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 #include "util/u_math.h"
40 struct lower_io_state
{
43 int (*type_size
)(const struct glsl_type
*type
, bool);
44 nir_variable_mode modes
;
45 nir_lower_io_options options
;
48 static nir_intrinsic_op
49 ssbo_atomic_for_deref(nir_intrinsic_op deref_op
)
52 #define OP(O) case nir_intrinsic_deref_##O: return nir_intrinsic_ssbo_##O;
69 unreachable("Invalid SSBO atomic");
73 static nir_intrinsic_op
74 global_atomic_for_deref(nir_intrinsic_op deref_op
)
77 #define OP(O) case nir_intrinsic_deref_##O: return nir_intrinsic_global_##O;
94 unreachable("Invalid SSBO atomic");
98 static nir_intrinsic_op
99 shared_atomic_for_deref(nir_intrinsic_op deref_op
)
102 #define OP(O) case nir_intrinsic_deref_##O: return nir_intrinsic_shared_##O;
116 OP(atomic_fcomp_swap
)
119 unreachable("Invalid shared atomic");
124 nir_assign_var_locations(struct exec_list
*var_list
, unsigned *size
,
125 int (*type_size
)(const struct glsl_type
*, bool))
127 unsigned location
= 0;
129 nir_foreach_variable(var
, var_list
) {
131 * UBOs have their own address spaces, so don't count them towards the
132 * number of global uniforms
134 if (var
->data
.mode
== nir_var_mem_ubo
|| var
->data
.mode
== nir_var_mem_ssbo
)
137 var
->data
.driver_location
= location
;
138 bool bindless_type_size
= var
->data
.mode
== nir_var_shader_in
||
139 var
->data
.mode
== nir_var_shader_out
||
141 location
+= type_size(var
->type
, bindless_type_size
);
148 * Return true if the given variable is a per-vertex input/output array.
149 * (such as geometry shader inputs).
152 nir_is_per_vertex_io(const nir_variable
*var
, gl_shader_stage stage
)
154 if (var
->data
.patch
|| !glsl_type_is_array(var
->type
))
157 if (var
->data
.mode
== nir_var_shader_in
)
158 return stage
== MESA_SHADER_GEOMETRY
||
159 stage
== MESA_SHADER_TESS_CTRL
||
160 stage
== MESA_SHADER_TESS_EVAL
;
162 if (var
->data
.mode
== nir_var_shader_out
)
163 return stage
== MESA_SHADER_TESS_CTRL
;
169 get_io_offset(nir_builder
*b
, nir_deref_instr
*deref
,
170 nir_ssa_def
**vertex_index
,
171 int (*type_size
)(const struct glsl_type
*, bool),
172 unsigned *component
, bool bts
)
175 nir_deref_path_init(&path
, deref
, NULL
);
177 assert(path
.path
[0]->deref_type
== nir_deref_type_var
);
178 nir_deref_instr
**p
= &path
.path
[1];
180 /* For per-vertex input arrays (i.e. geometry shader inputs), keep the
181 * outermost array index separate. Process the rest normally.
183 if (vertex_index
!= NULL
) {
184 assert((*p
)->deref_type
== nir_deref_type_array
);
185 *vertex_index
= nir_ssa_for_src(b
, (*p
)->arr
.index
, 1);
189 if (path
.path
[0]->var
->data
.compact
) {
190 assert((*p
)->deref_type
== nir_deref_type_array
);
191 assert(glsl_type_is_scalar((*p
)->type
));
193 /* We always lower indirect dereferences for "compact" array vars. */
194 const unsigned index
= nir_src_as_uint((*p
)->arr
.index
);
195 const unsigned total_offset
= *component
+ index
;
196 const unsigned slot_offset
= total_offset
/ 4;
197 *component
= total_offset
% 4;
198 return nir_imm_int(b
, type_size(glsl_vec4_type(), bts
) * slot_offset
);
201 /* Just emit code and let constant-folding go to town */
202 nir_ssa_def
*offset
= nir_imm_int(b
, 0);
205 if ((*p
)->deref_type
== nir_deref_type_array
) {
206 unsigned size
= type_size((*p
)->type
, bts
);
209 nir_amul_imm(b
, nir_ssa_for_src(b
, (*p
)->arr
.index
, 1), size
);
211 offset
= nir_iadd(b
, offset
, mul
);
212 } else if ((*p
)->deref_type
== nir_deref_type_struct
) {
213 /* p starts at path[1], so this is safe */
214 nir_deref_instr
*parent
= *(p
- 1);
216 unsigned field_offset
= 0;
217 for (unsigned i
= 0; i
< (*p
)->strct
.index
; i
++) {
218 field_offset
+= type_size(glsl_get_struct_field(parent
->type
, i
), bts
);
220 offset
= nir_iadd_imm(b
, offset
, field_offset
);
222 unreachable("Unsupported deref type");
226 nir_deref_path_finish(&path
);
232 emit_load(struct lower_io_state
*state
,
233 nir_ssa_def
*vertex_index
, nir_variable
*var
, nir_ssa_def
*offset
,
234 unsigned component
, unsigned num_components
, unsigned bit_size
,
237 nir_builder
*b
= &state
->builder
;
238 const nir_shader
*nir
= b
->shader
;
239 nir_variable_mode mode
= var
->data
.mode
;
240 nir_ssa_def
*barycentric
= NULL
;
244 case nir_var_shader_in
:
245 if (nir
->info
.stage
== MESA_SHADER_FRAGMENT
&&
246 nir
->options
->use_interpolated_input_intrinsics
&&
247 var
->data
.interpolation
!= INTERP_MODE_FLAT
) {
248 if (var
->data
.interpolation
== INTERP_MODE_EXPLICIT
) {
249 assert(vertex_index
!= NULL
);
250 op
= nir_intrinsic_load_input_vertex
;
252 assert(vertex_index
== NULL
);
254 nir_intrinsic_op bary_op
;
255 if (var
->data
.sample
||
256 (state
->options
& nir_lower_io_force_sample_interpolation
))
257 bary_op
= nir_intrinsic_load_barycentric_sample
;
258 else if (var
->data
.centroid
)
259 bary_op
= nir_intrinsic_load_barycentric_centroid
;
261 bary_op
= nir_intrinsic_load_barycentric_pixel
;
263 barycentric
= nir_load_barycentric(&state
->builder
, bary_op
,
264 var
->data
.interpolation
);
265 op
= nir_intrinsic_load_interpolated_input
;
268 op
= vertex_index
? nir_intrinsic_load_per_vertex_input
:
269 nir_intrinsic_load_input
;
272 case nir_var_shader_out
:
273 op
= vertex_index
? nir_intrinsic_load_per_vertex_output
:
274 nir_intrinsic_load_output
;
276 case nir_var_uniform
:
277 op
= nir_intrinsic_load_uniform
;
279 case nir_var_mem_shared
:
280 op
= nir_intrinsic_load_shared
;
283 unreachable("Unknown variable mode");
286 nir_intrinsic_instr
*load
=
287 nir_intrinsic_instr_create(state
->builder
.shader
, op
);
288 load
->num_components
= num_components
;
290 nir_intrinsic_set_base(load
, var
->data
.driver_location
);
291 if (mode
== nir_var_shader_in
|| mode
== nir_var_shader_out
)
292 nir_intrinsic_set_component(load
, component
);
294 if (load
->intrinsic
== nir_intrinsic_load_uniform
)
295 nir_intrinsic_set_range(load
,
296 state
->type_size(var
->type
, var
->data
.bindless
));
298 if (load
->intrinsic
== nir_intrinsic_load_input
||
299 load
->intrinsic
== nir_intrinsic_load_input_vertex
||
300 load
->intrinsic
== nir_intrinsic_load_uniform
)
301 nir_intrinsic_set_type(load
, type
);
304 load
->src
[0] = nir_src_for_ssa(vertex_index
);
305 load
