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
);
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
*))
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 location
+= type_size(var
->type
);
118 * Return true if the given variable is a per-vertex input/output array.
119 * (such as geometry shader inputs).
122 nir_is_per_vertex_io(const nir_variable
*var
, gl_shader_stage stage
)
124 if (var
->data
.patch
|| !glsl_type_is_array(var
->type
))
127 if (var
->data
.mode
== nir_var_shader_in
)
128 return stage
== MESA_SHADER_GEOMETRY
||
129 stage
== MESA_SHADER_TESS_CTRL
||
130 stage
== MESA_SHADER_TESS_EVAL
;
132 if (var
->data
.mode
== nir_var_shader_out
)
133 return stage
== MESA_SHADER_TESS_CTRL
;
139 get_io_offset(nir_builder
*b
, nir_deref_instr
*deref
,
140 nir_ssa_def
**vertex_index
,
141 int (*type_size
)(const struct glsl_type
*),
145 nir_deref_path_init(&path
, deref
, NULL
);
147 assert(path
.path
[0]->deref_type
== nir_deref_type_var
);
148 nir_deref_instr
**p
= &path
.path
[1];
150 /* For per-vertex input arrays (i.e. geometry shader inputs), keep the
151 * outermost array index separate. Process the rest normally.
153 if (vertex_index
!= NULL
) {
154 assert((*p
)->deref_type
== nir_deref_type_array
);
155 *vertex_index
= nir_ssa_for_src(b
, (*p
)->arr
.index
, 1);
159 if (path
.path
[0]->var
->data
.compact
) {
160 assert((*p
)->deref_type
== nir_deref_type_array
);
161 assert(glsl_type_is_scalar((*p
)->type
));
163 /* We always lower indirect dereferences for "compact" array vars. */
164 const unsigned index
= nir_src_as_uint((*p
)->arr
.index
);
165 const unsigned total_offset
= *component
+ index
;
166 const unsigned slot_offset
= total_offset
/ 4;
167 *component
= total_offset
% 4;
168 return nir_imm_int(b
, type_size(glsl_vec4_type()) * slot_offset
);
171 /* Just emit code and let constant-folding go to town */
172 nir_ssa_def
*offset
= nir_imm_int(b
, 0);
175 if ((*p
)->deref_type
== nir_deref_type_array
) {
176 unsigned size
= type_size((*p
)->type
);
179 nir_imul_imm(b
, nir_ssa_for_src(b
, (*p
)->arr
.index
, 1), size
);
181 offset
= nir_iadd(b
, offset
, mul
);
182 } else if ((*p
)->deref_type
== nir_deref_type_struct
) {
183 /* p starts at path[1], so this is safe */
184 nir_deref_instr
*parent
= *(p
- 1);
186 unsigned field_offset
= 0;
187 for (unsigned i
= 0; i
< (*p
)->strct
.index
; i
++) {
188 field_offset
+= type_size(glsl_get_struct_field(parent
->type
, i
));
190 offset
= nir_iadd_imm(b
, offset
, field_offset
);
192 unreachable("Unsupported deref type");
196 nir_deref_path_finish(&path
);
201 static nir_intrinsic_instr
*
202 lower_load(nir_intrinsic_instr
*intrin
, struct lower_io_state
*state
,
203 nir_ssa_def
*vertex_index
, nir_variable
*var
, nir_ssa_def
*offset
,
206 const nir_shader
*nir
= state
->builder
.shader
;
207 nir_variable_mode mode
= var
->data
.mode
;
208 nir_ssa_def
*barycentric
= NULL
;
212 case nir_var_shader_in
:
213 if (nir
->info
.stage
== MESA_SHADER_FRAGMENT
&&
214 nir
->options
->use_interpolated_input_intrinsics
&&
215 var
->data
.interpolation
!= INTERP_MODE_FLAT
) {
216 assert(vertex_index
== NULL
);
218 nir_intrinsic_op bary_op
;
219 if (var
->data
.sample
||
220 (state
->options
& nir_lower_io_force_sample_interpolation
))
221 bary_op
= nir_intrinsic_load_barycentric_sample
;
222 else if (var
->data
.centroid
)
223 bary_op
= nir_intrinsic_load_barycentric_centroid
;
225 bary_op
= nir_intrinsic_load_barycentric_pixel
;
227 barycentric
= nir_load_barycentric(&state
->builder
, bary_op
,
228 var
->data
.interpolation
);
229 op
= nir_intrinsic_load_interpolated_input
;
231 op
