2 * Copyright © 2015 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 DEALINGS
25 #include "nir_builder.h"
27 #include "util/hash_table.h"
29 /* This file contains various little helpers for doing simple linking in
30 * NIR. Eventually, we'll probably want a full-blown varying packing
31 * implementation in here. Right now, it just deletes unused things.
35 * Returns the bits in the inputs_read, outputs_written, or
36 * system_values_read bitfield corresponding to this variable.
39 get_variable_io_mask(nir_variable
*var
, gl_shader_stage stage
)
41 if (var
->data
.location
< 0)
44 unsigned location
= var
->data
.patch
?
45 var
->data
.location
- VARYING_SLOT_PATCH0
: var
->data
.location
;
47 assert(var
->data
.mode
== nir_var_shader_in
||
48 var
->data
.mode
== nir_var_shader_out
||
49 var
->data
.mode
== nir_var_system_value
);
50 assert(var
->data
.location
>= 0);
52 const struct glsl_type
*type
= var
->type
;
53 if (nir_is_per_vertex_io(var
, stage
)) {
54 assert(glsl_type_is_array(type
));
55 type
= glsl_get_array_element(type
);
58 unsigned slots
= glsl_count_attribute_slots(type
, false);
59 return ((1ull << slots
) - 1) << location
;
63 tcs_add_output_reads(nir_shader
*shader
, uint64_t *read
, uint64_t *patches_read
)
65 nir_foreach_function(function
, shader
) {
69 nir_foreach_block(block
, function
->impl
) {
70 nir_foreach_instr(instr
, block
) {
71 if (instr
->type
!= nir_instr_type_intrinsic
)
74 nir_intrinsic_instr
*intrin
= nir_instr_as_intrinsic(instr
);
75 if (intrin
->intrinsic
!= nir_intrinsic_load_deref
)
78 nir_deref_instr
*deref
= nir_src_as_deref(intrin
->src
[0]);
79 if (deref
->mode
!= nir_var_shader_out
)
82 nir_variable
*var
= nir_deref_instr_get_variable(deref
);
83 if (var
->data
.patch
) {
84 patches_read
[var
->data
.location_frac
] |=
85 get_variable_io_mask(var
, shader
->info
.stage
);
87 read
[var
->data
.location_frac
] |=
88 get_variable_io_mask(var
, shader
->info
.stage
);
96 * Helper for removing unused shader I/O variables, by demoting them to global
97 * variables (which may then by dead code eliminated).
101 * progress = nir_remove_unused_io_vars(producer,
102 * &producer->outputs,
103 * read, patches_read) ||
106 * The "used" should be an array of 4 uint64_ts (probably of VARYING_BIT_*)
107 * representing each .location_frac used. Note that for vector variables,
108 * only the first channel (.location_frac) is examined for deciding if the
112 nir_remove_unused_io_vars(nir_shader
*shader
, struct exec_list
*var_list
,
113 uint64_t *used_by_other_stage
,
114 uint64_t *used_by_other_stage_patches
)
116 bool progress
= false;
119 nir_foreach_variable_safe(var
, var_list
) {
121 used
= used_by_other_stage_patches
;
123 used
= used_by_other_stage
;
125 if (var
->data
.location
< VARYING_SLOT_VAR0
&& var
->data
.location
>= 0)
128 if (var
->data
.always_active_io
)
131 if (var
->data
.explicit_xfb_buffer
)
134 uint64_t other_stage
= used
[var
->data
.location_frac
];
136 if (!(other_stage
& get_variable_io_mask(var
, shader
->info
.stage
))) {
137 /* This one is invalid, make it a global variable instead */
138 var
->data
.location
= 0;
139 var
->data
.mode
= nir_var_shader_temp
;
141 exec_node_remove(&var
->node
);
142 exec_list_push_tail(&shader
->globals
, &var
->node
);
149 nir_fixup_deref_modes(shader
);
155 nir_remove_unused_varyings(nir_shader
*producer
, nir_shader
*consumer
)
157 assert(producer
->info
.stage
!= MESA_SHADER_FRAGMENT
);
158 assert(consumer
->info
.stage
!= MESA_SHADER_VERTEX
