1 /*
2 * Copyright © 2016 Broadcom
3 *
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:
10 *
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
13 * Software.
14 *
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
21 * IN THE SOFTWARE.
22 */
23
24 #include "nir.h"
25 #include "nir_builder.h"
26 #include "nir_deref.h"
27
28 /** @file nir_lower_io_to_scalar.c
29 *
30 * Replaces nir_load_input/nir_store_output operations with num_components !=
31 * 1 with individual per-channel operations.
32 */
33
34 static void
set_io_semantics(nir_intrinsic_instr * scalar_intr,nir_intrinsic_instr * vec_intr,unsigned component)35 set_io_semantics(nir_intrinsic_instr *scalar_intr,
36 nir_intrinsic_instr *vec_intr, unsigned component)
37 {
38 nir_io_semantics sem = nir_intrinsic_io_semantics(vec_intr);
39 sem.gs_streams = (sem.gs_streams >> (component * 2)) & 0x3;
40 nir_intrinsic_set_io_semantics(scalar_intr, sem);
41 }
42
43 static void
lower_load_input_to_scalar(nir_builder * b,nir_intrinsic_instr * intr)44 lower_load_input_to_scalar(nir_builder *b, nir_intrinsic_instr *intr)
45 {
46 b->cursor = nir_before_instr(&intr->instr);
47
48 nir_def *loads[NIR_MAX_VEC_COMPONENTS];
49
50 for (unsigned i = 0; i < intr->num_components; i++) {
51 bool is_64bit = (nir_intrinsic_instr_dest_type(intr) & NIR_ALU_TYPE_SIZE_MASK) == 64;
52 unsigned newi = is_64bit ? i * 2 : i;
53 unsigned newc = nir_intrinsic_component(intr);
54 nir_intrinsic_instr *chan_intr =
55 nir_intrinsic_instr_create(b->shader, intr->intrinsic);
56 nir_def_init(&chan_intr->instr, &chan_intr->def, 1,
57 intr->def.bit_size);
58 chan_intr->num_components = 1;
59
60 if (intr->name)
61 chan_intr->name = intr->name;
62 nir_intrinsic_set_base(chan_intr, nir_intrinsic_base(intr));
63 nir_intrinsic_set_component(chan_intr, (newc + newi) % 4);
64 nir_intrinsic_set_dest_type(chan_intr, nir_intrinsic_dest_type(intr));
65 set_io_semantics(chan_intr, intr, i);
66 /* offset and vertex (if needed) */
67 for (unsigned j = 0; j < nir_intrinsic_infos[intr->intrinsic].num_srcs; ++j)
68 chan_intr->src[j] = nir_src_for_ssa(intr->src[j].ssa);
69 if (newc + newi > 3) {
70 nir_src *src = nir_get_io_offset_src(chan_intr);
71 nir_def *offset = nir_iadd_imm(b, src->ssa, (newc + newi) / 4);
72 *src = nir_src_for_ssa(offset);
73 }
74
75 nir_builder_instr_insert(b, &chan_intr->instr);
76
77 loads[i] = &chan_intr->def;
78 }
79
80 nir_def_replace(&intr->def, nir_vec(b, loads, intr->num_components));
81 }
82
83 static void
lower_load_to_scalar(nir_builder * b,nir_intrinsic_instr * intr)84 lower_load_to_scalar(nir_builder *b, nir_intrinsic_instr *intr)
85 {
86 b->cursor = nir_before_instr(&intr->instr);
87
88 nir_def *loads[NIR_MAX_VEC_COMPONENTS];
89 nir_def *base_offset = nir_get_io_offset_src(intr)->ssa;
90