= vertex_index
? nir_intrinsic_load_per_vertex_input
:
232 nir_intrinsic_load_input
;
235 case nir_var_shader_out
:
236 op
= vertex_index
? nir_intrinsic_load_per_vertex_output
:
237 nir_intrinsic_load_output
;
239 case nir_var_uniform
:
240 op
= nir_intrinsic_load_uniform
;
242 case nir_var_mem_shared
:
243 op
= nir_intrinsic_load_shared
;
246 unreachable("Unknown variable mode");
249 nir_intrinsic_instr
*load
=
250 nir_intrinsic_instr_create(state
->builder
.shader
, op
);
251 load
->num_components
= intrin
->num_components
;
253 nir_intrinsic_set_base(load
, var
->data
.driver_location
);
254 if (mode
== nir_var_shader_in
|| mode
== nir_var_shader_out
)
255 nir_intrinsic_set_component(load
, component
);
257 if (load
->intrinsic
== nir_intrinsic_load_uniform
)
258 nir_intrinsic_set_range(load
, state
->type_size(var
->type
));
261 load
->src
[0] = nir_src_for_ssa(vertex_index
);
262 load
->src
[1] = nir_src_for_ssa(offset
);
263 } else if (barycentric
) {
264 load
->src
[0] = nir_src_for_ssa(barycentric
);
265 load
->src
[1] = nir_src_for_ssa(offset
);
267 load
->src
[0] = nir_src_for_ssa(offset
);
273 static nir_intrinsic_instr
*
274 lower_store(nir_intrinsic_instr
*intrin
, struct lower_io_state
*state
,
275 nir_ssa_def
*vertex_index
, nir_variable
*var
, nir_ssa_def
*offset
,
278 nir_variable_mode mode
= var
->data
.mode
;
281 if (mode
== nir_var_mem_shared
) {
282 op
= nir_intrinsic_store_shared
;
284 assert(mode
== nir_var_shader_out
);
285 op
= vertex_index
? nir_intrinsic_store_per_vertex_output
:
286 nir_intrinsic_store_output
;
289 nir_intrinsic_instr
*store
=
290 nir_intrinsic_instr_create(state
->builder
.shader
, op
);
291 store
->num_components
= intrin
->num_components
;
293 nir_src_copy(&store
->src
[0], &intrin
->src
[1], store
);
295 nir_intrinsic_set_base(store
, var
->data
.driver_location
);
297 if (mode
== nir_var_shader_out
)
298 nir_intrinsic_set_component(store
, component
);
300 nir_intrinsic_set_write_mask(store
, nir_intrinsic_write_mask(intrin
));
303 store
->src
[1] = nir_src_for_ssa(vertex_index
);
305 store
->src
[vertex_index
? 2 : 1] = nir_src_for_ssa(offset
);
310 static nir_intrinsic_instr
*
311 lower_atomic(nir_intrinsic_instr
*intrin
, struct lower_io_state
*state
,
312 nir_variable
*var
, nir_ssa_def
*offset
)
314 assert(var
->data
.mode
== nir_var_mem_shared
);
317 switch (intrin
->intrinsic
) {
318 #define OP(O) case nir_intrinsic_deref_##O: op = nir_intrinsic_shared_##O; break;
332 OP(atomic_fcomp_swap
)
335 unreachable("Invalid atomic");
338 nir_intrinsic_instr
*atomic
=
339 nir_intrinsic_instr_create(state
->builder
.shader
, op
);
341 nir_intrinsic_set_base(atomic
, var
->data
.driver_location
);
343 atomic
->src
[0] = nir_src_for_ssa(offset
);
344 assert(nir_intrinsic_infos
[intrin
->intrinsic
].num_srcs
==
345 nir_intrinsic_infos
[op
].num_srcs
);
346 for (unsigned i
= 1; i
< nir_intrinsic_infos
[op
].num_srcs
; i
++) {
347 nir_src_copy(&atomic
->src
[i
], &intrin
->src
[i
], atomic
);
353 static nir_intrinsic_instr
*
354 lower_interpolate_at(nir_intrinsic_instr
*intrin
, struct lower_io_state
*state
,
355 nir_variable
*var
, nir_ssa_def
*offset
, unsigned component
)
357 assert(var
->data
.mode
== nir_var_shader_in
);
359 /* Ignore interpolateAt() for flat variables - flat is flat. */
360 if (var
->data
.interpolation
== INTERP_MODE_FLAT
)
361 return lower_load(intrin
, state
, NULL
, var
, offset
, component
);
363 nir_intrinsic_op bary_op
;
364 switch (intrin
->intrinsic
) {
365 case nir_intrinsic_interp_deref_at_centroid
:
366 bary_op
= (state
->options
& nir_lower_io_force_sample_interpolation
) ?
