2 * Copyright © 2014 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
24 * Connor Abbott (cwabbott0@gmail.com)
25 * Jason Ekstrand (jason@jlekstrand.net)
30 * This lowering pass converts references to input/output variables with
31 * loads/stores to actual input/output intrinsics.
35 #include "nir_builder.h"
36 #include "nir_deref.h"
38 struct lower_io_state
{
41 int (*type_size
)(const struct glsl_type
*type
, bool);
42 nir_variable_mode modes
;
43 nir_lower_io_options options
;
46 static nir_intrinsic_op
47 ssbo_atomic_for_deref(nir_intrinsic_op deref_op
)
50 #define OP(O) case nir_intrinsic_deref_##O: return nir_intrinsic_ssbo_##O;
67 unreachable("Invalid SSBO atomic");
71 static nir_intrinsic_op
72 global_atomic_for_deref(nir_intrinsic_op deref_op
)
75 #define OP(O) case nir_intrinsic_deref_##O: return nir_intrinsic_global_##O;
92 unreachable("Invalid SSBO atomic");
97 nir_assign_var_locations(struct exec_list
*var_list
, unsigned *size
,
98 int (*type_size
)(const struct glsl_type
*, bool))
100 unsigned location
= 0;
102 nir_foreach_variable(var
, var_list
) {
104 * UBOs have their own address spaces, so don't count them towards the
105 * number of global uniforms
107 if (var
->data
.mode
== nir_var_mem_ubo
|| var
->data
.mode
== nir_var_mem_ssbo
)
110 var
->data
.driver_location
= location
;
111 bool bindless_type_size
= var
->data
.mode
== nir_var_shader_in
||
112 var
->data
.mode
== nir_var_shader_out
||
114 location
+= type_size(var
->type
, bindless_type_size
);
121 * Return true if the given variable is a per-vertex input/output array.
122 * (such as geometry shader inputs).
125 nir_is_per_vertex_io(const nir_variable
*var
, gl_shader_stage stage
)
127 if (var
->data
.patch
|| !glsl_type_is_array(var
->type
))
130 if (var
->data
.mode
== nir_var_shader_in
)
131 return stage
== MESA_SHADER_GEOMETRY
||
132 stage
== MESA_SHADER_TESS_CTRL
||
133 stage
== MESA_SHADER_TESS_EVAL
;
135 if (var
->data
.mode
== nir_var_shader_out
)
136 return stage
== MESA_SHADER_TESS_CTRL
;
142 get_io_offset(nir_builder
*b
, nir_deref_instr
*deref
,
143 nir_ssa_def
**vertex_index
,
144 int (*type_size
)(const struct glsl_type
*, bool),
145 unsigned *component
, bool bts
)
148 nir_deref_path_init(&path
, deref
, NULL
);
150 assert(path
.path
[0]->deref_type
== nir_deref_type_var
);
151 nir_deref_instr
**p
= &path
.path
[1];
153 /* For per-vertex input arrays (i.e. geometry shader inputs), keep the
154 * outermost array index separate. Process the rest normally.
156 if (vertex_index
!= NULL
) {
157 assert((*p
)->deref_type
== nir_deref_type_array
);
158 *vertex_index
= nir_ssa_for_src(b
, (*p
)->arr
.index
, 1);
162 if (path
.path
[0]->var
->data
.compact
) {
163 assert((*p
)->deref_type
== nir_deref_type_array
);
164 assert(glsl_type_is_scalar((*p
)->type
));
166 /* We always lower indirect dereferences for "compact" array vars. */
167 const unsigned index
= nir_src_as_uint((*p
)->arr
.index
);
168 const unsigned total_offset
= *component
+ index
;
169 const unsigned slot_offset
= total_offset
/ 4;
170 *component
= total_offset
% 4;
171 return nir_imm_int(b
, type_size(glsl_vec4_type(), bts
) * slot_offset
);
174 /* Just emit code and let constant-folding go to town */
175 nir_ssa_def
*offset
= nir_imm_int(b
, 0);
178 if ((*p
)->deref_type
== nir_deref_type_array
) {
179 unsigned size
= type_size((*p
)->type
, bts
);
182 nir_imul_imm(b
, nir_ssa_for_src(b
, (*p
)->arr
.index
, 1), size
);
184 offset
= nir_iadd(b
, offset
, mul
);
185 } else if ((*p
)->deref_type
== nir_deref_type_struct
) {
186 /* p starts at path[1], so this is safe */
187 nir_deref_instr
*parent
= *(p
- 1);
189 unsigned field_offset
= 0;
190 for (unsigned i
= 0; i
< (*p
)->strct
.index
; i
++) {
191 field_offset
+= type_size(glsl_get_struct_field(parent
->type
, i
), bts
);
193 offset
= nir_iadd_imm(b
, offset
, field_offset
);
195 unreachable("Unsupported deref type");
199 nir_deref_path_finish(&path
);
204 static nir_intrinsic_instr
*
205 lower_load(nir_intrinsic_instr
*intrin
, struct lower_io_state
*state
,
206 nir_ssa_def
*vertex_index
, nir_variable
*var
, nir_ssa_def
*offset
,
207 unsigned component
, const struct glsl_type
*type
)
209 const nir_shader
*nir
= state
->builder
.shader
;
210 nir_variable_mode mode
= var
->data
.mode
;
211 nir_ssa_def
*barycentric
= NULL
;
215 case nir_var_shader_in
:
216 if (nir
->info
.stage
== MESA_SHADER_FRAGMENT
&&
217 nir
->options
->use_interpolated_input_intrinsics
&&
218 var
->data
.interpolation
!= INTERP_MODE_FLAT
) {
219 assert(vertex_index
== NULL
);
