2 * Copyright © 2011 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 lower_varyings_to_packed.cpp
27 * This lowering pass generates GLSL code that manually packs varyings into
28 * vec4 slots, for the benefit of back-ends that don't support packed varyings
31 * For example, the following shader:
33 * out mat3x2 foo; // location=4, location_frac=0
34 * out vec3 bar[2]; // location=5, location_frac=2
45 * out vec4 packed4; // location=4, location_frac=0
46 * out vec4 packed5; // location=5, location_frac=0
47 * out vec4 packed6; // location=6, location_frac=0
52 * packed4.xy = foo[0];
53 * packed4.zw = foo[1];
54 * packed5.xy = foo[2];
55 * packed5.zw = bar[0].xy;
56 * packed6.x = bar[0].z;
57 * packed6.yzw = bar[1];
60 * This lowering pass properly handles "double parking" of a varying vector
61 * across two varying slots. For example, in the code above, two of the
62 * components of bar[0] are stored in packed5, and the remaining component is
65 * Note that in theory, the extra instructions may cause some loss of
66 * performance. However, hopefully in most cases the performance loss will
67 * either be absorbed by a later optimization pass, or it will be offset by
68 * memory bandwidth savings (because fewer varyings are used).
70 * This lowering pass also packs flat floats, ints, and uints together, by
71 * using ivec4 as the base type of flat "varyings", and using appropriate
72 * casts to convert floats and uints into ints.
74 * This lowering pass also handles varyings whose type is a struct or an array
75 * of struct. Structs are packed in order and with no gaps, so there may be a
76 * performance penalty due to structure elements being double-parked.
78 * Lowering of geometry shader inputs is slightly more complex, since geometry
79 * inputs are always arrays, so we need to lower arrays to arrays. For
80 * example, the following input:
86 * } arr[3]; // location=4, location_frac=0
88 * Would get lowered like this if it occurred in a fragment shader:
95 * in vec4 packed4; // location=4, location_frac=0
96 * in vec4 packed5; // location=5, location_frac=0
97 * in vec4 packed6; // location=6, location_frac=0
98 * in vec4 packed7; // location=7, location_frac=0
99 * in vec4 packed8; // location=8, location_frac=0
100 * in vec4 packed9; // location=9, location_frac=0
104 * arr[0].f = packed4.x;
105 * arr[0].v = packed4.yzw;
106 * arr[0].a[0] = packed5.xy;
107 * arr[0].a[1] = packed5.zw;
108 * arr[1].f = packed6.x;
109 * arr[1].v = packed6.yzw;
110 * arr[1].a[0] = packed7.xy;
111 * arr[1].a[1] = packed7.zw;
112 * arr[2].f = packed8.x;
113 * arr[2].v = packed8.yzw;
114 * arr[2].a[0] = packed9.xy;
115 * arr[2].a[1] = packed9.zw;
119 * But it would get lowered like this if it occurred in a geometry shader:
126 * in vec4 packed4[3]; // location=4, location_frac=0
127 * in vec4 packed5[3]; // location=5, location_frac=0
131 * arr[0].f = packed4[0].x;
132 * arr[0].v = packed4[0].yzw;
133 * arr[0].a[0] = packed5[0].xy;
134 * arr[0].a[1] = packed5[0].zw;
135 * arr[1].f = packed4[1].x;
136 * arr[1].v = packed4[1].yzw;
137 * arr[1].a[0] = packed5[1].xy;
138 * arr[1].a[1] = packed5[1].zw;
139 * arr[2].f = packed4[2].x;
140 * arr[2].v = packed4[2].yzw;
141 * arr[2].a[0] = packed5[2].xy;
142 * arr[2].a[1] = packed5[2].zw;
147 #include "glsl_symbol_table.h"
149 #include "ir_builder.h"
150 #include "ir_optimization.h"
151 #include "program/prog_instruction.h"
153 using namespace ir_builder
;
158 * Visitor that performs varying packing. For each varying declared in the
159 * shader, this visitor determines whether it needs to be packed. If so, it
160 * demotes it to an ordinary global, creates new packed varyings, and
161 * generates assignments to convert between the original varying and the
164 class lower_packed_varyings_visitor
167 lower_packed_varyings_visitor(void *mem_ctx
, unsigned locations_used
,
168 ir_variable_mode mode
,
169 unsigned gs_input_vertices
,
170 exec_list
*out_instructions
,
171 exec_list
*out_variables
);
173 void run(struct gl_shader
*shader
);
176 void bitwise_assign_pack(ir_rvalue
*lhs
, ir_rvalue
*rhs
);
177 void bitwise_assign_unpack(ir_rvalue
*lhs
, ir_rvalue
*rhs
);
178 unsigned lower_rvalue(ir_rvalue
*rvalue
, unsigned fine_location
,
179 ir_variable
*unpacked_var
, const char *name
,
180 bool gs_input_toplevel
, unsigned vertex_index
);
181 unsigned lower_arraylike(ir_rvalue
*rvalue
, unsigned array_size
,
182 unsigned fine_location
,
183 ir_variable
*unpacked_var
, const char *name
,
184 bool gs_input_toplevel
, unsigned vertex_index
);
185 ir_dereference
*get_packed_varying_deref(unsigned location
,
186 ir_variable
*unpacked_var
,
188 unsigned vertex_index
);
189 bool needs_lowering(ir_variable
*var
);
192 * Memory context used to allocate new instructions for the shader.