91 for (unsigned i = 0; i < intr->num_components; i++) {
92 nir_intrinsic_instr *chan_intr =
93 nir_intrinsic_instr_create(b->shader, intr->intrinsic);
94 nir_def_init(&chan_intr->instr, &chan_intr->def, 1,
95 intr->def.bit_size);
96 chan_intr->num_components = 1;
97
98 if (intr->name)
99 chan_intr->name = intr->name;
100 nir_intrinsic_set_align_offset(chan_intr,
101 (nir_intrinsic_align_offset(intr) +
102 i * (intr->def.bit_size / 8)) %
103 nir_intrinsic_align_mul(intr));
104 nir_intrinsic_set_align_mul(chan_intr, nir_intrinsic_align_mul(intr));
105 if (nir_intrinsic_has_access(intr))
106 nir_intrinsic_set_access(chan_intr, nir_intrinsic_access(intr));
107 if (nir_intrinsic_has_range(intr))
108 nir_intrinsic_set_range(chan_intr, nir_intrinsic_range(intr));
109 if (nir_intrinsic_has_range_base(intr))
110 nir_intrinsic_set_range_base(chan_intr, nir_intrinsic_range_base(intr));
111 if (nir_intrinsic_has_base(intr))
112 nir_intrinsic_set_base(chan_intr, nir_intrinsic_base(intr));
113 for (unsigned j = 0; j < nir_intrinsic_infos[intr->intrinsic].num_srcs - 1; j++)
114 chan_intr->src[j] = nir_src_for_ssa(intr->src[j].ssa);
115
116 /* increment offset per component */
117 nir_def *offset = nir_iadd_imm(b, base_offset, i * (intr->def.bit_size / 8));
118 *nir_get_io_offset_src(chan_intr) = nir_src_for_ssa(offset);
119
120 nir_builder_instr_insert(b, &chan_intr->instr);
121
122 loads[i] = &chan_intr->def;
123 }
124
125 nir_def_replace(&intr->def, nir_vec(b, loads, intr->num_components));
126 }
127
128 static void
lower_store_output_to_scalar(nir_builder * b,nir_intrinsic_instr * intr)129 lower_store_output_to_scalar(nir_builder *b, nir_intrinsic_instr *intr)
130 {
131 b->cursor = nir_before_instr(&intr->instr);
132
133 nir_def *value = intr->src[0].ssa;
134
135 for (unsigned i = 0; i < intr->num_components; i++) {
136 if (!(nir_intrinsic_write_mask(intr) & (1 << i)))
137 continue;
138
139 bool is_64bit = (nir_intrinsic_instr_src_type(intr, 0) & NIR_ALU_TYPE_SIZE_MASK) == 64;
140 unsigned newi = is_64bit ? i * 2 : i;
141 unsigned newc = nir_intrinsic_component(intr);
142 unsigned new_component = (newc + newi) % 4;
143 nir_io_semantics sem = nir_intrinsic_io_semantics(intr);
144 bool has_xfb = false;
145
146 if (nir_intrinsic_has_io_xfb(intr)) {
147 /* Find out which components are written via xfb. */
148 for (unsigned c = 0; c <= new_component; c++) {
149 nir_io_xfb xfb = c < 2 ? nir_intrinsic_io_xfb(intr) : nir_intrinsic_io_xfb2(intr);
150
151 if (new_component < c + xfb.out[c % 2].num_components) {
152 has_xfb = true;
153 break;
154 }
155 }
156 }
157
158 /* After scalarization, some channels might not write anywhere - i.e.
159 * they are not a sysval output, they don't feed the next shader, and
160 * they don't write xfb. Don't create such stores.
