1 /*
2 * Copyright © 2017 Timothy Arceri
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_arrays_to_elements.c
29 *
30 * Split arrays/matrices with direct indexing into individual elements. This
31 * will allow optimisation passes to better clean up unused elements.
32 *
33 */
34
35 static unsigned
get_io_offset(nir_builder * b,nir_deref_instr * deref,nir_variable * var,unsigned * element_index,unsigned * xfb_offset,nir_def ** array_index)36 get_io_offset(nir_builder *b, nir_deref_instr *deref, nir_variable *var,
37 unsigned *element_index, unsigned *xfb_offset,
38 nir_def **array_index)
39 {
40 nir_deref_path path;
41 nir_deref_path_init(&path, deref, NULL);
42
43 assert(path.path[0]->deref_type == nir_deref_type_var);
44 nir_deref_instr **p = &path.path[1];
45
46 /* For arrayed I/O (e.g., per-vertex input arrays in geometry shader
47 * inputs), skip the outermost array index. Process the rest normally.
48 */
49 if (nir_is_arrayed_io(var, b->shader->info.stage)) {
50 *array_index = (*p)->arr.index.ssa;
51 p++;
52 }
53
54 unsigned offset = 0;
55 *xfb_offset = 0;
56 for (; *p; p++) {
57 if ((*p)->deref_type == nir_deref_type_array) {
58 /* must not be indirect dereference */
59 unsigned index = nir_src_as_uint((*p)->arr.index);
60
61 unsigned size = glsl_count_attribute_slots((*p)->type, false);
62 offset += size * index;
63
64 *xfb_offset += index * glsl_get_component_slots((*p)->type) * 4;
65
66 unsigned num_elements = glsl_type_is_array((*p)->type) ? glsl_get_aoa_size((*p)->type) : 1;
67
68 num_elements *= glsl_type_is_matrix(glsl_without_array((*p)->type)) ? glsl_get_matrix_columns(glsl_without_array((*p)->type)) : 1;
69
70 *element_index += num_elements * index;
71 } else if ((*p)->deref_type == nir_deref_type_struct) {
72 /* TODO: we could also add struct splitting support to this pass */
73 break;
74 }
75 }
76
77 nir_deref_path_finish(&path);
78
79 return offset;
80 }
81
82 static nir_variable **
get_array_elements(struct hash_table * ht,nir_variable * var,gl_shader_stage stage)83 get_array_elements(struct hash_table *ht, nir_variable *var,
84 gl_shader_stage stage)
85 {
86 nir_variable **elements;
87 struct hash_entry *entry = _mesa_hash_table_search(ht, var);
88 if (!entry) {
89 const struct glsl_type *type = var->type;
90 if (nir_is_arrayed_io(var, stage)) {
91 assert(glsl_type_is_array(type));
92 type = glsl_get_array_element(type);
93 }
94
95 unsigned num_elements = glsl_type_is_array(type) ? glsl_get_aoa_size(type) : 1;
96
97 num_elements *= glsl_type_is_matrix(glsl_without_array(type)) ? glsl_get_matrix_columns(glsl_without_array(type)) : 1;
98
99 elements = (nir_variable **)calloc(num_elements, sizeof(nir_variable *));
100 _mesa_hash_table_insert(ht, var, elements);
101 } else {
102 elements = (nir_variable **)entry->data;
103 }
104
105 return elements;
106 }
107
108 static void
lower_array(nir_builder * b,nir_intrinsic_instr * intr,nir_variable * var,struct hash_table * varyings)109 lower_array(nir_builder *b, nir_intrinsic_instr *intr, nir_variable *var,
110 struct hash_table *varyings)
111 {
112 b->cursor = nir_before_instr(&intr->instr);
113
114 if (nir_deref_instr_is_known_out_of_bounds(nir_src_as_deref(intr->src[0]))) {
115 /* See Section 5.11 (Out-of-Bounds Accesses) of the GLSL 4.60 */
116 if (intr->intrinsic != nir_intrinsic_store_deref) {
117 nir_def *zero = nir_imm_zero(b, intr->def.num_components,
118 intr->def.bit_size);
119 nir_def_rewrite_uses(&intr->def,
120 zero);
121 }
122 nir_instr_remove(&intr->instr);
123 return;
124 }
125
126 nir_variable **elements =
127 get_array_elements(varyings, var, b->shader->info.stage);
128
129 nir_def *array_index = NULL;
130 unsigned elements_index = 0;