194 void * const mem_ctx
;
197 * Number of generic varying slots which are used by this shader. This is
198 * used to allocate temporary intermediate data structures. If any varying
199 * used by this shader has a location greater than or equal to
200 * VARYING_SLOT_VAR0 + locations_used, an assertion will fire.
202 const unsigned locations_used
;
205 * Array of pointers to the packed varyings that have been created for each
206 * generic varying slot. NULL entries in this array indicate varying slots
207 * for which a packed varying has not been created yet.
209 ir_variable
**packed_varyings
;
212 * Type of varying which is being lowered in this pass (either
213 * ir_var_shader_in or ir_var_shader_out).
215 const ir_variable_mode mode
;
218 * If we are currently lowering geometry shader inputs, the number of input
219 * vertices the geometry shader accepts. Otherwise zero.
221 const unsigned gs_input_vertices
;
224 * Exec list into which the visitor should insert the packing instructions.
225 * Caller provides this list; it should insert the instructions into the
226 * appropriate place in the shader once the visitor has finished running.
228 exec_list
*out_instructions
;
231 * Exec list into which the visitor should insert any new variables.
233 exec_list
*out_variables
;
236 } /* anonymous namespace */
238 lower_packed_varyings_visitor::lower_packed_varyings_visitor(
239 void *mem_ctx
, unsigned locations_used
, ir_variable_mode mode
,
240 unsigned gs_input_vertices
, exec_list
*out_instructions
,
241 exec_list
*out_variables
)
243 locations_used(locations_used
),
244 packed_varyings((ir_variable
**)
245 rzalloc_array_size(mem_ctx
, sizeof(*packed_varyings
),
248 gs_input_vertices(gs_input_vertices
),
249 out_instructions(out_instructions
),
250 out_variables(out_variables
)
255 lower_packed_varyings_visitor::run(struct gl_shader
*shader
)
257 foreach_in_list(ir_instruction
, node
, shader
->ir
) {
258 ir_variable
*var
= node
->as_variable();
262 if (var
->data
.mode
!= this->mode
||
263 var
->data
.location
< VARYING_SLOT_VAR0
||
264 !this->needs_lowering(var
))
267 /* This lowering pass is only capable of packing floats and ints
268 * together when their interpolation mode is "flat". Therefore, to be
269 * safe, caller should ensure that integral varyings always use flat
270 * interpolation, even when this is not required by GLSL.
272 assert(var
->data
.interpolation
== INTERP_QUALIFIER_FLAT
||
273 !var
->type
->contains_integer());
275 /* Clone the variable for program resource list before
276 * it gets modified and lost.
278 if (!shader
->packed_varyings
)
279 shader
->packed_varyings
= new (shader
) exec_list
;
281 shader
->packed_varyings
->push_tail(var
->clone(shader
, NULL
));
283 /* Change the old varying into an ordinary global. */
284 assert(var
->data
.mode
!= ir_var_temporary
);
285 var
->data
.mode
= ir_var_auto
;
287 /* Create a reference to the old varying. */
288 ir_dereference_variable
*deref
289 = new(this->mem_ctx
) ir_dereference_variable(var
);
291 /* Recursively pack or unpack it. */
292 this->lower_rvalue(deref
, var
->data
.location
* 4 + var
->data
.location_frac
, var
,
293 var
->name
, this->gs_input_vertices
!= 0, 0);
297 #define SWIZZLE_ZWZW MAKE_SWIZZLE4(SWIZZLE_Z, SWIZZLE_W, SWIZZLE_Z, SWIZZLE_W)
300 * Make an ir_assignment from \c rhs to \c lhs, performing appropriate
301 * bitcasts if necessary to match up types.