->src
[1] = nir_src_for_ssa(offset
);
306 } else if (barycentric
) {
307 load
->src
[0] = nir_src_for_ssa(barycentric
);
308 load
->src
[1] = nir_src_for_ssa(offset
);
310 load
->src
[0] = nir_src_for_ssa(offset
);
313 nir_ssa_dest_init(&load
->instr
, &load
->dest
,
314 num_components
, bit_size
, NULL
);
315 nir_builder_instr_insert(b
, &load
->instr
);
317 return &load
->dest
.ssa
;
321 lower_load(nir_intrinsic_instr
*intrin
, struct lower_io_state
*state
,
322 nir_ssa_def
*vertex_index
, nir_variable
*var
, nir_ssa_def
*offset
,
323 unsigned component
, const struct glsl_type
*type
)
325 assert(intrin
->dest
.is_ssa
);
326 if (intrin
->dest
.ssa
.bit_size
== 64 &&
327 (state
->options
& nir_lower_io_lower_64bit_to_32
)) {
328 nir_builder
*b
= &state
->builder
;
330 const unsigned slot_size
= state
->type_size(glsl_dvec_type(2), false);
332 nir_ssa_def
*comp64
[4];
333 assert(component
== 0 || component
== 2);
334 unsigned dest_comp
= 0;
335 while (dest_comp
< intrin
->dest
.ssa
.num_components
) {
336 const unsigned num_comps
=
337 MIN2(intrin
->dest
.ssa
.num_components
- dest_comp
,
338 (4 - component
) / 2);
340 nir_ssa_def
*data32
=
341 emit_load(state
, vertex_index
, var
, offset
, component
,
342 num_comps
* 2, 32, nir_type_uint32
);
343 for (unsigned i
= 0; i
< num_comps
; i
++) {
344 comp64
[dest_comp
+ i
] =
345 nir_pack_64_2x32(b
, nir_channels(b
, data32
, 3 << (i
* 2)));
348 /* Only the first store has a component offset */
350 dest_comp
+= num_comps
;
351 offset
= nir_iadd_imm(b
, offset
, slot_size
);
354 return nir_vec(b
, comp64
, intrin
->dest
.ssa
.num_components
);
356 return emit_load(state
, vertex_index
, var
, offset
, component
,
357 intrin
->dest
.ssa
.num_components
,
358 intrin
->dest
.ssa
.bit_size
,
359 nir_get_nir_type_for_glsl_type(type
));
364 emit_store(struct lower_io_state
*state
, nir_ssa_def
*data
,
365 nir_ssa_def
*vertex_index
, nir_variable
*var
, nir_ssa_def
*offset
,
366 unsigned component
, unsigned num_components
,
367 nir_component_mask_t write_mask
, nir_alu_type type
)
369 nir_builder
*b
= &state
->builder
;
370 nir_variable_mode mode
= var
->data
.mode
;
373 if (mode
== nir_var_mem_shared
) {
374 op
= nir_intrinsic_store_shared
;
376 assert(mode
== nir_var_shader_out
);
377 op
= vertex_index
? nir_intrinsic_store_per_vertex_output
:
378 nir_intrinsic_store_output
;
381 nir_intrinsic_instr
*store
=
382 nir_intrinsic_instr_create(state
->builder
.shader
, op
);
383 store
->num_components
= num_components
;
385 store
->src
[0] = nir_src_for_ssa(data
);
387 nir_intrinsic_set_base(store
, var
->data
.driver_location
);
389 if (mode
== nir_var_shader_out
)
390 nir_intrinsic_set_component(store
, component
);
392 if (store
->intrinsic
== nir_intrinsic_store_output
)
393 nir_intrinsic_set_type(store
, type
);
395 nir_intrinsic_set_write_mask(store
, write_mask
);
398 store
->src
[1] = nir_src_for_ssa(vertex_index
);
400 store
->src
[vertex_index
? 2 : 1] = nir_src_for_ssa(offset
);
402 nir_builder_instr_insert(b
, &store
->instr
);
406 lower_store(nir_intrinsic_instr
*intrin
, struct lower_io_state
*state
,
407 nir_ssa_def
*vertex_index
, nir_variable
*var
, nir_ssa_def
*offset
,
408 unsigned component
, const struct glsl_type
*type
)
410 assert(intrin
->src
[1].is_ssa
);
411 if (intrin
->src
[1].ssa
->bit_size
== 64 &&
412 (state
->options
& nir_lower_io_lower_64bit_to_32
)) {
413 nir_builder
*b
= &state
->builder
;
415 const unsigned slot_size
= state
->type_size(glsl_dvec_type(2), false);
417 assert(component
== 0 || component
== 2);
418 unsigned src_comp
= 0;
419 nir_component_mask_t write_mask
= nir_intrinsic_write_mask(intrin
);
420 while (src_comp
< intrin
->num_components
) {
421 const unsigned num_comps
=
422 MIN2(intrin
->num_components
- src_comp
,
423 (4 - component
) / 2);
425 if (write_mask
& BITFIELD_MASK(num_comps
)) {
427 nir_channels(b
, intrin
->src
[1].ssa
,
428 BITFIELD_RANGE(src_comp
, num_comps
));
429 nir_ssa_def
*data32
= nir_bitcast_vector(b
, data
, 32);
431 nir_component_mask_t write_mask32
= 0;
432 for (unsigned i
= 0; i
< num_comps
; i
++) {
433 if (write_mask
& BITFIELD_MASK(num_comps
) & (1 << i
))
434 write_mask32
|= 3 << (i
* 2);
437 emit_store(state
, data32
, vertex_index
, var
, offset
,
438 component
, data32
->num_components
, write_mask32
,
442 /* Only the first store has a component offset */
444 src_comp
+= num_comps
;
445 write_mask
>>= num_comps
;
446 offset
= nir_iadd_imm(b
, offset
, slot_size
);
449 emit_store(state
, intrin
->src
[1].ssa
, vertex_index
, var
, offset
,
450 component
, intrin
->num_components
,
451 nir_intrinsic_write_mask(intrin
),
452 nir_get_nir_type_for_glsl_type(type
));
457 lower_atomic(nir_intrinsic_instr
*intrin
, struct lower_io_state
*state
,
458 nir_variable
*var
, nir_ssa_def
*offset
)
460 nir_builder
*b
= &state
->builder
;
461 assert(var
->data
.mode
== nir_var_mem_shared
);
463 nir_intrinsic_op op
= shared_atomic_for_deref(intrin
->intrinsic
);
465 nir_intrinsic_instr
*atomic
=
466 nir_intrinsic_instr_create(state
->builder
.shader
, op
);
468 nir_intrinsic_set_base(atomic
, var
->data
.driver_location
);
470 atomic
->src
[0] = nir_src_for_ssa(offset
);
471 assert(nir_intrinsic_infos
[intrin
->intrinsic
].num_srcs
==
472 nir_intrinsic_infos
[op
].num_srcs
);
473 for (unsigned i
= 1; i
< nir_intrinsic_infos
[op
].num_srcs
; i
++) {
474 nir_src_copy(&atomic
->src
[i
], &intrin
->src
[i
], atomic
);
477 if (nir_intrinsic_infos
[op
].has_dest
) {
478 assert(intrin
->dest
.is_ssa
);
479 assert(nir_intrinsic_infos
[intrin
->intrinsic
].has_dest
);
480 nir_ssa_dest_init(&atomic
->instr
, &atomic
->dest
,
481 intrin
->dest
.ssa
.num_components
,
482 intrin
->dest
.ssa
.bit_size
, NULL
);
485 nir_builder_instr_insert(b
, &atomic
->instr
);
487 return nir_intrinsic_infos
[op
].has_dest
? &atomic
->dest
.ssa
: NULL
;
491 lower_interpolate_at(nir_intrinsic_instr
*intrin
, struct lower_io_state
*state
,
492 nir_variable
*var
, nir_ssa_def
*offset
, unsigned component
,
493 const struct glsl_type
*type
)
495 nir_builder
*b
= &state
->builder
;
496 assert(var
->data
.mode
== nir_var_shader_in
);
498 /* Ignore interpolateAt() for flat variables - flat is flat. Lower
499 * interpolateAtVertex() for explicit variables.