221 nir_intrinsic_op bary_op
;
222 if (var
->data
.sample
||
223 (state
->options
& nir_lower_io_force_sample_interpolation
))
224 bary_op
= nir_intrinsic_load_barycentric_sample
;
225 else if (var
->data
.centroid
)
226 bary_op
= nir_intrinsic_load_barycentric_centroid
;
228 bary_op
= nir_intrinsic_load_barycentric_pixel
;
230 barycentric
= nir_load_barycentric(&state
->builder
, bary_op
,
231 var
->data
.interpolation
);
232 op
= nir_intrinsic_load_interpolated_input
;
234 op
= vertex_index
? nir_intrinsic_load_per_vertex_input
:
235 nir_intrinsic_load_input
;
238 case nir_var_shader_out
:
239 op
= vertex_index
? nir_intrinsic_load_per_vertex_output
:
240 nir_intrinsic_load_output
;
242 case nir_var_uniform
:
243 op
= nir_intrinsic_load_uniform
;
245 case nir_var_mem_shared
:
246 op
= nir_intrinsic_load_shared
;
249 unreachable("Unknown variable mode");
252 nir_intrinsic_instr
*load
=
253 nir_intrinsic_instr_create(state
->builder
.shader
, op
);
254 load
->num_components
= intrin
->num_components
;
256 nir_intrinsic_set_base(load
, var
->data
.driver_location
);
257 if (mode
== nir_var_shader_in
|| mode
== nir_var_shader_out
)
258 nir_intrinsic_set_component(load
, component
);
260 if (load
->intrinsic
== nir_intrinsic_load_uniform
)
261 nir_intrinsic_set_range(load
,
262 state
->type_size(var
->type
, var
->data
.bindless
));
264 if (load
->intrinsic
== nir_intrinsic_load_input
||
265 load
->intrinsic
== nir_intrinsic_load_uniform
)
266 nir_intrinsic_set_type(load
, nir_get_nir_type_for_glsl_type(type
));
269 load
->src
[0] = nir_src_for_ssa(vertex_index
);
270 load
->src
[1] = nir_src_for_ssa(offset
);
271 } else if (barycentric
) {
272 load
->src
[0] = nir_src_for_ssa(barycentric
);
273 load
->src
[1] = nir_src_for_ssa(offset
);
275 load
->src
[0] = nir_src_for_ssa(offset
);
281 static nir_intrinsic_instr
*
282 lower_store(nir_intrinsic_instr
*intrin
, struct lower_io_state
*state
,
283 nir_ssa_def
*vertex_index
, nir_variable
*var
, nir_ssa_def
*offset
,
284 unsigned component
, const struct glsl_type
*type
)
286 nir_variable_mode mode
= var
->data
.mode
;
289 if (mode
== nir_var_mem_shared
) {
290 op
= nir_intrinsic_store_shared
;
292 assert(mode
== nir_var_shader_out
);
293 op
= vertex_index
? nir_intrinsic_store_per_vertex_output
:
294 nir_intrinsic_store_output
;
297 nir_intrinsic_instr
*store
=
298 nir_intrinsic_instr_create(state
->builder
.shader
, op
);
299 store
->num_components
= intrin
->num_components
;
301 nir_src_copy(&store
->src
[0], &intrin
->src
[1], store
);
303 nir_intrinsic_set_base(store
, var
->data
.driver_location
);
305 if (mode
== nir_var_shader_out
)
306 nir_intrinsic_set_component(store
, component
);
308 if (store
->intrinsic
== nir_intrinsic_store_output
)
309 nir_intrinsic_set_type(store
, nir_get_nir_type_for_glsl_type(type
));
311 nir_intrinsic_set_write_mask(store
, nir_intrinsic_write_mask(intrin
));
314 store
->src
[1] = nir_src_for_ssa(vertex_index
);
316 store
->src
[vertex_index
? 2 : 1] = nir_src_for_ssa(offset
);
321 static nir_intrinsic_instr
*
322 lower_atomic(nir_intrinsic_instr
*intrin
, struct lower_io_state
*state
,
323 nir_variable
*var
, nir_ssa_def
*offset
)
325 assert(var
->data
.mode
== nir_var_mem_shared
);
328 switch (intrin
->intrinsic
) {
329 #define OP(O) case nir_intrinsic_deref_##O: op = nir_intrinsic_shared_##O; break;
343 OP(atomic_fcomp_swap
)
346 unreachable("Invalid atomic");
349 nir_intrinsic_instr
*atomic
=
350 nir_intrinsic_instr_create(state
->builder
.shader
, op
);
352 nir_intrinsic_set_base(atomic
, var
->data
.driver_location
);
354 atomic
->src
[0] = nir_src_for_ssa(offset
);
355 assert(nir_intrinsic_infos
[intrin
->intrinsic
].num_srcs
==
356 nir_intrinsic_infos
[op
].num_srcs
);
357 for (unsigned i
= 1; i
< nir_intrinsic_infos
[op
].num_srcs
; i
++) {
358 nir_src_copy(&atomic
->src
[i
], &intrin
->src
[i
], atomic
);
364 static nir_intrinsic_instr
*
365 lower_interpolate_at(nir_intrinsic_instr
*intrin
, struct lower_io_state
*state
,
366 nir_variable
*var
, nir_ssa_def
*offset
, unsigned component
,
367 const struct glsl_type
*type
)
369 assert(var
->data
.mode
== nir_var_shader_in
);
371 /* Ignore interpolateAt() for flat variables - flat is flat. */
372 if (var
->data
.interpolation
== INTERP_MODE_FLAT
)
373 return lower_load(intrin
, state
, NULL
, var
, offset
, component
, type
);
375 nir_intrinsic_op bary_op
;
376 switch (intrin
->intrinsic
) {
377 case nir_intrinsic_interp_deref_at_centroid
:
378 bary_op
= (state
->options
& nir_lower_io_force_sample_interpolation
) ?
