, so we ignore them.
727 assert(!input
->data
.assigned
);
728 if (input
->data
.used
&& !input
->get_interface_type() &&
729 !input
->data
.explicit_location
&& !prog
->SeparateShader
)
731 "%s shader input `%s' "
732 "has no matching output in the previous stage\n",
733 _mesa_shader_stage_to_string(consumer
->Stage
),
741 * Demote shader inputs and outputs that are not used in other stages, and
742 * remove them via dead code elimination.
745 remove_unused_shader_inputs_and_outputs(bool is_separate_shader_object
,
746 gl_linked_shader
*sh
,
747 enum ir_variable_mode mode
)
749 if (is_separate_shader_object
)
752 foreach_in_list(ir_instruction
, node
, sh
->ir
) {
753 ir_variable
*const var
= node
->as_variable();
755 if (var
== NULL
|| var
->data
.mode
!= int(mode
))
758 /* A shader 'in' or 'out' variable is only really an input or output if
759 * its value is used by other shader stages. This will cause the
760 * variable to have a location assigned.
762 if (var
->data
.is_unmatched_generic_inout
&& !var
->data
.is_xfb_only
) {
763 assert(var
->data
.mode
!= ir_var_temporary
);
765 /* Assign zeros to demoted inputs to allow more optimizations. */
766 if (var
->data
.mode
== ir_var_shader_in
&& !var
->constant_value
)
767 var
->constant_value
= ir_constant::zero(var
, var
->type
);
769 var
->data
.mode
= ir_var_auto
;
773 /* Eliminate code that is now dead due to unused inputs/outputs being
776 while (do_dead_code(sh
->ir
, false))
782 * Initialize this object based on a string that was passed to
783 * glTransformFeedbackVaryings.
785 * If the input is mal-formed, this call still succeeds, but it sets
786 * this->var_name to a mal-formed input, so tfeedback_decl::find_output_var()
787 * will fail to find any matching variable.
790 tfeedback_decl::init(struct gl_context
*ctx
, const void *mem_ctx
,
793 /* We don't have to be pedantic about what is a valid GLSL variable name,
794 * because any variable with an invalid name can't exist in the IR anyway.
798 this->orig_name
= input
;
799 this->lowered_builtin_array_variable
= none
;
800 this->skip_components
= 0;
801 this->next_buffer_separator
= false;
802 this->matched_candidate
= NULL
;
807 if (ctx
->Extensions
.ARB_transform_feedback3
) {
808 /* Parse gl_NextBuffer. */
809 if (strcmp(input
, "gl_NextBuffer") == 0) {
810 this->next_buffer_separator
= true;
814 /* Parse gl_SkipComponents. */
815 if (strcmp(input
, "gl_SkipComponents1") == 0)
816 this->skip_components
= 1;
817 else if (strcmp(input
, "gl_SkipComponents2") == 0)
818 this->skip_components
= 2;
819 else if (strcmp(input
, "gl_SkipComponents3") == 0)
820 this->skip_components
= 3;
821 else if (strcmp(input
, "gl_SkipComponents4") == 0)
822 this->skip_components
= 4;
824 if (this->skip_components
)
828 /* Parse a declaration. */
829 const char *base_name_end
;
830 long subscript
= parse_program_resource_name(input
, &base_name_end
);
831 this->var_name
= ralloc_strndup(mem_ctx
, input
, base_name_end
- input
);
832 if (this->var_name
== NULL
) {
833 _mesa_error_no_memory(__func__
);
837 if (subscript
>= 0) {
838 this->array_subscript
= subscript
;
839 this->is_subscripted
= true;
841 this->is_subscripted
= false;
844 /* For drivers that lower gl_ClipDistance to gl_ClipDistanceMESA, this
845 * class must behave specially to account for the fact that gl_ClipDistance
846 * is converted from a float[8] to a vec4[2].
848 if (ctx
->Const
.ShaderCompilerOptions
[MESA_SHADER_VERTEX
].LowerCombinedClipCullDistance
&&
849 strcmp(this->var_name
, "gl_ClipDistance") == 0) {
850 this->lowered_builtin_array_variable
= clip_distance
;
852 if (ctx
->Const
.ShaderCompilerOptions
[MESA_SHADER_VERTEX
].LowerCombinedClipCullDistance
&&
853 strcmp(this->var_name
, "gl_CullDistance") == 0) {
854 this->lowered_builtin_array_variable
= cull_distance
;
857 if (ctx
->Const
.LowerTessLevel
&&
858 (strcmp(this->var_name
, "gl_TessLevelOuter") == 0))
859 this->lowered_builtin_array_variable
= tess_level_outer
;
860 if (ctx
->Const
.LowerTessLevel
&&
861 (strcmp(this->var_name
, "gl_TessLevelInner") == 0))
862 this->lowered_builtin_array_variable
= tess_level_inner
;
867 * Determine whether two tfeedback_decl objects refer to the same variable and
868 * array index (if applicable).
871 tfeedback_decl::is_same(const tfeedback_decl
&x
, const tfeedback_decl
&y
)
873 assert(x
.is_varying() && y
.is_varying());
875 if (strcmp(x
.var_name
, y
.var_name
) != 0)
877 if (x
.is_subscripted
!= y
.is_subscripted
)
879 if (x
.is_subscripted
&& x
.array_subscript
!= y
.array_subscript
)
886 * Assign a location and stream ID for this tfeedback_decl object based on the
887 * transform feedback candidate found by find_candidate.
889 * If an error occurs, the error is reported through linker_error() and false
893 tfeedback_decl::assign_location(struct gl_context
*ctx
,
894 struct gl_shader_program
*prog
)
896 assert(this->is_varying());
898 unsigned fine_location
899 = this->matched_candidate
->toplevel_var
->data
.location
* 4
900 + this->matched_candidate
->toplevel_var
->data
.location_frac
901 + this->matched_candidate
->offset
;
902 const unsigned dmul
=
903 this->matched_candidate
->type
->without_array()->is_64bit() ? 2 : 1;
905 if (this->matched_candidate
->type
->is_array()) {
907 const unsigned matrix_cols
=
908 this->matched_candidate
->type
->fields
.array
->matrix_columns
;
909 const unsigned vector_elements
=
910 this->matched_candidate
->type
->fields
.array
->vector_elements
;
911 unsigned actual_array_size
;
912 switch (this->lowered_builtin_array_variable
) {
914 actual_array_size
= prog
->last_vert_prog
?
915 prog
->last_vert_prog
->info
.clip_distance_array_size
: 0;
918 actual_array_size
= prog
->last_vert_prog
?
919 prog
->last_vert_prog
->info
.cull_distance_array_size
: 0;
921 case tess_level_outer
:
922 actual_array_size
= 4;
924 case tess_level_inner
:
925 actual_array_size
= 2;
929 actual_array_size
= this->matched_candidate
->type
->array_size();
933 if (this->is_subscripted
) {
934 /* Check array bounds. */
935 if (this->array_subscript
>= actual_array_size
) {
936 linker_error(prog
, "Transform feedback varying %s has index "
937 "%i, but the array size is %u.",
938 this->orig_name
, this->array_subscript
,
942 unsigned array_elem_size
= this->lowered_builtin_array_variable
?
943 1 : vector_elements
* matrix_cols
* dmul
;
944 fine_location
+= array_elem_size
* this->array_subscript
;
947 this->size
= actual_array_size
;
949 this->vector_elements
= vector_elements
;
950 this->matrix_columns
= matrix_cols
;
951 if (this->lowered_builtin_array_variable
)
952 this->type
= GL_FLOAT
;
954 this->type
= this->matched_candidate
->type
->fields
.array
->gl_type
;
956 /* Regular variable (scalar, vector, or matrix) */
957 if (this->is_subscripted
) {
958 linker_error(prog
, "Transform feedback varying %s requested, "
959 "but %s is not an array.",
960 this->orig_name
, this->var_name
);
964 this->vector_elements
= this->matched_candidate
->type
->vector_elements
;
965 this->matrix_columns
= this->matched_candidate
->type
->matrix_columns
;
966 this->type
= this->matched_candidate
->type
->gl_type
;
968 this->location
= fine_location
/ 4;
969 this->location_frac
= fine_location
% 4;
971 /* From GL_EXT_transform_feedback:
972 * A program will fail to link if:
974 * * the total number of components to capture in any varying
975 * variable in <varyings> is greater than the constant
976 * MAX_TRANSFORM_FEEDBACK_SEPARATE_COMPONENTS_EXT and the
977 * buffer mode is SEPARATE_ATTRIBS_EXT;
979 if (prog
->TransformFeedback
.BufferMode
== GL_SEPARATE_ATTRIBS
&&
980 this->num_components() >
981 ctx
->Const
.MaxTransformFeedbackSeparateComponents
) {
982 linker_error(prog
, "Transform feedback varying %s exceeds "
983 "MAX_TRANSFORM_FEEDBACK_SEPARATE_COMPONENTS.",
988 /* Only transform feedback varyings can be assigned to non-zero streams,
989 * so assign the stream id here.