);
160 uint64_t read
[4] = { 0 }, written
[4] = { 0 };
161 uint64_t patches_read
[4] = { 0 }, patches_written
[4] = { 0 };
163 nir_foreach_variable(var
, &producer
->outputs
) {
164 if (var
->data
.patch
) {
165 patches_written
[var
->data
.location_frac
] |=
166 get_variable_io_mask(var
, producer
->info
.stage
);
168 written
[var
->data
.location_frac
] |=
169 get_variable_io_mask(var
, producer
->info
.stage
);
173 nir_foreach_variable(var
, &consumer
->inputs
) {
174 if (var
->data
.patch
) {
175 patches_read
[var
->data
.location_frac
] |=
176 get_variable_io_mask(var
, consumer
->info
.stage
);
178 read
[var
->data
.location_frac
] |=
179 get_variable_io_mask(var
, consumer
->info
.stage
);
183 /* Each TCS invocation can read data written by other TCS invocations,
184 * so even if the outputs are not used by the TES we must also make
185 * sure they are not read by the TCS before demoting them to globals.
187 if (producer
->info
.stage
== MESA_SHADER_TESS_CTRL
)
188 tcs_add_output_reads(producer
, read
, patches_read
);
190 bool progress
= false;
191 progress
= nir_remove_unused_io_vars(producer
, &producer
->outputs
, read
,
194 progress
= nir_remove_unused_io_vars(consumer
, &consumer
->inputs
, written
,
195 patches_written
) || progress
;
201 get_interp_type(nir_variable
*var
, const struct glsl_type
*type
,
202 bool default_to_smooth_interp
)
204 if (glsl_type_is_integer(type
))
205 return INTERP_MODE_FLAT
;
206 else if (var
->data
.interpolation
!= INTERP_MODE_NONE
)
207 return var
->data
.interpolation
;
208 else if (default_to_smooth_interp
)
209 return INTERP_MODE_SMOOTH
;
211 return INTERP_MODE_NONE
;
214 #define INTERPOLATE_LOC_SAMPLE 0
215 #define INTERPOLATE_LOC_CENTROID 1
216 #define INTERPOLATE_LOC_CENTER 2
219 get_interp_loc(nir_variable
*var
)
221 if (var
->data
.sample
)
222 return INTERPOLATE_LOC_SAMPLE
;
223 else if (var
->data
.centroid
)
224 return INTERPOLATE_LOC_CENTROID
;
226 return INTERPOLATE_LOC_CENTER
;
230 is_packing_supported_for_type(const struct glsl_type
*type
)
232 /* We ignore complex types such as arrays, matrices, structs and bitsizes
233 * other then 32bit. All other vector types should have been split into
234 * scalar variables by the lower_io_to_scalar pass. The only exception
235 * should be OpenGL xfb varyings.
236 * TODO: add support for more complex types?
238 return glsl_type_is_scalar(type
) && glsl_type_is_32bit(type
);
241 struct assigned_comps
248 /* Packing arrays and dual slot varyings is difficult so to avoid complex
249 * algorithms this function just assigns them their existing location for now.
250 * TODO: allow better packing of complex types.
253 get_unmoveable_components_masks(struct exec_list
*var_list
,
254 struct assigned_comps
*comps
,
255 gl_shader_stage stage
,
256 bool default_to_smooth_interp
)
258 nir_foreach_variable_safe(var
, var_list
) {
259 assert(var
->data
.location
>= 0);
261 /* Only remap things that aren't built-ins. */
262 if (var
->data
.location
>= VARYING_SLOT_VAR0
&&
263 var
->data
.location
- VARYING_SLOT_VAR0
< MAX_VARYINGS_INCL_PATCH
) {
265 const struct glsl_type
*type
= var
->type
;
266 if (nir_is_per_vertex_io(var
, stage
)) {
267 assert(glsl_type_is_array(type
));
268 type
= glsl_get_array_element(type
);
271 /* If we can pack this varying then don't mark the components as
274 if (is_packing_supported_for_type(type
))
277 unsigned location
= var
->data
.location
- VARYING_SLOT_VAR0
;
280 glsl_type_is_vector_or_scalar(glsl_without_array(type
)) ?