379 nir_intrinsic_load_barycentric_sample
:
380 nir_intrinsic_load_barycentric_centroid
;
382 case nir_intrinsic_interp_deref_at_sample
:
383 bary_op
= nir_intrinsic_load_barycentric_at_sample
;
385 case nir_intrinsic_interp_deref_at_offset
:
386 bary_op
= nir_intrinsic_load_barycentric_at_offset
;
389 unreachable("Bogus interpolateAt() intrinsic.");
392 nir_intrinsic_instr
*bary_setup
=
393 nir_intrinsic_instr_create(state
->builder
.shader
, bary_op
);
395 nir_ssa_dest_init(&bary_setup
->instr
, &bary_setup
->dest
, 2, 32, NULL
);
396 nir_intrinsic_set_interp_mode(bary_setup
, var
->data
.interpolation
);
398 if (intrin
->intrinsic
== nir_intrinsic_interp_deref_at_sample
||
399 intrin
->intrinsic
== nir_intrinsic_interp_deref_at_offset
)
400 nir_src_copy(&bary_setup
->src
[0], &intrin
->src
[1], bary_setup
);
402 nir_builder_instr_insert(&state
->builder
, &bary_setup
->instr
);
404 nir_intrinsic_instr
*load
=
405 nir_intrinsic_instr_create(state
->builder
.shader
,
406 nir_intrinsic_load_interpolated_input
);
407 load
->num_components
= intrin
->num_components
;
409 nir_intrinsic_set_base(load
, var
->data
.driver_location
);
410 nir_intrinsic_set_component(load
, component
);
412 load
->src
[0] = nir_src_for_ssa(&bary_setup
->dest
.ssa
);
413 load
->src
[1] = nir_src_for_ssa(offset
);
419 nir_lower_io_block(nir_block
*block
,
420 struct lower_io_state
*state
)
422 nir_builder
*b
= &state
->builder
;
423 const nir_shader_compiler_options
*options
= b
->shader
->options
;
424 bool progress
= false;
426 nir_foreach_instr_safe(instr
, block
) {
427 if (instr
->type
!= nir_instr_type_intrinsic
)
430 nir_intrinsic_instr
*intrin
= nir_instr_as_intrinsic(instr
);
432 switch (intrin
->intrinsic
) {
433 case nir_intrinsic_load_deref
:
434 case nir_intrinsic_store_deref
:
435 case nir_intrinsic_deref_atomic_add
:
436 case nir_intrinsic_deref_atomic_imin
:
437 case nir_intrinsic_deref_atomic_umin
:
438 case nir_intrinsic_deref_atomic_imax
:
439 case nir_intrinsic_deref_atomic_umax
:
440 case nir_intrinsic_deref_atomic_and
:
441 case nir_intrinsic_deref_atomic_or
:
442 case nir_intrinsic_deref_atomic_xor
:
443 case nir_intrinsic_deref_atomic_exchange
:
444 case nir_intrinsic_deref_atomic_comp_swap
:
445 case nir_intrinsic_deref_atomic_fadd
:
446 case nir_intrinsic_deref_atomic_fmin
:
447 case nir_intrinsic_deref_atomic_fmax
:
448 case nir_intrinsic_deref_atomic_fcomp_swap
:
449 /* We can lower the io for this nir instrinsic */
451 case nir_intrinsic_interp_deref_at_centroid
:
452 case nir_intrinsic_interp_deref_at_sample
:
453 case nir_intrinsic_interp_deref_at_offset
:
454 /* We can optionally lower these to load_interpolated_input */
455 if (options
->use_interpolated_input_intrinsics
)
458 /* We can't lower the io for this nir instrinsic, so skip it */
462 nir_deref_instr
*deref
= nir_src_as_deref(intrin
->src
[0]);
464 nir_variable
*var
= nir_deref_instr_get_variable(deref
);
465 nir_variable_mode mode
= var
->data
.mode
;
467 if ((state
->modes
& mode
) == 0)
470 if (mode
!= nir_var_shader_in
&&
471 mode
!= nir_var_shader_out
&&
472 mode
!= nir_var_mem_shared
&&
473 mode
!= nir_var_uniform
)
476 b
->cursor
= nir_before_instr(instr
);
478 const bool per_vertex
= nir_is_per_vertex_io(var
, b
->shader
->info
.stage
);
481 nir_ssa_def
*vertex_index
= NULL
;
482 unsigned component_offset
= var
->data
.location_frac
;
483 bool bindless_type_size
= mode
== nir_var_shader_in
||
484 mode
== nir_var_shader_out
||
487 offset
= get_io_offset(b
, deref
, per_vertex
? &vertex_index
: NULL
,
488 state
->type_size
, &component_offset
,
491 nir_intrinsic_instr
*replacement
;
493 switch (intrin
->intrinsic
) {
494 case nir_intrinsic_load_deref
:
495 replacement
= lower_load(intrin
, state
, vertex_index
, var
, offset
,
496 component_offset
, deref
->type
);
499 case nir_intrinsic_store_deref
:
500 replacement
= lower_store(intrin
, state
, vertex_index
, var
, offset
,
501 component_offset
, deref
->type
);
504 case nir_intrinsic_deref_atomic_add
:
505 case nir_intrinsic_deref_atomic_imin
:
506 case nir_intrinsic_deref_atomic_umin
:
507 case nir_intrinsic_deref_atomic_imax
:
508 case nir_intrinsic_deref_atomic_umax
:
509 case nir_intrinsic_deref_atomic_and
:
510 case nir_intrinsic_deref_atomic_or
:
511 case nir_intrinsic_deref_atomic_xor
:
512 case nir_intrinsic_deref_atomic_exchange
:
513 case nir_intrinsic_deref_atomic_comp_swap
:
514 case nir_intrinsic_deref_atomic_fadd
:
515 case nir_intrinsic_deref_atomic_fmin
:
516 case nir_intrinsic_deref_atomic_fmax
:
517 case nir_intrinsic_deref_atomic_fcomp_swap
:
518 assert(vertex_index
== NULL
);
519 replacement
= lower_atomic(intrin
, state
, var
, offset
);
522 case nir_intrinsic_interp_deref_at_centroid
:
523 case nir_intrinsic_interp_deref_at_sample
:
524 case nir_intrinsic_interp_deref_at_offset
:
525 assert(vertex_index
== NULL
);
526 replacement
= lower_interpolate_at(intrin
, state
, var
, offset
,
527 component_offset
, deref
->type
);
534 if (nir_intrinsic_infos
[intrin
->intrinsic
].has_dest
) {
535 if (intrin
->dest
.is_ssa
) {
536 nir_ssa_dest_init(&replacement
->instr
, &replacement
->dest
,
537 intrin
->dest
.ssa
.num_components
,
538 intrin
->dest
.ssa
.bit_size
, NULL
);
539 nir_ssa_def_rewrite_uses(&intrin
->dest
.ssa
,
540 nir_src_for_ssa(&replacement
->dest
.ssa
));
542 nir_dest_copy(&replacement
->dest
, &intrin
->dest
, &intrin
->instr
);
546 nir_instr_insert_before(&intrin
->instr
, &replacement
->instr
);
547 nir_instr_remove(&intrin
->instr
);
555 nir_lower_io_impl(nir_function_impl
*impl
,
556 nir_variable_mode modes
,
557 int (*type_size
)(const struct glsl_type