303 * This function is called when packing varyings.
306 lower_packed_varyings_visitor::bitwise_assign_pack(ir_rvalue
*lhs
,
309 if (lhs
->type
->base_type
!= rhs
->type
->base_type
) {
310 /* Since we only mix types in flat varyings, and we always store flat
311 * varyings as type ivec4, we need only produce conversions from (uint
314 assert(lhs
->type
->base_type
== GLSL_TYPE_INT
);
315 switch (rhs
->type
->base_type
) {
317 rhs
= new(this->mem_ctx
)
318 ir_expression(ir_unop_u2i
, lhs
->type
, rhs
);
320 case GLSL_TYPE_FLOAT
:
321 rhs
= new(this->mem_ctx
)
322 ir_expression(ir_unop_bitcast_f2i
, lhs
->type
, rhs
);
324 case GLSL_TYPE_DOUBLE
:
325 assert(rhs
->type
->vector_elements
<= 2);
326 if (rhs
->type
->vector_elements
== 2) {
327 ir_variable
*t
= new(mem_ctx
) ir_variable(lhs
->type
, "pack", ir_var_temporary
);
329 assert(lhs
->type
->vector_elements
== 4);
330 this->out_variables
->push_tail(t
);
331 this->out_instructions
->push_tail(
332 assign(t
, u2i(expr(ir_unop_unpack_double_2x32
, swizzle_x(rhs
->clone(mem_ctx
, NULL
)))), 0x3));
333 this->out_instructions
->push_tail(
334 assign(t
, u2i(expr(ir_unop_unpack_double_2x32
, swizzle_y(rhs
))), 0xc));
337 rhs
= u2i(expr(ir_unop_unpack_double_2x32
, rhs
));
341 assert(!"Unexpected type conversion while lowering varyings");
345 this->out_instructions
->push_tail(new (this->mem_ctx
) ir_assignment(lhs
, rhs
));
350 * Make an ir_assignment from \c rhs to \c lhs, performing appropriate
351 * bitcasts if necessary to match up types.
353 * This function is called when unpacking varyings.
356 lower_packed_varyings_visitor::bitwise_assign_unpack(ir_rvalue
*lhs
,
359 if (lhs
->type
->base_type
!= rhs
->type
->base_type
) {
360 /* Since we only mix types in flat varyings, and we always store flat
361 * varyings as type ivec4, we need only produce conversions from int to
364 assert(rhs
->type
->base_type
== GLSL_TYPE_INT
);
365 switch (lhs
->type
->base_type
) {
367 rhs
= new(this->mem_ctx
)
368 ir_expression(ir_unop_i2u
, lhs
->type
, rhs
);
370 case GLSL_TYPE_FLOAT
:
371 rhs
= new(this->mem_ctx
)
372 ir_expression(ir_unop_bitcast_i2f
, lhs
->type
, rhs
);
374 case GLSL_TYPE_DOUBLE
:
375 assert(lhs
->type
->vector_elements
<= 2);
376 if (lhs
->type
->vector_elements
== 2) {
377 ir_variable
*t
= new(mem_ctx
) ir_variable(lhs
->type
, "unpack", ir_var_temporary
);
378 assert(rhs
->type
->vector_elements
== 4);
379 this->out_variables
->push_tail(t
);
380 this->out_instructions
->push_tail(
381 assign(t
, expr(ir_unop_pack_double_2x32
, i2u(swizzle_xy(rhs
->clone(mem_ctx
, NULL
)))), 0x1));
382 this->out_instructions
->push_tail(
383 assign(t
, expr(ir_unop_pack_double_2x32
, i2u(swizzle(rhs
->clone(mem_ctx
, NULL
), SWIZZLE_ZWZW
, 2))), 0x2));
386 rhs
= expr(ir_unop_pack_double_2x32
, i2u(rhs
));
390 assert(!"Unexpected type conversion while lowering varyings");
394 this->out_instructions
->push_tail(new(this->mem_ctx
) ir_assignment(lhs
, rhs
));
399 * Recursively pack or unpack the given varying (or portion of a varying) by
400 * traversing all of its constituent vectors.