991 this->stream_id
= this->matched_candidate
->toplevel_var
->data
.stream
;
993 unsigned array_offset
= this->array_subscript
* 4 * dmul
;
994 unsigned struct_offset
= this->matched_candidate
->offset
* 4 * dmul
;
995 this->buffer
= this->matched_candidate
->toplevel_var
->data
.xfb_buffer
;
996 this->offset
= this->matched_candidate
->toplevel_var
->data
.offset
+
997 array_offset
+ struct_offset
;
1004 tfeedback_decl::get_num_outputs() const
1006 if (!this->is_varying()) {
1009 return (this->num_components() + this->location_frac
+ 3)/4;
1014 * Update gl_transform_feedback_info to reflect this tfeedback_decl.
1016 * If an error occurs, the error is reported through linker_error() and false
1020 tfeedback_decl::store(struct gl_context
*ctx
, struct gl_shader_program
*prog
,
1021 struct gl_transform_feedback_info
*info
,
1022 unsigned buffer
, unsigned buffer_index
,
1023 const unsigned max_outputs
, bool *explicit_stride
,
1024 bool has_xfb_qualifiers
) const
1026 unsigned xfb_offset
= 0;
1027 unsigned size
= this->size
;
1028 /* Handle gl_SkipComponents. */
1029 if (this->skip_components
) {
1030 info
->Buffers
[buffer
].Stride
+= this->skip_components
;
1031 size
= this->skip_components
;
1035 if (this->next_buffer_separator
) {
1040 if (has_xfb_qualifiers
) {
1041 xfb_offset
= this->offset
/ 4;
1043 xfb_offset
= info
->Buffers
[buffer
].Stride
;
1045 info
->Varyings
[info
->NumVarying
].Offset
= xfb_offset
* 4;
1048 unsigned location
= this->location
;
1049 unsigned location_frac
= this->location_frac
;
1050 unsigned num_components
= this->num_components();
1051 while (num_components
> 0) {
1052 unsigned output_size
= MIN2(num_components
, 4 - location_frac
);
1053 assert((info
->NumOutputs
== 0 && max_outputs
== 0) ||
1054 info
->NumOutputs
< max_outputs
);
1056 /* From the ARB_enhanced_layouts spec:
1058 * "If such a block member or variable is not written during a shader
1059 * invocation, the buffer contents at the assigned offset will be
1060 * undefined. Even if there are no static writes to a variable or
1061 * member that is assigned a transform feedback offset, the space is
1062 * still allocated in the buffer and still affects the stride."
1064 if (this->is_varying_written()) {
1065 info
->Outputs
[info
->NumOutputs
].ComponentOffset
= location_frac
;
1066 info
->Outputs
[info
->NumOutputs
].OutputRegister
= location
;
1067 info
->Outputs
[info
->NumOutputs
].NumComponents
= output_size
;
1068 info
->Outputs
[info
->NumOutputs
].StreamId
= stream_id
;
1069 info
->Outputs
[info
->NumOutputs
].OutputBuffer
= buffer
;
1070 info
->Outputs
[info
->NumOutputs
].DstOffset
= xfb_offset
;
1073 info
->Buffers
[buffer
].Stream
= this->stream_id
;
1074 xfb_offset
+= output_size
;
1076 num_components
-= output_size
;
1082 if (explicit_stride
&& explicit_stride
[buffer
]) {
1083 if (this->is_64bit() && info
->Buffers
[buffer
].Stride
% 2) {
1084 linker_error(prog
, "invalid qualifier xfb_stride=%d must be a "
1085 "multiple of 8 as its applied to a type that is or "
1086 "contains a double.",
1087 info
->Buffers
[buffer
].Stride
* 4);
1091 if ((this->offset
/ 4) / info
->Buffers
[buffer
].Stride
!=
1092 (xfb_offset
- 1) / info
->Buffers
[buffer
].Stride
) {
1093 linker_error(prog
, "xfb_offset (%d) overflows xfb_stride (%d) for "
1094 "buffer (%d)", xfb_offset
* 4,
1095 info
->Buffers
[buffer
].Stride
* 4, buffer
);
1099 info
->Buffers
[buffer
].Stride
= xfb_offset
;
1102 /* From GL_EXT_transform_feedback:
1103 * A program will fail to link if:
1105 * * the total number of components to capture is greater than
1106 * the constant MAX_TRANSFORM_FEEDBACK_INTERLEAVED_COMPONENTS_EXT
1107 * and the buffer mode is INTERLEAVED_ATTRIBS_EXT.
1109 * From GL_ARB_enhanced_layouts:
1111 * "The resulting stride (implicit or explicit) must be less than or
1112 * equal to the implementation-dependent constant
1113 * gl_MaxTransformFeedbackInterleavedComponents."
1115 if ((prog
->TransformFeedback
.BufferMode
== GL_INTERLEAVED_ATTRIBS
||
1116 has_xfb_qualifiers
) &&
1117 info
->Buffers
[buffer
].Stride
>
1118 ctx
->Const
.MaxTransformFeedbackInterleavedComponents
) {
1119 linker_error(prog
, "The MAX_TRANSFORM_FEEDBACK_INTERLEAVED_COMPONENTS "
1120 "limit has been exceeded.");
1125 info
->Varyings
[info
->NumVarying
].Name
= ralloc_strdup(prog
,
1127 info
->Varyings
[info
->NumVarying
].Type
= this->type
;
1128 info
->Varyings
[info
->NumVarying
].Size
= size
;
1129 info
->Varyings
[info
->NumVarying
].BufferIndex
= buffer_index
;
1131 info
->Buffers
[buffer
].NumVaryings
++;
1137 const tfeedback_candidate
*
1138 tfeedback_decl::find_candidate(gl_shader_program
*prog
,
1139 hash_table
*tfeedback_candidates
)
1141 const char *name
= this->var_name
;
1142 switch (this->lowered_builtin_array_variable
) {
1144 name
= this->var_name
;
1147 name
= "gl_ClipDistanceMESA";
1150 name
= "gl_CullDistanceMESA";
1152 case tess_level_outer
:
1153 name
= "gl_TessLevelOuterMESA";
1155 case tess_level_inner
:
1156 name
= "gl_TessLevelInnerMESA";
1159 hash_entry
*entry
= _mesa_hash_table_search(tfeedback_candidates
, name
);
1161 this->matched_candidate
= entry
?
1162 (const tfeedback_candidate
*) entry
->data
: NULL
;
1164 if (!this->matched_candidate
) {
1165 /* From GL_EXT_transform_feedback:
1166 * A program will fail to link if:
1168 * * any variable name specified in the <varyings> array is not
1169 * declared as an output in the geometry shader (if present) or
1170 * the vertex shader (if no geometry shader is present);
1172 linker_error(prog
, "Transform feedback varying %s undeclared.",
1176 return this->matched_candidate
;
1181 * Parse all the transform feedback declarations that were passed to
1182 * glTransformFeedbackVaryings() and store them in tfeedback_decl objects.
1184 * If an error occurs, the error is reported through linker_error() and false
1188 parse_tfeedback_decls(struct gl_context
*ctx
, struct gl_shader_program
*prog
,
1189 const void *mem_ctx
, unsigned num_names
,
1190 char **varying_names
, tfeedback_decl
*decls
)
1192 for (unsigned i
= 0; i
< num_names
; ++i
) {
1193 decls
[i
].init(ctx
, mem_ctx
, varying_names
[i
]);
1195 if (!decls
[i
].is_varying())
1198 /* From GL_EXT_transform_feedback:
1199 * A program will fail to link if:
1201 * * any two entries in the <varyings> array specify the same varying
1204 * We interpret this to mean "any two entries in the <varyings> array
1205 * specify the same varying variable and array index", since transform
1206 * feedback of arrays would be useless otherwise.
1208 for (unsigned j
= 0; j
< i
; ++j
) {
1209 if (!decls
[j
].is_varying())
1212 if (tfeedback_decl::is_same(decls
[i
], decls
[j
])) {
1213 linker_error(prog
, "Transform feedback varying %s specified "
1214 "more than once.", varying_names
[i
]);
1224 cmp_xfb_offset(const void * x_generic
, const void * y_generic
)
1226 tfeedback_decl
*x
= (tfeedback_decl
*) x_generic
;
1227 tfeedback_decl
*y
= (tfeedback_decl
*) y_generic
;
1229 if (x
->get_buffer() != y
->get_buffer())
1230 return x
->get_buffer() - y
->get_buffer();
1231 return x
->get_offset() - y
->get_offset();
1235 * Store transform feedback location assignments into
1236 * prog->sh.LinkedTransformFeedback based on the data stored in
1239 * If an error occurs, the error is reported through linker_error() and false
1243 store_tfeedback_info(struct gl_context
*ctx
, struct gl_shader_program
*prog
,
1244 unsigned num_tfeedback_decls
,
1245 tfeedback_decl
*tfeedback_decls
, bool has_xfb_qualifiers
)
1247 if (!prog
->last_vert_prog
)
1250 /* Make sure MaxTransformFeedbackBuffers is less than 32 so the bitmask for
1251 * tracking the number of buffers doesn't overflow.
1253 assert(ctx
->Const
.MaxTransformFeedbackBuffers
< 32);
1255 bool separate_attribs_mode
=
1256 prog
->TransformFeedback
.BufferMode
== GL_SEPARATE_ATTRIBS
;
1258 struct gl_program
*xfb_prog
= prog
->last_vert_prog
;
1259 xfb_prog
->sh
.LinkedTransformFeedback
=
1260 rzalloc(xfb_prog
, struct gl_transform_feedback_info
);
1262 /* The xfb_offset qualifier does not have to be used in increasing order
1263 * however some drivers expect to receive the list of transform feedback
1264 * declarations in order so sort it now for convenience.
