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
,
168 unsigned locations_used
,
169 const uint8_t *components
,
170 ir_variable_mode mode
,
171 unsigned gs_input_vertices
,
172 exec_list
*out_instructions
,
173 exec_list
*out_variables
,
174 bool disable_varying_packing
,
177 void run(struct gl_linked_shader
*shader
);
180 void bitwise_assign_pack(ir_rvalue
*lhs
, ir_rvalue
*rhs
);
181 void bitwise_assign_unpack(ir_rvalue
*lhs
, ir_rvalue
*rhs
);
182 unsigned lower_rvalue(ir_rvalue
*rvalue
, unsigned fine_location
,
183 ir_variable
*unpacked_var
, const char *name
,
184 bool gs_input_toplevel
, unsigned vertex_index
);
185 unsigned lower_arraylike(ir_rvalue
*rvalue
, unsigned array_size
,
186 unsigned fine_location
,
187 ir_variable
*unpacked_var
, const char *name
,
188 bool gs_input_toplevel
, unsigned vertex_index
);
189 ir_dereference
*get_packed_varying_deref(unsigned location
,
190 ir_variable
*unpacked_var
,
192 unsigned vertex_index
);
193 bool needs_lowering(ir_variable
*var
);
196 * Memory context used to allocate new instructions for the shader.
198 void * const mem_ctx
;
201 * Number of generic varying slots which are used by this shader. This is
202 * used to allocate temporary intermediate data structures. If any varying
203 * used by this shader has a location greater than or equal to
204 * VARYING_SLOT_VAR0 + locations_used, an assertion will fire.
206 const unsigned locations_used
;
208 const uint8_t* components
;
211 * Array of pointers to the packed varyings that have been created for each
212 * generic varying slot. NULL entries in this array indicate varying slots
213 * for which a packed varying has not been created yet.
215 ir_variable
**packed_varyings
;
218 * Type of varying which is being lowered in this pass (either
219 * ir_var_shader_in or ir_var_shader_out).
221 const ir_variable_mode mode
;
224 * If we are currently lowering geometry shader inputs, the number of input
225 * vertices the geometry shader accepts. Otherwise zero.
227 const unsigned gs_input_vertices
;
230 * Exec list into which the visitor should insert the packing instructions.
231 * Caller provides this list; it should insert the instructions into the
232 * appropriate place in the shader once the visitor has finished running.
234 exec_list
*out_instructions
;
237 * Exec list into which the visitor should insert any new variables.
239 exec_list
*out_variables
;
241 bool disable_varying_packing
;
245 } /* anonymous namespace */
247 lower_packed_varyings_visitor::lower_packed_varyings_visitor(
248 void *mem_ctx
, unsigned locations_used
, const uint8_t *components
,
249 ir_variable_mode mode
,
250 unsigned gs_input_vertices
, exec_list
*out_instructions
,
251 exec_list
*out_variables
, bool disable_varying_packing
,
254 locations_used(locations_used
),
255 components(components
),
256 packed_varyings((ir_variable
**)
257 rzalloc_array_size(mem_ctx
, sizeof(*packed_varyings
),
260 gs_input_vertices(gs_input_vertices
),
261 out_instructions(out_instructions
),
262 out_variables(out_variables
),
263 disable_varying_packing(disable_varying_packing
),
264 xfb_enabled(xfb_enabled
)
269 lower_packed_varyings_visitor::run(struct gl_linked_shader
*shader
)
271 foreach_in_list(ir_instruction
, node
, shader
->ir
) {
272 ir_variable
*var
= node
->as_variable();
276 if (var
->data
.mode
!= this->mode
||
277 var
->data
.location
< VARYING_SLOT_VAR0
||
278 !this->needs_lowering(var
))
281 /* This lowering pass is only capable of packing floats and ints
282 * together when their interpolation mode is "flat". Treat integers as
283 * being flat when the interpolation mode is none.
285 assert(var
->data
.interpolation
== INTERP_MODE_FLAT
||
286 var
->data
.interpolation
== INTERP_MODE_NONE
||
287 !var
->type
->contains_integer());
289 /* Clone the variable for program resource list before
290 * it gets modified and lost.