*, bool),
558 nir_lower_io_options options
)
560 struct lower_io_state state
;
561 bool progress
= false;
563 nir_builder_init(&state
.builder
, impl
);
564 state
.dead_ctx
= ralloc_context(NULL
);
566 state
.type_size
= type_size
;
567 state
.options
= options
;
569 nir_foreach_block(block
, impl
) {
570 progress
|= nir_lower_io_block(block
, &state
);
573 ralloc_free(state
.dead_ctx
);
575 nir_metadata_preserve(impl
, nir_metadata_block_index
|
576 nir_metadata_dominance
);
581 nir_lower_io(nir_shader
*shader
, nir_variable_mode modes
,
582 int (*type_size
)(const struct glsl_type
*, bool),
583 nir_lower_io_options options
)
585 bool progress
= false;
587 nir_foreach_function(function
, shader
) {
588 if (function
->impl
) {
589 progress
|= nir_lower_io_impl(function
->impl
, modes
,
598 type_scalar_size_bytes(const struct glsl_type
*type
)
600 assert(glsl_type_is_vector_or_scalar(type
) ||
601 glsl_type_is_matrix(type
));
602 return glsl_type_is_boolean(type
) ? 4 : glsl_get_bit_size(type
) / 8;
606 build_addr_iadd(nir_builder
*b
, nir_ssa_def
*addr
,
607 nir_address_format addr_format
, nir_ssa_def
*offset
)
609 assert(offset
->num_components
== 1);
610 assert(addr
->bit_size
== offset
->bit_size
);
612 switch (addr_format
) {
613 case nir_address_format_32bit_global
:
614 case nir_address_format_64bit_global
:
615 case nir_address_format_32bit_offset
:
616 assert(addr
->num_components
== 1);
617 return nir_iadd(b
, addr
, offset
);
619 case nir_address_format_64bit_bounded_global
:
620 assert(addr
->num_components
== 4);
621 return nir_vec4(b
, nir_channel(b
, addr
, 0),
622 nir_channel(b
, addr
, 1),
623 nir_channel(b
, addr
, 2),
624 nir_iadd(b
, nir_channel(b
, addr
, 3), offset
));
626 case nir_address_format_32bit_index_offset
:
627 assert(addr
->num_components
== 2);
628 return nir_vec2(b
, nir_channel(b
, addr
, 0),
629 nir_iadd(b
, nir_channel(b
, addr
, 1), offset
));
630 case nir_address_format_logical
:
631 unreachable("Unsupported address format");
633 unreachable("Invalid address format");
637 build_addr_iadd_imm(nir_builder
*b
, nir_ssa_def
*addr
,
638 nir_address_format addr_format
, int64_t offset
)
640 return build_addr_iadd(b
, addr
, addr_format
,
641 nir_imm_intN_t(b
, offset
, addr
->bit_size
));
645 addr_to_index(nir_builder
*b
, nir_ssa_def
*addr
,
646 nir_address_format addr_format
)
648 assert(addr_format
== nir_address_format_32bit_index_offset
);
649 assert(addr
->num_components
== 2);
650 return nir_channel(b
, addr
, 0);
654 addr_to_offset(nir_builder
*b
, nir_ssa_def
*addr
,
655 nir_address_format addr_format
)
657 assert(addr_format
== nir_address_format_32bit_index_offset
);
658 assert(addr
->num_components
== 2);
659 return nir_channel(b
, addr
, 1);
662 /** Returns true if the given address format resolves to a global address */
664 addr_format_is_global(nir_address_format addr_format
)
666 return addr_format
== nir_address_format_32bit_global
||
667 addr_format
== nir_address_format_64bit_global
||
668 addr_format
== nir_address_format_64bit_bounded_global
;
672 addr_to_global(nir_builder
*b
, nir_ssa_def
*addr
,
673 nir_address_format addr_format
)
675 switch (addr_format
) {
676 case nir_address_format_32bit_global
:
677 case nir_address_format_64bit_global
:
678 assert(addr
->num_components
== 1);
681 case nir_address_format_64bit_bounded_global
:
682 assert(addr
->num_components
== 4);
683 return nir_iadd(b
, nir_pack_64_2x32(b
, nir_channels(b
, addr
, 0x3)),
684 nir_u2u64(b
, nir_channel(b
, addr
, 3)));
686 case nir_address_format_32bit_index_offset
:
687 case nir_address_format_32bit_offset
:
688 case nir_address_format_logical
:
689 unreachable("Cannot get a 64-bit address with this address format");
692 unreachable("Invalid address format");
696 addr_format_needs_bounds_check(nir_address_format addr_format
)
698 return addr_format
== nir_address_format_64bit_bounded_global
;
702 addr_is_in_bounds(nir_builder
*b
, nir_ssa_def
*addr
,
703 nir_address_format addr_format
, unsigned size
)
705 assert(addr_format
== nir_address_format_64bit_bounded_global
);
706 assert(addr
->num_components
== 4);
707 return nir_ige(b
, nir_channel(b
, addr
, 2),
708 nir_iadd_imm(b
, nir_channel(b
, addr
, 3), size
));
712 build_explicit_io_load(nir_builder
*b
, nir_intrinsic_instr
*intrin
,
713 nir_ssa_def
*addr
, nir_address_format addr_format
,
714 unsigned num_components
)
716 nir_variable_mode mode
= nir_src_as_deref(intrin
->src
[0])->mode
;
720 case nir_var_mem_ubo
:
721 op
= nir_intrinsic_load_ubo
;
723 case nir_var_mem_ssbo
:
724 if (addr_format_is_global(addr_format
))
725 op
= nir_intrinsic_load_global
;
727 op
= nir_intrinsic_load_ssbo
;
729 case nir_var_mem_global
:
730 assert(addr_format_is_global(addr_format
));
731 op
= nir_intrinsic_load_global
;
733 case nir_var_shader_in
:
734 assert(addr_format_is_global(addr_format
));
735 op
= nir_intrinsic_load_kernel_input
;
738 unreachable("Unsupported explicit IO variable mode");
741 nir_intrinsic_instr
*load
= nir_intrinsic_instr_create(b
->shader
, op
);
743 if (addr_format_is_global(addr_format
)) {
744 load
->src
[0] = nir_src_for_ssa(addr_to_global(b
, addr
, addr_format
));
746 load
->src
[0] = nir_src_for_ssa(addr_to_index(b
, addr
, addr_format
));
747 load
->src
[1] = nir_src_for_ssa(addr_to_offset(b
, addr
, addr_format
));
750 if (mode
!= nir_var_mem_ubo
&& mode
!= nir_var_shader_in
)
751 nir_intrinsic_set_access(load
, nir_intrinsic_access(intrin
));
753 /* TODO: We should try and provide a better alignment. For OpenCL, we need
754 * to plumb the alignment through from SPIR-V when we have one.