1266 if (has_xfb_qualifiers
)
1267 qsort(tfeedback_decls
, num_tfeedback_decls
, sizeof(*tfeedback_decls
),
1270 xfb_prog
->sh
.LinkedTransformFeedback
->Varyings
=
1271 rzalloc_array(xfb_prog
, struct gl_transform_feedback_varying_info
,
1272 num_tfeedback_decls
);
1274 unsigned num_outputs
= 0;
1275 for (unsigned i
= 0; i
< num_tfeedback_decls
; ++i
) {
1276 if (tfeedback_decls
[i
].is_varying_written())
1277 num_outputs
+= tfeedback_decls
[i
].get_num_outputs();
1280 xfb_prog
->sh
.LinkedTransformFeedback
->Outputs
=
1281 rzalloc_array(xfb_prog
, struct gl_transform_feedback_output
,
1284 unsigned num_buffers
= 0;
1285 unsigned buffers
= 0;
1287 if (!has_xfb_qualifiers
&& separate_attribs_mode
) {
1288 /* GL_SEPARATE_ATTRIBS */
1289 for (unsigned i
= 0; i
< num_tfeedback_decls
; ++i
) {
1290 if (!tfeedback_decls
[i
].store(ctx
, prog
,
1291 xfb_prog
->sh
.LinkedTransformFeedback
,
1292 num_buffers
, num_buffers
, num_outputs
,
1293 NULL
, has_xfb_qualifiers
))
1296 buffers
|= 1 << num_buffers
;
1301 /* GL_INVERLEAVED_ATTRIBS */
1302 int buffer_stream_id
= -1;
1304 num_tfeedback_decls
? tfeedback_decls
[0].get_buffer() : 0;
1305 bool explicit_stride
[MAX_FEEDBACK_BUFFERS
] = { false };
1307 /* Apply any xfb_stride global qualifiers */
1308 if (has_xfb_qualifiers
) {
1309 for (unsigned j
= 0; j
< MAX_FEEDBACK_BUFFERS
; j
++) {
1310 if (prog
->TransformFeedback
.BufferStride
[j
]) {
1312 explicit_stride
[j
] = true;
1313 xfb_prog
->sh
.LinkedTransformFeedback
->Buffers
[j
].Stride
=
1314 prog
->TransformFeedback
.BufferStride
[j
] / 4;
1319 for (unsigned i
= 0; i
< num_tfeedback_decls
; ++i
) {
1320 if (has_xfb_qualifiers
&&
1321 buffer
!= tfeedback_decls
[i
].get_buffer()) {
1322 /* we have moved to the next buffer so reset stream id */
1323 buffer_stream_id
= -1;
1327 if (tfeedback_decls
[i
].is_next_buffer_separator()) {
1328 if (!tfeedback_decls
[i
].store(ctx
, prog
,
1329 xfb_prog
->sh
.LinkedTransformFeedback
,
1330 buffer
, num_buffers
, num_outputs
,
1331 explicit_stride
, has_xfb_qualifiers
))
1334 buffer_stream_id
= -1;
1336 } else if (tfeedback_decls
[i
].is_varying()) {
1337 if (buffer_stream_id
== -1) {
1338 /* First varying writing to this buffer: remember its stream */
1339 buffer_stream_id
= (int) tfeedback_decls
[i
].get_stream_id();
1340 } else if (buffer_stream_id
!=
1341 (int) tfeedback_decls
[i
].get_stream_id()) {
1342 /* Varying writes to the same buffer from a different stream */
1344 "Transform feedback can't capture varyings belonging "
1345 "to different vertex streams in a single buffer. "
1346 "Varying %s writes to buffer from stream %u, other "
1347 "varyings in the same buffer write from stream %u.",
1348 tfeedback_decls
[i
].name(),
1349 tfeedback_decls
[i
].get_stream_id(),
1355 if (has_xfb_qualifiers
) {
1356 buffer
= tfeedback_decls
[i
].get_buffer();
1358 buffer
= num_buffers
;
1360 buffers
|= 1 << buffer
;
1362 if (!tfeedback_decls
[i
].store(ctx
, prog
,
1363 xfb_prog
->sh
.LinkedTransformFeedback
,
1364 buffer
, num_buffers
, num_outputs
,
1365 explicit_stride
, has_xfb_qualifiers
))
1370 assert(xfb_prog
->sh
.LinkedTransformFeedback
->NumOutputs
== num_outputs
);
1372 xfb_prog
->sh
.LinkedTransformFeedback
->ActiveBuffers
= buffers
;
1379 * Data structure recording the relationship between outputs of one shader
1380 * stage (the "producer") and inputs of another (the "consumer").
1382 class varying_matches
1385 varying_matches(bool disable_varying_packing
, bool xfb_enabled
,
1386 bool enhanced_layouts_enabled
,
1387 gl_shader_stage producer_stage
,
1388 gl_shader_stage consumer_stage
);
1390 void record(ir_variable
*producer_var
, ir_variable
*consumer_var
);
1391 unsigned assign_locations(struct gl_shader_program
*prog
,
1392 uint8_t *components
,
1393 uint64_t reserved_slots
);
1394 void store_locations() const;
1397 bool is_varying_packing_safe(const glsl_type
*type
,
1398 const ir_variable
*var
);
1401 * If true, this driver disables varying packing, so all varyings need to
1402 * be aligned on slot boundaries, and take up a number of slots equal to
1403 * their number of matrix columns times their array size.
1405 * Packing may also be disabled because our current packing method is not
1406 * safe in SSO or versions of OpenGL where interpolation qualifiers are not
1407 * guaranteed to match across stages.
1409 const bool disable_varying_packing
;
1412 * If true, this driver has transform feedback enabled. The transform
1413 * feedback code requires at least some packing be done even when varying
1414 * packing is disabled, fortunately where transform feedback requires
1415 * packing it's safe to override the disabled setting. See
1416 * is_varying_packing_safe().
1418 const bool xfb_enabled
;
1420 const bool enhanced_layouts_enabled
;
1423 * Enum representing the order in which varyings are packed within a
1426 * Currently we pack vec4's first, then vec2's, then scalar values, then
1427 * vec3's. This order ensures that the only vectors that are at risk of
1428 * having to be "double parked" (split between two adjacent varying slots)
1431 enum packing_order_enum
{
1434 PACKING_ORDER_SCALAR
,
1438 static unsigned compute_packing_class(const ir_variable
*var
);
1439 static packing_order_enum
compute_packing_order(const ir_variable
*var
);
1440 static int match_comparator(const void *x_generic
, const void *y_generic
);
1441 static int xfb_comparator(const void *x_generic
, const void *y_generic
);
1444 * Structure recording the relationship between a single producer output
1445 * and a single consumer input.
1449 * Packing class for this varying, computed by compute_packing_class().
1451 unsigned packing_class
;
1454 * Packing order for this varying, computed by compute_packing_order().
1456 packing_order_enum packing_order
;
1457 unsigned num_components
;
1460 * The output variable in the producer stage.
1462 ir_variable
*producer_var
;
1465 * The input variable in the consumer stage.
1467 ir_variable
*consumer_var
;
1470 * The location which has been assigned for this varying. This is
1471 * expressed in multiples of a float, with the first generic varying
1472 * (i.e. the one referred to by VARYING_SLOT_VAR0) represented by the
1475 unsigned generic_location
;
1479 * The number of elements in the \c matches array that are currently in
1482 unsigned num_matches
;
1485 * The number of elements that were set aside for the \c matches array when
1488 unsigned matches_capacity
;
1490 gl_shader_stage producer_stage
;
1491 gl_shader_stage consumer_stage
;
1494 } /* anonymous namespace */
1496 varying_matches::varying_matches(bool disable_varying_packing
,
1498 bool enhanced_layouts_enabled
,
1499 gl_shader_stage producer_stage
,
1500 gl_shader_stage consumer_stage
)
1501 : disable_varying_packing(disable_varying_packing
),
1502 xfb_enabled(xfb_enabled
),
1503 enhanced_layouts_enabled(enhanced_layouts_enabled
),
1504 producer_stage(producer_stage
),
1505 consumer_stage(consumer_stage
)
1507 /* Note: this initial capacity is rather arbitrarily chosen to be large
1508 * enough for many cases without wasting an unreasonable amount of space.
1509 * varying_matches::record() will resize the array if there are more than
1510 * this number of varyings.
1512 this->matches_capacity
= 8;
1513 this->matches
= (match
*)
1514 malloc(sizeof(*this->matches
) * this->matches_capacity
);
1515 this->num_matches
= 0;
1519 varying_matches::~varying_matches()
1521 free(this->matches
);
1526 * Packing is always safe on individual arrays, structures, and matrices. It
1527 * is also safe if the varying is only used for transform feedback.