367 nir_intrinsic_load_barycentric_sample
:
368 nir_intrinsic_load_barycentric_centroid
;
370 case nir_intrinsic_interp_deref_at_sample
:
371 bary_op
= nir_intrinsic_load_barycentric_at_sample
;
373 case nir_intrinsic_interp_deref_at_offset
:
374 bary_op
= nir_intrinsic_load_barycentric_at_offset
;
377 unreachable("Bogus interpolateAt() intrinsic.");
380 nir_intrinsic_instr
*bary_setup
=
381 nir_intrinsic_instr_create(state
->builder
.shader
, bary_op
);
383 nir_ssa_dest_init(&bary_setup
->instr
, &bary_setup
->dest
, 2, 32, NULL
);
384 nir_intrinsic_set_interp_mode(bary_setup
, var
->data
.interpolation
);
386 if (intrin
->intrinsic
== nir_intrinsic_interp_deref_at_sample
||
387 intrin
->intrinsic
== nir_intrinsic_interp_deref_at_offset
)
388 nir_src_copy(&bary_setup
->src
[0], &intrin
->src
[1], bary_setup
);
390 nir_builder_instr_insert(&state
->builder
, &bary_setup
->instr
);
392 nir_intrinsic_instr
*load
=
393 nir_intrinsic_instr_create(state
->builder
.shader
,
394 nir_intrinsic_load_interpolated_input
);
395 load
->num_components
= intrin
->num_components
;
397 nir_intrinsic_set_base(load
, var
->data
.driver_location
);
398 nir_intrinsic_set_component(load
, component
);
400 load
->src
[0] = nir_src_for_ssa(&bary_setup
->dest
.ssa
);
401 load
->src
[1] = nir_src_for_ssa(offset
);
407 nir_lower_io_block(nir_block
*block
,
408 struct lower_io_state
*state
)
410 nir_builder
*b
= &state
->builder
;
411 const nir_shader_compiler_options
*options
= b
->shader
->options
;
412 bool progress
= false;
414 nir_foreach_instr_safe(instr
, block
) {
415 if (instr
->type
!= nir_instr_type_intrinsic
)
418 nir_intrinsic_instr
*intrin
= nir_instr_as_intrinsic(instr
);
420 switch (intrin
->intrinsic
) {
421 case nir_intrinsic_load_deref
:
422 case nir_intrinsic_store_deref
:
423 case nir_intrinsic_deref_atomic_add
:
424 case nir_intrinsic_deref_atomic_imin
:
425 case nir_intrinsic_deref_atomic_umin
:
426 case nir_intrinsic_deref_atomic_imax
:
427 case nir_intrinsic_deref_atomic_umax
:
428 case nir_intrinsic_deref_atomic_and
:
429 case nir_intrinsic_deref_atomic_or
:
430 case nir_intrinsic_deref_atomic_xor
:
431 case nir_intrinsic_deref_atomic_exchange
:
432 case nir_intrinsic_deref_atomic_comp_swap
:
433 case nir_intrinsic_deref_atomic_fadd
:
434 case nir_intrinsic_deref_atomic_fmin
:
435 case nir_intrinsic_deref_atomic_fmax
:
436 case nir_intrinsic_deref_atomic_fcomp_swap
:
437 /* We can lower the io for this nir instrinsic */
439 case nir_intrinsic_interp_deref_at_centroid
:
440 case nir_intrinsic_interp_deref_at_sample
:
441 case nir_intrinsic_interp_deref_at_offset
:
442 /* We can optionally lower these to load_interpolated_input */
443 if (options
->use_interpolated_input_intrinsics
)
446 /* We can't lower the io for this nir instrinsic, so skip it */
450 nir_deref_instr
*deref
= nir_src_as_deref(intrin
->src
[0]);
452 nir_variable
*var
= nir_deref_instr_get_variable(deref
);
453 nir_variable_mode mode
= var
->data
.mode
;
455 if ((state
->modes
& mode
) == 0)
458 if (mode
!= nir_var_shader_in
&&
459 mode
!= nir_var_shader_out
&&
460 mode
!= nir_var_mem_shared
&&
461 mode
!= nir_var_uniform
)
464 b
->cursor
= nir_before_instr(instr
);
466 const bool per_vertex
= nir_is_per_vertex_io(var
, b
->shader
->info
.stage
);
469 nir_ssa_def
*vertex_index
= NULL
;
470 unsigned component_offset
= var
->data
.location_frac
;
472 offset
= get_io_offset(b
, deref
, per_vertex
? &vertex_index
: NULL
,
473 state
->type_size
, &component_offset
);
475 nir_intrinsic_instr
*replacement
;
477 switch (intrin
->intrinsic
) {
478 case nir_intrinsic_load_deref
:
479 replacement
= lower_load(intrin
, state
, vertex_index
, var
, offset
,
483 case nir_intrinsic_store_deref
:
484 replacement
= lower_store(intrin
, state
, vertex_index
, var
, offset
,