402 * \param fine_location is the location where the first constituent vector
403 * should be packed--the word "fine" indicates that this location is expressed
404 * in multiples of a float, rather than multiples of a vec4 as is used
407 * \param gs_input_toplevel should be set to true if we are lowering geometry
408 * shader inputs, and we are currently lowering the whole input variable
409 * (i.e. we are lowering the array whose index selects the vertex).
411 * \param vertex_index: if we are lowering geometry shader inputs, and the
412 * level of the array that we are currently lowering is *not* the top level,
413 * then this indicates which vertex we are currently lowering. Otherwise it
416 * \return the location where the next constituent vector (after this one)
420 lower_packed_varyings_visitor::lower_rvalue(ir_rvalue
*rvalue
,
421 unsigned fine_location
,
422 ir_variable
*unpacked_var
,
424 bool gs_input_toplevel
,
425 unsigned vertex_index
)
427 unsigned dmul
= rvalue
->type
->is_double() ? 2 : 1;
428 /* When gs_input_toplevel is set, we should be looking at a geometry shader
431 assert(!gs_input_toplevel
|| rvalue
->type
->is_array());
433 if (rvalue
->type
->is_record()) {
434 for (unsigned i
= 0; i
< rvalue
->type
->length
; i
++) {
436 rvalue
= rvalue
->clone(this->mem_ctx
, NULL
);
437 const char *field_name
= rvalue
->type
->fields
.structure
[i
].name
;
438 ir_dereference_record
*dereference_record
= new(this->mem_ctx
)
439 ir_dereference_record(rvalue
, field_name
);
441 = ralloc_asprintf(this->mem_ctx
, "%s.%s", name
, field_name
);
442 fine_location
= this->lower_rvalue(dereference_record
, fine_location
,
443 unpacked_var
, deref_name
, false,
446 return fine_location
;
447 } else if (rvalue
->type
->is_array()) {
448 /* Arrays are packed/unpacked by considering each array element in
451 return this->lower_arraylike(rvalue
, rvalue
->type
->array_size(),
452 fine_location
, unpacked_var
, name
,
453 gs_input_toplevel
, vertex_index
);
454 } else if (rvalue
->type
->is_matrix()) {
455 /* Matrices are packed/unpacked by considering each column vector in
458 return this->lower_arraylike(rvalue
, rvalue
->type
->matrix_columns
,
459 fine_location
, unpacked_var
, name
,
460 false, vertex_index
);
461 } else if (rvalue
->type
->vector_elements
* dmul
+
462 fine_location
% 4 > 4) {
463 /* This vector is going to be "double parked" across two varying slots,
464 * so handle it as two separate assignments. For doubles, a dvec3/dvec4
465 * can end up being spread over 3 slots. However the second splitting
466 * will happen later, here we just always want to split into 2.
468 unsigned left_components
, right_components
;
469 unsigned left_swizzle_values
[4] = { 0, 0, 0, 0 };
470 unsigned right_swizzle_values
[4] = { 0, 0, 0, 0 };
471 char left_swizzle_name
[4] = { 0, 0, 0, 0 };
472 char right_swizzle_name
[4] = { 0, 0, 0, 0 };
474 left_components
= 4 - fine_location
% 4;
475 if (rvalue
->type
->is_double()) {
476 /* We might actually end up with 0 left components! */
477 left_components
/= 2;
479 right_components
= rvalue
->type
->vector_elements
- left_components
;
481 for (unsigned i
= 0; i
< left_components
; i
++) {
482 left_swizzle_values
[i
] = i
;
483 left_swizzle_name
[i
] = "xyzw"[i
];
485 for (unsigned i
= 0; i
< right_components
; i
++) {
486 right_swizzle_values
[i
] = i
+ left_components
;
487 right_swizzle_name
[i
] = "xyzw"[i
+ left_components
];
489 ir_swizzle
*left_swizzle
= new(this->mem_ctx
)
490 ir_swizzle(rvalue
, left_swizzle_values
, left_components
);
491 ir_swizzle
*right_swizzle
= new(this->mem_ctx
)
492 ir_swizzle(rvalue
->clone(this->mem_ctx
, NULL
), right_swizzle_values
,
495 = ralloc_asprintf(this->mem_ctx
, "%s.%s", name
, left_swizzle_name
);
497 = ralloc_asprintf(this->mem_ctx
, "%s.%s", name
, right_swizzle_name
);
499 fine_location
= this->lower_rvalue(left_swizzle
, fine_location
,
500 unpacked_var
, left_name
, false,
503 /* Top up the fine location to the next slot */
505 return this->lower_rvalue(right_swizzle
, fine_location
, unpacked_var
,
506 right_name
, false, vertex_index
);
508 /* No special handling is necessary; pack the rvalue into the
511 unsigned swizzle_values
[4] = { 0, 0, 0, 0 };
512 unsigned components
= rvalue
->type
->vector_elements
* dmul
;
513 unsigned location
= fine_location
/ 4;
514 unsigned location_frac
= fine_location
% 4;
515 for (unsigned i
= 0; i
< components
; ++i
)
516 swizzle_values
[i
] = i
+ location_frac
;
517 ir_dereference
*packed_deref
=
518 this->get_packed_varying_deref(location
, unpacked_var
, name
,
520 ir_swizzle
*swizzle
= new(this->mem_ctx
)
521 ir_swizzle(packed_deref
, swizzle_values
, components
);
522 if (this->mode
== ir_var_shader_out
) {
523 this->bitwise_assign_pack(swizzle
, rvalue
);
525 this->bitwise_assign_unpack(rvalue
, swizzle
);
527 return fine_location
+ components
;
532 * Recursively pack or unpack a varying for which we need to iterate over its
533 * constituent elements, accessing each one using an ir_dereference_array.