501 if (var
->data
.interpolation
== INTERP_MODE_FLAT
||
502 var
->data
.interpolation
== INTERP_MODE_EXPLICIT
) {
503 nir_ssa_def
*vertex_index
= NULL
;
505 if (var
->data
.interpolation
== INTERP_MODE_EXPLICIT
) {
506 assert(intrin
->intrinsic
== nir_intrinsic_interp_deref_at_vertex
);
507 vertex_index
= intrin
->src
[1].ssa
;
510 return lower_load(intrin
, state
, vertex_index
, var
, offset
, component
, type
);
513 /* None of the supported APIs allow interpolation on 64-bit things */
514 assert(intrin
->dest
.is_ssa
&& intrin
->dest
.ssa
.bit_size
<= 32);
516 nir_intrinsic_op bary_op
;
517 switch (intrin
->intrinsic
) {
518 case nir_intrinsic_interp_deref_at_centroid
:
519 bary_op
= (state
->options
& nir_lower_io_force_sample_interpolation
) ?
520 nir_intrinsic_load_barycentric_sample
:
521 nir_intrinsic_load_barycentric_centroid
;
523 case nir_intrinsic_interp_deref_at_sample
:
524 bary_op
= nir_intrinsic_load_barycentric_at_sample
;
526 case nir_intrinsic_interp_deref_at_offset
:
527 bary_op
= nir_intrinsic_load_barycentric_at_offset
;
530 unreachable("Bogus interpolateAt() intrinsic.");
533 nir_intrinsic_instr
*bary_setup
=
534 nir_intrinsic_instr_create(state
->builder
.shader
, bary_op
);
536 nir_ssa_dest_init(&bary_setup
->instr
, &bary_setup
->dest
, 2, 32, NULL
);
537 nir_intrinsic_set_interp_mode(bary_setup
, var
->data
.interpolation
);
539 if (intrin
->intrinsic
== nir_intrinsic_interp_deref_at_sample
||
540 intrin
->intrinsic
== nir_intrinsic_interp_deref_at_offset
||
541 intrin
->intrinsic
== nir_intrinsic_interp_deref_at_vertex
)
542 nir_src_copy(&bary_setup
->src
[0], &intrin
->src
[1], bary_setup
);
544 nir_builder_instr_insert(b
, &bary_setup
->instr
);
546 nir_intrinsic_instr
*load
=
547 nir_intrinsic_instr_create(state
->builder
.shader
,
548 nir_intrinsic_load_interpolated_input
);
549 load
->num_components
= intrin
->num_components
;
551 nir_intrinsic_set_base(load
, var
->data
.driver_location
);
552 nir_intrinsic_set_component(load
, component
);
554 load
->src
[0] = nir_src_for_ssa(&bary_setup
->dest
.ssa
);
555 load
->src
[1] = nir_src_for_ssa(offset
);
557 assert(intrin
->dest
.is_ssa
);
558 nir_ssa_dest_init(&load
->instr
, &load
->dest
,
559 intrin
->dest
.ssa
.num_components
,
560 intrin
->dest
.ssa
.bit_size
, NULL
);
561 nir_builder_instr_insert(b
, &load
->instr
);
563 return &load
->dest
.ssa
;
567 nir_lower_io_block(nir_block
*block
,
568 struct lower_io_state
*state
)
570 nir_builder
*b
= &state
->builder
;
571 const nir_shader_compiler_options
*options
= b
->shader
->options
;
572 bool progress
= false;
574 nir_foreach_instr_safe(instr
, block
) {
575 if (instr
->type
!= nir_instr_type_intrinsic
)
578 nir_intrinsic_instr
*intrin
= nir_instr_as_intrinsic(instr
);
580 switch (intrin
->intrinsic
) {
581 case nir_intrinsic_load_deref
:
582 case nir_intrinsic_store_deref
:
583 case nir_intrinsic_deref_atomic_add
:
584 case nir_intrinsic_deref_atomic_imin
:
585 case nir_intrinsic_deref_atomic_umin
:
586 case nir_intrinsic_deref_atomic_imax
:
587 case nir_intrinsic_deref_atomic_umax
:
588 case nir_intrinsic_deref_atomic_and
:
589 case nir_intrinsic_deref_atomic_or
:
590 case nir_intrinsic_deref_atomic_xor
:
591 case nir_intrinsic_deref_atomic_exchange
:
592 case nir_intrinsic_deref_atomic_comp_swap
:
593 case nir_intrinsic_deref_atomic_fadd
:
594 case nir_intrinsic_deref_atomic_fmin
:
595 case nir_intrinsic_deref_atomic_fmax
:
596 case nir_intrinsic_deref_atomic_fcomp_swap
:
597 /* We can lower the io for this nir instrinsic */
599 case nir_intrinsic_interp_deref_at_centroid
:
600 case nir_intrinsic_interp_deref_at_sample
:
601 case nir_intrinsic_interp_deref_at_offset
:
602 case nir_intrinsic_interp_deref_at_vertex
:
603 /* We can optionally lower these to load_interpolated_input */
604 if (options
->use_interpolated_input_intrinsics
)
607 /* We can't lower the io for this nir instrinsic, so skip it */
611 nir_deref_instr
*deref
= nir_src_as_deref(intrin
->src
[0]);
613 nir_variable_mode mode
= deref
->mode
;
615 if ((state
->modes
& mode
) == 0)
618 if (mode
!= nir_var_shader_in
&&
619 mode
!= nir_var_shader_out
&&
620 mode
!= nir_var_mem_shared
&&
621 mode
!= nir_var_uniform
)
624 nir_variable
*var
= nir_deref_instr_get_variable(deref
);
626 b
->cursor
= nir_before_instr(instr
);
628 const bool per_vertex
= nir_is_per_vertex_io(var
, b
->shader
->info
.stage
);
631 nir_ssa_def
*vertex_index
= NULL
;
632 unsigned component_offset
= var
->data
.location_frac
;
633 bool bindless_type_size
= mode
== nir_var_shader_in
||
634 mode
== nir_var_shader_out
||
637 offset
= get_io_offset(b
, deref
, per_vertex
? &vertex_index
: NULL
,
638 state
->type_size
, &component_offset
,
641 nir_ssa_def
*replacement
= NULL
;
643 switch (intrin
->intrinsic
) {
644 case nir_intrinsic_load_deref
:
645 replacement
= lower_load(intrin
, state
, vertex_index
, var
, offset
,
646 component_offset
, deref
->type
);
649 case nir_intrinsic_store_deref
:
650 lower_store(intrin
, state
, vertex_index
, var
, offset
,
651 component_offset
, deref
->type
);
654 case nir_intrinsic_deref_atomic_add
:
655 case nir_intrinsic_deref_atomic_imin
:
656 case nir_intrinsic_deref_atomic_umin
:
657 case nir_intrinsic_deref_atomic_imax
:
658 case nir_intrinsic_deref_atomic_umax
:
659 case nir_intrinsic_deref_atomic_and
:
660 case nir_intrinsic_deref_atomic_or
:
661 case nir_intrinsic_deref_atomic_xor
:
662 case nir_intrinsic_deref_atomic_exchange
:
663 case nir_intrinsic_deref_atomic_comp_swap
:
664 case nir_intrinsic_deref_atomic_fadd
:
665 case nir_intrinsic_deref_atomic_fmin
:
666 case nir_intrinsic_deref_atomic_fmax
:
667 case nir_intrinsic_deref_atomic_fcomp_swap
:
668 assert(vertex_index
== NULL
);
669 replacement
= lower_atomic(intrin
, state
, var
, offset
);
672 case nir_intrinsic_interp_deref_at_centroid
:
673 case nir_intrinsic_interp_deref_at_sample
:
674 case nir_intrinsic_interp_deref_at_offset
:
675 case nir_intrinsic_interp_deref_at_vertex
:
676 assert(vertex_index
== NULL
);
677 replacement
= lower_interpolate_at(intrin
, state
, var
, offset
,
678 component_offset
, deref
->type
);
686 nir_ssa_def_rewrite_uses(&intrin
->dest
.ssa
,
687 nir_src_for_ssa(replacement