488 case nir_intrinsic_deref_atomic_add
:
489 case nir_intrinsic_deref_atomic_imin
:
490 case nir_intrinsic_deref_atomic_umin
:
491 case nir_intrinsic_deref_atomic_imax
:
492 case nir_intrinsic_deref_atomic_umax
:
493 case nir_intrinsic_deref_atomic_and
:
494 case nir_intrinsic_deref_atomic_or
:
495 case nir_intrinsic_deref_atomic_xor
:
496 case nir_intrinsic_deref_atomic_exchange
:
497 case nir_intrinsic_deref_atomic_comp_swap
:
498 case nir_intrinsic_deref_atomic_fadd
:
499 case nir_intrinsic_deref_atomic_fmin
:
500 case nir_intrinsic_deref_atomic_fmax
:
501 case nir_intrinsic_deref_atomic_fcomp_swap
:
502 assert(vertex_index
== NULL
);
503 replacement
= lower_atomic(intrin
, state
, var
, offset
);
506 case nir_intrinsic_interp_deref_at_centroid
:
507 case nir_intrinsic_interp_deref_at_sample
:
508 case nir_intrinsic_interp_deref_at_offset
:
509 assert(vertex_index
== NULL
);
510 replacement
= lower_interpolate_at(intrin
, state
, var
, offset
,
518 if (nir_intrinsic_infos
[intrin
->intrinsic
].has_dest
) {
519 if (intrin
->dest
.is_ssa
) {
520 nir_ssa_dest_init(&replacement
->instr
, &replacement
->dest
,
521 intrin
->dest
.ssa
.num_components
,
522 intrin
->dest
.ssa
.bit_size
, NULL
);
523 nir_ssa_def_rewrite_uses(&intrin
->dest
.ssa
,
524 nir_src_for_ssa(&replacement
->dest
.ssa
));
526 nir_dest_copy(&replacement
->dest
, &intrin
->dest
, &intrin
->instr
);
530 nir_instr_insert_before(&intrin
->instr
, &replacement
->instr
);
531 nir_instr_remove(&intrin
->instr
);
539 nir_lower_io_impl(nir_function_impl
*impl
,
540 nir_variable_mode modes
,
541 int (*type_size
)(const struct glsl_type
*),
542 nir_lower_io_options options
)
544 struct lower_io_state state
;
545 bool progress
= false;
547 nir_builder_init(&state
.builder
, impl
);
548 state
.dead_ctx
= ralloc_context(NULL
);
550 state
.type_size
= type_size
;
551 state
.options
= options
;
553 nir_foreach_block(block
, impl
) {
554 progress
|= nir_lower_io_block(block
, &state
);
557 ralloc_free(state
.dead_ctx
);
559 nir_metadata_preserve(impl
, nir_metadata_block_index
|
560 nir_metadata_dominance
);
565 nir_lower_io(nir_shader
*shader
, nir_variable_mode modes
,
566 int (*type_size
)(const struct glsl_type
*),
567 nir_lower_io_options options
)
569 bool progress
= false;
571 nir_foreach_function(function
, shader
) {
572 if (function
->impl
) {
573 progress
|= nir_lower_io_impl(function
->impl
, modes
,
582 type_scalar_size_bytes(const struct glsl_type
*type
)
584 assert(glsl_type_is_vector_or_scalar(type
) ||
585 glsl_type_is_matrix(type
));
586 return glsl_type_is_boolean(type
) ? 4 : glsl_get_bit_size(type
) / 8;
590 build_addr_iadd(nir_builder
*b
, nir_ssa_def
*addr
,
591 nir_address_format addr_format
, nir_ssa_def
*offset
)
593 assert(offset
->num_components
== 1);
594 assert(addr
->bit_size
== offset
->bit_size
);
596 switch (addr_format
) {
597 case nir_address_format_32bit_global
:
598 case nir_address_format_64bit_global
:
599 assert(addr
->num_components
== 1);
600 return nir_iadd(b
, addr
, offset
);
602 case nir_address_format_64bit_bounded_global
:
603 assert(addr
->num_components
== 4);
604 return nir_vec4(b
, nir_channel(b
, addr
, 0),
605 nir_channel(b
, addr
, 1),
606 nir_channel(b
, addr
, 2),
607 nir_iadd(b
, nir_channel(b
, addr
, 3), offset
));
609 case nir_address_format_32bit_index_offset
:
610 assert(addr
->num_components
== 2);
611 return nir_vec2(b
, nir_channel(b
, addr
, 0),
612 nir_iadd(b
, nir_channel(b
, addr
, 1), offset
));
614 unreachable("Invalid address format");
618 build_addr_iadd_imm(nir_builder
*b
, nir_ssa_def
*addr
,
619 nir_address_format addr_format
, int64_t offset
)
621 return build_addr_iadd(b
, addr
, addr_format
,
622 nir_imm_intN_t(b
, offset
, addr
->bit_size
));
626 addr_to_index(nir_builder
*b
, nir_ssa_def
*addr
,