756 nir_intrinsic_set_align(load
, intrin
->dest
.ssa
.bit_size
/ 8, 0);
758 assert(intrin
->dest
.is_ssa
);
759 load
->num_components
= num_components
;
760 nir_ssa_dest_init(&load
->instr
, &load
->dest
, num_components
,
761 intrin
->dest
.ssa
.bit_size
, intrin
->dest
.ssa
.name
);
763 assert(load
->dest
.ssa
.bit_size
% 8 == 0);
765 if (addr_format_needs_bounds_check(addr_format
)) {
766 /* The Vulkan spec for robustBufferAccess gives us quite a few options
767 * as to what we can do with an OOB read. Unfortunately, returning
768 * undefined values isn't one of them so we return an actual zero.
770 nir_ssa_def
*zero
= nir_imm_zero(b
, load
->num_components
,
771 load
->dest
.ssa
.bit_size
);
773 const unsigned load_size
=
774 (load
->dest
.ssa
.bit_size
/ 8) * load
->num_components
;
775 nir_push_if(b
, addr_is_in_bounds(b
, addr
, addr_format
, load_size
));
777 nir_builder_instr_insert(b
, &load
->instr
);
781 return nir_if_phi(b
, &load
->dest
.ssa
, zero
);
783 nir_builder_instr_insert(b
, &load
->instr
);
784 return &load
->dest
.ssa
;
789 build_explicit_io_store(nir_builder
*b
, nir_intrinsic_instr
*intrin
,
790 nir_ssa_def
*addr
, nir_address_format addr_format
,
791 nir_ssa_def
*value
, nir_component_mask_t write_mask
)
793 nir_variable_mode mode
= nir_src_as_deref(intrin
->src
[0])->mode
;
797 case nir_var_mem_ssbo
:
798 if (addr_format_is_global(addr_format
))
799 op
= nir_intrinsic_store_global
;
801 op
= nir_intrinsic_store_ssbo
;
803 case nir_var_mem_global
:
804 assert(addr_format_is_global(addr_format
));
805 op
= nir_intrinsic_store_global
;
808 unreachable("Unsupported explicit IO variable mode");
811 nir_intrinsic_instr
*store
= nir_intrinsic_instr_create(b
->shader
, op
);
813 store
->src
[0] = nir_src_for_ssa(value
);
814 if (addr_format_is_global(addr_format
)) {
815 store
->src
[1] = nir_src_for_ssa(addr_to_global(b
, addr
, addr_format
));
817 store
->src
[1] = nir_src_for_ssa(addr_to_index(b
, addr
, addr_format
));
818 store
->src
[2] = nir_src_for_ssa(addr_to_offset(b
, addr
, addr_format
));
821 nir_intrinsic_set_write_mask(store
, write_mask
);
823 nir_intrinsic_set_access(store
, nir_intrinsic_access(intrin
));
825 /* TODO: We should try and provide a better alignment. For OpenCL, we need
826 * to plumb the alignment through from SPIR-V when we have one.
828 nir_intrinsic_set_align(store
, value
->bit_size
/ 8, 0);
830 assert(value
->num_components
== 1 ||
831 value
->num_components
== intrin
->num_components
);
832 store
->num_components
= value
->num_components
;
834 assert(value
->bit_size
% 8 == 0);
836 if (addr_format_needs_bounds_check(addr_format
)) {
837 const unsigned store_size
= (value
->bit_size
/ 8) * store
->num_components
;
838 nir_push_if(b
, addr_is_in_bounds(b
, addr
, addr_format
, store_size
));
840 nir_builder_instr_insert(b
, &store
->instr
);
844 nir_builder_instr_insert(b
, &store
->instr
);
849 build_explicit_io_atomic(nir_builder
*b
, nir_intrinsic_instr
*intrin
,
850 nir_ssa_def
*addr
, nir_address_format addr_format
)
852 nir_variable_mode mode
= nir_src_as_deref(intrin
->src
[0])->mode
;
853 const unsigned num_data_srcs
=
854 nir_intrinsic_infos
[intrin
->intrinsic
].num_srcs
- 1;
858 case nir_var_mem_ssbo
:
859 if (addr_format_is_global(addr_format
))
860 op
= global_atomic_for_deref(intrin
->intrinsic
);
862 op
= ssbo_atomic_for_deref(intrin
->intrinsic
);
864 case nir_var_mem_global
:
865 assert(addr_format_is_global(addr_format
));
866 op
= global_atomic_for_deref(intrin
->intrinsic
);
869 unreachable("Unsupported explicit IO variable mode");
872 nir_intrinsic_instr
*atomic
= nir_intrinsic_instr_create(b
->shader
, op
);
875 if (addr_format_is_global(addr_format
)) {
876 atomic
->src
[src
++] = nir_src_for_ssa(addr_to_global(b
, addr
, addr_format
));
878 atomic
->src
[src
++] = nir_src_for_ssa(addr_to_index(b
, addr
, addr_format
));
879 atomic
->src
[src
++] = nir_src_for_ssa(addr_to_offset(b
, addr
, addr_format
));
881 for (unsigned i
= 0; i
< num_data_srcs
; i
++) {
882 atomic
->src
[src
++] = nir_src_for_ssa(intrin
->src
[1 + i
].ssa
);
885 /* Global atomics don't have access flags because they assume that the
886 * address may be non-uniform.