161 */
162 if ((sem.no_sysval_output ||
163 !nir_slot_is_sysval_output(sem.location, MESA_SHADER_NONE)) &&
164 (sem.no_varying ||
165 !nir_slot_is_varying(sem.location, MESA_SHADER_NONE)) &&
166 !has_xfb)
167 continue;
168
169 nir_intrinsic_instr *chan_intr =
170 nir_intrinsic_instr_create(b->shader, intr->intrinsic);
171 chan_intr->num_components = 1;
172
173 if (intr->name)
174 chan_intr->name = intr->name;
175 nir_intrinsic_set_base(chan_intr, nir_intrinsic_base(intr));
176 nir_intrinsic_set_write_mask(chan_intr, 0x1);
177 nir_intrinsic_set_component(chan_intr, new_component);
178 nir_intrinsic_set_src_type(chan_intr, nir_intrinsic_src_type(intr));
179 set_io_semantics(chan_intr, intr, i);
180
181 if (nir_intrinsic_has_io_xfb(intr)) {
182 /* Scalarize transform feedback info. */
183 for (unsigned c = 0; c <= new_component; c++) {
184 nir_io_xfb xfb = c < 2 ? nir_intrinsic_io_xfb(intr) : nir_intrinsic_io_xfb2(intr);
185
186 if (new_component < c + xfb.out[c % 2].num_components) {
187 nir_io_xfb scalar_xfb;
188
189 memset(&scalar_xfb, 0, sizeof(scalar_xfb));
190 scalar_xfb.out[new_component % 2].num_components = is_64bit ? 2 : 1;
191 scalar_xfb.out[new_component % 2].buffer = xfb.out[c % 2].buffer;
192 scalar_xfb.out[new_component % 2].offset = xfb.out[c % 2].offset +
193 new_component - c;
194 if (new_component < 2)
195 nir_intrinsic_set_io_xfb(chan_intr, scalar_xfb);
196 else
197 nir_intrinsic_set_io_xfb2(chan_intr, scalar_xfb);
198 break;
199 }
200 }
201 }
202
203 /* value */
204 chan_intr->src[0] = nir_src_for_ssa(nir_channel(b, value, i));
205 /* offset and vertex (if needed) */
206 for (unsigned j = 1; j < nir_intrinsic_infos[intr->intrinsic].num_srcs; ++j)
207 chan_intr->src[j] = nir_src_for_ssa(intr->src[j].ssa);
208 if (newc + newi > 3) {
209 nir_src *src = nir_get_io_offset_src(chan_intr);
210 nir_def *offset = nir_iadd_imm(b, src->ssa, (newc + newi) / 4);
211 *src = nir_src_for_ssa(offset);
212 }
213
214 nir_builder_instr_insert(b, &chan_intr->instr);
215 }
216
217 nir_instr_remove(&intr->instr);
218 }
219
220 static void
lower_store_to_scalar(nir_builder * b,nir_intrinsic_instr * intr)221 lower_store_to_scalar(nir_builder *b, nir_intrinsic_instr *intr)
222 {
223 b->cursor = nir_before_instr(&intr->instr);
224
225 nir_def *value = intr->src[0].ssa;
226 nir_def *base_offset = nir_get_io_offset_src(intr)->ssa;
227
228 /* iterate wrmask instead of num_components to handle split components */
229 u_foreach_bit(i, nir_intrinsic_write_mask(intr)) {
230 nir_intrinsic_instr *chan_intr =
231 nir_intrinsic_instr_create(b->shader, intr->intrinsic);
232 chan_intr->num_components = 1;
233
234 if (intr->name)
235 chan_intr->name = intr->name;
236 nir_intrinsic_set_write_mask(chan_intr, 0x1);
237 nir_intrinsic_set_align_offset(chan_intr,
238 (nir_intrinsic_align_offset(intr) +
239 i * (value->bit_size / 8)) %
240 nir_intrinsic_align_mul(intr));
241 nir_intrinsic_set_align_mul(chan_intr, nir_intrinsic_align_mul(intr));
242 if (nir_intrinsic_has_access(intr))
243 nir_intrinsic_set_access(chan_intr, nir_intrinsic_access(intr));
244 if (nir_intrinsic_has_base(intr))
245 nir_intrinsic_set_base(chan_intr, nir_intrinsic_base(intr));
246
247 /* value */
248 chan_intr->src[0] = nir_src_for_ssa(nir_channel(b, value, i));
249 for (unsigned j = 1; j < nir_intrinsic_infos[intr->intrinsic].num_srcs - 1; j++)
250 chan_intr->src[j] = nir_src_for_ssa(intr->src[j].ssa);