627 nir_address_format addr_format
)
629 assert(addr_format
== nir_address_format_32bit_index_offset
);
630 assert(addr
->num_components
== 2);
631 return nir_channel(b
, addr
, 0);
635 addr_to_offset(nir_builder
*b
, nir_ssa_def
*addr
,
636 nir_address_format addr_format
)
638 assert(addr_format
== nir_address_format_32bit_index_offset
);
639 assert(addr
->num_components
== 2);
640 return nir_channel(b
, addr
, 1);
643 /** Returns true if the given address format resolves to a global address */
645 addr_format_is_global(nir_address_format addr_format
)
647 return addr_format
== nir_address_format_32bit_global
||
648 addr_format
== nir_address_format_64bit_global
||
649 addr_format
== nir_address_format_64bit_bounded_global
;
653 addr_to_global(nir_builder
*b
, nir_ssa_def
*addr
,
654 nir_address_format addr_format
)
656 switch (addr_format
) {
657 case nir_address_format_32bit_global
:
658 case nir_address_format_64bit_global
:
659 assert(addr
->num_components
== 1);
662 case nir_address_format_64bit_bounded_global
:
663 assert(addr
->num_components
== 4);
664 return nir_iadd(b
, nir_pack_64_2x32(b
, nir_channels(b
, addr
, 0x3)),
665 nir_u2u64(b
, nir_channel(b
, addr
, 3)));
667 case nir_address_format_32bit_index_offset
:
668 unreachable("Cannot get a 64-bit address with this address format");
671 unreachable("Invalid address format");
675 addr_format_needs_bounds_check(nir_address_format addr_format
)
677 return addr_format
== nir_address_format_64bit_bounded_global
;
681 addr_is_in_bounds(nir_builder
*b
, nir_ssa_def
*addr
,
682 nir_address_format addr_format
, unsigned size
)
684 assert(addr_format
== nir_address_format_64bit_bounded_global
);
685 assert(addr
->num_components
== 4);
686 return nir_ige(b
, nir_channel(b
, addr
, 2),
687 nir_iadd_imm(b
, nir_channel(b
, addr
, 3), size
));
691 build_explicit_io_load(nir_builder
*b
, nir_intrinsic_instr
*intrin
,
692 nir_ssa_def
*addr
, nir_address_format addr_format
,
693 unsigned num_components
)
695 nir_variable_mode mode
= nir_src_as_deref(intrin
->src
[0])->mode
;
699 case nir_var_mem_ubo
:
700 op
= nir_intrinsic_load_ubo
;
702 case nir_var_mem_ssbo
:
703 if (addr_format_is_global(addr_format
))
704 op
= nir_intrinsic_load_global
;
706 op
= nir_intrinsic_load_ssbo
;
708 case nir_var_mem_global
:
709 assert(addr_format_is_global(addr_format
));
710 op
= nir_intrinsic_load_global
;
712 case nir_var_shader_in
:
713 assert(addr_format_is_global(addr_format
));
714 op
= nir_intrinsic_load_kernel_input
;
717 unreachable("Unsupported explicit IO variable mode");
720 nir_intrinsic_instr
*load
= nir_intrinsic_instr_create(b
->shader
, op
);
722 if (addr_format_is_global(addr_format
)) {
723 load
->src
[0] = nir_src_for_ssa(addr_to_global(b
, addr
, addr_format
));
725 load
->src
[0] = nir_src_for_ssa(addr_to_index(b
, addr
, addr_format
));
726 load
->src
[1] = nir_src_for_ssa(addr_to_offset(b
, addr
, addr_format
));
729 if (mode
!= nir_var_mem_ubo
&& mode
!= nir_var_shader_in
)
730 nir_intrinsic_set_access(load
, nir_intrinsic_access(intrin
));
732 /* TODO: We should try and provide a better alignment. For OpenCL, we need
733 * to plumb the alignment through from SPIR-V when we have one.
735 nir_intrinsic_set_align(load
, intrin
->dest
.ssa
.bit_size
/ 8, 0);
737 assert(intrin
->dest
.is_ssa
);
738 load
->num_components
= num_components
;
739 nir_ssa_dest_init(&load
->instr
, &load
->dest
, num_components
,
740 intrin
->dest
.ssa
.bit_size
, intrin
->dest
.ssa
.name
);
742 assert(load
->dest
.ssa
.bit_size
% 8 == 0);
744 if (addr_format_needs_bounds_check(addr_format
)) {
745 /* The Vulkan spec for robustBufferAccess gives us quite a few options
746 * as to what we can do with an OOB read. Unfortunately, returning
747 * undefined values isn't one of them so we return an actual zero.