281 glsl_get_vector_elements(glsl_without_array(type
)) : 4;
283 bool dual_slot
= glsl_type_is_dual_slot(glsl_without_array(type
));
284 unsigned slots
= glsl_count_attribute_slots(type
, false);
285 unsigned dmul
= glsl_type_is_64bit(glsl_without_array(type
)) ? 2 : 1;
286 unsigned comps_slot2
= 0;
287 for (unsigned i
= 0; i
< slots
; i
++) {
290 comps
[location
+ i
].comps
|= ((1 << comps_slot2
) - 1);
292 unsigned num_comps
= 4 - var
->data
.location_frac
;
293 comps_slot2
= (elements
* dmul
) - num_comps
;
295 /* Assume ARB_enhanced_layouts packing rules for doubles */
296 assert(var
->data
.location_frac
== 0 ||
297 var
->data
.location_frac
== 2);
298 assert(comps_slot2
<= 4);
300 comps
[location
+ i
].comps
|=
301 ((1 << num_comps
) - 1) << var
->data
.location_frac
;
304 comps
[location
+ i
].comps
|=
305 ((1 << (elements
* dmul
)) - 1) << var
->data
.location_frac
;
308 comps
[location
+ i
].interp_type
=
309 get_interp_type(var
, type
, default_to_smooth_interp
);
310 comps
[location
+ i
].interp_loc
= get_interp_loc(var
);
323 mark_all_used_slots(nir_variable
*var
, uint64_t *slots_used
,
324 uint64_t slots_used_mask
, unsigned num_slots
)
326 unsigned loc_offset
= var
->data
.patch
? VARYING_SLOT_PATCH0
: 0;
328 slots_used
[var
->data
.patch
? 1 : 0] |= slots_used_mask
&
329 BITFIELD64_RANGE(var
->data
.location
- loc_offset
, num_slots
);
333 mark_used_slot(nir_variable
*var
, uint64_t *slots_used
, unsigned offset
)
335 unsigned loc_offset
= var
->data
.patch
? VARYING_SLOT_PATCH0
: 0;
337 slots_used
[var
->data
.patch
? 1 : 0] |=
338 BITFIELD64_BIT(var
->data
.location
- loc_offset
+ offset
);
342 remap_slots_and_components(struct exec_list
*var_list
, gl_shader_stage stage
,
343 struct varying_loc (*remap
)[4],
344 uint64_t *slots_used
, uint64_t *out_slots_read
,
345 uint32_t *p_slots_used
, uint32_t *p_out_slots_read
)
347 uint64_t out_slots_read_tmp
[2] = {0};
348 uint64_t slots_used_tmp
[2] = {0};
350 /* We don't touch builtins so just copy the bitmask */
351 slots_used_tmp
[0] = *slots_used
& BITFIELD64_RANGE(0, VARYING_SLOT_VAR0
);
353 nir_foreach_variable(var
, var_list
) {
354 assert(var
->data
.location
>= 0);
356 /* Only remap things that aren't built-ins */
357 if (var
->data
.location
>= VARYING_SLOT_VAR0
&&
358 var
->data
.location
- VARYING_SLOT_VAR0
< MAX_VARYINGS_INCL_PATCH
) {
360 const struct glsl_type
*type
= var
->type
;
361 if (nir_is_per_vertex_io(var
, stage
)) {
362 assert(glsl_type_is_array(type
));
363 type
= glsl_get_array_element(type
);
366 unsigned num_slots
= glsl_count_attribute_slots(type
, false);
367 bool used_across_stages
= false;
368 bool outputs_read
= false;