292 if (!shader
->packed_varyings
)
293 shader
->packed_varyings
= new (shader
) exec_list
;
295 shader
->packed_varyings
->push_tail(var
->clone(shader
, NULL
));
297 /* Change the old varying into an ordinary global. */
298 assert(var
->data
.mode
!= ir_var_temporary
);
299 var
->data
.mode
= ir_var_auto
;
301 /* Create a reference to the old varying. */
302 ir_dereference_variable
*deref
303 = new(this->mem_ctx
) ir_dereference_variable(var
);
305 /* Recursively pack or unpack it. */
306 this->lower_rvalue(deref
, var
->data
.location
* 4 + var
->data
.location_frac
, var
,
307 var
->name
, this->gs_input_vertices
!= 0, 0);
311 #define SWIZZLE_ZWZW MAKE_SWIZZLE4(SWIZZLE_Z, SWIZZLE_W, SWIZZLE_Z, SWIZZLE_W)
314 * Make an ir_assignment from \c rhs to \c lhs, performing appropriate
315 * bitcasts if necessary to match up types.
317 * This function is called when packing varyings.
320 lower_packed_varyings_visitor::bitwise_assign_pack(ir_rvalue
*lhs
,
323 if (lhs
->type
->base_type
!= rhs
->type
->base_type
) {
324 /* Since we only mix types in flat varyings, and we always store flat
325 * varyings as type ivec4, we need only produce conversions from (uint
328 assert(lhs
->type
->base_type
== GLSL_TYPE_INT
);
329 switch (rhs
->type
->base_type
) {
331 rhs
= new(this->mem_ctx
)
332 ir_expression(ir_unop_u2i
, lhs
->type
, rhs
);
334 case GLSL_TYPE_FLOAT
:
335 rhs
= new(this->mem_ctx
)
336 ir_expression(ir_unop_bitcast_f2i
, lhs
->type
, rhs
);
338 case GLSL_TYPE_DOUBLE
:
339 assert(rhs
->type
->vector_elements
<= 2);
340 if (rhs
->type
->vector_elements
== 2) {
341 ir_variable
*t
= new(mem_ctx
) ir_variable(lhs
->type
, "pack", ir_var_temporary
);
343 assert(lhs
->type
->vector_elements
== 4);
344 this->out_variables
->push_tail(t
);
345 this->out_instructions
->push_tail(
346 assign(t
, u2i(expr(ir_unop_unpack_double_2x32
, swizzle_x(rhs
->clone(mem_ctx
, NULL
)))), 0x3));
347 this->out_instructions
->push_tail(
348 assign(t
, u2i(expr(ir_unop_unpack_double_2x32
, swizzle_y(rhs
))), 0xc));
351 rhs
= u2i(expr(ir_unop_unpack_double_2x32
, rhs
));
354 case GLSL_TYPE_INT64
:
355 assert(rhs
->type
->vector_elements
<= 2);
356 if (rhs
->type
->vector_elements
== 2) {
357 ir_variable
*t
= new(mem_ctx
) ir_variable(lhs
->type
, "pack", ir_var_temporary
);
359 assert(lhs
->type
->vector_elements
== 4);
360 this->out_variables
->push_tail(t
);
361 this->out_instructions
->push_tail(
362 assign(t
, expr(ir_unop_unpack_int_2x32
, swizzle_x(rhs
->clone(mem_ctx
, NULL
))), 0x3));
363 this->out_instructions
->push_tail(
364 assign(t
, expr(ir_unop_unpack_int_2x32
, swizzle_y(rhs
)), 0xc));
367 rhs
= expr(ir_unop_unpack_int_2x32
, rhs
);
370 case GLSL_TYPE_UINT64
:
371 assert(rhs
->type
->vector_elements
<= 2);
372 if (rhs
->type
->vector_elements
== 2) {
373 ir_variable
*t
= new(mem_ctx
) ir_variable(lhs
->type
, "pack", ir_var_temporary
);
375 assert(lhs
->type
->vector_elements
== 4);
376 this->out_variables
->push_tail(t
);
377 this->out_instructions
->push_tail(
378 assign(t
, u2i(expr(ir_unop_unpack_uint_2x32
, swizzle_x(rhs
->clone(mem_ctx
, NULL
)))), 0x3));
379 this->out_instructions
->push_tail(
380 assign(t
, u2i(expr(ir_unop_unpack_uint_2x32
, swizzle_y(rhs
))), 0xc));
383 rhs
= u2i(expr(ir_unop_unpack_uint_2x32
, rhs
));
387 assert(!"Unexpected type conversion while lowering varyings");
391 this->out_instructions
->push_tail(new (this->mem_ctx
) ir_assignment(lhs
, rhs
));
396 * Make an ir_assignment from \c rhs to \c lhs, performing appropriate
397 * bitcasts if necessary to match up types.