1530 varying_matches::is_varying_packing_safe(const glsl_type
*type
,
1531 const ir_variable
*var
)
1533 if (consumer_stage
== MESA_SHADER_TESS_EVAL
||
1534 consumer_stage
== MESA_SHADER_TESS_CTRL
||
1535 producer_stage
== MESA_SHADER_TESS_CTRL
)
1538 return xfb_enabled
&& (type
->is_array() || type
->is_record() ||
1539 type
->is_matrix() || var
->data
.is_xfb_only
);
1544 * Record the given producer/consumer variable pair in the list of variables
1545 * that should later be assigned locations.
1547 * It is permissible for \c consumer_var to be NULL (this happens if a
1548 * variable is output by the producer and consumed by transform feedback, but
1549 * not consumed by the consumer).
1551 * If \c producer_var has already been paired up with a consumer_var, or
1552 * producer_var is part of fixed pipeline functionality (and hence already has
1553 * a location assigned), this function has no effect.
1555 * Note: as a side effect this function may change the interpolation type of
1556 * \c producer_var, but only when the change couldn't possibly affect
1560 varying_matches::record(ir_variable
*producer_var
, ir_variable
*consumer_var
)
1562 assert(producer_var
!= NULL
|| consumer_var
!= NULL
);
1564 if ((producer_var
&& (!producer_var
->data
.is_unmatched_generic_inout
||
1565 producer_var
->data
.explicit_location
)) ||
1566 (consumer_var
&& (!consumer_var
->data
.is_unmatched_generic_inout
||
1567 consumer_var
->data
.explicit_location
))) {
1568 /* Either a location already exists for this variable (since it is part
1569 * of fixed functionality), or it has already been recorded as part of a
1575 bool needs_flat_qualifier
= consumer_var
== NULL
&&
1576 (producer_var
->type
->contains_integer() ||
1577 producer_var
->type
->contains_double());
1579 if (!disable_varying_packing
&&
1580 (needs_flat_qualifier
||
1581 (consumer_stage
!= MESA_SHADER_NONE
&& consumer_stage
!= MESA_SHADER_FRAGMENT
))) {
1582 /* Since this varying is not being consumed by the fragment shader, its
1583 * interpolation type varying cannot possibly affect rendering.
1584 * Also, this variable is non-flat and is (or contains) an integer
1586 * If the consumer stage is unknown, don't modify the interpolation
1587 * type as it could affect rendering later with separate shaders.
1589 * lower_packed_varyings requires all integer varyings to flat,
1590 * regardless of where they appear. We can trivially satisfy that
1591 * requirement by changing the interpolation type to flat here.
1594 producer_var
->data
.centroid
= false;
1595 producer_var
->data
.sample
= false;
1596 producer_var
->data
.interpolation
= INTERP_MODE_FLAT
;
1600 consumer_var
->data
.centroid
= false;
1601 consumer_var
->data
.sample
= false;
1602 consumer_var
->data
.interpolation
= INTERP_MODE_FLAT
;
1606 if (this->num_matches
== this->matches_capacity
) {
1607 this->matches_capacity
*= 2;
1608 this->matches
= (match
*)
1609 realloc(this->matches
,
1610 sizeof(*this->matches
) * this->matches_capacity
);
1613 /* We must use the consumer to compute the packing class because in GL4.4+
1614 * there is no guarantee interpolation qualifiers will match across stages.
1616 * From Section 4.5 (Interpolation Qualifiers) of the GLSL 4.30 spec:
1618 * "The type and presence of interpolation qualifiers of variables with
1619 * the same name declared in all linked shaders for the same cross-stage
1620 * interface must match, otherwise the link command will fail.
1622 * When comparing an output from one stage to an input of a subsequent
1623 * stage, the input and output don't match if their interpolation
1624 * qualifiers (or lack thereof) are not the same."
1626 * This text was also in at least revison 7 of the 4.40 spec but is no
1627 * longer in revision 9 and not in the 4.50 spec.
1629 const ir_variable
*const var
= (consumer_var
!= NULL
)
1630 ? consumer_var
: producer_var
;
1631 const gl_shader_stage stage
= (consumer_var
!= NULL
)
1632 ? consumer_stage
: producer_stage
;
1633 const glsl_type
*type
= get_varying_type(var
, stage
);
1635 if (producer_var
&& consumer_var
&&
1636 consumer_var
->data
.must_be_shader_input
) {
1637 producer_var
->data
.must_be_shader_input
= 1;
1640 this->matches
[this->num_matches
].packing_class
1641 = this->compute_packing_class(var
);
1642 this->matches
[this->num_matches
].packing_order
1643 = this->compute_packing_order(var
);
1644 if ((this->disable_varying_packing
&& !is_varying_packing_safe(type
, var
)) ||
1645 var
->data
.must_be_shader_input
) {
1646 unsigned slots
= type
->count_attribute_slots(false);
1647 this->matches
[this->num_matches
].num_components
= slots
* 4;
1649 this->matches
[this->num_matches
].num_components
1650 = type
->component_slots();
1653 this->matches
[this->num_matches
].producer_var
= producer_var
;
1654 this->matches
[this->num_matches
].consumer_var
= consumer_var
;
1655 this->num_matches
++;
1657 producer_var
->data
.is_unmatched_generic_inout
= 0;
1659 consumer_var
->data
.is_unmatched_generic_inout
= 0;
1664 * Choose locations for all of the variable matches that were previously
1665 * passed to varying_matches::record().
1668 varying_matches::assign_locations(struct gl_shader_program
*prog
,
1669 uint8_t *components
,
1670 uint64_t reserved_slots
)
1672 /* If packing has been disabled then we cannot safely sort the varyings by
1673 * class as it may mean we are using a version of OpenGL where
1674 * interpolation qualifiers are not guaranteed to be matching across
1675 * shaders, sorting in this case could result in mismatching shader
1677 * When packing is disabled the sort orders varyings used by transform
1678 * feedback first, but also depends on *undefined behaviour* of qsort to
1679 * reverse the order of the varyings. See: xfb_comparator().
1681 if (!this->disable_varying_packing
) {
1682 /* Sort varying matches into an order that makes them easy to pack. */
1683 qsort(this->matches
, this->num_matches
, sizeof(*this->matches
),
1684 &varying_matches::match_comparator
);
1686 /* Only sort varyings that are only used by transform feedback. */
1687 qsort(this->matches
, this->num_matches
, sizeof(*this->matches
),
1688 &varying_matches::xfb_comparator
);
1691 unsigned generic_location
= 0;
1692 unsigned generic_patch_location
= MAX_VARYING
*4;
1693 bool previous_var_xfb_only
= false;
1695 for (unsigned i
= 0; i
< this->num_matches
; i
++) {
1696 unsigned *location
= &generic_location
;
1698 const ir_variable
*var
;
1699 const glsl_type
*type
;
1700 bool is_vertex_input
= false;
1701 if (matches
[i
].consumer_var
) {
1702 var
= matches
[i
].consumer_var
;
1703 type
= get_varying_type(var
, consumer_stage
);
1704 if (consumer_stage
== MESA_SHADER_VERTEX
)
1705 is_vertex_input
= true;
1707 var
= matches
[i
].producer_var
;
1708 type
= get_varying_type(var
, producer_stage
);
1711 if (var
->data
.patch
)
1712 location
= &generic_patch_location
;
1714 /* Advance to the next slot if this varying has a different packing
1715 * class than the previous one, and we're not already on a slot
1718 * Also advance to the next slot if packing is disabled. This makes sure
1719 * we don't assign varyings the same locations which is possible
1720 * because we still pack individual arrays, records and matrices even
1721 * when packing is disabled. Note we don't advance to the next slot if
1722 * we can pack varyings together that are only used for transform
1725 if (var
->data
.must_be_shader_input
||
1726 (this->disable_varying_packing
&&
1727 !(previous_var_xfb_only
&& var
->data
.is_xfb_only
)) ||
1728 (i
> 0 && this->matches
[i
- 1].packing_class
1729 != this->matches
[i
].packing_class
)) {
1730 *location
= ALIGN(*location
, 4);
1733 previous_var_xfb_only
= var
->data
.is_xfb_only
;
1735 /* The number of components taken up by this variable. For vertex shader
1736 * inputs, we use the number of slots * 4, as they have different
1739 unsigned num_components
= is_vertex_input
?
1740 type
->count_attribute_slots(is_vertex_input
) * 4 :
1741 this->matches
[i
].num_components
;
1743 /* The last slot for this variable, inclusive. */
1744 unsigned slot_end
= *location
+ num_components
- 1;
1746 /* FIXME: We could be smarter in the below code and loop back over
1747 * trying to fill any locations that we skipped because we couldn't pack
1748 * the varying between an explicit location. For now just let the user
1749 * hit the linking error if we run out of room and suggest they use
1750 * explicit locations.