370 unsigned location
= var
->data
.location
- VARYING_SLOT_VAR0
;
371 struct varying_loc
*new_loc
= &remap
[location
][var
->data
.location_frac
];
373 unsigned loc_offset
= var
->data
.patch
? VARYING_SLOT_PATCH0
: 0;
374 uint64_t used
= var
->data
.patch
? *p_slots_used
: *slots_used
;
376 var
->data
.patch
? *p_out_slots_read
: *out_slots_read
;
378 BITFIELD64_RANGE(var
->data
.location
- loc_offset
, num_slots
);
381 used_across_stages
= true;
383 if (slots
& outs_used
)
386 if (new_loc
->location
) {
387 var
->data
.location
= new_loc
->location
;
388 var
->data
.location_frac
= new_loc
->component
;
391 if (var
->data
.always_active_io
) {
392 /* We can't apply link time optimisations (specifically array
393 * splitting) to these so we need to copy the existing mask
394 * otherwise we will mess up the mask for things like partially
397 if (used_across_stages
)
398 mark_all_used_slots(var
, slots_used_tmp
, used
, num_slots
);
401 mark_all_used_slots(var
, out_slots_read_tmp
, outs_used
,
405 for (unsigned i
= 0; i
< num_slots
; i
++) {
406 if (used_across_stages
)
407 mark_used_slot(var
, slots_used_tmp
, i
);
410 mark_used_slot(var
, out_slots_read_tmp
, i
);
416 *slots_used
= slots_used_tmp
[0];
417 *out_slots_read
= out_slots_read_tmp
[0];
418 *p_slots_used
= slots_used_tmp
[1];
419 *p_out_slots_read
= out_slots_read_tmp
[1];
422 struct varying_component
{
431 cmp_varying_component(const void *comp1_v
, const void *comp2_v
)
433 struct varying_component
*comp1
= (struct varying_component
*) comp1_v
;
434 struct varying_component
*comp2
= (struct varying_component
*) comp2_v
;
436 /* We want patches to be order at the end of the array */
437 if (comp1
->is_patch
!= comp2
->is_patch
)
438 return comp1
->is_patch
? 1 : -1;
440 /* We can only pack varyings with matching interpolation types so group
443 if (comp1
->interp_type
!= comp2
->interp_type
)
444 return comp1
->interp_type
- comp2
->interp_type
;
446 /* Interpolation loc must match also. */
447 if (comp1
->interp_loc
!= comp2
->interp_loc
)
448 return comp1
->interp_loc
- comp2
->interp_loc
;
450 /* If everything else matches just use the original location to sort */
451 return comp1
->var
->data
.location
- comp2
->var
->data
.location
;
455 gather_varying_component_info(nir_shader
*consumer
,
456 struct varying_component
**varying_comp_info
,
457 unsigned *varying_comp_info_size
,
458 bool default_to_smooth_interp
)
460 unsigned store_varying_info_idx
[MAX_VARYINGS_INCL_PATCH
][4] = {0};
461 unsigned num_of_comps_to_pack
= 0;
463 /* Count the number of varying that can be packed and create a mapping
464 * of those varyings to the array we will pass to qsort.