));
689 nir_instr_remove(&intrin
->instr
);
697 nir_lower_io_impl(nir_function_impl
*impl
,
698 nir_variable_mode modes
,
699 int (*type_size
)(const struct glsl_type
*, bool),
700 nir_lower_io_options options
)
702 struct lower_io_state state
;
703 bool progress
= false;
705 nir_builder_init(&state
.builder
, impl
);
706 state
.dead_ctx
= ralloc_context(NULL
);
708 state
.type_size
= type_size
;
709 state
.options
= options
;
711 nir_foreach_block(block
, impl
) {
712 progress
|= nir_lower_io_block(block
, &state
);
715 ralloc_free(state
.dead_ctx
);
717 nir_metadata_preserve(impl
, nir_metadata_block_index
|
718 nir_metadata_dominance
);
723 nir_lower_io(nir_shader
*shader
, nir_variable_mode modes
,
724 int (*type_size
)(const struct glsl_type
*, bool),
725 nir_lower_io_options options
)
727 bool progress
= false;
729 nir_foreach_function(function
, shader
) {
730 if (function
->impl
) {
731 progress
|= nir_lower_io_impl(function
->impl
, modes
,
740 type_scalar_size_bytes(const struct glsl_type
*type
)
742 assert(glsl_type_is_vector_or_scalar(type
) ||
743 glsl_type_is_matrix(type
));
744 return glsl_type_is_boolean(type
) ? 4 : glsl_get_bit_size(type
) / 8;
748 build_addr_iadd(nir_builder
*b
, nir_ssa_def
*addr
,
749 nir_address_format addr_format
, nir_ssa_def
*offset
)
751 assert(offset
->num_components
== 1);
752 assert(addr
->bit_size
== offset
->bit_size
);
754 switch (addr_format
) {
755 case nir_address_format_32bit_global
:
756 case nir_address_format_64bit_global
:
757 case nir_address_format_32bit_offset
:
758 assert(addr
->num_components
== 1);
759 return nir_iadd(b
, addr
, offset
);
761 case nir_address_format_64bit_bounded_global
:
762 assert(addr
->num_components
== 4);
763 return nir_vec4(b
, nir_channel(b
, addr
, 0),
764 nir_channel(b
, addr
, 1),
765 nir_channel(b
, addr
, 2),
766 nir_iadd(b
, nir_channel(b
, addr
, 3), offset
));
768 case nir_address_format_32bit_index_offset
:
769 assert(addr
->num_components
== 2);
770 return nir_vec2(b
, nir_channel(b
, addr
, 0),
771 nir_iadd(b
, nir_channel(b
, addr
, 1), offset
));
772 case nir_address_format_logical
:
773 unreachable("Unsupported address format");
775 unreachable("Invalid address format");
779 build_addr_iadd_imm(nir_builder
*b
, nir_ssa_def
*addr
,
780 nir_address_format addr_format
, int64_t offset
)
782 return build_addr_iadd(b
, addr
, addr_format
,
783 nir_imm_intN_t(b
, offset
, addr
->bit_size
));
787 addr_to_index(nir_builder
*b
, nir_ssa_def
*addr
,
788 nir_address_format addr_format
)
790 assert(addr_format
== nir_address_format_32bit_index_offset
);
791 assert(addr
->num_components
== 2);
792 return nir_channel(b
, addr
, 0);
796 addr_to_offset(nir_builder
*b
, nir_ssa_def
*addr
,
797 nir_address_format addr_format
)
799 assert(addr_format
== nir_address_format_32bit_index_offset
);
800 assert(addr
->num_components
== 2);
801 return nir_channel(b
, addr
, 1);
804 /** Returns true if the given address format resolves to a global address */
806 addr_format_is_global(nir_address_format addr_format
)
808 return addr_format
== nir_address_format_32bit_global
||
809 addr_format
== nir_address_format_64bit_global
||
810 addr_format
== nir_address_format_64bit_bounded_global
;
814 addr_to_global(nir_builder
*b
, nir_ssa_def
*addr
,
815 nir_address_format addr_format
)
817 switch (addr_format
) {
818 case nir_address_format_32bit_global
:
819 case nir_address_format_64bit_global
:
820 assert(addr
->num_components
== 1);
823 case nir_address_format_64bit_bounded_global
:
824 assert(addr
->num_components
== 4);
825 return nir_iadd(b
, nir_pack_64_2x32(b
, nir_channels(b
, addr
, 0x3)),
826 nir_u2u64(b
, nir_channel(b
, addr
, 3)));
828 case nir_address_format_32bit_index_offset
:
829 case nir_address_format_32bit_offset
:
830 case nir_address_format_logical
:
831 unreachable("Cannot get a 64-bit address with this address format");
834 unreachable("Invalid address format");
838 addr_format_needs_bounds_check(nir_address_format addr_format
)
840 return addr_format
== nir_address_format_64bit_bounded_global
;
844 addr_is_in_bounds(nir_builder
*b
, nir_ssa_def
*addr
,
845 nir_address_format addr_format
, unsigned size
)
847 assert(addr_format
== nir_address_format_64bit_bounded_global
);
848 assert(addr
->num_components
== 4);
849 return nir_ige(b
, nir_channel(b
, addr
, 2),
850 nir_iadd_imm(b
, nir_channel(b
, addr
, 3), size
));
854 build_explicit_io_load(nir_builder
*b
, nir_intrinsic_instr
*intrin
,
855 nir_ssa_def
*addr
, nir_address_format addr_format
,
856 unsigned num_components
)
858 nir_variable_mode mode
= nir_src_as_deref(intrin
->src
[0])->mode
;
862 case nir_var_mem_ubo
:
863 op
= nir_intrinsic_load_ubo
;
865 case nir_var_mem_ssbo
:
866 if (addr_format_is_global(addr_format
))
867 op
= nir_intrinsic_load_global
;
869 op
= nir_intrinsic_load_ssbo
;
871 case nir_var_mem_global
:
872 assert(addr_format_is_global(addr_format
));
873 op
= nir_intrinsic_load_global
;
875 case nir_var_shader_in
:
876 assert(addr_format_is_global(addr_format
));
877 op
= nir_intrinsic_load_kernel_input
;
879 case nir_var_mem_shared
:
880 assert(addr_format
== nir_address_format_32bit_offset
);
881 op
= nir_intrinsic_load_shared
;
884 unreachable("Unsupported explicit IO variable mode");
887 nir_intrinsic_instr
*load
= nir_intrinsic_instr_create(b
->shader
, op
);
889 if (addr_format_is_global(addr_format
)) {
890 load
->src
[0] = nir_src_for_ssa(addr_to_global(b
, addr
, addr_format
));
891 } else if (addr_format
== nir_address_format_32bit_offset
) {
892 assert(addr
->num_components
== 1);
893 load
->src
[0] = nir_src_for_ssa(addr
);
895 load
->src
[0] = nir_src_for_ssa(addr_to_index(b
, addr
, addr_format
));
896 load
->src
[1] = nir_src_for_ssa(addr_to_offset(b
, addr
, addr_format
));
899 if (mode
!= nir_var_mem_ubo
&& mode
!= nir_var_shader_in
&& mode
!= nir_var_mem_shared
)
900 nir_intrinsic_set_access(load
, nir_intrinsic_access(intrin
));
902 unsigned bit_size
= intrin
->dest
.ssa
.bit_size
;
904 /* TODO: Make the native bool bit_size an option. */
908 /* TODO: We should try and provide a better alignment. For OpenCL, we need
909 * to plumb the alignment through from SPIR-V when we have one.