534 * This takes care of both arrays and matrices, since ir_dereference_array
535 * treats a matrix like an array of its column vectors.
537 * \param gs_input_toplevel should be set to true if we are lowering geometry
538 * shader inputs, and we are currently lowering the whole input variable
539 * (i.e. we are lowering the array whose index selects the vertex).
541 * \param vertex_index: if we are lowering geometry shader inputs, and the
542 * level of the array that we are currently lowering is *not* the top level,
543 * then this indicates which vertex we are currently lowering. Otherwise it
547 lower_packed_varyings_visitor::lower_arraylike(ir_rvalue
*rvalue
,
549 unsigned fine_location
,
550 ir_variable
*unpacked_var
,
552 bool gs_input_toplevel
,
553 unsigned vertex_index
)
555 for (unsigned i
= 0; i
< array_size
; i
++) {
557 rvalue
= rvalue
->clone(this->mem_ctx
, NULL
);
558 ir_constant
*constant
= new(this->mem_ctx
) ir_constant(i
);
559 ir_dereference_array
*dereference_array
= new(this->mem_ctx
)
560 ir_dereference_array(rvalue
, constant
);
561 if (gs_input_toplevel
) {
562 /* Geometry shader inputs are a special case. Instead of storing
563 * each element of the array at a different location, all elements
564 * are at the same location, but with a different vertex index.
566 (void) this->lower_rvalue(dereference_array
, fine_location
,
567 unpacked_var
, name
, false, i
);
569 char *subscripted_name
570 = ralloc_asprintf(this->mem_ctx
, "%s[%d]", name
, i
);
572 this->lower_rvalue(dereference_array
, fine_location
,
573 unpacked_var
, subscripted_name
,
574 false, vertex_index
);
577 return fine_location
;
581 * Retrieve the packed varying corresponding to the given varying location.
582 * If no packed varying has been created for the given varying location yet,
583 * create it and add it to the shader before returning it.
585 * The newly created varying inherits its interpolation parameters from \c
586 * unpacked_var. Its base type is ivec4 if we are lowering a flat varying,
589 * \param vertex_index: if we are lowering geometry shader inputs, then this
590 * indicates which vertex we are currently lowering. Otherwise it is ignored.
593 lower_packed_varyings_visitor::get_packed_varying_deref(
594 unsigned location
, ir_variable
*unpacked_var
, const char *name
,
595 unsigned vertex_index
)
597 unsigned slot
= location
- VARYING_SLOT_VAR0
;
598 assert(slot
< locations_used
);
599 if (this->packed_varyings
[slot
] == NULL
) {
600 char *packed_name
= ralloc_asprintf(this->mem_ctx
, "packed:%s", name
);
601 const glsl_type
*packed_type
;
602 if (unpacked_var
->data
.interpolation
== INTERP_QUALIFIER_FLAT
)
603 packed_type
= glsl_type::ivec4_type
;
605 packed_type
= glsl_type::vec4_type
;
606 if (this->gs_input_vertices
!= 0) {
608 glsl_type::get_array_instance(packed_type
,
609 this->gs_input_vertices
);
611 ir_variable
*packed_var
= new(this->mem_ctx
)
612 ir_variable(packed_type
, packed_name
, this->mode
);
613 if (this->gs_input_vertices
!= 0) {
614 /* Prevent update_array_sizes() from messing with the size of the
617 packed_var
->data
.max_array_access
= this->gs_input_vertices
- 1;
619 packed_var
->data
.centroid
= unpacked_var
->data
.centroid
;
620 packed_var
->data
.sample
= unpacked_var
->data
.sample
;
621 packed_var
->data
.patch
= unpacked_var
->data
.patch
;
622 packed_var
->data
.interpolation
= unpacked_var
->data
.interpolation
;
623 packed_var
->data
.location
= location
;
624 packed_var
->data
.precision
= unpacked_var
->data
.precision
;
625 unpacked_var
->insert_before(packed_var
);
626 this->packed_varyings
[slot
] = packed_var
;
628 /* For geometry shader inputs, only update the packed variable name the
629 * first time we visit each component.