399 * This function is called when unpacking varyings.
402 lower_packed_varyings_visitor::bitwise_assign_unpack(ir_rvalue
*lhs
,
405 if (lhs
->type
->base_type
!= rhs
->type
->base_type
) {
406 /* Since we only mix types in flat varyings, and we always store flat
407 * varyings as type ivec4, we need only produce conversions from int to
410 assert(rhs
->type
->base_type
== GLSL_TYPE_INT
);
411 switch (lhs
->type
->base_type
) {
413 rhs
= new(this->mem_ctx
)
414 ir_expression(ir_unop_i2u
, lhs
->type
, rhs
);
416 case GLSL_TYPE_FLOAT
:
417 rhs
= new(this->mem_ctx
)
418 ir_expression(ir_unop_bitcast_i2f
, lhs
->type
, rhs
);
420 case GLSL_TYPE_DOUBLE
:
421 assert(lhs
->type
->vector_elements
<= 2);
422 if (lhs
->type
->vector_elements
== 2) {
423 ir_variable
*t
= new(mem_ctx
) ir_variable(lhs
->type
, "unpack", ir_var_temporary
);
424 assert(rhs
->type
->vector_elements
== 4);
425 this->out_variables
->push_tail(t
);
426 this->out_instructions
->push_tail(
427 assign(t
, expr(ir_unop_pack_double_2x32
, i2u(swizzle_xy(rhs
->clone(mem_ctx
, NULL
)))), 0x1));
428 this->out_instructions
->push_tail(
429 assign(t
, expr(ir_unop_pack_double_2x32
, i2u(swizzle(rhs
->clone(mem_ctx
, NULL
), SWIZZLE_ZWZW
, 2))), 0x2));
432 rhs
= expr(ir_unop_pack_double_2x32
, i2u(rhs
));
435 case GLSL_TYPE_INT64
:
436 assert(lhs
->type
->vector_elements
<= 2);
437 if (lhs
->type
->vector_elements
== 2) {
438 ir_variable
*t
= new(mem_ctx
) ir_variable(lhs
->type
, "unpack", ir_var_temporary
);
439 assert(rhs
->type
->vector_elements
== 4);
440 this->out_variables
->push_tail(t
);
441 this->out_instructions
->push_tail(
442 assign(t
, expr(ir_unop_pack_int_2x32
, swizzle_xy(rhs
->clone(mem_ctx
, NULL
))), 0x1));
443 this->out_instructions
->push_tail(
444 assign(t
, expr(ir_unop_pack_int_2x32
, swizzle(rhs
->clone(mem_ctx
, NULL
), SWIZZLE_ZWZW
, 2)), 0x2));
447 rhs
= expr(ir_unop_pack_int_2x32
, rhs
);
450 case GLSL_TYPE_UINT64
:
451 assert(lhs
->type
->vector_elements
<= 2);
452 if (lhs
->type
->vector_elements
== 2) {
453 ir_variable
*t
= new(mem_ctx
) ir_variable(lhs
->type
, "unpack", ir_var_temporary
);
454 assert(rhs
->type
->vector_elements
== 4);
455 this->out_variables
->push_tail(t
);
456 this->out_instructions
->push_tail(
457 assign(t
, expr(ir_unop_pack_uint_2x32
, i2u(swizzle_xy(rhs
->clone(mem_ctx
, NULL
)))), 0x1));
458 this->out_instructions
->push_tail(
459 assign(t
, expr(ir_unop_pack_uint_2x32
, i2u(swizzle(rhs
->clone(mem_ctx
, NULL
), SWIZZLE_ZWZW
, 2))), 0x2));
462 rhs
= expr(ir_unop_pack_uint_2x32
, i2u(rhs
));
466 assert(!"Unexpected type conversion while lowering varyings");
470 this->out_instructions
->push_tail(new(this->mem_ctx
) ir_assignment(lhs
, rhs
));
475 * Recursively pack or unpack the given varying (or portion of a varying) by
476 * traversing all of its constituent vectors.