466 nir_foreach_variable(var
, &consumer
->inputs
) {
468 /* Only remap things that aren't builtins. */
469 if (var
->data
.location
>= VARYING_SLOT_VAR0
&&
470 var
->data
.location
- VARYING_SLOT_VAR0
< MAX_VARYINGS_INCL_PATCH
) {
472 /* We can't repack xfb varyings. */
473 if (var
->data
.always_active_io
)
476 const struct glsl_type
*type
= var
->type
;
477 if (nir_is_per_vertex_io(var
, consumer
->info
.stage
)) {
478 assert(glsl_type_is_array(type
));
479 type
= glsl_get_array_element(type
);
482 if (!is_packing_supported_for_type(type
))
485 unsigned loc
= var
->data
.location
- VARYING_SLOT_VAR0
;
486 store_varying_info_idx
[loc
][var
->data
.location_frac
] =
487 ++num_of_comps_to_pack
;
491 *varying_comp_info_size
= num_of_comps_to_pack
;
492 *varying_comp_info
= rzalloc_array(NULL
, struct varying_component
,
493 num_of_comps_to_pack
);
495 nir_function_impl
*impl
= nir_shader_get_entrypoint(consumer
);
497 /* Walk over the shader and populate the varying component info array */
498 nir_foreach_block(block
, impl
) {
499 nir_foreach_instr(instr
, block
) {
500 if (instr
->type
!= nir_instr_type_intrinsic
)
503 nir_intrinsic_instr
*intr
= nir_instr_as_intrinsic(instr
);
504 if (intr
->intrinsic
!= nir_intrinsic_load_deref
&&
505 intr
->intrinsic
!= nir_intrinsic_interp_deref_at_centroid
&&
506 intr
->intrinsic
!= nir_intrinsic_interp_deref_at_sample
&&
507 intr
->intrinsic
!= nir_intrinsic_interp_deref_at_offset
)
510 nir_deref_instr
*deref
= nir_src_as_deref(intr
->src
[0]);
511 if (deref
->mode
!= nir_var_shader_in
)
514 /* We only remap things that aren't builtins. */
515 nir_variable
*in_var
= nir_deref_instr_get_variable(deref
);
516 if (in_var
->data
.location
< VARYING_SLOT_VAR0
)
519 unsigned location
= in_var
->data
.location
- VARYING_SLOT_VAR0
;
520 if (location
>= MAX_VARYINGS_INCL_PATCH
)
523 unsigned var_info_idx
=
524 store_varying_info_idx
[location
][in_var
->data
.location_frac
];
528 struct varying_component
*vc_info
=
529 &(*varying_comp_info
)[var_info_idx
-1];
531 if (!vc_info
->initialised
) {
532 const struct glsl_type
*type
= in_var
->type
;
533 if (nir_is_per_vertex_io(in_var
, consumer
->info
.stage
)) {
534 assert(glsl_type_is_array(type
));
535 type
= glsl_get_array_element(type
);
538 vc_info
->var
= in_var
;
539 vc_info
->interp_type
=
540 get_interp_type(in_var
, type
, default_to_smooth_interp
);
541 vc_info
->interp_loc
= get_interp_loc(in_var
);
542 vc_info
->is_patch
= in_var
->data
.patch
;
549 assign_remap_locations(struct varying_loc (*remap
)[4],
550 struct assigned_comps
*assigned_comps
,
551 struct varying_component
*info
,
552 unsigned *cursor
, unsigned *comp
,
553 unsigned max_location
)
555 unsigned tmp_cursor
= *cursor
;
556 unsigned tmp_comp
= *comp
;
558 for (; tmp_cursor
< max_location
; tmp_cursor
++) {
560 if (assigned_comps
[tmp_cursor
].comps
) {
561 /* We can only pack varyings with matching interpolation types,
562 * interpolation loc must match also.
563 * TODO: i965 can handle interpolation locations that don't match,
564 * but the radeonsi nir backend handles everything as vec4s and so
565 * expects this to be the same for all components. We could make this
566 * check driver specfific or drop it if NIR ever become the only
569 if (assigned_comps
[tmp_cursor
].interp_type
!= info
->interp_type
||
570 assigned_comps
[tmp_cursor
].interp_loc
!= info
->interp_loc
) {
575 while (tmp_comp
< 4 &&
576 (assigned_comps
[tmp_cursor
].comps
& (1 << tmp_comp
))) {
586 unsigned location
= info
->var
->data
.location
- VARYING_SLOT_VAR0
;
588 /* Once we have assigned a location mark it as used */
589 assigned_comps
[tmp_cursor
].comps
|= (1 << tmp_comp
);
590 assigned_comps
[tmp_cursor
].interp_type
= info
->interp_type
;
591 assigned_comps
[tmp_cursor
].interp_loc
= info
->interp_loc
;
593 /* Assign remap location */
594 remap
[location
][info
->var
->data
.location_frac
].component
= tmp_comp
++;
595 remap
[location
][info
->var
->data
.location_frac
].location
=
596 tmp_cursor
+ VARYING_SLOT_VAR0
;
601 *cursor
= tmp_cursor
;
605 /* If there are empty components in the slot compact the remaining components
606 * as close to component 0 as possible. This will make it easier to fill the
607 * empty components with components from a different slot in a following pass.