1752 while (slot_end
< MAX_VARYING
* 4u) {
1753 const unsigned slots
= (slot_end
/ 4u) - (*location
/ 4u) + 1;
1754 const uint64_t slot_mask
= ((1ull << slots
) - 1) << (*location
/ 4u);
1757 if (reserved_slots
& slot_mask
) {
1758 *location
= ALIGN(*location
+ 1, 4);
1759 slot_end
= *location
+ num_components
- 1;
1766 if (!var
->data
.patch
&& slot_end
>= MAX_VARYING
* 4u) {
1767 linker_error(prog
, "insufficient contiguous locations available for "
1768 "%s it is possible an array or struct could not be "
1769 "packed between varyings with explicit locations. Try "
1770 "using an explicit location for arrays and structs.",
1774 if (slot_end
< MAX_VARYINGS_INCL_PATCH
* 4u) {
1775 for (unsigned j
= *location
/ 4u; j
< slot_end
/ 4u; j
++)
1777 components
[slot_end
/ 4u] = (slot_end
& 3) + 1;
1780 this->matches
[i
].generic_location
= *location
;
1782 *location
= slot_end
+ 1;
1785 return (generic_location
+ 3) / 4;
1790 * Update the producer and consumer shaders to reflect the locations
1791 * assignments that were made by varying_matches::assign_locations().
1794 varying_matches::store_locations() const
1796 /* Check is location needs to be packed with lower_packed_varyings() or if
1797 * we can just use ARB_enhanced_layouts packing.
1799 bool pack_loc
[MAX_VARYINGS_INCL_PATCH
] = { 0 };
1800 const glsl_type
*loc_type
[MAX_VARYINGS_INCL_PATCH
][4] = { {NULL
, NULL
} };
1802 for (unsigned i
= 0; i
< this->num_matches
; i
++) {
1803 ir_variable
*producer_var
= this->matches
[i
].producer_var
;
1804 ir_variable
*consumer_var
= this->matches
[i
].consumer_var
;
1805 unsigned generic_location
= this->matches
[i
].generic_location
;
1806 unsigned slot
= generic_location
/ 4;
1807 unsigned offset
= generic_location
% 4;
1810 producer_var
->data
.location
= VARYING_SLOT_VAR0
+ slot
;
1811 producer_var
->data
.location_frac
= offset
;
1815 assert(consumer_var
->data
.location
== -1);
1816 consumer_var
->data
.location
= VARYING_SLOT_VAR0
+ slot
;
1817 consumer_var
->data
.location_frac
= offset
;
1820 /* Find locations suitable for native packing via
1821 * ARB_enhanced_layouts.
1823 if (producer_var
&& consumer_var
) {
1824 if (enhanced_layouts_enabled
) {
1825 const glsl_type
*type
=
1826 get_varying_type(producer_var
, producer_stage
);
1827 if (type
->is_array() || type
->is_matrix() || type
->is_record() ||
1828 type
->is_double()) {
1829 unsigned comp_slots
= type
->component_slots() + offset
;
1830 unsigned slots
= comp_slots
/ 4;
1834 for (unsigned j
= 0; j
< slots
; j
++) {
1835 pack_loc
[slot
+ j
] = true;
1837 } else if (offset
+ type
->vector_elements
> 4) {
1838 pack_loc
[slot
] = true;
1839 pack_loc
[slot
+ 1] = true;
1841 loc_type
[slot
][offset
] = type
;
1847 /* Attempt to use ARB_enhanced_layouts for more efficient packing if
1850 if (enhanced_layouts_enabled
) {
1851 for (unsigned i
= 0; i
< this->num_matches
; i
++) {
1852 ir_variable
*producer_var
= this->matches
[i
].producer_var
;
1853 ir_variable
*consumer_var
= this->matches
[i
].consumer_var
;
1854 unsigned generic_location
= this->matches
[i
].generic_location
;
1855 unsigned slot
= generic_location
/ 4;
1857 if (pack_loc
[slot
] || !producer_var
|| !consumer_var
)
1860 const glsl_type
*type
=
1861 get_varying_type(producer_var
, producer_stage
);
1862 bool type_match
= true;
1863 for (unsigned j
= 0; j
< 4; j
++) {
1864 if (loc_type
[slot
][j
]) {
1865 if (type
->base_type
!= loc_type
[slot
][j
]->base_type
)
1871 producer_var
->data
.explicit_location
= 1;
1872 consumer_var
->data
.explicit_location
= 1;
1873 producer_var
->data
.explicit_component
= 1;
1874 consumer_var
->data
.explicit_component
= 1;
1882 * Compute the "packing class" of the given varying. This is an unsigned
1883 * integer with the property that two variables in the same packing class can
1884 * be safely backed into the same vec4.
1887 varying_matches::compute_packing_class(const ir_variable
*var
)
1889 /* Without help from the back-end, there is no way to pack together
1890 * variables with different interpolation types, because
1891 * lower_packed_varyings must choose exactly one interpolation type for
1892 * each packed varying it creates.
1894 * However, we can safely pack together floats, ints, and uints, because:
1896 * - varyings of base type "int" and "uint" must use the "flat"
1897 * interpolation type, which can only occur in GLSL 1.30 and above.
1899 * - On platforms that support GLSL 1.30 and above, lower_packed_varyings
1900 * can store flat floats as ints without losing any information (using
1901 * the ir_unop_bitcast_* opcodes).
1903 * Therefore, the packing class depends only on the interpolation type.
1905 unsigned packing_class
= var
->data
.centroid
| (var
->data
.sample
<< 1) |
1906 (var
->data
.patch
<< 2) |
1907 (var
->data
.must_be_shader_input
<< 3);
1909 packing_class
+= var
->is_interpolation_flat()
1910 ? unsigned(INTERP_MODE_FLAT
) : var
->data
.interpolation
;
1911 return packing_class
;
1916 * Compute the "packing order" of the given varying. This is a sort key we
1917 * use to determine when to attempt to pack the given varying relative to
1918 * other varyings in the same packing class.
1920 varying_matches::packing_order_enum
1921 varying_matches::compute_packing_order(const ir_variable
*var
)
1923 const glsl_type
*element_type
= var
->type
;
1925 while (element_type
->is_array()) {
1926 element_type
= element_type
->fields
.array
;
1929 switch (element_type
->component_slots() % 4) {
1930 case 1: return PACKING_ORDER_SCALAR
;
1931 case 2: return PACKING_ORDER_VEC2
;
1932 case 3: return PACKING_ORDER_VEC3
;
1933 case 0: return PACKING_ORDER_VEC4
;
1935 assert(!"Unexpected value of vector_elements");
1936 return PACKING_ORDER_VEC4
;
1942 * Comparison function passed to qsort() to sort varyings by packing_class and
1943 * then by packing_order.
1946 varying_matches::match_comparator(const void *x_generic
, const void *y_generic
)
1948 const match
*x
= (const match
*) x_generic
;
1949 const match
*y
= (const match
*) y_generic
;
1951 if (x
->packing_class
!= y
->packing_class
)
1952 return x
->packing_class
- y
->packing_class
;
1953 return x
->packing_order
- y
->packing_order
;
1958 * Comparison function passed to qsort() to sort varyings used only by
1959 * transform feedback when packing of other varyings is disabled.
1962 varying_matches::xfb_comparator(const void *x_generic
, const void *y_generic
)
1964 const match
*x
= (const match
*) x_generic
;
1966 if (x
->producer_var
!= NULL
&& x
->producer_var
->data
.is_xfb_only
)
1967 return match_comparator(x_generic
, y_generic
);
1969 /* FIXME: When the comparator returns 0 it means the elements being
1970 * compared are equivalent. However the qsort documentation says:
1972 * "The order of equivalent elements is undefined."
1974 * In practice the sort ends up reversing the order of the varyings which
1975 * means locations are also assigned in this reversed order and happens to
1976 * be what we want. This is also whats happening in
1977 * varying_matches::match_comparator().
1984 * Is the given variable a varying variable to be counted against the
1985 * limit in ctx->Const.MaxVarying?
1986 * This includes variables such as texcoords, colors and generic
1987 * varyings, but excludes variables such as gl_FrontFacing and gl_FragCoord.
1990 var_counts_against_varying_limit(gl_shader_stage stage
, const ir_variable
*var
)
1992 /* Only fragment shaders will take a varying variable as an input */
1993 if (stage
== MESA_SHADER_FRAGMENT
&&
1994 var
->data
.mode
== ir_var_shader_in
) {
1995 switch (var
->data
.location
) {
1996 case VARYING_SLOT_POS
:
1997 case VARYING_SLOT_FACE
:
1998 case VARYING_SLOT_PNTC
:
2009 * Visitor class that generates tfeedback_candidate structs describing all
2010 * possible targets of transform feedback.
2012 * tfeedback_candidate structs are stored in the hash table
2013 * tfeedback_candidates, which is passed to the constructor. This hash table
2014 * maps varying names to instances of the tfeedback_candidate struct.
2016 class tfeedback_candidate_generator
: public program_resource_visitor
2019 tfeedback_candidate_generator(void *mem_ctx
,
2020 hash_table
*tfeedback_candidates
)
2022 tfeedback_candidates(tfeedback_candidates
),
2028 void process(ir_variable
*var
)
2030 /* All named varying interface blocks should be flattened by now */
2031 assert(!var
->is_interface_instance());
2033 this->toplevel_var
= var
;
2034 this->varying_floats
= 0;
2035 program_resource_visitor::process(var
, false);
2039 virtual void visit_field(const glsl_type
*type
, const char *name
,
2040 bool /* row_major */,
2041 const glsl_type
* /* record_type */,
2042 const enum glsl_interface_packing
,
2043 bool /* last_field */)
2045 assert(!type
->without_array()->is_record());
2046 assert(!type
->without_array()->is_interface());
2048 tfeedback_candidate
*candidate
2049 = rzalloc(this->mem_ctx
, tfeedback_candidate
);
2050 candidate
->toplevel_var
= this->toplevel_var
;
2051 candidate
->type
= type
;
2052 candidate
->offset
= this->varying_floats
;
2053 _mesa_hash_table_insert(this->tfeedback_candidates
,
2054 ralloc_strdup(this->mem_ctx
, name
),
2056 this->varying_floats
+= type
->component_slots();
2060 * Memory context used to allocate hash table keys and values.