478 * \param fine_location is the location where the first constituent vector
479 * should be packed--the word "fine" indicates that this location is expressed
480 * in multiples of a float, rather than multiples of a vec4 as is used
483 * \param gs_input_toplevel should be set to true if we are lowering geometry
484 * shader inputs, and we are currently lowering the whole input variable
485 * (i.e. we are lowering the array whose index selects the vertex).
487 * \param vertex_index: if we are lowering geometry shader inputs, and the
488 * level of the array that we are currently lowering is *not* the top level,
489 * then this indicates which vertex we are currently lowering. Otherwise it
492 * \return the location where the next constituent vector (after this one)
496 lower_packed_varyings_visitor::lower_rvalue(ir_rvalue
*rvalue
,
497 unsigned fine_location
,
498 ir_variable
*unpacked_var
,
500 bool gs_input_toplevel
,
501 unsigned vertex_index
)
503 unsigned dmul
= rvalue
->type
->is_64bit() ? 2 : 1;
504 /* When gs_input_toplevel is set, we should be looking at a geometry shader
507 assert(!gs_input_toplevel
|| rvalue
->type
->is_array());
509 if (rvalue
->type
->is_record()) {
510 for (unsigned i
= 0; i
< rvalue
->type
->length
; i
++) {
512 rvalue
= rvalue
->clone(this->mem_ctx
, NULL
);
513 const char *field_name
= rvalue
->type
->fields
.structure
[i
].name
;
514 ir_dereference_record
*dereference_record
= new(this->mem_ctx
)
515 ir_dereference_record(rvalue
, field_name
);
517 = ralloc_asprintf(this->mem_ctx
, "%s.%s", name
, field_name
);
518 fine_location
= this->lower_rvalue(dereference_record
, fine_location
,
519 unpacked_var
, deref_name
, false,
522 return fine_location
;
523 } else if (rvalue
->type
->is_array()) {
524 /* Arrays are packed/unpacked by considering each array element in
527 return this->lower_arraylike(rvalue
, rvalue
->type
->array_size(),
528 fine_location
, unpacked_var
, name
,
529 gs_input_toplevel
, vertex_index
);
530 } else if (rvalue
->type
->is_matrix()) {
531 /* Matrices are packed/unpacked by considering each column vector in
534 return this->lower_arraylike(rvalue
, rvalue
->type
->matrix_columns
,
535 fine_location
, unpacked_var
, name
,
536 false, vertex_index
);
537 } else if (rvalue
->type
->vector_elements
* dmul
+
538 fine_location
% 4 > 4) {
539 /* This vector is going to be "double parked" across two varying slots,
540 * so handle it as two separate assignments. For doubles, a dvec3/dvec4
541 * can end up being spread over 3 slots. However the second splitting
542 * will happen later, here we just always want to split into 2.