749 nir_const_value zero_val
;
750 memset(&zero_val
, 0, sizeof(zero_val
));
751 nir_ssa_def
*zero
= nir_build_imm(b
, load
->num_components
,
752 load
->dest
.ssa
.bit_size
, zero_val
);
754 const unsigned load_size
=
755 (load
->dest
.ssa
.bit_size
/ 8) * load
->num_components
;
756 nir_push_if(b
, addr_is_in_bounds(b
, addr
, addr_format
, load_size
));
758 nir_builder_instr_insert(b
, &load
->instr
);
762 return nir_if_phi(b
, &load
->dest
.ssa
, zero
);
764 nir_builder_instr_insert(b
, &load
->instr
);
765 return &load
->dest
.ssa
;
770 build_explicit_io_store(nir_builder
*b
, nir_intrinsic_instr
*intrin
,
771 nir_ssa_def
*addr
, nir_address_format addr_format
,
772 nir_ssa_def
*value
, nir_component_mask_t write_mask
)
774 nir_variable_mode mode
= nir_src_as_deref(intrin
->src
[0])->mode
;
778 case nir_var_mem_ssbo
:
779 if (addr_format_is_global(addr_format
))
780 op
= nir_intrinsic_store_global
;
782 op
= nir_intrinsic_store_ssbo
;
784 case nir_var_mem_global
:
785 assert(addr_format_is_global(addr_format
));
786 op
= nir_intrinsic_store_global
;
789 unreachable("Unsupported explicit IO variable mode");
792 nir_intrinsic_instr
*store
= nir_intrinsic_instr_create(b
->shader
, op
);
794 store
->src
[0] = nir_src_for_ssa(value
);
795 if (addr_format_is_global(addr_format
)) {
796 store
->src
[1] = nir_src_for_ssa(addr_to_global(b
, addr
, addr_format
));
798 store
->src
[1] = nir_src_for_ssa(addr_to_index(b
, addr
, addr_format
));
799 store
->src
[2] = nir_src_for_ssa(addr_to_offset(b
, addr
, addr_format
));
802 nir_intrinsic_set_write_mask(store
, write_mask
);
804 nir_intrinsic_set_access(store
, nir_intrinsic_access(intrin
));
806 /* TODO: We should try and provide a better alignment. For OpenCL, we need
807 * to plumb the alignment through from SPIR-V when we have one.
809 nir_intrinsic_set_align(store
, value
->bit_size
/ 8, 0);
811 assert(value
->num_components
== 1 ||
812 value
->num_components
== intrin
->num_components
);
813 store
->num_components
= value
->num_components
;
815 assert(value
->bit_size
% 8 == 0);
817 if (addr_format_needs_bounds_check(addr_format
)) {
818 const unsigned store_size
= (value
->bit_size
/ 8) * store
->num_components
;
819 nir_push_if(b
, addr_is_in_bounds(b
, addr
, addr_format
, store_size
));
821 nir_builder_instr_insert(b
, &store
->instr
);
825 nir_builder_instr_insert(b
, &store
->instr
);
830 build_explicit_io_atomic(nir_builder
*b
, nir_intrinsic_instr
*intrin
,
831 nir_ssa_def
*addr
, nir_address_format addr_format
)
833 nir_variable_mode mode
= nir_src_as_deref(intrin
->src
[0])->mode
;
834 const unsigned num_data_srcs
=
835 nir_intrinsic_infos
[intrin
->intrinsic
].num_srcs
- 1;
839 case nir_var_mem_ssbo
:
840 if (addr_format_is_global(addr_format
))
841 op
= global_atomic_for_deref(intrin
->intrinsic
);
843 op
= ssbo_atomic_for_deref(intrin
->intrinsic
);
845 case nir_var_mem_global
:
846 assert(addr_format_is_global(addr_format
));
847 op
= global_atomic_for_deref(intrin
->intrinsic
);
850 unreachable("Unsupported explicit IO variable mode");
853 nir_intrinsic_instr
*atomic
= nir_intrinsic_instr_create(b
->shader
, op
);
856 if (addr_format_is_global(addr_format
)) {
857 atomic
->src
[src
++] = nir_src_for_ssa(addr_to_global(b
, addr
, addr_format
));
859 atomic
->src
[src
++] = nir_src_for_ssa(addr_to_index(b
, addr
, addr_format
));
860 atomic
->src
[src
++] = nir_src_for_ssa(addr_to_offset(b
, addr
, addr_format
));
862 for (unsigned i
= 0; i
< num_data_srcs
; i
++) {
863 atomic
->src
[src
++] = nir_src_for_ssa(intrin
->src
[1 + i
].ssa
);
866 assert(intrin
->dest
.ssa
.num_components
== 1);
867 nir_ssa_dest_init(&atomic
->instr
, &atomic
->dest
,
868 1, intrin
->dest
.ssa
.bit_size
, intrin
->dest
.ssa
.name
);
870 assert(atomic
->dest
.ssa
.bit_size
% 8 == 0);
872 if (addr_format_needs_bounds_check(addr_format
)) {
873 const unsigned atomic_size
= atomic
->dest
.ssa
.bit_size
/ 8;
874 nir_push_if(b
, addr_is_in_bounds(b
, addr
, addr_format
, atomic_size
));
876 nir_builder_instr_insert(b
, &atomic
->instr
);
879 return nir_if_phi(b
, &atomic
->dest
.ssa
,
880 nir_ssa_undef(b
, 1, atomic
->dest
.ssa
.bit_size
));
882 nir_builder_instr_insert(b
, &atomic
->instr
);
883 return &atomic
->dest
.ssa
;
888 lower_explicit_io_deref(nir_builder
*b
, nir_deref_instr
*deref
,
889 nir_address_format addr_format
)
891 /* Just delete the deref if it's not used. We can't use