2062 void * const mem_ctx
;
2065 * Hash table in which tfeedback_candidate objects should be stored.
2067 hash_table
* const tfeedback_candidates
;
2070 * Pointer to the toplevel variable that is being traversed.
2072 ir_variable
*toplevel_var
;
2075 * Total number of varying floats that have been visited so far. This is
2076 * used to determine the offset to each varying within the toplevel
2079 unsigned varying_floats
;
2086 populate_consumer_input_sets(void *mem_ctx
, exec_list
*ir
,
2087 hash_table
*consumer_inputs
,
2088 hash_table
*consumer_interface_inputs
,
2089 ir_variable
*consumer_inputs_with_locations
[VARYING_SLOT_TESS_MAX
])
2091 memset(consumer_inputs_with_locations
,
2093 sizeof(consumer_inputs_with_locations
[0]) * VARYING_SLOT_TESS_MAX
);
2095 foreach_in_list(ir_instruction
, node
, ir
) {
2096 ir_variable
*const input_var
= node
->as_variable();
2098 if (input_var
!= NULL
&& input_var
->data
.mode
== ir_var_shader_in
) {
2099 /* All interface blocks should have been lowered by this point */
2100 assert(!input_var
->type
->is_interface());
2102 if (input_var
->data
.explicit_location
) {
2103 /* assign_varying_locations only cares about finding the
2104 * ir_variable at the start of a contiguous location block.
2106 * - For !producer, consumer_inputs_with_locations isn't used.
2108 * - For !consumer, consumer_inputs_with_locations is empty.
2110 * For consumer && producer, if you were trying to set some
2111 * ir_variable to the middle of a location block on the other side
2112 * of producer/consumer, cross_validate_outputs_to_inputs() should
2113 * be link-erroring due to either type mismatch or location
2114 * overlaps. If the variables do match up, then they've got a
2115 * matching data.location and you only looked at
2116 * consumer_inputs_with_locations[var->data.location], not any
2117 * following entries for the array/structure.
2119 consumer_inputs_with_locations
[input_var
->data
.location
] =
2121 } else if (input_var
->get_interface_type() != NULL
) {
2122 char *const iface_field_name
=
2123 ralloc_asprintf(mem_ctx
, "%s.%s",
2124 input_var
->get_interface_type()->without_array()->name
,
2126 _mesa_hash_table_insert(consumer_interface_inputs
,
2127 iface_field_name
, input_var
);
2129 _mesa_hash_table_insert(consumer_inputs
,
2130 ralloc_strdup(mem_ctx
, input_var
->name
),
2138 * Find a variable from the consumer that "matches" the specified variable
2140 * This function only finds inputs with names that match. There is no
2141 * validation (here) that the types, etc. are compatible.
2144 get_matching_input(void *mem_ctx
,
2145 const ir_variable
*output_var
,
2146 hash_table
*consumer_inputs
,
2147 hash_table
*consumer_interface_inputs
,
2148 ir_variable
*consumer_inputs_with_locations
[VARYING_SLOT_TESS_MAX
])
2150 ir_variable
*input_var
;
2152 if (output_var
->data
.explicit_location
) {
2153 input_var
= consumer_inputs_with_locations
[output_var
->data
.location
];
2154 } else if (output_var
->get_interface_type() != NULL
) {
2155 char *const iface_field_name
=
2156 ralloc_asprintf(mem_ctx
, "%s.%s",
2157 output_var
->get_interface_type()->without_array()->name
,
2159 hash_entry
*entry
= _mesa_hash_table_search(consumer_interface_inputs
, iface_field_name
);
2160 input_var
= entry
? (ir_variable
*) entry
->data
: NULL
;
2162 hash_entry
*entry
= _mesa_hash_table_search(consumer_inputs
, output_var
->name
);
2163 input_var
= entry
? (ir_variable
*) entry
->data
: NULL
;
2166 return (input_var
== NULL
|| input_var
->data
.mode
!= ir_var_shader_in
)
2173 io_variable_cmp(const void *_a
, const void *_b
)
2175 const ir_variable
*const a
= *(const ir_variable
**) _a
;
2176 const ir_variable
*const b
= *(const ir_variable
**) _b
;
2178 if (a
->data
.explicit_location
&& b
->data
.explicit_location
)
2179 return b
->data
.location
- a
->data
.location
;
2181 if (a
->data
.explicit_location
&& !b
->data
.explicit_location
)
2184 if (!a
->data
.explicit_location
&& b
->data
.explicit_location
)
2187 return -strcmp(a
->name
, b
->name
);
2191 * Sort the shader IO variables into canonical order
2194 canonicalize_shader_io(exec_list
*ir
, enum ir_variable_mode io_mode
)
2196 ir_variable
*var_table
[MAX_PROGRAM_OUTPUTS
* 4];
2197 unsigned num_variables
= 0;
2199 foreach_in_list(ir_instruction
, node
, ir
) {
2200 ir_variable
*const var
= node
->as_variable();
2202 if (var
== NULL
|| var
->data
.mode
!= io_mode
)
2205 /* If we have already encountered more I/O variables that could
2206 * successfully link, bail.
2208 if (num_variables
== ARRAY_SIZE(var_table
))
2211 var_table
[num_variables
++] = var
;
2214 if (num_variables
== 0)
2217 /* Sort the list in reverse order (io_variable_cmp handles this). Later
2218 * we're going to push the variables on to the IR list as a stack, so we
2219 * want the last variable (in canonical order) to be first in the list.
2221 qsort(var_table
, num_variables
, sizeof(var_table
[0]), io_variable_cmp
);
2223 /* Remove the variable from it's current location in the IR, and put it at
2226 for (unsigned i
= 0; i
< num_variables
; i
++) {
2227 var_table
[i
]->remove();
2228 ir
->push_head(var_table
[i
]);
2233 * Generate a bitfield map of the explicit locations for shader varyings.
2235 * Note: For Tessellation shaders we are sitting right on the limits of the
2236 * 64 bit map. Per-vertex and per-patch both have separate location domains
2237 * with a max of MAX_VARYING.
2240 reserved_varying_slot(struct gl_linked_shader
*stage
,
2241 ir_variable_mode io_mode
)
2243 assert(io_mode
== ir_var_shader_in
|| io_mode
== ir_var_shader_out
);
2244 /* Avoid an overflow of the returned value */
2245 assert(MAX_VARYINGS_INCL_PATCH
<= 64);
2253 foreach_in_list(ir_instruction
, node
, stage
->ir
) {
2254 ir_variable
*const var
= node
->as_variable();
2256 if (var
== NULL
|| var
->data
.mode
!= io_mode
||
2257 !var
->data
.explicit_location
||
2258 var
->data
.location
< VARYING_SLOT_VAR0
)
2261 var_slot
= var
->data
.location
- VARYING_SLOT_VAR0
;
2263 unsigned num_elements
= get_varying_type(var
, stage
->Stage
)
2264 ->count_attribute_slots(io_mode
== ir_var_shader_in
&&
2265 stage
->Stage
== MESA_SHADER_VERTEX
);
2266 for (unsigned i
= 0; i
< num_elements
; i
++) {
2267 if (var_slot
>= 0 && var_slot
< MAX_VARYINGS_INCL_PATCH
)
2268 slots
|= UINT64_C(1) << var_slot
;
2278 * Assign locations for all variables that are produced in one pipeline stage
2279 * (the "producer") and consumed in the next stage (the "consumer").
2281 * Variables produced by the producer may also be consumed by transform
2284 * \param num_tfeedback_decls is the number of declarations indicating
2285 * variables that may be consumed by transform feedback.
2287 * \param tfeedback_decls is a pointer to an array of tfeedback_decl objects
2288 * representing the result of parsing the strings passed to
2289 * glTransformFeedbackVaryings(). assign_location() will be called for
2290 * each of these objects that matches one of the outputs of the
2293 * When num_tfeedback_decls is nonzero, it is permissible for the consumer to
2294 * be NULL. In this case, varying locations are assigned solely based on the
2295 * requirements of transform feedback.
2298 assign_varying_locations(struct gl_context
*ctx
,
2300 struct gl_shader_program
*prog
,
2301 gl_linked_shader
*producer
,
2302 gl_linked_shader
*consumer
,
2303 unsigned num_tfeedback_decls
,
2304 tfeedback_decl
*tfeedback_decls
,
2305 const uint64_t reserved_slots
)
2307 /* Tessellation shaders treat inputs and outputs as shared memory and can
2308 * access inputs and outputs of other invocations.
2309 * Therefore, they can't be lowered to temps easily (and definitely not
2312 bool unpackable_tess
=
2313 (consumer
&& consumer
->Stage
== MESA_SHADER_TESS_EVAL
) ||
2314 (consumer
&& consumer
->Stage
== MESA_SHADER_TESS_CTRL
) ||
2315 (producer
&& producer
->Stage
== MESA_SHADER_TESS_CTRL
);
2317 /* Transform feedback code assumes varying arrays are packed, so if the
2318 * driver has disabled varying packing, make sure to at least enable
2319 * packing required by transform feedback.