544 unsigned left_components
, right_components
;
545 unsigned left_swizzle_values
[4] = { 0, 0, 0, 0 };
546 unsigned right_swizzle_values
[4] = { 0, 0, 0, 0 };
547 char left_swizzle_name
[4] = { 0, 0, 0, 0 };
548 char right_swizzle_name
[4] = { 0, 0, 0, 0 };
550 left_components
= 4 - fine_location
% 4;
551 if (rvalue
->type
->is_64bit()) {
552 /* We might actually end up with 0 left components! */
553 left_components
/= 2;
555 right_components
= rvalue
->type
->vector_elements
- left_components
;
557 for (unsigned i
= 0; i
< left_components
; i
++) {
558 left_swizzle_values
[i
] = i
;
559 left_swizzle_name
[i
] = "xyzw"[i
];
561 for (unsigned i
= 0; i
< right_components
; i
++) {
562 right_swizzle_values
[i
] = i
+ left_components
;
563 right_swizzle_name
[i
] = "xyzw"[i
+ left_components
];
565 ir_swizzle
*left_swizzle
= new(this->mem_ctx
)
566 ir_swizzle(rvalue
, left_swizzle_values
, left_components
);
567 ir_swizzle
*right_swizzle
= new(this->mem_ctx
)
568 ir_swizzle(rvalue
->clone(this->mem_ctx
, NULL
), right_swizzle_values
,
571 = ralloc_asprintf(this->mem_ctx
, "%s.%s", name
, left_swizzle_name
);
573 = ralloc_asprintf(this->mem_ctx
, "%s.%s", name
, right_swizzle_name
);
575 fine_location
= this->lower_rvalue(left_swizzle
, fine_location
,
576 unpacked_var
, left_name
, false,
579 /* Top up the fine location to the next slot */
581 return this->lower_rvalue(right_swizzle
, fine_location
, unpacked_var
,
582 right_name
, false, vertex_index
);
584 /* No special handling is necessary; pack the rvalue into the
587 unsigned swizzle_values
[4] = { 0, 0, 0, 0 };
588 unsigned components
= rvalue
->type
->vector_elements
* dmul
;
589 unsigned location
= fine_location
/ 4;
590 unsigned location_frac
= fine_location
% 4;
591 for (unsigned i
= 0; i
< components
; ++i
)
592 swizzle_values
[i
] = i
+ location_frac
;
593 ir_dereference
*packed_deref
=
594 this->get_packed_varying_deref(location
, unpacked_var
, name
,
596 if (unpacked_var
->data
.stream
!= 0) {
597 assert(unpacked_var
->data
.stream
< 4);
598 ir_variable
*packed_var
= packed_deref
->variable_referenced();
599 for (unsigned i
= 0; i
< components
; ++i
) {
600 packed_var
->data
.stream
|=
601 unpacked_var
->data
.stream
<< (2 * (location_frac
+ i
));
604 ir_swizzle
*swizzle
= new(this->mem_ctx
)
605 ir_swizzle(packed_deref
, swizzle_values
, components
);
606 if (this->mode
== ir_var_shader_out
) {
607 this->bitwise_assign_pack(swizzle
, rvalue
);
609 this->bitwise_assign_unpack(rvalue
, swizzle
);
611 return fine_location
+ components
;
616 * Recursively pack or unpack a varying for which we need to iterate over its
617 * constituent elements, accessing each one using an ir_dereference_array.
618 * This takes care of both arrays and matrices, since ir_dereference_array
619 * treats a matrix like an array of its column vectors.
621 * \param gs_input_toplevel should be set to true if we are lowering geometry
622 * shader inputs, and we are currently lowering the whole input variable
623 * (i.e. we are lowering the array whose index selects the vertex).
625 * \param vertex_index: if we are lowering geometry shader inputs, and the
626 * level of the array that we are currently lowering is *not* the top level,
627 * then this indicates which vertex we are currently lowering. Otherwise it
631 lower_packed_varyings_visitor::lower_arraylike(ir_rvalue
*rvalue
,
633 unsigned fine_location
,
634 ir_variable
*unpacked_var
,
636 bool gs_input_toplevel
,
637 unsigned vertex_index
)
639 for (unsigned i
= 0; i
< array_size
; i
++) {
641 rvalue
= rvalue
->clone(this->mem_ctx
, NULL
);
642 ir_constant
*constant
= new(this->mem_ctx
) ir_constant(i
);
643 ir_dereference_array
*dereference_array
= new(this->mem_ctx
)
644 ir_dereference_array(rvalue
, constant
);
645 if (gs_input_toplevel
) {
646 /* Geometry shader inputs are a special case. Instead of storing
647 * each element of the array at a different location, all elements
648 * are at the same location, but with a different vertex index.
650 (void) this->lower_rvalue(dereference_array
, fine_location
,
651 unpacked_var
, name
, false, i
);
653 char *subscripted_name
654 = ralloc_asprintf(this->mem_ctx
, "%s[%d]", name
, i
);
656 this->lower_rvalue(dereference_array
, fine_location
,
657 unpacked_var
, subscripted_name
,
658 false, vertex_index
);
661 return fine_location
;
665 * Retrieve the packed varying corresponding to the given varying location.