892 * nir_deref_instr_remove_if_unused here because it may remove more than
893 * one deref which could break our list walking since we walk the list
896 assert(list_empty(&deref
->dest
.ssa
.if_uses
));
897 if (list_empty(&deref
->dest
.ssa
.uses
)) {
898 nir_instr_remove(&deref
->instr
);
902 b
->cursor
= nir_after_instr(&deref
->instr
);
904 nir_ssa_def
*parent_addr
= NULL
;
905 if (deref
->deref_type
!= nir_deref_type_var
) {
906 assert(deref
->parent
.is_ssa
);
907 parent_addr
= deref
->parent
.ssa
;
911 nir_ssa_def
*addr
= NULL
;
912 assert(deref
->dest
.is_ssa
);
913 switch (deref
->deref_type
) {
914 case nir_deref_type_var
:
915 assert(deref
->mode
== nir_var_shader_in
);
916 addr
= nir_imm_intN_t(b
, deref
->var
->data
.driver_location
,
917 deref
->dest
.ssa
.bit_size
);
920 case nir_deref_type_array
: {
921 nir_deref_instr
*parent
= nir_deref_instr_parent(deref
);
923 unsigned stride
= glsl_get_explicit_stride(parent
->type
);
924 if ((glsl_type_is_matrix(parent
->type
) &&
925 glsl_matrix_type_is_row_major(parent
->type
)) ||
926 (glsl_type_is_vector(parent
->type
) && stride
== 0))
927 stride
= type_scalar_size_bytes(parent
->type
);
931 nir_ssa_def
*index
= nir_ssa_for_src(b
, deref
->arr
.index
, 1);
932 index
= nir_i2i(b
, index
, parent_addr
->bit_size
);
933 addr
= build_addr_iadd(b
, parent_addr
, addr_format
,
934 nir_imul_imm(b
, index
, stride
));
938 case nir_deref_type_ptr_as_array
: {
939 nir_ssa_def
*index
= nir_ssa_for_src(b
, deref
->arr
.index
, 1);
940 index
= nir_i2i(b
, index
, parent_addr
->bit_size
);
941 unsigned stride
= nir_deref_instr_ptr_as_array_stride(deref
);
942 addr
= build_addr_iadd(b
, parent_addr
, addr_format
,
943 nir_imul_imm(b
, index
, stride
));
947 case nir_deref_type_array_wildcard
:
948 unreachable("Wildcards should be lowered by now");
951 case nir_deref_type_struct
: {
952 nir_deref_instr
*parent
= nir_deref_instr_parent(deref
);
953 int offset
= glsl_get_struct_field_offset(parent
->type
,
956 addr
= build_addr_iadd_imm(b
, parent_addr
, addr_format
, offset
);
960 case nir_deref_type_cast
:
961 /* Nothing to do here */
966 nir_instr_remove(&deref
->instr
);
967 nir_ssa_def_rewrite_uses(&deref
->dest
.ssa
, nir_src_for_ssa(addr
));
971 lower_explicit_io_access(nir_builder
*b
, nir_intrinsic_instr
*intrin
,
972 nir_address_format addr_format
)
974 b
->cursor
= nir_after_instr(&intrin
->instr
);
976 nir_deref_instr
*deref
= nir_src_as_deref(intrin
->src
[0]);
977 unsigned vec_stride
= glsl_get_explicit_stride(deref
->type
);
978 unsigned scalar_size
= type_scalar_size_bytes(deref
->type
);
979 assert(vec_stride
== 0 || glsl_type_is_vector(deref
->type
));
980 assert(vec_stride
== 0 || vec_stride
>= scalar_size
);
982 nir_ssa_def
*addr
= &deref
->dest
.ssa
;
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_array_length(nir_builder
*b
, nir_intrinsic_instr
*intrin
,
1028 nir_address_format addr_format
)
1030 b
->cursor
= nir_after_instr(&intrin
->instr
);
1032 nir_deref_instr
*deref
= nir_src_as_deref(intrin
->src
[0]);
1034 assert(glsl_type_is_array(deref
->type
));
1035 assert(glsl_get_length(deref
->type
) == 0);
1036 unsigned stride
= glsl_get_explicit_stride(deref
->type
);
1039 assert(addr_format
== nir_address_format_32bit_index_offset
);
1040 nir_ssa_def
*addr
= &deref
->dest
.ssa
;
1041 nir_ssa_def
*index
= addr_to_index(b
, addr
, addr_format
);
1042 nir_ssa_def
*offset
= addr_to_offset(b
, addr
, addr_format
);
1044 nir_intrinsic_instr
*bsize
=
1045 nir_intrinsic_instr_create(b
->shader
, nir_intrinsic_get_buffer_size
);
1046 bsize
->src
[0] = nir_src_for_ssa(index
);
1047 nir_ssa_dest_init(&bsize
->instr
, &bsize
->dest
, 1, 32, NULL
);
1048 nir_builder_instr_insert(b
, &bsize
->instr
);
1050 nir_ssa_def
*arr_size
=
1051 nir_idiv(b
, nir_isub(b
, &bsize
->dest
.ssa
, offset
),
1052 nir_imm_int(b
, stride
));
1054 nir_ssa_def_rewrite_uses(&intrin
->dest
.ssa
, nir_src_for_ssa(arr_size
));
1055 nir_instr_remove(&intrin
->instr
);
1059 nir_lower_explicit_io_impl(nir_function_impl
*impl
, nir_variable_mode modes
,
1060 nir_address_format addr_format
)
1062 bool progress
= false;
1065 nir_builder_init(&b
, impl
);
1067 /* Walk in reverse order so that we can see the full deref chain when we
1068 * lower the access operations. We lower them assuming that the derefs
1069 * will be turned into address calculations later.