251
252 /* increment offset per component */
253 nir_def *offset = nir_iadd_imm(b, base_offset, i * (value->bit_size / 8));
254 *nir_get_io_offset_src(chan_intr) = nir_src_for_ssa(offset);
255
256 nir_builder_instr_insert(b, &chan_intr->instr);
257 }
258
259 nir_instr_remove(&intr->instr);
260 }
261
262 struct scalarize_state {
263 nir_variable_mode mask;
264 nir_instr_filter_cb filter;
265 void *filter_data;
266 };
267
268 static bool
nir_lower_io_to_scalar_instr(nir_builder * b,nir_instr * instr,void * data)269 nir_lower_io_to_scalar_instr(nir_builder *b, nir_instr *instr, void *data)
270 {
271 struct scalarize_state *state = data;
272
273 if (instr->type != nir_instr_type_intrinsic)
274 return false;
275
276 nir_intrinsic_instr *intr = nir_instr_as_intrinsic(instr);
277
278 if (intr->num_components == 1)
279 return false;
280
281 if ((intr->intrinsic == nir_intrinsic_load_input ||
282 intr->intrinsic == nir_intrinsic_load_per_primitive_input ||
283 intr->intrinsic == nir_intrinsic_load_per_vertex_input ||
284 intr->intrinsic == nir_intrinsic_load_interpolated_input ||
285 intr->intrinsic == nir_intrinsic_load_input_vertex) &&
286 (state->mask & nir_var_shader_in) &&
287 (!state->filter || state->filter(instr, state->filter_data))) {
288 lower_load_input_to_scalar(b, intr);
289 return true;
290 }
291
292 if ((intr->intrinsic == nir_intrinsic_load_output ||
293 intr->intrinsic == nir_intrinsic_load_per_vertex_output ||
294 intr->intrinsic == nir_intrinsic_load_per_view_output ||
295 intr->intrinsic == nir_intrinsic_load_per_primitive_output) &&
296 (state->mask & nir_var_shader_out) &&
297 (!state->filter || state->filter(instr, state->filter_data))) {
298 lower_load_input_to_scalar(b, intr);
299 return true;
300 }
301
302 if (((intr->intrinsic == nir_intrinsic_load_ubo && (state->mask & nir_var_mem_ubo)) ||
303 (intr->intrinsic == nir_intrinsic_load_ssbo && (state->mask & nir_var_mem_ssbo)) ||
304 (intr->intrinsic == nir_intrinsic_load_global && (state->mask & nir_var_mem_global)) ||
305 (intr->intrinsic == nir_intrinsic_load_shared && (state->mask & nir_var_mem_shared))) &&
306 (!state->filter || state->filter(instr, state->filter_data))) {
307 lower_load_to_scalar(b, intr);
308 return true;
309 }
310
311 if ((intr->intrinsic == nir_intrinsic_store_output ||
312 intr->intrinsic == nir_intrinsic_store_per_vertex_output ||
313 intr->intrinsic == nir_intrinsic_store_per_view_output ||
314 intr->intrinsic == nir_intrinsic_store_per_primitive_output) &&
315 state->mask & nir_var_shader_out &&
316 (!state->filter || state->filter(instr, state->filter_data))) {
317 lower_store_output_to_scalar(b, intr);
318 return true;
319 }
320
321 if (((intr->intrinsic == nir_intrinsic_store_ssbo && (state->mask & nir_var_mem_ssbo)) ||
322 (intr->intrinsic == nir_intrinsic_store_global && (state->mask & nir_var_mem_global)) ||
323 (intr->intrinsic == nir_intrinsic_store_shared && (state->mask & nir_var_mem_shared))) &&
324 (!state->filter || state->filter(instr, state->filter_data))) {
325 lower_store_to_scalar(b, intr);
326 return true;
327 }
328
329 return false;
330 }
331
332 bool
nir_lower_io_to_scalar(nir_shader * shader,nir_variable_mode mask,nir_instr_filter_cb filter,void * filter_data)333 nir_lower_io_to_scalar(nir_shader *shader, nir_variable_mode mask, nir_instr_filter_cb filter, void *filter_data)
334 {