666 * If no packed varying has been created for the given varying location yet,
667 * create it and add it to the shader before returning it.
669 * The newly created varying inherits its interpolation parameters from \c
670 * unpacked_var. Its base type is ivec4 if we are lowering a flat varying,
673 * \param vertex_index: if we are lowering geometry shader inputs, then this
674 * indicates which vertex we are currently lowering. Otherwise it is ignored.
677 lower_packed_varyings_visitor::get_packed_varying_deref(
678 unsigned location
, ir_variable
*unpacked_var
, const char *name
,
679 unsigned vertex_index
)
681 unsigned slot
= location
- VARYING_SLOT_VAR0
;
682 assert(slot
< locations_used
);
683 if (this->packed_varyings
[slot
] == NULL
) {
684 char *packed_name
= ralloc_asprintf(this->mem_ctx
, "packed:%s", name
);
685 const glsl_type
*packed_type
;
686 assert(components
[slot
] != 0);
687 if (unpacked_var
->is_interpolation_flat())
688 packed_type
= glsl_type::get_instance(GLSL_TYPE_INT
, components
[slot
], 1);
690 packed_type
= glsl_type::get_instance(GLSL_TYPE_FLOAT
, components
[slot
], 1);
691 if (this->gs_input_vertices
!= 0) {
693 glsl_type::get_array_instance(packed_type
,
694 this->gs_input_vertices
);
696 ir_variable
*packed_var
= new(this->mem_ctx
)
697 ir_variable(packed_type
, packed_name
, this->mode
);
698 if (this->gs_input_vertices
!= 0) {
699 /* Prevent update_array_sizes() from messing with the size of the
702 packed_var
->data
.max_array_access
= this->gs_input_vertices
- 1;
704 packed_var
->data
.centroid
= unpacked_var
->data
.centroid
;
705 packed_var
->data
.sample
= unpacked_var
->data
.sample
;
706 packed_var
->data
.patch
= unpacked_var
->data
.patch
;
707 packed_var
->data
.interpolation
= packed_type
== glsl_type::ivec4_type
708 ? unsigned(INTERP_MODE_FLAT
) : unpacked_var
->data
.interpolation
;
709 packed_var
->data
.location
= location
;
710 packed_var
->data
.precision
= unpacked_var
->data
.precision
;
711 packed_var
->data
.always_active_io
= unpacked_var
->data
.always_active_io
;
712 packed_var
->data
.stream
= 1u << 31;
713 unpacked_var
->insert_before(packed_var
);
714 this->packed_varyings
[slot
] = packed_var
;
716 /* For geometry shader inputs, only update the packed variable name the
717 * first time we visit each component.
719 if (this->gs_input_vertices
== 0 || vertex_index
== 0) {
720 ir_variable
*var
= this->packed_varyings
[slot
];
722 if (var
->is_name_ralloced())
723 ralloc_asprintf_append((char **) &var
->name
, ",%s", name
);
725 var
->name
= ralloc_asprintf(var
, "%s,%s", var
->name
, name
);
729 ir_dereference
*deref
= new(this->mem_ctx
)
730 ir_dereference_variable(this->packed_varyings
[slot
]);
731 if (this->gs_input_vertices
!= 0) {
732 /* When lowering GS inputs, the packed variable is an array, so we need
733 * to dereference it using vertex_index.
735 ir_constant
*constant
= new(this->mem_ctx
) ir_constant(vertex_index
);
736 deref
= new(this->mem_ctx
) ir_dereference_array(deref
, constant
);
742 lower_packed_varyings_visitor::needs_lowering(ir_variable
*var
)
744 /* Things composed of vec4's and varyings with explicitly assigned
745 * locations don't need lowering. Everything else does.
747 if (var
->data
.explicit_location
)
750 /* Override disable_varying_packing if the var is only used by transform
751 * feedback. Also override it if transform feedback is enabled and the
752 * variable is an array, struct or matrix as the elements of these types
753 * will always has the same interpolation and therefore asre safe to pack.
755 const glsl_type
*type
= var
->type
;
756 if (disable_varying_packing
&& !var
->data
.is_xfb_only
&&
757 !((type
->is_array() || type
->is_record() || type
->is_matrix()) &&
761 type
= type
->without_array();
762 if (type
->vector_elements
== 4 && !type
->is_64bit())
769 * Visitor that splices varying packing code before every use of EmitVertex()
770 * in a geometry shader.