1071 nir_foreach_block_reverse(block
, impl
) {
1072 nir_foreach_instr_reverse_safe(instr
, block
) {
1073 switch (instr
->type
) {
1074 case nir_instr_type_deref
: {
1075 nir_deref_instr
*deref
= nir_instr_as_deref(instr
);
1076 if (deref
->mode
& modes
) {
1077 lower_explicit_io_deref(&b
, deref
, addr_format
);
1083 case nir_instr_type_intrinsic
: {
1084 nir_intrinsic_instr
*intrin
= nir_instr_as_intrinsic(instr
);
1085 switch (intrin
->intrinsic
) {
1086 case nir_intrinsic_load_deref
:
1087 case nir_intrinsic_store_deref
:
1088 case nir_intrinsic_deref_atomic_add
:
1089 case nir_intrinsic_deref_atomic_imin
:
1090 case nir_intrinsic_deref_atomic_umin
:
1091 case nir_intrinsic_deref_atomic_imax
:
1092 case nir_intrinsic_deref_atomic_umax
:
1093 case nir_intrinsic_deref_atomic_and
:
1094 case nir_intrinsic_deref_atomic_or
:
1095 case nir_intrinsic_deref_atomic_xor
:
1096 case nir_intrinsic_deref_atomic_exchange
:
1097 case nir_intrinsic_deref_atomic_comp_swap
:
1098 case nir_intrinsic_deref_atomic_fadd
:
1099 case nir_intrinsic_deref_atomic_fmin
:
1100 case nir_intrinsic_deref_atomic_fmax
:
1101 case nir_intrinsic_deref_atomic_fcomp_swap
: {
1102 nir_deref_instr
*deref
= nir_src_as_deref(intrin
->src
[0]);
1103 if (deref
->mode
& modes
) {
1104 lower_explicit_io_access(&b
, intrin
, addr_format
);
1110 case nir_intrinsic_deref_buffer_array_length
: {
1111 nir_deref_instr
*deref
= nir_src_as_deref(intrin
->src
[0]);
1112 if (deref
->mode
& modes
) {
1113 lower_explicit_io_array_length(&b
, intrin
, addr_format
);
1133 nir_metadata_preserve(impl
, nir_metadata_block_index
|
1134 nir_metadata_dominance
);
1141 nir_lower_explicit_io(nir_shader
*shader
, nir_variable_mode modes
,
1142 nir_address_format addr_format
)
1144 bool progress
= false;
1146 nir_foreach_function(function
, shader
) {
1147 if (function
->impl
&&
1148 nir_lower_explicit_io_impl(function
->impl
, modes
, addr_format
))
1156 * Return the offset source for a load/store intrinsic.
1159 nir_get_io_offset_src(nir_intrinsic_instr
*instr
)
1161 switch (instr
->intrinsic
) {
1162 case nir_intrinsic_load_input
:
1163 case nir_intrinsic_load_output
:
1164 case nir_intrinsic_load_shared
:
1165 case nir_intrinsic_load_uniform
:
1166 case nir_intrinsic_load_global
:
1167 return &instr
->src
[0];
1168 case nir_intrinsic_load_ubo
:
1169 case nir_intrinsic_load_ssbo
:
1170 case nir_intrinsic_load_per_vertex_input
:
1171 case nir_intrinsic_load_per_vertex_output
:
1172 case nir_intrinsic_load_interpolated_input
:
1173 case nir_intrinsic_store_output
:
1174 case nir_intrinsic_store_shared
:
1175 case nir_intrinsic_store_global
:
1176 return &instr
->src
[1];
1177 case nir_intrinsic_store_ssbo
:
1178 case nir_intrinsic_store_per_vertex_output
:
1179 return &instr
->src
[2];
1186 * Return the vertex index source for a load/store per_vertex intrinsic.
1189 nir_get_io_vertex_index_src(nir_intrinsic_instr
*instr
)
1191 switch (instr
->intrinsic
) {
1192 case nir_intrinsic_load_per_vertex_input
:
1193 case nir_intrinsic_load_per_vertex_output
:
1194 return &instr
->src
[0];
1195 case nir_intrinsic_store_per_vertex_output
:
1196 return &instr
->src
[1];