2322 ctx
->Extensions
.EXT_transform_feedback
&& !unpackable_tess
;
2324 /* Disable packing on outward facing interfaces for SSO because in ES we
2325 * need to retain the unpacked varying information for draw time
2328 * Packing is still enabled on individual arrays, structs, and matrices as
2329 * these are required by the transform feedback code and it is still safe
2330 * to do so. We also enable packing when a varying is only used for
2331 * transform feedback and its not a SSO.
2333 bool disable_varying_packing
=
2334 ctx
->Const
.DisableVaryingPacking
|| unpackable_tess
;
2335 if (prog
->SeparateShader
&& (producer
== NULL
|| consumer
== NULL
))
2336 disable_varying_packing
= true;
2338 varying_matches
matches(disable_varying_packing
, xfb_enabled
,
2339 ctx
->Extensions
.ARB_enhanced_layouts
,
2340 producer
? producer
->Stage
: MESA_SHADER_NONE
,
2341 consumer
? consumer
->Stage
: MESA_SHADER_NONE
);
2342 hash_table
*tfeedback_candidates
=
2343 _mesa_hash_table_create(NULL
, _mesa_key_hash_string
,
2344 _mesa_key_string_equal
);
2345 hash_table
*consumer_inputs
=
2346 _mesa_hash_table_create(NULL
, _mesa_key_hash_string
,
2347 _mesa_key_string_equal
);
2348 hash_table
*consumer_interface_inputs
=
2349 _mesa_hash_table_create(NULL
, _mesa_key_hash_string
,
2350 _mesa_key_string_equal
);
2351 ir_variable
*consumer_inputs_with_locations
[VARYING_SLOT_TESS_MAX
] = {
2355 unsigned consumer_vertices
= 0;
2356 if (consumer
&& consumer
->Stage
== MESA_SHADER_GEOMETRY
)
2357 consumer_vertices
= prog
->Geom
.VerticesIn
;
2359 /* Operate in a total of four passes.
2361 * 1. Sort inputs / outputs into a canonical order. This is necessary so
2362 * that inputs / outputs of separable shaders will be assigned
2363 * predictable locations regardless of the order in which declarations
2364 * appeared in the shader source.
2366 * 2. Assign locations for any matching inputs and outputs.
2368 * 3. Mark output variables in the producer that do not have locations as
2369 * not being outputs. This lets the optimizer eliminate them.
2371 * 4. Mark input variables in the consumer that do not have locations as
2372 * not being inputs. This lets the optimizer eliminate them.
2375 canonicalize_shader_io(consumer
->ir
, ir_var_shader_in
);
2378 canonicalize_shader_io(producer
->ir
, ir_var_shader_out
);
2381 linker::populate_consumer_input_sets(mem_ctx
, consumer
->ir
,
2383 consumer_interface_inputs
,
2384 consumer_inputs_with_locations
);
2387 foreach_in_list(ir_instruction
, node
, producer
->ir
) {
2388 ir_variable
*const output_var
= node
->as_variable();
2390 if (output_var
== NULL
|| output_var
->data
.mode
!= ir_var_shader_out
)
2393 /* Only geometry shaders can use non-zero streams */
2394 assert(output_var
->data
.stream
== 0 ||
2395 (output_var
->data
.stream
< MAX_VERTEX_STREAMS
&&
2396 producer
->Stage
== MESA_SHADER_GEOMETRY
));
2398 if (num_tfeedback_decls
> 0) {
2399 tfeedback_candidate_generator
g(mem_ctx
, tfeedback_candidates
);
2400 g
.process(output_var
);
2403 ir_variable
*const input_var
=
2404 linker::get_matching_input(mem_ctx
, output_var
, consumer_inputs
,
2405 consumer_interface_inputs
,
2406 consumer_inputs_with_locations
);
2408 /* If a matching input variable was found, add this output (and the
2409 * input) to the set. If this is a separable program and there is no
2410 * consumer stage, add the output.
2412 * Always add TCS outputs. They are shared by all invocations
2413 * within a patch and can be used as shared memory.
2415 if (input_var
|| (prog
->SeparateShader
&& consumer
== NULL
) ||
2416 producer
->Stage
== MESA_SHADER_TESS_CTRL
) {
2417 matches
.record(output_var
, input_var
);
2420 /* Only stream 0 outputs can be consumed in the next stage */
2421 if (input_var
&& output_var
->data
.stream
!= 0) {
2422 linker_error(prog
, "output %s is assigned to stream=%d but "
2423 "is linked to an input, which requires stream=0",
2424 output_var
->name
, output_var
->data
.stream
);
2429 /* If there's no producer stage, then this must be a separable program.
2430 * For example, we may have a program that has just a fragment shader.
2431 * Later this program will be used with some arbitrary vertex (or
2432 * geometry) shader program. This means that locations must be assigned
2433 * for all the inputs.
2435 foreach_in_list(ir_instruction
, node
, consumer
->ir
) {
2436 ir_variable
*const input_var
= node
->as_variable();
2438 if (input_var
== NULL
|| input_var
->data
.mode
!= ir_var_shader_in
)
2441 matches
.record(NULL
, input_var
);
2445 for (unsigned i
= 0; i
< num_tfeedback_decls
; ++i
) {
2446 if (!tfeedback_decls
[i
].is_varying())
2449 const tfeedback_candidate
*matched_candidate
2450 = tfeedback_decls
[i
].find_candidate(prog
, tfeedback_candidates
);
2452 if (matched_candidate
== NULL
) {
2453 _mesa_hash_table_destroy(tfeedback_candidates
, NULL
);
2457 /* Mark xfb varyings as always active */
2458 matched_candidate
->toplevel_var
->data
.always_active_io
= 1;
2460 /* Mark any corresponding inputs as always active also. We must do this
2461 * because we have a NIR pass that lowers vectors to scalars and another
2462 * that removes unused varyings.
2463 * We don't split varyings marked as always active because there is no
2464 * point in doing so. This means we need to mark both sides of the
2465 * interface as always active otherwise we will have a mismatch and
2466 * start removing things we shouldn't.
2468 ir_variable
*const input_var
=
2469 linker::get_matching_input(mem_ctx
, matched_candidate
->toplevel_var
,
2471 consumer_interface_inputs
,
2472 consumer_inputs_with_locations
);
2474 input_var
->data
.always_active_io
= 1;
2476 if (matched_candidate
->toplevel_var
->data
.is_unmatched_generic_inout
) {
2477 matched_candidate
->toplevel_var
->data
.is_xfb_only
= 1;
2478 matches
.record(matched_candidate
->toplevel_var
, NULL
);
2482 _mesa_hash_table_destroy(consumer_inputs
, NULL
);
2483 _mesa_hash_table_destroy(consumer_interface_inputs
, NULL
);
2485 uint8_t components
[MAX_VARYINGS_INCL_PATCH
] = {0};
2486 const unsigned slots_used
= matches
.assign_locations(
2487 prog
, components
, reserved_slots
);
2488 matches
.store_locations();
2490 for (unsigned i
= 0; i
< num_tfeedback_decls
; ++i
) {
2491 if (!tfeedback_decls
[i
].is_varying())
2494 if (!tfeedback_decls
[i
].assign_location(ctx
, prog
)) {
2495 _mesa_hash_table_destroy(tfeedback_candidates
, NULL
);
2499 _mesa_hash_table_destroy(tfeedback_candidates
, NULL
);
2501 if (consumer
&& producer
) {
2502 foreach_in_list(ir_instruction
, node
, consumer
->ir
) {
2503 ir_variable
*const var
= node
->as_variable();
2505 if (var
&& var
->data
.mode
== ir_var_shader_in
&&
2506 var
->data
.is_unmatched_generic_inout
) {
2507 if (!prog
->IsES
&& prog
->data
->Version
<= 120) {
2508 /* On page 25 (page 31 of the PDF) of the GLSL 1.20 spec:
2510 * Only those varying variables used (i.e. read) in
2511 * the fragment shader executable must be written to
2512 * by the vertex shader executable; declaring
2513 * superfluous varying variables in a vertex shader is
2516 * We interpret this text as meaning that the VS must
2517 * write the variable for the FS to read it. See
2518 * "glsl1-varying read but not written" in piglit.
2520 linker_error(prog
, "%s shader varying %s not written "
2522 _mesa_shader_stage_to_string(consumer
->Stage
),
2524 _mesa_shader_stage_to_string(producer
->Stage
));
2526 linker_warning(prog
, "%s shader varying %s not written "
2528 _mesa_shader_stage_to_string(consumer
->Stage
),
2530 _mesa_shader_stage_to_string(producer
->Stage
));
2535 /* Now that validation is done its safe to remove unused varyings. As
2536 * we have both a producer and consumer its safe to remove unused
2537 * varyings even if the program is a SSO because the stages are being
2538 * linked together i.e. we have a multi-stage SSO.