335 struct scalarize_state state = {
336 mask,
337 filter,
338 filter_data
339 };
340 return nir_shader_instructions_pass(shader,
341 nir_lower_io_to_scalar_instr,
342 nir_metadata_control_flow,
343 &state);
344 }
345
346 static nir_variable **
get_channel_variables(struct hash_table * ht,nir_variable * var)347 get_channel_variables(struct hash_table *ht, nir_variable *var)
348 {
349 nir_variable **chan_vars;
350 struct hash_entry *entry = _mesa_hash_table_search(ht, var);
351 if (!entry) {
352 chan_vars = (nir_variable **)calloc(4, sizeof(nir_variable *));
353 _mesa_hash_table_insert(ht, var, chan_vars);
354 } else {
355 chan_vars = (nir_variable **)entry->data;
356 }
357
358 return chan_vars;
359 }
360
361 /*
362 * Note that the src deref that we are cloning is the head of the
363 * chain of deref instructions from the original intrinsic, but
364 * the dst we are cloning to is the tail (because chains of deref
365 * instructions are created back to front)
366 */
367
368 static nir_deref_instr *
clone_deref_array(nir_builder * b,nir_deref_instr * dst_tail,const nir_deref_instr * src_head)369 clone_deref_array(nir_builder *b, nir_deref_instr *dst_tail,
370 const nir_deref_instr *src_head)
371 {
372 const nir_deref_instr *parent = nir_deref_instr_parent(src_head);
373
374 if (!parent)
375 return dst_tail;
376
377 assert(src_head->deref_type == nir_deref_type_array);
378
379 dst_tail = clone_deref_array(b, dst_tail, parent);
380
381 return nir_build_deref_array(b, dst_tail,
382 src_head->arr.index.ssa);
383 }
384
385 static void
lower_load_to_scalar_early(nir_builder * b,nir_intrinsic_instr * intr,nir_variable * var,struct hash_table * split_inputs,struct hash_table * split_outputs)386 lower_load_to_scalar_early(nir_builder *b, nir_intrinsic_instr *intr,
387 nir_variable *var, struct hash_table *split_inputs,
388 struct hash_table *split_outputs)
389 {
390 b->cursor = nir_before_instr(&intr->instr);
391
392 nir_def *loads[NIR_MAX_VEC_COMPONENTS];
393
394 nir_variable **chan_vars;
395 if (var->data.mode == nir_var_shader_in) {
396 chan_vars = get_channel_variables(split_inputs, var);
397 } else {
398 chan_vars = get_channel_variables(split_outputs, var);
399 }
400
401 for (unsigned i = 0; i < intr->num_components; i++) {
402 nir_variable *chan_var = chan_vars[var->data.location_frac + i];
403 if (!chan_vars[var->data.location_frac + i]) {
404 chan_var = nir_variable_clone(var, b->shader);
405 chan_var->data.location_frac = var->data.location_frac + i;
406 chan_var->type = glsl_channel_type(chan_var->type);
407
408 chan_vars[var->data.location_frac + i] = chan_var;
409
410 nir_shader_add_variable(b->shader, chan_var);
411 }
412
413 nir_intrinsic_instr *chan_intr =
414 nir_intrinsic_instr_create(b->shader, intr->intrinsic);
415 nir_def_init(&chan_intr->instr, &chan_intr->def, 1,
416 intr->def.bit_size);
417 chan_intr->num_components = 1;
418
419 nir_deref_instr *deref = nir_build_deref_var(b, chan_var);
420
421 deref = clone_deref_array(b, deref, nir_src_as_deref(intr->src[0]));
422
423 chan_intr->src[0] = nir_src_for_ssa(&deref->def);
424
425 if (intr->intrinsic == nir_intrinsic_interp_deref_at_offset ||
426 intr->intrinsic == nir_intrinsic_interp_deref_at_sample ||
427 intr->intrinsic == nir_intrinsic_interp_deref_at_vertex)
428 chan_intr->src[1] = nir_src_for_ssa(intr->src[1].ssa);
429
430 nir_builder_instr_insert(b, &chan_intr->instr);
431