911 nir_intrinsic_set_align(load
, bit_size
/ 8, 0);
913 assert(intrin
->dest
.is_ssa
);
914 load
->num_components
= num_components
;
915 nir_ssa_dest_init(&load
->instr
, &load
->dest
, num_components
,
916 bit_size
, intrin
->dest
.ssa
.name
);
918 assert(bit_size
% 8 == 0);
921 if (addr_format_needs_bounds_check(addr_format
)) {
922 /* The Vulkan spec for robustBufferAccess gives us quite a few options
923 * as to what we can do with an OOB read. Unfortunately, returning
924 * undefined values isn't one of them so we return an actual zero.
926 nir_ssa_def
*zero
= nir_imm_zero(b
, load
->num_components
, bit_size
);
928 const unsigned load_size
= (bit_size
/ 8) * load
->num_components
;
929 nir_push_if(b
, addr_is_in_bounds(b
, addr
, addr_format
, load_size
));
931 nir_builder_instr_insert(b
, &load
->instr
);
935 result
= nir_if_phi(b
, &load
->dest
.ssa
, zero
);
937 nir_builder_instr_insert(b
, &load
->instr
);
938 result
= &load
->dest
.ssa
;
941 if (intrin
->dest
.ssa
.bit_size
== 1)
942 result
= nir_i2b(b
, result
);
948 build_explicit_io_store(nir_builder
*b
, nir_intrinsic_instr
*intrin
,
949 nir_ssa_def
*addr
, nir_address_format addr_format
,
950 nir_ssa_def
*value
, nir_component_mask_t write_mask
)
952 nir_variable_mode mode
= nir_src_as_deref(intrin
->src
[0])->mode
;
956 case nir_var_mem_ssbo
:
957 if (addr_format_is_global(addr_format
))
958 op
= nir_intrinsic_store_global
;
960 op
= nir_intrinsic_store_ssbo
;
962 case nir_var_mem_global
:
963 assert(addr_format_is_global(addr_format
));
964 op
= nir_intrinsic_store_global
;
966 case nir_var_mem_shared
:
967 assert(addr_format
== nir_address_format_32bit_offset
);
968 op
= nir_intrinsic_store_shared
;
971 unreachable("Unsupported explicit IO variable mode");
974 nir_intrinsic_instr
*store
= nir_intrinsic_instr_create(b
->shader
, op
);
976 if (value
->bit_size
== 1) {
977 /* TODO: Make the native bool bit_size an option. */
978 value
= nir_b2i(b
, value
, 32);
981 store
->src
[0] = nir_src_for_ssa(value
);
982 if (addr_format_is_global(addr_format
)) {
983 store
->src
[1] = nir_src_for_ssa(addr_to_global(b
, addr
, addr_format
));
984 } else if (addr_format
== nir_address_format_32bit_offset
) {
985 assert(addr
->num_components
== 1);
986 store
->src
[1] = nir_src_for_ssa(addr
);
988 store
->src
[1] = nir_src_for_ssa(addr_to_index(b
, addr
, addr_format
));
989 store
->src
[2] = nir_src_for_ssa(addr_to_offset(b
, addr
, addr_format
));
992 nir_intrinsic_set_write_mask(store
, write_mask
);
994 if (mode
!= nir_var_mem_shared
)
995 nir_intrinsic_set_access(store
, nir_intrinsic_access(intrin
));
997 /* TODO: We should try and provide a better alignment. For OpenCL, we need
998 * to plumb the alignment through from SPIR-V when we have one.
1000 nir_intrinsic_set_align(store
, value
->bit_size
/ 8, 0);
1002 assert(value
->num_components
== 1 ||
1003 value
->num_components
== intrin
->num_components
);
1004 store
->num_components
= value
->num_components
;
1006 assert(value
->bit_size
% 8 == 0);
1008 if (addr_format_needs_bounds_check(addr_format
)) {
1009 const unsigned store_size
= (value
->bit_size
/ 8) * store
->num_components
;
1010 nir_push_if(b
, addr_is_in_bounds(b
, addr
, addr_format
, store_size
));
1012 nir_builder_instr_insert(b
, &store
->instr
);
1014 nir_pop_if(b
, NULL
);
1016 nir_builder_instr_insert(b
, &store
->instr
);
1020 static nir_ssa_def
*
1021 build_explicit_io_atomic(nir_builder
*b
, nir_intrinsic_instr
*intrin
,
1022 nir_ssa_def
*addr
, nir_address_format addr_format
)
1024 nir_variable_mode mode
= nir_src_as_deref(intrin
->src
[0])->mode
;
1025 const unsigned num_data_srcs
=
1026 nir_intrinsic_infos
[intrin
->intrinsic
].num_srcs
- 1;
1028 nir_intrinsic_op op
;
1030 case nir_var_mem_ssbo
:
1031 if (addr_format_is_global(addr_format
))
1032 op
= global_atomic_for_deref(intrin
->intrinsic
);
1034 op
= ssbo_atomic_for_deref(intrin
->intrinsic
);
1036 case nir_var_mem_global
:
1037 assert(addr_format_is_global(addr_format
));
1038 op
= global_atomic_for_deref(intrin
->intrinsic
);
1040 case nir_var_mem_shared
:
1041 assert(addr_format
== nir_address_format_32bit_offset
);
1042 op
= shared_atomic_for_deref(intrin
->intrinsic
);
1045 unreachable("Unsupported explicit IO variable mode");
1048 nir_intrinsic_instr
*atomic
= nir_intrinsic_instr_create(b
->shader
, op
);
1051 if (addr_format_is_global(addr_format
)) {
1052 atomic
->src
[src
++] = nir_src_for_ssa(addr_to_global(b
, addr
, addr_format
));
1053 } else if (addr_format
== nir_address_format_32bit_offset
) {
1054 assert(addr
->num_components
== 1);
1055 atomic
->src
[src
++] = nir_src_for_ssa(addr
);
1057 atomic
->src
[src
++] = nir_src_for_ssa(addr_to_index(b
, addr
, addr_format
));
1058 atomic
->src
[src
++] = nir_src_for_ssa(addr_to_offset(b
, addr
, addr_format
));
1060 for (unsigned i
= 0; i
< num_data_srcs
; i
++) {
1061 atomic
->src
[src
++] = nir_src_for_ssa(intrin
->src
[1 + i
].ssa
);
1064 /* Global atomics don't have access flags because they assume that the
1065 * address may be non-uniform.