888 if (!addr_format_is_global(addr_format
))
889 nir_intrinsic_set_access(atomic
, nir_intrinsic_access(intrin
));
891 assert(intrin
->dest
.ssa
.num_components
== 1);
892 nir_ssa_dest_init(&atomic
->instr
, &atomic
->dest
,
893 1, intrin
->dest
.ssa
.bit_size
, intrin
->dest
.ssa
.name
);
895 assert(atomic
->dest
.ssa
.bit_size
% 8 == 0);
897 if (addr_format_needs_bounds_check(addr_format
)) {
898 const unsigned atomic_size
= atomic
->dest
.ssa
.bit_size
/ 8;
899 nir_push_if(b
, addr_is_in_bounds(b
, addr
, addr_format
, atomic_size
));
901 nir_builder_instr_insert(b
, &atomic
->instr
);
904 return nir_if_phi(b
, &atomic
->dest
.ssa
,
905 nir_ssa_undef(b
, 1, atomic
->dest
.ssa
.bit_size
));
907 nir_builder_instr_insert(b
, &atomic
->instr
);
908 return &atomic
->dest
.ssa
;
913 nir_explicit_io_address_from_deref(nir_builder
*b
, nir_deref_instr
*deref
,
914 nir_ssa_def
*base_addr
,
915 nir_address_format addr_format
)
917 assert(deref
->dest
.is_ssa
);
918 switch (deref
->deref_type
) {
919 case nir_deref_type_var
:
920 assert(deref
->mode
== nir_var_shader_in
);
921 return nir_imm_intN_t(b
, deref
->var
->data
.driver_location
,
922 deref
->dest
.ssa
.bit_size
);
924 case nir_deref_type_array
: {
925 nir_deref_instr
*parent
= nir_deref_instr_parent(deref
);
927 unsigned stride
= glsl_get_explicit_stride(parent
->type
);
928 if ((glsl_type_is_matrix(parent
->type
) &&
929 glsl_matrix_type_is_row_major(parent
->type
)) ||
930 (glsl_type_is_vector(parent
->type
) && stride
== 0))
931 stride
= type_scalar_size_bytes(parent
->type
);
935 nir_ssa_def
*index
= nir_ssa_for_src(b
, deref
->arr
.index
, 1);
936 index
= nir_i2i(b
, index
, base_addr
->bit_size
);
937 return build_addr_iadd(b
, base_addr
, addr_format
,
938 nir_imul_imm(b
, index
, stride
));
941 case nir_deref_type_ptr_as_array
: {
942 nir_ssa_def
*index
= nir_ssa_for_src(b
, deref
->arr
.index
, 1);
943 index
= nir_i2i(b
, index
, base_addr
->bit_size
);
944 unsigned stride
= nir_deref_instr_ptr_as_array_stride(deref
);
945 return build_addr_iadd(b
, base_addr
, addr_format
,
946 nir_imul_imm(b
, index
, stride
));
949 case nir_deref_type_array_wildcard
:
950 unreachable("Wildcards should be lowered by now");
953 case nir_deref_type_struct
: {
954 nir_deref_instr
*parent
= nir_deref_instr_parent(deref
);
955 int offset
= glsl_get_struct_field_offset(parent
->type
,
958 return build_addr_iadd_imm(b
, base_addr
, addr_format
, offset
);
961 case nir_deref_type_cast
:
962 /* Nothing to do here */
966 unreachable("Invalid NIR deref type");
970 nir_lower_explicit_io_instr(nir_builder
*b
,
971 nir_intrinsic_instr
*intrin
,
973 nir_address_format addr_format
)
975 b
->cursor
= nir_after_instr(&intrin
->instr
);
977 nir_deref_instr
*deref
= nir_src_as_deref(intrin
->src
[0]);
978 unsigned vec_stride
= glsl_get_explicit_stride(deref
->type
);
979 unsigned scalar_size
= type_scalar_size_bytes(deref
->type
);
980 assert(vec_stride
== 0 || glsl_type_is_vector(deref
->type
));
981 assert(vec_stride
== 0 || vec_stride
>= scalar_size
);
983 if (intrin
->intrinsic
== nir_intrinsic_load_deref
) {
985 if (vec_stride
> scalar_size
) {
986 nir_ssa_def
*comps
[4] = { NULL
, };
987 for (unsigned i
= 0; i
< intrin
->num_components
; i
++) {
988 nir_ssa_def
*comp_addr
= build_addr_iadd_imm(b
, addr
, addr_format
,
990 comps
[i
] = build_explicit_io_load(b
, intrin
, comp_addr
,
993 value
= nir_vec(b
, comps
, intrin
->num_components
);
995 value
= build_explicit_io_load(b
, intrin
, addr
, addr_format
,
996 intrin
->num_components
);
998 nir_ssa_def_rewrite_uses(&intrin
->dest
.ssa
, nir_src_for_ssa(value
));
999 } else if (intrin
->intrinsic
== nir_intrinsic_store_deref
) {
1000 assert(intrin
->src
[1].is_ssa
);
1001 nir_ssa_def
*value
= intrin
->src
[1].ssa
;
1002 nir_component_mask_t write_mask
= nir_intrinsic_write_mask(intrin
);
1003 if (vec_stride
> scalar_size
) {
1004 for (unsigned i
= 0; i
< intrin
->num_components
; i
++) {
1005 if (!(write_mask
& (1 << i
)))
1008 nir_ssa_def
*comp_addr
= build_addr_iadd_imm(b
, addr
, addr_format
,
1010 build_explicit_io_store(b
, intrin
, comp_addr
, addr_format
,
1011 nir_channel(b
, value
, i
), 1);
1014 build_explicit_io_store(b
, intrin
, addr
, addr_format
,
1018 nir_ssa_def
*value
=
1019 build_explicit_io_atomic(b
, intrin
, addr
, addr_format
);
1020 nir_ssa_def_rewrite_uses(&intrin
->dest
.ssa
, nir_src_for_ssa(value
));
1023 nir_instr_remove(&intrin
->instr
);
1027 lower_explicit_io_deref(nir_builder
*b
, nir_deref_instr
*deref
,
1028 nir_address_format addr_format
)
1030 /* Just delete the deref if it's not used. We can't use
1031 * nir_deref_instr_remove_if_unused here because it may remove more than
1032 * one deref which could break our list walking since we walk the list
1035 assert(list_empty(&deref
->dest
.ssa
.if_uses
));
1036 if (list_empty(&deref
->dest
.ssa
.uses
)) {
1037 nir_instr_remove(&deref
->instr
);
1041 b
->cursor
= nir_after_instr(&deref
->instr
);
1043 nir_ssa_def
*base_addr
= NULL
;
1044 if (deref
->deref_type
!= nir_deref_type_var
) {
1045 assert(deref