432 loads[i] = &chan_intr->def;
433 }
434
435 nir_def_replace(&intr->def, nir_vec(b, loads, intr->num_components));
436 }
437
438 static void
lower_store_output_to_scalar_early(nir_builder * b,nir_intrinsic_instr * intr,nir_variable * var,struct hash_table * split_outputs)439 lower_store_output_to_scalar_early(nir_builder *b, nir_intrinsic_instr *intr,
440 nir_variable *var,
441 struct hash_table *split_outputs)
442 {
443 b->cursor = nir_before_instr(&intr->instr);
444
445 nir_def *value = intr->src[1].ssa;
446
447 nir_variable **chan_vars = get_channel_variables(split_outputs, var);
448 for (unsigned i = 0; i < intr->num_components; i++) {
449 if (!(nir_intrinsic_write_mask(intr) & (1 << i)))
450 continue;
451
452 nir_variable *chan_var = chan_vars[var->data.location_frac + i];
453 if (!chan_vars[var->data.location_frac + i]) {
454 chan_var = nir_variable_clone(var, b->shader);
455 chan_var->data.location_frac = var->data.location_frac + i;
456 chan_var->type = glsl_channel_type(chan_var->type);
457
458 chan_vars[var->data.location_frac + i] = chan_var;
459
460 nir_shader_add_variable(b->shader, chan_var);
461 }
462
463 nir_intrinsic_instr *chan_intr =
464 nir_intrinsic_instr_create(b->shader, intr->intrinsic);
465 chan_intr->num_components = 1;
466
467 nir_intrinsic_set_write_mask(chan_intr, 0x1);
468
469 nir_deref_instr *deref = nir_build_deref_var(b, chan_var);
470
471 deref = clone_deref_array(b, deref, nir_src_as_deref(intr->src[0]));
472
473 chan_intr->src[0] = nir_src_for_ssa(&deref->def);
474 chan_intr->src[1] = nir_src_for_ssa(nir_channel(b, value, i));
475
476 nir_builder_instr_insert(b, &chan_intr->instr);
477 }
478
479 /* Remove the old store intrinsic */
480 nir_instr_remove(&intr->instr);
481 }
482
483 struct io_to_scalar_early_state {
484 struct hash_table *split_inputs, *split_outputs;
485 nir_variable_mode mask;
486 };
487
488 static bool
nir_lower_io_to_scalar_early_instr(nir_builder * b,nir_instr * instr,void * data)489 nir_lower_io_to_scalar_early_instr(nir_builder *b, nir_instr *instr, void *data)
490 {
491 struct io_to_scalar_early_state *state = data;
492
493 if (instr->type != nir_instr_type_intrinsic)
494 return false;
495
496 nir_intrinsic_instr *intr = nir_instr_as_intrinsic(instr);
497
498 if (intr->num_components == 1)
499 return false;
500
501 if (intr->intrinsic != nir_intrinsic_load_deref &&
502 intr->intrinsic != nir_intrinsic_store_deref &&
503 intr->intrinsic != nir_intrinsic_interp_deref_at_centroid &&
504 intr->intrinsic != nir_intrinsic_interp_deref_at_sample &&
505 intr->intrinsic != nir_intrinsic_interp_deref_at_offset &&
506 intr->intrinsic != nir_intrinsic_interp_deref_at_vertex)
507 return false;
508
509 nir_deref_instr *deref = nir_src_as_deref(intr->src[0]);
510 if (!nir_deref_mode_is_one_of(deref, state->mask))
511 return false;
512
513 nir_variable *var = nir_deref_instr_get_variable(deref);
514 nir_variable_mode mode = var->data.mode;
515
516 /* TODO: add patch support */
517 if (var->data.patch)
518 return false;
519
520 /* TODO: add doubles support */
521 if (glsl_type_is_64bit(glsl_without_array(var->type)))
522 return false;
523
524 if (!(b->shader->info.stage == MESA_SHADER_VERTEX &&
525 mode == nir_var_shader_in) &&
526 var->data.location < VARYING_SLOT_VAR0 &&
527 var->data.location >= 0)
528 return false;
529
530 /* Don't bother splitting if we can't opt away any unused
531 * components.