1067 if (!addr_format_is_global(addr_format
) && mode
!= nir_var_mem_shared
)
1068 nir_intrinsic_set_access(atomic
, nir_intrinsic_access(intrin
));
1070 assert(intrin
->dest
.ssa
.num_components
== 1);
1071 nir_ssa_dest_init(&atomic
->instr
, &atomic
->dest
,
1072 1, intrin
->dest
.ssa
.bit_size
, intrin
->dest
.ssa
.name
);
1074 assert(atomic
->dest
.ssa
.bit_size
% 8 == 0);
1076 if (addr_format_needs_bounds_check(addr_format
)) {
1077 const unsigned atomic_size
= atomic
->dest
.ssa
.bit_size
/ 8;
1078 nir_push_if(b
, addr_is_in_bounds(b
, addr
, addr_format
, atomic_size
));
1080 nir_builder_instr_insert(b
, &atomic
->instr
);
1082 nir_pop_if(b
, NULL
);
1083 return nir_if_phi(b
, &atomic
->dest
.ssa
,
1084 nir_ssa_undef(b
, 1, atomic
->dest
.ssa
.bit_size
));
1086 nir_builder_instr_insert(b
, &atomic
->instr
);
1087 return &atomic
->dest
.ssa
;
1092 nir_explicit_io_address_from_deref(nir_builder
*b
, nir_deref_instr
*deref
,
1093 nir_ssa_def
*base_addr
,
1094 nir_address_format addr_format
)
1096 assert(deref
->dest
.is_ssa
);
1097 switch (deref
->deref_type
) {
1098 case nir_deref_type_var
:
1099 assert(deref
->mode
& (nir_var_shader_in
| nir_var_mem_shared
));
1100 return nir_imm_intN_t(b
, deref
->var
->data
.driver_location
,
1101 deref
->dest
.ssa
.bit_size
);
1103 case nir_deref_type_array
: {
1104 nir_deref_instr
*parent
= nir_deref_instr_parent(deref
);
1106 unsigned stride
= glsl_get_explicit_stride(parent
->type
);
1107 if ((glsl_type_is_matrix(parent
->type
) &&
1108 glsl_matrix_type_is_row_major(parent
->type
)) ||
1109 (glsl_type_is_vector(parent
->type
) && stride
== 0))
1110 stride
= type_scalar_size_bytes(parent
->type
);
1114 nir_ssa_def
*index
= nir_ssa_for_src(b
, deref
->arr
.index
, 1);
1115 index
= nir_i2i(b
, index
, base_addr
->bit_size
);
1116 return build_addr_iadd(b
, base_addr
, addr_format
,
1117 nir_amul_imm(b
, index
, stride
));
1120 case nir_deref_type_ptr_as_array
: {
1121 nir_ssa_def
*index
= nir_ssa_for_src(b
, deref
->arr
.index
, 1);
1122 index
= nir_i2i(b
, index
, base_addr
->bit_size
);
1123 unsigned stride
= nir_deref_instr_ptr_as_array_stride(deref
);
1124 return build_addr_iadd(b
, base_addr
, addr_format
,
1125 nir_amul_imm(b
, index
, stride
));
1128 case nir_deref_type_array_wildcard
:
1129 unreachable("Wildcards should be lowered by now");
1132 case nir_deref_type_struct
: {
1133 nir_deref_instr
*parent
= nir_deref_instr_parent(deref
);
1134 int offset
= glsl_get_struct_field_offset(parent
->type
,
1135 deref
->strct
.index
);
1136 assert(offset
>= 0);
1137 return build_addr_iadd_imm(b
, base_addr
, addr_format
, offset
);
1140 case nir_deref_type_cast
:
1141 /* Nothing to do here */
1145 unreachable("Invalid NIR deref type");
1149 nir_lower_explicit_io_instr(nir_builder
*b
,
1150 nir_intrinsic_instr
*intrin
,
1152 nir_address_format addr_format
)
1154 b
->cursor
= nir_after_instr(&intrin
->instr
);
1156 nir_deref_instr
*deref
= nir_src_as_deref(intrin
->src
[0]);
1157 unsigned vec_stride
= glsl_get_explicit_stride(deref
->type
);
1158 unsigned scalar_size
= type_scalar_size_bytes(deref
->type
);
1159 assert(vec_stride
== 0 || glsl_type_is_vector(deref
->type
));
1160 assert(vec_stride
== 0 || vec_stride
>= scalar_size
);
1162 if (intrin
->intrinsic
== nir_intrinsic_load_deref
) {
1164 if (vec_stride
> scalar_size
) {
1165 nir_ssa_def
*comps
[4] = { NULL
, };
1166 for (unsigned i
= 0; i
< intrin
->num_components
; i
++) {
1167 nir_ssa_def
*comp_addr
= build_addr_iadd_imm(b
, addr
, addr_format
,
1169 comps
[i
] = build_explicit_io_load(b
, intrin
, comp_addr
,
1172 value
= nir_vec(b
, comps
, intrin
->num_components
);
1174 value
= build_explicit_io_load(b
, intrin
, addr
, addr_format
,
1175 intrin
->num_components
);
1177 nir_ssa_def_rewrite_uses(&intrin
->dest
.ssa
, nir_src_for_ssa(value
));
1178 } else if (intrin
->intrinsic
== nir_intrinsic_store_deref
) {
1179 assert(intrin
->src
[1].is_ssa
);
1180 nir_ssa_def
*value
= intrin
->src
[1].ssa
;
1181 nir_component_mask_t write_mask
= nir_intrinsic_write_mask(intrin
);
1182 if (vec_stride
> scalar_size
) {
1183 for (unsigned i
= 0; i
< intrin
->num_components
; i
++) {
1184 if (!(write_mask
& (1 << i
)))
1187 nir_ssa_def
*comp_addr
= build_addr_iadd_imm(b
, addr
, addr_format
,
1189 build_explicit_io_store(b
, intrin
, comp_addr
, addr_format
,
1190 nir_channel(b
, value
, i
), 1);
1193 build_explicit_io_store(b
, intrin
, addr
, addr_format
,
1197 nir_ssa_def
*value
=
1198 build_explicit_io_atomic(b
, intrin
, addr
, addr_format
);
1199 nir_ssa_def_rewrite_uses(&intrin
->dest
.ssa
, nir_src_for_ssa(value
));
1202 nir_instr_remove(&intrin
->instr
);
1206 lower_explicit_io_deref(nir_builder
*b
, nir_deref_instr
*deref
,
1207 nir_address_format addr_format
)
1209 /* Just delete the deref if it's not used. We can't use
1210 * nir_deref_instr_remove_if_unused here because it may remove more than
1211 * one deref which could break our list walking since we walk the list
1214 assert(list_is_empty(&deref
->dest
.ssa
.if_uses
));
1215 if (list_is_empty(&deref
->dest
.ssa
.uses
)) {
1216 nir_instr_remove(&deref
->instr
);
1220 b
->cursor
= nir_after_instr(&deref
->instr
);
1222 nir_ssa_def
*base_addr
= NULL
;
1223 if (deref
->deref_type
!= nir_deref_type_var
) {
1224 assert(deref
->parent
.is_ssa
);
1225 base_addr
= deref
->parent
.ssa
;
1228 nir_ssa_def
*addr
= nir_explicit_io_address_from_deref(b
, deref
, base_addr
,
1231 nir_instr_remove(&deref
->instr
);
1232 nir_ssa_def_rewrite_uses(&deref
->dest
.ssa
, nir_src_for_ssa(addr
));
1236 lower_explicit_io_access(nir_builder
*b
, nir_intrinsic_instr
*intrin
,
1237 nir_address_format addr_format
)
1239 assert(intrin
->src
[0].is_ssa
);
1240 nir_lower_explicit_io_instr(b
, intrin
, intrin
->src
[0].ssa
, addr_format
);
1244 lower_explicit_io_array_length(nir_builder
*b
, nir_intrinsic_instr
*intrin
,
1245 nir_address_format addr_format
)
1247 b
->cursor
= nir_after_instr(&intrin
->instr
);
1249 nir_deref_instr
*deref
= nir_src_as_deref(intrin
->src
[0]);
1251 assert(glsl_type_is_array(deref
->type
));
1252 assert(glsl_get_length(deref
->type
) == 0);
1253 unsigned stride
= glsl_get_explicit_stride(deref
->type
);
1256 assert(addr_format
== nir_address_format_32bit_index_offset
);
1257 nir_ssa_def
*addr
= &deref
->dest
.ssa
;
1258 nir_ssa_def
*index
= addr_to_index(b
, addr
, addr_format
);
1259 nir_ssa_def
*offset
= addr_to_offset(b
, addr
, addr_format
);
1261 nir_intrinsic_instr
*bsize
=
1262 nir_intrinsic_instr_create(b
->shader
, nir_intrinsic_get_buffer_size
);
1263 bsize
->src
[0] = nir_src_for_ssa(index
);
1264 nir_ssa_dest_init(&bsize
->instr
, &bsize
->dest
, 1, 32, NULL
);
1265 nir_builder_instr_insert(b
, &bsize
->instr
);
1267 nir_ssa_def
*arr_size
=
1268 nir_idiv(b
, nir_isub(b
, &bsize
->dest
.ssa
, offset
),
1269 nir_imm_int(b
, stride
));
1271 nir_ssa_def_rewrite_uses(&intrin
->dest
.ssa
, nir_src_for_ssa(arr_size
));
1272 nir_instr_remove(&intrin
->instr
);
1276 nir_lower_explicit_io_impl(nir_function_impl
*impl
, nir_variable_mode modes
,
1277 nir_address_format addr_format
)
1279 bool progress
= false;
1282 nir_builder_init(&b
, impl
);
1284 /* Walk in reverse order so that we can see the full deref chain when we
1285 * lower the access operations. We lower them assuming that the derefs
1286 * will be turned into address calculations later.