610 compact_components(nir_shader
*producer
, nir_shader
*consumer
,
611 struct assigned_comps
*assigned_comps
,
612 bool default_to_smooth_interp
)
614 struct exec_list
*input_list
= &consumer
->inputs
;
615 struct exec_list
*output_list
= &producer
->outputs
;
616 struct varying_loc remap
[MAX_VARYINGS_INCL_PATCH
][4] = {{{0}, {0}}};
617 struct varying_component
*varying_comp_info
;
618 unsigned varying_comp_info_size
;
620 /* Gather varying component info */
621 gather_varying_component_info(consumer
, &varying_comp_info
,
622 &varying_comp_info_size
,
623 default_to_smooth_interp
);
625 /* Sort varying components. */
626 qsort(varying_comp_info
, varying_comp_info_size
,
627 sizeof(struct varying_component
), cmp_varying_component
);
632 /* Set the remap array based on the sorted components */
633 for (unsigned i
= 0; i
< varying_comp_info_size
; i
++ ) {
634 struct varying_component
*info
= &varying_comp_info
[i
];
636 assert(info
->is_patch
|| cursor
< MAX_VARYING
);
637 if (info
->is_patch
) {
638 /* The list should be sorted with all non-patch inputs first followed
639 * by patch inputs. When we hit our first patch input, we need to
640 * reset the cursor to MAX_VARYING so we put them in the right slot.
642 if (cursor
< MAX_VARYING
) {
643 cursor
= MAX_VARYING
;
647 assign_remap_locations(remap
, assigned_comps
, info
,
648 &cursor
, &comp
, MAX_VARYINGS_INCL_PATCH
);
650 assign_remap_locations(remap
, assigned_comps
, info
,
651 &cursor
, &comp
, MAX_VARYING
);
653 /* Check if we failed to assign a remap location. This can happen if
654 * for example there are a bunch of unmovable components with
655 * mismatching interpolation types causing us to skip over locations
656 * that would have been useful for packing later components.
657 * The solution is to iterate over the locations again (this should
658 * happen very rarely in practice).
660 if (cursor
== MAX_VARYING
) {
663 assign_remap_locations(remap
, assigned_comps
, info
,
664 &cursor
, &comp
, MAX_VARYING
);
669 ralloc_free(varying_comp_info
);
673 remap_slots_and_components(input_list
, consumer
->info
.stage
, remap
,
674 &consumer
->info
.inputs_read
, &zero
,
675 &consumer
->info
.patch_inputs_read
, &zero32
);
676 remap_slots_and_components(output_list
, producer
->info
.stage
, remap
,
677 &producer
->info
.outputs_written
,
678 &producer
->info
.outputs_read
,
679 &producer
->info
.patch_outputs_written
,
680 &producer
->info
.patch_outputs_read
);
683 /* We assume that this has been called more-or-less directly after
684 * remove_unused_varyings. At this point, all of the varyings that we
685 * aren't going to be using have been completely removed and the
686 * inputs_read and outputs_written fields in nir_shader_info reflect
687 * this. Therefore, the total set of valid slots is the OR of the two
688 * sets of varyings; this accounts for varyings which one side may need
689 * to read/write even if the other doesn't. This can happen if, for
690 * instance, an array is used indirectly from one side causing it to be
691 * unsplittable but directly from the other.