772 class lower_packed_varyings_gs_splicer
: public ir_hierarchical_visitor
775 explicit lower_packed_varyings_gs_splicer(void *mem_ctx
,
776 const exec_list
*instructions
);
778 virtual ir_visitor_status
visit_leave(ir_emit_vertex
*ev
);
782 * Memory context used to allocate new instructions for the shader.
784 void * const mem_ctx
;
787 * Instructions that should be spliced into place before each EmitVertex()
790 const exec_list
*instructions
;
794 lower_packed_varyings_gs_splicer::lower_packed_varyings_gs_splicer(
795 void *mem_ctx
, const exec_list
*instructions
)
796 : mem_ctx(mem_ctx
), instructions(instructions
)
802 lower_packed_varyings_gs_splicer::visit_leave(ir_emit_vertex
*ev
)
804 foreach_in_list(ir_instruction
, ir
, this->instructions
) {
805 ev
->insert_before(ir
->clone(this->mem_ctx
, NULL
));
807 return visit_continue
;
811 * Visitor that splices varying packing code before every return.
813 class lower_packed_varyings_return_splicer
: public ir_hierarchical_visitor
816 explicit lower_packed_varyings_return_splicer(void *mem_ctx
,
817 const exec_list
*instructions
);
819 virtual ir_visitor_status
visit_leave(ir_return
*ret
);
823 * Memory context used to allocate new instructions for the shader.
825 void * const mem_ctx
;
828 * Instructions that should be spliced into place before each return.
830 const exec_list
*instructions
;
834 lower_packed_varyings_return_splicer::lower_packed_varyings_return_splicer(
835 void *mem_ctx
, const exec_list
*instructions
)
836 : mem_ctx(mem_ctx
), instructions(instructions
)
842 lower_packed_varyings_return_splicer::visit_leave(ir_return
*ret
)
844 foreach_in_list(ir_instruction
, ir
, this->instructions
) {
845 ret
->insert_before(ir
->clone(this->mem_ctx
, NULL
));
847 return visit_continue
;
851 lower_packed_varyings(void *mem_ctx
, unsigned locations_used
,
852 const uint8_t *components
,
853 ir_variable_mode mode
, unsigned gs_input_vertices
,
854 gl_linked_shader
*shader
, bool disable_varying_packing
,
857 exec_list
*instructions
= shader
->ir
;
858 ir_function
*main_func
= shader
->symbols
->get_function("main");
859 exec_list void_parameters
;
860 ir_function_signature
*main_func_sig
861 = main_func
->matching_signature(NULL
, &void_parameters
, false);
862 exec_list new_instructions
, new_variables
;
863 lower_packed_varyings_visitor
visitor(mem_ctx
,
870 disable_varying_packing
,
873 if (mode
== ir_var_shader_out
) {
874 if (shader
->Stage
== MESA_SHADER_GEOMETRY
) {
875 /* For geometry shaders, outputs need to be lowered before each call
878 lower_packed_varyings_gs_splicer
splicer(mem_ctx
, &new_instructions
);
880 /* Add all the variables in first. */
881 main_func_sig
->body
.get_head_raw()->insert_before(&new_variables
);
883 /* Now update all the EmitVertex instances */
884 splicer
.run(instructions
);
886 /* For other shader types, outputs need to be lowered before each
887 * return statement and at the end of main()
890 lower_packed_varyings_return_splicer
splicer(mem_ctx
, &new_instructions
);
892 main_func_sig
->body
.get_head_raw()->insert_before(&new_variables
);
894 splicer
.run(instructions
);
896 /* Lower outputs at the end of main() if the last instruction is not
899 if (((ir_instruction
*)instructions
->get_tail())->ir_type
!= ir_type_return
) {
900 main_func_sig
->body
.append_list(&new_instructions
);
904 /* Shader inputs need to be lowered at the beginning of main() */
905 main_func_sig
->body
.get_head_raw()->insert_before(&new_instructions
);
906 main_func_sig
->body
.get_head_raw()->insert_before(&new_variables
);