1288 nir_foreach_block_reverse(block
, impl
) {
1289 nir_foreach_instr_reverse_safe(instr
, block
) {
1290 switch (instr
->type
) {
1291 case nir_instr_type_deref
: {
1292 nir_deref_instr
*deref
= nir_instr_as_deref(instr
);
1293 if (deref
->mode
& modes
) {
1294 lower_explicit_io_deref(&b
, deref
, addr_format
);
1300 case nir_instr_type_intrinsic
: {
1301 nir_intrinsic_instr
*intrin
= nir_instr_as_intrinsic(instr
);
1302 switch (intrin
->intrinsic
) {
1303 case nir_intrinsic_load_deref
:
1304 case nir_intrinsic_store_deref
:
1305 case nir_intrinsic_deref_atomic_add
:
1306 case nir_intrinsic_deref_atomic_imin
:
1307 case nir_intrinsic_deref_atomic_umin
:
1308 case nir_intrinsic_deref_atomic_imax
:
1309 case nir_intrinsic_deref_atomic_umax
:
1310 case nir_intrinsic_deref_atomic_and
:
1311 case nir_intrinsic_deref_atomic_or
:
1312 case nir_intrinsic_deref_atomic_xor
:
1313 case nir_intrinsic_deref_atomic_exchange
:
1314 case nir_intrinsic_deref_atomic_comp_swap
:
1315 case nir_intrinsic_deref_atomic_fadd
:
1316 case nir_intrinsic_deref_atomic_fmin
:
1317 case nir_intrinsic_deref_atomic_fmax
:
1318 case nir_intrinsic_deref_atomic_fcomp_swap
: {
1319 nir_deref_instr
*deref
= nir_src_as_deref(intrin
->src
[0]);
1320 if (deref
->mode
& modes
) {
1321 lower_explicit_io_access(&b
, intrin
, addr_format
);
1327 case nir_intrinsic_deref_buffer_array_length
: {
1328 nir_deref_instr
*deref
= nir_src_as_deref(intrin
->src
[0]);
1329 if (deref
->mode
& modes
) {
1330 lower_explicit_io_array_length(&b
, intrin
, addr_format
);
1350 nir_metadata_preserve(impl
, nir_metadata_block_index
|
1351 nir_metadata_dominance
);
1358 nir_lower_explicit_io(nir_shader
*shader
, nir_variable_mode modes
,
1359 nir_address_format addr_format
)
1361 bool progress
= false;
1363 nir_foreach_function(function
, shader
) {
1364 if (function
->impl
&&
1365 nir_lower_explicit_io_impl(function
->impl
, modes
, addr_format
))
1373 nir_lower_vars_to_explicit_types_impl(nir_function_impl
*impl
,
1374 nir_variable_mode modes
,
1375 glsl_type_size_align_func type_info
)
1377 bool progress
= false;
1379 nir_foreach_block(block
, impl
) {
1380 nir_foreach_instr(instr
, block
) {
1381 if (instr
->type
!= nir_instr_type_deref
)
1384 nir_deref_instr
*deref
= nir_instr_as_deref(instr
);
1385 if (!(deref
->mode
& modes
))
1388 unsigned size
, alignment
;
1389 const struct glsl_type
*new_type
=
1390 glsl_get_explicit_type_for_size_align(deref
->type
, type_info
, &size
, &alignment
);
1391 if (new_type
!= deref
->type
) {
1393 deref
->type
= new_type
;
1395 if (deref
->deref_type
== nir_deref_type_cast
) {
1396 /* See also glsl_type::get_explicit_type_for_size_align() */
1397 unsigned new_stride
= align(size
, alignment
);
1398 if (new_stride
!= deref
->cast
.ptr_stride
) {
1399 deref
->cast
.ptr_stride
= new_stride
;
1407 nir_metadata_preserve(impl
, nir_metadata_block_index
|
1408 nir_metadata_dominance
|
1409 nir_metadata_live_ssa_defs
|
1410 nir_metadata_loop_analysis
);
1417 lower_vars_to_explicit(nir_shader
*shader
,
1418 struct exec_list
*vars
, nir_variable_mode mode
,
1419 glsl_type_size_align_func type_info
)
1421 bool progress
= false;
1422 unsigned offset
= 0;
1423 nir_foreach_variable(var
, vars
) {
1424 unsigned size
, align
;
1425 const struct glsl_type
*explicit_type
=
1426 glsl_get_explicit_type_for_size_align(var
->type
, type_info
, &size
, &align
);
1428 if (explicit_type
!= var
->type
) {
1430 var
->type
= explicit_type
;
1433 var
->data
.driver_location
= ALIGN_POT(offset
, align
);
1434 offset
= var
->data
.driver_location
+ size
;
1437 if (mode
== nir_var_mem_shared
) {
1438 shader
->info
.cs
.shared_size
= offset
;
1439 shader
->num_shared
= offset
;
1446 nir_lower_vars_to_explicit_types(nir_shader
*shader
,
1447 nir_variable_mode modes
,
1448 glsl_type_size_align_func type_info
)
1450 /* TODO: Situations which need to be handled to support more modes:
1451 * - row-major matrices
1452 * - compact shader inputs/outputs
1455 ASSERTED nir_variable_mode supported
= nir_var_mem_shared
|
1456 nir_var_shader_temp
| nir_var_function_temp
;
1457 assert(!(modes
& ~supported
) && "unsupported");
1459 bool progress
= false;
1461 if (modes
& nir_var_mem_shared
)
1462 progress
|= lower_vars_to_explicit(shader
, &shader
->shared
, nir_var_mem_shared
, type_info
);
1463 if (modes
& nir_var_shader_temp
)
1464 progress
|= lower_vars_to_explicit(shader
, &shader
->globals
, nir_var_shader_temp
, type_info
);
1466 nir_foreach_function(function
, shader
) {
1467 if (function
->impl
) {
1468 if (modes
& nir_var_function_temp
)
1469 progress
|= lower_vars_to_explicit(shader
, &function
->impl
->locals
, nir_var_function_temp
, type_info
);
1471 progress
|= nir_lower_vars_to_explicit_types_impl(function
->impl
, modes
, type_info
);
1479 * Return the offset source for a load/store intrinsic.