694 nir_compact_varyings(nir_shader
*producer
, nir_shader
*consumer
,
695 bool default_to_smooth_interp
)
697 assert(producer
->info
.stage
!= MESA_SHADER_FRAGMENT
);
698 assert(consumer
->info
.stage
!= MESA_SHADER_VERTEX
);
700 struct assigned_comps assigned_comps
[MAX_VARYINGS_INCL_PATCH
] = {0};
702 get_unmoveable_components_masks(&producer
->outputs
, assigned_comps
,
703 producer
->info
.stage
,
704 default_to_smooth_interp
);
705 get_unmoveable_components_masks(&consumer
->inputs
, assigned_comps
,
706 consumer
->info
.stage
,
707 default_to_smooth_interp
);
709 compact_components(producer
, consumer
, assigned_comps
,
710 default_to_smooth_interp
);
714 * Mark XFB varyings as always_active_io in the consumer so the linking opts
718 nir_link_xfb_varyings(nir_shader
*producer
, nir_shader
*consumer
)
720 nir_variable
*input_vars
[MAX_VARYING
] = { 0 };
722 nir_foreach_variable(var
, &consumer
->inputs
) {
723 if (var
->data
.location
>= VARYING_SLOT_VAR0
&&
724 var
->data
.location
- VARYING_SLOT_VAR0
< MAX_VARYING
) {
726 unsigned location
= var
->data
.location
- VARYING_SLOT_VAR0
;
727 input_vars
[location
] = var
;
731 nir_foreach_variable(var
, &producer
->outputs
) {
732 if (var
->data
.location
>= VARYING_SLOT_VAR0
&&
733 var
->data
.location
- VARYING_SLOT_VAR0
< MAX_VARYING
) {
735 if (!var
->data
.always_active_io
)
738 unsigned location
= var
->data
.location
- VARYING_SLOT_VAR0
;
739 if (input_vars
[location
]) {
740 input_vars
[location
]->data
.always_active_io
= true;
747 does_varying_match(nir_variable
*out_var
, nir_variable
*in_var
)
749 return in_var
->data
.location
== out_var
->data
.location
&&
750 in_var
->data
.location_frac
== out_var
->data
.location_frac
;
753 static nir_variable
*
754 get_matching_input_var(nir_shader
*consumer
, nir_variable
*out_var
)
756 nir_foreach_variable(var
, &consumer
->inputs
) {
757 if (does_varying_match(out_var
, var
))
765 can_replace_varying(nir_variable
*out_var
)
767 /* Skip types that require more complex handling.
768 * TODO: add support for these types.
770 if (glsl_type_is_array(out_var
->type
) ||
771 glsl_type_is_dual_slot(out_var
->type
) ||
772 glsl_type_is_matrix(out_var
->type
) ||
773 glsl_type_is_struct_or_ifc(out_var
->type
))
776 /* Limit this pass to scalars for now to keep things simple. Most varyings
777 * should have been lowered to scalars at this point anyway.
779 if (!glsl_type_is_scalar(out_var
->type
))
782 if (out_var
->data
.location
< VARYING_SLOT_VAR0
||
783 out_var
->data
.location
- VARYING_SLOT_VAR0
>= MAX_VARYING
)
790 replace_constant_input(nir_shader
*shader
, nir_intrinsic_instr
*store_intr
)
792 nir_function_impl
*impl
= nir_shader_get_entrypoint(shader
);
795 nir_builder_init(&b
, impl
);
797 nir_variable
*out_var
=
798 nir_deref_instr_get_variable(nir_src_as_deref(store_intr
->src
[0]));
800 bool progress
= false;
801 nir_foreach_block(block
, impl
) {
802 nir_foreach_instr(instr
, block
) {
803 if (instr
->type
!= nir_instr_type_intrinsic
)
806 nir_intrinsic_instr
*intr
= nir_instr_as_intrinsic(instr
);
807 if (intr
->intrinsic
!= nir_intrinsic_load_deref
)
810 nir_deref_instr
*in_deref
= nir_src_as_deref(intr
->src
[0]);
811 if (in_deref
->mode
!= nir_var_shader_in
)
814 nir_variable