532 */
533 if (var->data.always_active_io)
534 return false;
535
536 if (var->data.must_be_shader_input)
537 return false;
538
539 /* Skip types we cannot split */
540 if (glsl_type_is_matrix(glsl_without_array(var->type)) ||
541 glsl_type_is_struct_or_ifc(glsl_without_array(var->type)))
542 return false;
543
544 switch (intr->intrinsic) {
545 case nir_intrinsic_interp_deref_at_centroid:
546 case nir_intrinsic_interp_deref_at_sample:
547 case nir_intrinsic_interp_deref_at_offset:
548 case nir_intrinsic_interp_deref_at_vertex:
549 case nir_intrinsic_load_deref:
550 if ((state->mask & nir_var_shader_in && mode == nir_var_shader_in) ||
551 (state->mask & nir_var_shader_out && mode == nir_var_shader_out)) {
552 lower_load_to_scalar_early(b, intr, var, state->split_inputs,
553 state->split_outputs);
554 return true;
555 }
556 break;
557 case nir_intrinsic_store_deref:
558 if (state->mask & nir_var_shader_out &&
559 mode == nir_var_shader_out) {
560 lower_store_output_to_scalar_early(b, intr, var, state->split_outputs);
561 return true;
562 }
563 break;
564 default:
565 break;
566 }
567
568 return false;
569 }
570
571 /*
572 * This function is intended to be called earlier than nir_lower_io_to_scalar()
573 * i.e. before nir_lower_io() is called.
574 */
575 bool
nir_lower_io_to_scalar_early(nir_shader * shader,nir_variable_mode mask)576 nir_lower_io_to_scalar_early(nir_shader *shader, nir_variable_mode mask)
577 {
578 struct io_to_scalar_early_state state = {
579 .split_inputs = _mesa_pointer_hash_table_create(NULL),
580 .split_outputs = _mesa_pointer_hash_table_create(NULL),
581 .mask = mask
582 };
583
584 bool progress = nir_shader_instructions_pass(shader,
585 nir_lower_io_to_scalar_early_instr,
586 nir_metadata_control_flow,
587 &state);
588
589 /* Remove old input from the shaders inputs list */
590 hash_table_foreach(state.split_inputs, entry) {
591 nir_variable *var = (nir_variable *)entry->key;
592 exec_node_remove(&var->node);
593
594 free(entry->data);
595 }
596
597 /* Remove old output from the shaders outputs list */
598 hash_table_foreach(state.split_outputs, entry) {
599 nir_variable *var = (nir_variable *)entry->key;
600 exec_node_remove(&var->node);
601
602 free(entry->data);
603 }
604
605 _mesa_hash_table_destroy(state.split_inputs, NULL);
606 _mesa_hash_table_destroy(state.split_outputs, NULL);
607
608 nir_remove_dead_derefs(shader);
609
610 return progress;
611 }
612