1482 nir_get_io_offset_src(nir_intrinsic_instr
*instr
)
1484 switch (instr
->intrinsic
) {
1485 case nir_intrinsic_load_input
:
1486 case nir_intrinsic_load_output
:
1487 case nir_intrinsic_load_shared
:
1488 case nir_intrinsic_load_uniform
:
1489 case nir_intrinsic_load_global
:
1490 case nir_intrinsic_load_scratch
:
1491 case nir_intrinsic_load_fs_input_interp_deltas
:
1492 return &instr
->src
[0];
1493 case nir_intrinsic_load_ubo
:
1494 case nir_intrinsic_load_ssbo
:
1495 case nir_intrinsic_load_per_vertex_input
:
1496 case nir_intrinsic_load_per_vertex_output
:
1497 case nir_intrinsic_load_interpolated_input
:
1498 case nir_intrinsic_store_output
:
1499 case nir_intrinsic_store_shared
:
1500 case nir_intrinsic_store_global
:
1501 case nir_intrinsic_store_scratch
:
1502 return &instr
->src
[1];
1503 case nir_intrinsic_store_ssbo
:
1504 case nir_intrinsic_store_per_vertex_output
:
1505 return &instr
->src
[2];
1512 * Return the vertex index source for a load/store per_vertex intrinsic.
1515 nir_get_io_vertex_index_src(nir_intrinsic_instr
*instr
)
1517 switch (instr
->intrinsic
) {
1518 case nir_intrinsic_load_per_vertex_input
:
1519 case nir_intrinsic_load_per_vertex_output
:
1520 return &instr
->src
[0];
1521 case nir_intrinsic_store_per_vertex_output
:
1522 return &instr
->src
[1];
1529 * Return the numeric constant that identify a NULL pointer for each address
1532 const nir_const_value
*
1533 nir_address_format_null_value(nir_address_format addr_format
)
1535 const static nir_const_value null_values
[][NIR_MAX_VEC_COMPONENTS
] = {
1536 [nir_address_format_32bit_global
] = {{0}},
1537 [nir_address_format_64bit_global
] = {{0}},
1538 [nir_address_format_64bit_bounded_global
] = {{0}},
1539 [nir_address_format_32bit_index_offset
] = {{.u32
= ~0}, {.u32
= ~0}},
1540 [nir_address_format_32bit_offset
] = {{.u32
= ~0}},
1541 [nir_address_format_logical
] = {{.u32
= ~0}},
1544 assert(addr_format
< ARRAY_SIZE(null_values
));
1545 return null_values
[addr_format
];
1549 nir_build_addr_ieq(nir_builder
*b
, nir_ssa_def
*addr0
, nir_ssa_def
*addr1
,
1550 nir_address_format addr_format
)
1552 switch (addr_format
) {
1553 case nir_address_format_32bit_global
:
1554 case nir_address_format_64bit_global
:
1555 case nir_address_format_64bit_bounded_global
:
1556 case nir_address_format_32bit_index_offset
:
1557 case nir_address_format_32bit_offset
:
1558 return nir_ball_iequal(b
, addr0
, addr1
);
1560 case nir_address_format_logical
:
1561 unreachable("Unsupported address format");
1564 unreachable("Invalid address format");
1568 nir_build_addr_isub(nir_builder
*b
, nir_ssa_def
*addr0
, nir_ssa_def
*addr1
,
1569 nir_address_format addr_format
)
1571 switch (addr_format
) {
1572 case nir_address_format_32bit_global
:
1573 case nir_address_format_64bit_global
:
1574 case nir_address_format_32bit_offset
:
1575 assert(addr0
->num_components
== 1);
1576 assert(addr1
->num_components
== 1);
1577 return nir_isub(b
, addr0
, addr1
);
1579 case nir_address_format_64bit_bounded_global
:
1580 return nir_isub(b
, addr_to_global(b
, addr0
, addr_format
),
1581 addr_to_global(b
, addr1
, addr_format
));
1583 case nir_address_format_32bit_index_offset
:
1584 assert(addr0
->num_components
== 2);
1585 assert(addr1
->num_components
== 2);
1586 /* Assume the same buffer index. */
1587 return nir_isub(b
, nir_channel(b
, addr0
, 1), nir_channel(b
, addr1
, 1));
1589 case nir_address_format_logical
:
1590 unreachable("Unsupported address format");
1593 unreachable("Invalid address format");
1597 is_input(nir_intrinsic_instr
*intrin
)
1599 return intrin
->intrinsic
== nir_intrinsic_load_input
||
1600 intrin
->intrinsic
== nir_intrinsic_load_per_vertex_input
||
1601 intrin
->intrinsic
== nir_intrinsic_load_interpolated_input
||
1602 intrin
->intrinsic
== nir_intrinsic_load_fs_input_interp_deltas
;
1606 is_output(nir_intrinsic_instr
*intrin
)
1608 return intrin
->intrinsic
== nir_intrinsic_load_output
||
1609 intrin
->intrinsic
== nir_intrinsic_load_per_vertex_output
||
1610 intrin
->intrinsic
== nir_intrinsic_store_output
||
1611 intrin
->intrinsic
== nir_intrinsic_store_per_vertex_output
;
1616 * This pass adds constant offsets to instr->const_index[0] for input/output
1617 * intrinsics, and resets the offset source to 0. Non-constant offsets remain
1618 * unchanged - since we don't know what part of a compound variable is
1619 * accessed, we allocate storage for the entire thing. For drivers that use
1620 * nir_lower_io_to_temporaries() before nir_lower_io(), this guarantees that
1621 * the offset source will be 0, so that they don't have to add it in manually.
1625 add_const_offset_to_base_block(nir_block
*block
, nir_builder
*b
,
1626 nir_variable_mode mode
)
1628 bool progress
= false;
1629 nir_foreach_instr_safe(instr
, block
) {
1630 if (instr
->type
!= nir_instr_type_intrinsic
)
1633 nir_intrinsic_instr
*intrin
= nir_instr_as_intrinsic(instr
);
1635 if ((mode
== nir_var_shader_in
&& is_input(intrin
)) ||
1636 (mode
== nir_var_shader_out
&& is_output(intrin
))) {
1637 nir_src
*offset
= nir_get_io_offset_src(intrin
);
1639 if (nir_src_is_const(*offset
)) {
1640 intrin
->const_index
[0] += nir_src_as_uint(*offset
);
1641 b
->cursor
= nir_before_instr(&intrin
->instr
);
1642 nir_instr_rewrite_src(&intrin
->instr
, offset
,
1643 nir_src_for_ssa(nir_imm_int(b
, 0)));
1653 nir_io_add_const_offset_to_base(nir_shader
*nir
, nir_variable_mode mode
)
1655 bool progress
= false;
1657 nir_foreach_function(f
, nir
) {
1660 nir_builder_init(&b
, f
->impl
);
1661 nir_foreach_block(block
, f
->impl
) {
1662 progress
|= add_const_offset_to_base_block(block
, &b
, mode
);