631 if (this->gs_input_vertices
== 0 || vertex_index
== 0) {
632 ralloc_asprintf_append((char **) &this->packed_varyings
[slot
]->name
,
637 ir_dereference
*deref
= new(this->mem_ctx
)
638 ir_dereference_variable(this->packed_varyings
[slot
]);
639 if (this->gs_input_vertices
!= 0) {
640 /* When lowering GS inputs, the packed variable is an array, so we need
641 * to dereference it using vertex_index.
643 ir_constant
*constant
= new(this->mem_ctx
) ir_constant(vertex_index
);
644 deref
= new(this->mem_ctx
) ir_dereference_array(deref
, constant
);
650 lower_packed_varyings_visitor::needs_lowering(ir_variable
*var
)
652 /* Things composed of vec4's and varyings with explicitly assigned
653 * locations don't need lowering. Everything else does.
655 if (var
->data
.explicit_location
)
658 const glsl_type
*type
= var
->type
->without_array();
659 if (type
->vector_elements
== 4 && !type
->is_double())
666 * Visitor that splices varying packing code before every use of EmitVertex()
667 * in a geometry shader.
669 class lower_packed_varyings_gs_splicer
: public ir_hierarchical_visitor
672 explicit lower_packed_varyings_gs_splicer(void *mem_ctx
,
673 const exec_list
*instructions
);
675 virtual ir_visitor_status
visit_leave(ir_emit_vertex
*ev
);
679 * Memory context used to allocate new instructions for the shader.
681 void * const mem_ctx
;
684 * Instructions that should be spliced into place before each EmitVertex()
687 const exec_list
*instructions
;
691 lower_packed_varyings_gs_splicer::lower_packed_varyings_gs_splicer(
692 void *mem_ctx
, const exec_list
*instructions
)
693 : mem_ctx(mem_ctx
), instructions(instructions
)
699 lower_packed_varyings_gs_splicer::visit_leave(ir_emit_vertex
*ev
)
701 foreach_in_list(ir_instruction
, ir
, this->instructions
) {
702 ev
->insert_before(ir
->clone(this->mem_ctx
, NULL
));
704 return visit_continue
;
709 lower_packed_varyings(void *mem_ctx
, unsigned locations_used
,
710 ir_variable_mode mode
, unsigned gs_input_vertices
,
713 exec_list
*instructions
= shader
->ir
;
714 ir_function
*main_func
= shader
->symbols
->get_function("main");
715 exec_list void_parameters
;
716 ir_function_signature
*main_func_sig
717 = main_func
->matching_signature(NULL
, &void_parameters
, false);
718 exec_list new_instructions
, new_variables
;
719 lower_packed_varyings_visitor
visitor(mem_ctx
, locations_used
, mode
,
724 if (mode
== ir_var_shader_out
) {
725 if (shader
->Stage
== MESA_SHADER_GEOMETRY
) {
726 /* For geometry shaders, outputs need to be lowered before each call
729 lower_packed_varyings_gs_splicer
splicer(mem_ctx
, &new_instructions
);
731 /* Add all the variables in first. */
732 main_func_sig
->body
.head
->insert_before(&new_variables
);
734 /* Now update all the EmitVertex instances */
735 splicer
.run(instructions
);
737 /* For other shader types, outputs need to be lowered at the end of
740 main_func_sig
->body
.append_list(&new_variables
);
741 main_func_sig
->body
.append_list(&new_instructions
);
744 /* Shader inputs need to be lowered at the beginning of main() */
745 main_func_sig
->body
.head
->insert_before(&new_instructions
);
746 main_func_sig
->body
.head
->insert_before(&new_variables
);