->parent
.is_ssa
);
1046 base_addr
= deref
->parent
.ssa
;
1049 nir_ssa_def
*addr
= nir_explicit_io_address_from_deref(b
, deref
, base_addr
,
1052 nir_instr_remove(&deref
->instr
);
1053 nir_ssa_def_rewrite_uses(&deref
->dest
.ssa
, nir_src_for_ssa(addr
));
1057 lower_explicit_io_access(nir_builder
*b
, nir_intrinsic_instr
*intrin
,
1058 nir_address_format addr_format
)
1060 assert(intrin
->src
[0].is_ssa
);
1061 nir_lower_explicit_io_instr(b
, intrin
, intrin
->src
[0].ssa
, addr_format
);
1065 lower_explicit_io_array_length(nir_builder
*b
, nir_intrinsic_instr
*intrin
,
1066 nir_address_format addr_format
)
1068 b
->cursor
= nir_after_instr(&intrin
->instr
);
1070 nir_deref_instr
*deref
= nir_src_as_deref(intrin
->src
[0]);
1072 assert(glsl_type_is_array(deref
->type
));
1073 assert(glsl_get_length(deref
->type
) == 0);
1074 unsigned stride
= glsl_get_explicit_stride(deref
->type
);
1077 assert(addr_format
== nir_address_format_32bit_index_offset
);
1078 nir_ssa_def
*addr
= &deref
->dest
.ssa
;
1079 nir_ssa_def
*index
= addr_to_index(b
, addr
, addr_format
);
1080 nir_ssa_def
*offset
= addr_to_offset(b
, addr
, addr_format
);
1082 nir_intrinsic_instr
*bsize
=
1083 nir_intrinsic_instr_create(b
->shader
, nir_intrinsic_get_buffer_size
);
1084 bsize
->src
[0] = nir_src_for_ssa(index
);
1085 nir_ssa_dest_init(&bsize
->instr
, &bsize
->dest
, 1, 32, NULL
);
1086 nir_builder_instr_insert(b
, &bsize
->instr
);
1088 nir_ssa_def
*arr_size
=
1089 nir_idiv(b
, nir_isub(b
, &bsize
->dest
.ssa
, offset
),
1090 nir_imm_int(b
, stride
));
1092 nir_ssa_def_rewrite_uses(&intrin
->dest
.ssa
, nir_src_for_ssa(arr_size
));
1093 nir_instr_remove(&intrin
->instr
);
1097 nir_lower_explicit_io_impl(nir_function_impl
*impl
, nir_variable_mode modes
,
1098 nir_address_format addr_format
)
1100 bool progress
= false;
1103 nir_builder_init(&b
, impl
);
1105 /* Walk in reverse order so that we can see the full deref chain when we
1106 * lower the access operations. We lower them assuming that the derefs
1107 * will be turned into address calculations later.
1109 nir_foreach_block_reverse(block
, impl
) {
1110 nir_foreach_instr_reverse_safe(instr
, block
) {
1111 switch (instr
->type
) {
1112 case nir_instr_type_deref
: {
1113 nir_deref_instr
*deref
= nir_instr_as_deref(instr
);
1114 if (deref
->mode
& modes
) {
1115 lower_explicit_io_deref(&b
, deref
, addr_format
);
1121 case nir_instr_type_intrinsic
: {
1122 nir_intrinsic_instr
*intrin
= nir_instr_as_intrinsic(instr
);
1123 switch (intrin
->intrinsic
) {
1124 case nir_intrinsic_load_deref
:
1125 case nir_intrinsic_store_deref
:
1126 case nir_intrinsic_deref_atomic_add
:
1127 case nir_intrinsic_deref_atomic_imin
:
1128 case nir_intrinsic_deref_atomic_umin
:
1129 case nir_intrinsic_deref_atomic_imax
:
1130 case nir_intrinsic_deref_atomic_umax
:
1131 case nir_intrinsic_deref_atomic_and
:
1132 case nir_intrinsic_deref_atomic_or
:
1133 case nir_intrinsic_deref_atomic_xor
:
1134 case nir_intrinsic_deref_atomic_exchange
:
1135 case nir_intrinsic_deref_atomic_comp_swap
:
1136 case nir_intrinsic_deref_atomic_fadd
:
1137 case nir_intrinsic_deref_atomic_fmin
:
1138 case nir_intrinsic_deref_atomic_fmax
:
1139 case nir_intrinsic_deref_atomic_fcomp_swap
: {
1140 nir_deref_instr
*deref
= nir_src_as_deref(intrin
->src
[0]);
1141 if (deref
->mode
& modes
) {
1142 lower_explicit_io_access(&b
, intrin
, addr_format
);
1148 case nir_intrinsic_deref_buffer_array_length
: {
1149 nir_deref_instr
*deref
= nir_src_as_deref(intrin
->src
[0]);
1150 if (deref
->mode
& modes
) {
1151 lower_explicit_io_array_length(&b
, intrin
, addr_format
);
1171 nir_metadata_preserve(impl
, nir_metadata_block_index
|
1172 nir_metadata_dominance
);
1179 nir_lower_explicit_io(nir_shader
*shader
, nir_variable_mode modes
,
1180 nir_address_format addr_format
)
1182 bool progress
= false;
1184 nir_foreach_function(function
, shader
) {
1185 if (function
->impl
&&
1186 nir_lower_explicit_io_impl(function
->impl
, modes
, addr_format
))
1194 * Return the offset source for a load/store intrinsic.
1197 nir_get_io_offset_src(nir_intrinsic_instr
*instr
)
1199 switch (instr
->intrinsic
) {
1200 case nir_intrinsic_load_input
:
1201 case nir_intrinsic_load_output
:
1202 case nir_intrinsic_load_shared
:
1203 case nir_intrinsic_load_uniform
:
1204 case nir_intrinsic_load_global
:
1205 case nir_intrinsic_load_scratch
:
1206 case nir_intrinsic_load_fs_input_interp_deltas
:
1207 return &instr
->src
[0];
1208 case nir_intrinsic_load_ubo
:
1209 case nir_intrinsic_load_ssbo
:
1210 case nir_intrinsic_load_per_vertex_input
:
1211 case nir_intrinsic_load_per_vertex_output
:
1212 case nir_intrinsic_load_interpolated_input
:
1213 case nir_intrinsic_store_output
:
1214 case nir_intrinsic_store_shared
:
1215 case nir_intrinsic_store_global
:
1216 case nir_intrinsic_store_scratch
:
1217 return &instr
->src
[1];
1218 case nir_intrinsic_store_ssbo
:
1219 case nir_intrinsic_store_per_vertex_output
:
1220 return &instr
->src
[2];
1227 * Return the vertex index source for a load/store per_vertex intrinsic.