2540 remove_unused_shader_inputs_and_outputs(false, producer
,
2542 remove_unused_shader_inputs_and_outputs(false, consumer
,
2547 lower_packed_varyings(mem_ctx
, slots_used
, components
, ir_var_shader_out
,
2548 0, producer
, disable_varying_packing
,
2553 lower_packed_varyings(mem_ctx
, slots_used
, components
, ir_var_shader_in
,
2554 consumer_vertices
, consumer
,
2555 disable_varying_packing
, xfb_enabled
);
2562 check_against_output_limit(struct gl_context
*ctx
,
2563 struct gl_shader_program
*prog
,
2564 gl_linked_shader
*producer
,
2565 unsigned num_explicit_locations
)
2567 unsigned output_vectors
= num_explicit_locations
;
2569 foreach_in_list(ir_instruction
, node
, producer
->ir
) {
2570 ir_variable
*const var
= node
->as_variable();
2572 if (var
&& !var
->data
.explicit_location
&&
2573 var
->data
.mode
== ir_var_shader_out
&&
2574 var_counts_against_varying_limit(producer
->Stage
, var
)) {
2575 /* outputs for fragment shader can't be doubles */
2576 output_vectors
+= var
->type
->count_attribute_slots(false);
2580 assert(producer
->Stage
!= MESA_SHADER_FRAGMENT
);
2581 unsigned max_output_components
=
2582 ctx
->Const
.Program
[producer
->Stage
].MaxOutputComponents
;
2584 const unsigned output_components
= output_vectors
* 4;
2585 if (output_components
> max_output_components
) {
2586 if (ctx
->API
== API_OPENGLES2
|| prog
->IsES
)
2587 linker_error(prog
, "%s shader uses too many output vectors "
2589 _mesa_shader_stage_to_string(producer
->Stage
),
2591 max_output_components
/ 4);
2593 linker_error(prog
, "%s shader uses too many output components "
2595 _mesa_shader_stage_to_string(producer
->Stage
),
2597 max_output_components
);
2606 check_against_input_limit(struct gl_context
*ctx
,
2607 struct gl_shader_program
*prog
,
2608 gl_linked_shader
*consumer
,
2609 unsigned num_explicit_locations
)
2611 unsigned input_vectors
= num_explicit_locations
;
2613 foreach_in_list(ir_instruction
, node
, consumer
->ir
) {
2614 ir_variable
*const var
= node
->as_variable();
2616 if (var
&& !var
->data
.explicit_location
&&
2617 var
->data
.mode
== ir_var_shader_in
&&
2618 var_counts_against_varying_limit(consumer
->Stage
, var
)) {
2619 /* vertex inputs aren't varying counted */
2620 input_vectors
+= var
->type
->count_attribute_slots(false);
2624 assert(consumer
->Stage
!= MESA_SHADER_VERTEX
);
2625 unsigned max_input_components
=
2626 ctx
->Const
.Program
[consumer
->Stage
].MaxInputComponents
;
2628 const unsigned input_components
= input_vectors
* 4;
2629 if (input_components
> max_input_components
) {
2630 if (ctx
->API
== API_OPENGLES2
|| prog
->IsES
)
2631 linker_error(prog
, "%s shader uses too many input vectors "
2633 _mesa_shader_stage_to_string(consumer
->Stage
),
2635 max_input_components
/ 4);
2637 linker_error(prog
, "%s shader uses too many input components "
2639 _mesa_shader_stage_to_string(consumer
->Stage
),
2641 max_input_components
);
2650 link_varyings(struct gl_shader_program
*prog
, unsigned first
, unsigned last
,
2651 struct gl_context
*ctx
, void *mem_ctx
)
2653 bool has_xfb_qualifiers
= false;
2654 unsigned num_tfeedback_decls
= 0;
2655 char **varying_names
= NULL
;
2656 tfeedback_decl
*tfeedback_decls
= NULL
;
2658 /* From the ARB_enhanced_layouts spec:
2660 * "If the shader used to record output variables for transform feedback
2661 * varyings uses the "xfb_buffer", "xfb_offset", or "xfb_stride" layout
2662 * qualifiers, the values specified by TransformFeedbackVaryings are
2663 * ignored, and the set of variables captured for transform feedback is
2664 * instead derived from the specified layout qualifiers."
2666 for (int i
= MESA_SHADER_FRAGMENT
- 1; i
>= 0; i
--) {
2667 /* Find last stage before fragment shader */
2668 if (prog
->_LinkedShaders
[i
]) {
2669 has_xfb_qualifiers
=
2670 process_xfb_layout_qualifiers(mem_ctx
, prog
->_LinkedShaders
[i
],
2671 prog
, &num_tfeedback_decls
,
2677 if (!has_xfb_qualifiers
) {
2678 num_tfeedback_decls
= prog
->TransformFeedback
.NumVarying
;
2679 varying_names
= prog
->TransformFeedback
.VaryingNames
;
2682 if (num_tfeedback_decls
!= 0) {
2683 /* From GL_EXT_transform_feedback:
2684 * A program will fail to link if:
2686 * * the <count> specified by TransformFeedbackVaryingsEXT is
2687 * non-zero, but the program object has no vertex or geometry
2690 if (first
>= MESA_SHADER_FRAGMENT
) {
2691 linker_error(prog
, "Transform feedback varyings specified, but "
2692 "no vertex, tessellation, or geometry shader is "
2697 tfeedback_decls
= rzalloc_array(mem_ctx
, tfeedback_decl
,
2698 num_tfeedback_decls
);
2699 if (!parse_tfeedback_decls(ctx
, prog
, mem_ctx
, num_tfeedback_decls
,
2700 varying_names
, tfeedback_decls
))
2704 /* If there is no fragment shader we need to set transform feedback.
2706 * For SSO we also need to assign output locations. We assign them here
2707 * because we need to do it for both single stage programs and multi stage
2710 if (last
< MESA_SHADER_FRAGMENT
&&
2711 (num_tfeedback_decls
!= 0 || prog
->SeparateShader
)) {
2712 const uint64_t reserved_out_slots
=
2713 reserved_varying_slot(prog
->_LinkedShaders
[last
], ir_var_shader_out
);
2714 if (!assign_varying_locations(ctx
, mem_ctx
, prog
,
2715 prog
->_LinkedShaders
[last
], NULL
,
2716 num_tfeedback_decls
, tfeedback_decls
,
2717 reserved_out_slots
))
2721 if (last
<= MESA_SHADER_FRAGMENT
) {
2722 /* Remove unused varyings from the first/last stage unless SSO */
2723 remove_unused_shader_inputs_and_outputs(prog
->SeparateShader
,
2724 prog
->_LinkedShaders
[first
],
2726 remove_unused_shader_inputs_and_outputs(prog
->SeparateShader
,
2727 prog
->_LinkedShaders
[last
],
2730 /* If the program is made up of only a single stage */
2731 if (first
== last
) {
2732 gl_linked_shader
*const sh
= prog
->_LinkedShaders
[last
];
2734 do_dead_builtin_varyings(ctx
, NULL
, sh
, 0, NULL
);
2735 do_dead_builtin_varyings(ctx
, sh
, NULL
, num_tfeedback_decls
,
2738 if (prog
->SeparateShader
) {
2739 const uint64_t reserved_slots
=
2740 reserved_varying_slot(sh
, ir_var_shader_in
);
2742 /* Assign input locations for SSO, output locations are already
2745 if (!assign_varying_locations(ctx
, mem_ctx
, prog
,
2746 NULL
/* producer */,
2748 0 /* num_tfeedback_decls */,
2749 NULL
/* tfeedback_decls */,
2754 /* Linking the stages in the opposite order (from fragment to vertex)
2755 * ensures that inter-shader outputs written to in an earlier stage
2756 * are eliminated if they are (transitively) not used in a later
2760 for (int i
= next
- 1; i
>= 0; i
--) {
2761 if (prog
->_LinkedShaders
[i
] == NULL
&& i
!= 0)
2764 gl_linked_shader
*const sh_i
= prog
->_LinkedShaders
[i
];
2765 gl_linked_shader
*const sh_next
= prog
->_LinkedShaders
[next
];
2767 const uint64_t reserved_out_slots
=
2768 reserved_varying_slot(sh_i
, ir_var_shader_out
);
2769 const uint64_t reserved_in_slots
=
2770 reserved_varying_slot(sh_next
, ir_var_shader_in
);
2772 do_dead_builtin_varyings(ctx
, sh_i
, sh_next
,
2773 next
== MESA_SHADER_FRAGMENT
? num_tfeedback_decls
: 0,
2776 if (!assign_varying_locations(ctx
, mem_ctx
, prog
, sh_i
, sh_next
,
2777 next
== MESA_SHADER_FRAGMENT
? num_tfeedback_decls
: 0,
2779 reserved_out_slots
| reserved_in_slots
))
2782 /* This must be done after all dead varyings are eliminated. */
2784 unsigned slots_used
= _mesa_bitcount_64(reserved_out_slots
);
2785 if (!check_against_output_limit(ctx
, prog
, sh_i
, slots_used
)) {
2790 unsigned slots_used
= _mesa_bitcount_64(reserved_in_slots
);
2791 if (!check_against_input_limit(ctx
, prog
, sh_next
, slots_used
))
2799 if (!store_tfeedback_info(ctx
, prog
, num_tfeedback_decls
, tfeedback_decls
,
2800 has_xfb_qualifiers
))