*in_var
= nir_deref_instr_get_variable(in_deref
);
816 if (!does_varying_match(out_var
, in_var
))
819 b
.cursor
= nir_before_instr(instr
);
821 nir_load_const_instr
*out_const
=
822 nir_instr_as_load_const(store_intr
->src
[1].ssa
->parent_instr
);
824 /* Add new const to replace the input */
825 nir_ssa_def
*nconst
= nir_build_imm(&b
, store_intr
->num_components
,
826 intr
->dest
.ssa
.bit_size
,
829 nir_ssa_def_rewrite_uses(&intr
->dest
.ssa
, nir_src_for_ssa(nconst
));
839 replace_duplicate_input(nir_shader
*shader
, nir_variable
*input_var
,
840 nir_intrinsic_instr
*dup_store_intr
)
844 nir_function_impl
*impl
= nir_shader_get_entrypoint(shader
);
847 nir_builder_init(&b
, impl
);
849 nir_variable
*dup_out_var
=
850 nir_deref_instr_get_variable(nir_src_as_deref(dup_store_intr
->src
[0]));
852 bool progress
= false;
853 nir_foreach_block(block
, impl
) {
854 nir_foreach_instr(instr
, block
) {
855 if (instr
->type
!= nir_instr_type_intrinsic
)
858 nir_intrinsic_instr
*intr
= nir_instr_as_intrinsic(instr
);
859 if (intr
->intrinsic
!= nir_intrinsic_load_deref
)
862 nir_deref_instr
*in_deref
= nir_src_as_deref(intr
->src
[0]);
863 if (in_deref
->mode
!= nir_var_shader_in
)
866 nir_variable
*in_var
= nir_deref_instr_get_variable(in_deref
);
868 if (!does_varying_match(dup_out_var
, in_var
) ||
869 in_var
->data
.interpolation
!= input_var
->data
.interpolation
||
870 get_interp_loc(in_var
) != get_interp_loc(input_var
))
873 b
.cursor
= nir_before_instr(instr
);
875 nir_ssa_def
*load
= nir_load_var(&b
, input_var
);
876 nir_ssa_def_rewrite_uses(&intr
->dest
.ssa
, nir_src_for_ssa(load
));
886 nir_link_opt_varyings(nir_shader
*producer
, nir_shader
*consumer
)
888 /* TODO: Add support for more shader stage combinations */
889 if (consumer
->info
.stage
!= MESA_SHADER_FRAGMENT
||
890 (producer
->info
.stage
!= MESA_SHADER_VERTEX
&&
891 producer
->info
.stage
!= MESA_SHADER_TESS_EVAL
))
894 bool progress
= false;
896 nir_function_impl
*impl
= nir_shader_get_entrypoint(producer
);
898 struct hash_table
*varying_values
= _mesa_pointer_hash_table_create(NULL
);
900 /* If we find a store in the last block of the producer we can be sure this
901 * is the only possible value for this output.
903 nir_block
*last_block
= nir_impl_last_block(impl
);
904 nir_foreach_instr_reverse(instr
, last_block
) {
905 if (instr
->type
!= nir_instr_type_intrinsic
)
908 nir_intrinsic_instr
*intr
= nir_instr_as_intrinsic(instr
);
910 if (intr
->intrinsic
!= nir_intrinsic_store_deref
)
913 nir_deref_instr
*out_deref
= nir_src_as_deref(intr
->src
[0]);
914 if (out_deref
->mode
!= nir_var_shader_out
)
917 nir_variable
*out_var
= nir_deref_instr_get_variable(out_deref
);
918 if (!can_replace_varying(out_var
))
921 if (intr
->src
[1].ssa
->parent_instr
->type
== nir_instr_type_load_const
) {
922 progress
|= replace_constant_input(consumer
, intr
);
924 struct hash_entry
*entry
=
925 _mesa_hash_table_search(varying_values
, intr
->src
[1].ssa
);
927 progress
|= replace_duplicate_input(consumer
,
928 (nir_variable
*) entry
->data
,
931 nir_variable
*in_var
= get_matching_input_var(consumer
, out_var
);
933 _mesa_hash_table_insert(varying_values
, intr
->src
[1].ssa
,
940 _mesa_hash_table_destroy(varying_values
, NULL
);