1230 nir_get_io_vertex_index_src(nir_intrinsic_instr
*instr
)
1232 switch (instr
->intrinsic
) {
1233 case nir_intrinsic_load_per_vertex_input
:
1234 case nir_intrinsic_load_per_vertex_output
:
1235 return &instr
->src
[0];
1236 case nir_intrinsic_store_per_vertex_output
:
1237 return &instr
->src
[1];
1244 * Return the numeric constant that identify a NULL pointer for each address
1247 const nir_const_value
*
1248 nir_address_format_null_value(nir_address_format addr_format
)
1250 const static nir_const_value null_values
[][NIR_MAX_VEC_COMPONENTS
] = {
1251 [nir_address_format_32bit_global
] = {{0}},
1252 [nir_address_format_64bit_global
] = {{0}},
1253 [nir_address_format_64bit_bounded_global
] = {{0}},
1254 [nir_address_format_32bit_index_offset
] = {{.u32
= ~0}, {.u32
= ~0}},
1255 [nir_address_format_32bit_offset
] = {{.u32
= ~0}},
1256 [nir_address_format_logical
] = {{.u32
= ~0}},
1259 assert(addr_format
< ARRAY_SIZE(null_values
));
1260 return null_values
[addr_format
];
1264 nir_build_addr_ieq(nir_builder
*b
, nir_ssa_def
*addr0
, nir_ssa_def
*addr1
,
1265 nir_address_format addr_format
)
1267 switch (addr_format
) {
1268 case nir_address_format_32bit_global
:
1269 case nir_address_format_64bit_global
:
1270 case nir_address_format_64bit_bounded_global
:
1271 case nir_address_format_32bit_index_offset
:
1272 case nir_address_format_32bit_offset
:
1273 return nir_ball_iequal(b
, addr0
, addr1
);
1275 case nir_address_format_logical
:
1276 unreachable("Unsupported address format");
1279 unreachable("Invalid address format");
1283 nir_build_addr_isub(nir_builder
*b
, nir_ssa_def
*addr0
, nir_ssa_def
*addr1
,
1284 nir_address_format addr_format
)
1286 switch (addr_format
) {
1287 case nir_address_format_32bit_global
:
1288 case nir_address_format_64bit_global
:
1289 case nir_address_format_32bit_offset
:
1290 assert(addr0
->num_components
== 1);
1291 assert(addr1
->num_components
== 1);
1292 return nir_isub(b
, addr0
, addr1
);
1294 case nir_address_format_64bit_bounded_global
:
1295 return nir_isub(b
, addr_to_global(b
, addr0
, addr_format
),
1296 addr_to_global(b
, addr1
, addr_format
));
1298 case nir_address_format_32bit_index_offset
:
1299 assert(addr0
->num_components
== 2);
1300 assert(addr1
->num_components
== 2);
1301 /* Assume the same buffer index. */
1302 return nir_isub(b
, nir_channel(b
, addr0
, 1), nir_channel(b
, addr1
, 1));
1304 case nir_address_format_logical
:
1305 unreachable("Unsupported address format");
1308 unreachable("Invalid address format");
1312 is_input(nir_intrinsic_instr
*intrin
)
1314 return intrin
->intrinsic
== nir_intrinsic_load_input
||
1315 intrin
->intrinsic
== nir_intrinsic_load_per_vertex_input
||
1316 intrin
->intrinsic
== nir_intrinsic_load_interpolated_input
||
1317 intrin
->intrinsic
== nir_intrinsic_load_fs_input_interp_deltas
;
1321 is_output(nir_intrinsic_instr
*intrin
)
1323 return intrin
->intrinsic
== nir_intrinsic_load_output
||
1324 intrin
->intrinsic
== nir_intrinsic_load_per_vertex_output
||
1325 intrin
->intrinsic
== nir_intrinsic_store_output
||
1326 intrin
->intrinsic
== nir_intrinsic_store_per_vertex_output
;
1331 * This pass adds constant offsets to instr->const_index[0] for input/output
1332 * intrinsics, and resets the offset source to 0. Non-constant offsets remain
1333 * unchanged - since we don't know what part of a compound variable is
1334 * accessed, we allocate storage for the entire thing. For drivers that use
1335 * nir_lower_io_to_temporaries() before nir_lower_io(), this guarantees that
1336 * the offset source will be 0, so that they don't have to add it in manually.
1340 add_const_offset_to_base_block(nir_block
*block
, nir_builder
*b
,
1341 nir_variable_mode mode
)
1343 bool progress
= false;
1344 nir_foreach_instr_safe(instr
, block
) {
1345 if (instr
->type
!= nir_instr_type_intrinsic
)
1348 nir_intrinsic_instr
*intrin
= nir_instr_as_intrinsic(instr
);
1350 if ((mode
== nir_var_shader_in
&& is_input(intrin
)) ||
1351 (mode
== nir_var_shader_out
&& is_output(intrin
))) {
1352 nir_src
*offset
= nir_get_io_offset_src(intrin
);
1354 if (nir_src_is_const(*offset
)) {
1355 intrin
->const_index
[0] += nir_src_as_uint(*offset
);
1356 b
->cursor
= nir_before_instr(&intrin
->instr
);
1357 nir_instr_rewrite_src(&intrin
->instr
, offset
,
1358 nir_src_for_ssa(nir_imm_int(b
, 0)));
1368 nir_io_add_const_offset_to_base(nir_shader
*nir
, nir_variable_mode mode
)
1370 bool progress
= false;
1372 nir_foreach_function(f
, nir
) {
1375 nir_builder_init(&b
, f
->impl
);
1376 nir_foreach_block(block
, f
->impl
) {
1377 progress
|= add_const_offset_to_base_block(block
, &b
, mode
);