131 unsigned xfb_offset = 0;
132 unsigned io_offset = get_io_offset(b, nir_src_as_deref(intr->src[0]),
133 var, &elements_index, &xfb_offset,
134 &array_index);
135
136 nir_variable *element = elements[elements_index];
137 if (!element) {
138 element = nir_variable_clone(var, b->shader);
139 element->data.location = var->data.location + io_offset;
140
141 if (var->data.explicit_offset)
142 element->data.offset = var->data.offset + xfb_offset;
143
144 const struct glsl_type *type = glsl_without_array(element->type);
145
146 /* This pass also splits matrices so we need give them a new type. */
147 if (glsl_type_is_matrix(type))
148 type = glsl_get_column_type(type);
149
150 if (nir_is_arrayed_io(var, b->shader->info.stage)) {
151 type = glsl_array_type(type, glsl_get_length(element->type),
152 glsl_get_explicit_stride(element->type));
153 }
154
155 element->type = type;
156 elements[elements_index] = element;
157
158 nir_shader_add_variable(b->shader, element);
159 }
160
161 nir_deref_instr *element_deref = nir_build_deref_var(b, element);
162
163 if (nir_is_arrayed_io(var, b->shader->info.stage)) {
164 assert(array_index);
165 element_deref = nir_build_deref_array(b, element_deref, array_index);
166 }
167
168 nir_intrinsic_instr *element_intr =
169 nir_intrinsic_instr_create(b->shader, intr->intrinsic);
170 element_intr->num_components = intr->num_components;
171 element_intr->src[0] = nir_src_for_ssa(&element_deref->def);
172
173 if (intr->intrinsic != nir_intrinsic_store_deref) {
174 nir_def_init(&element_intr->instr, &element_intr->def,
175 intr->num_components, intr->def.bit_size);
176
177 if (intr->intrinsic == nir_intrinsic_interp_deref_at_offset ||
178 intr->intrinsic == nir_intrinsic_interp_deref_at_sample ||
179 intr->intrinsic == nir_intrinsic_interp_deref_at_vertex) {
180 element_intr->src[1] = nir_src_for_ssa(intr->src[1].ssa);
181 }
182
183 nir_def_rewrite_uses(&intr->def,
184 &element_intr->def);
185 } else {
186 nir_intrinsic_set_write_mask(element_intr,
187 nir_intrinsic_write_mask(intr));
188 element_intr->src[1] = nir_src_for_ssa(intr->src[1].ssa);
189 }
190
191 nir_builder_instr_insert(b, &element_intr->instr);
192
193 /* Remove the old load intrinsic */
194 nir_instr_remove(&intr->instr);
195 }
196
197 static bool
deref_has_indirect(nir_builder * b,nir_variable * var,nir_deref_path * path)198 deref_has_indirect(nir_builder *b, nir_variable *var, nir_deref_path *path)
199 {
200 assert(path->path[0]->deref_type == nir_deref_type_var);
201 nir_deref_instr **p = &path->path[1];
202
203 if (nir_is_arrayed_io(var, b->shader->info.stage)) {
204 p++;
205 }
206
207 for (; *p; p++) {
208 if ((*p)->deref_type != nir_deref_type_array)
209 continue;
210
211 if (!nir_src_is_const((*p)->arr.index))
212 return true;
213 }
214
215 return false;
216 }
217
218 /* Creates a mask of locations that contains arrays that are indexed via
219 * indirect indexing.
220 */
221 static void
create_indirects_mask(nir_shader * shader,BITSET_WORD * indirects,nir_variable_mode mode)222 create_indirects_mask(nir_shader *shader,
223 BITSET_WORD *indirects, nir_variable_mode mode)
224 {
225 nir_foreach_function_impl(impl, shader) {
226 nir_builder b = nir_builder_create(impl);
227
228 nir_foreach_block(block, impl) {
229 nir_foreach_instr_safe(instr, block) {
230
231 if (instr->type != nir_instr_type_intrinsic)
232 continue;
233
234 nir_intrinsic_instr *intr = nir_instr_as_intrinsic(instr);
235
236 if (intr->intrinsic != nir_intrinsic_load_deref &&
237 intr->intrinsic != nir_intrinsic_store_deref &&
238 intr->intrinsic != nir_intrinsic_interp_deref_at_centroid &&
239 intr->intrinsic != nir_intrinsic_interp_deref_at_sample &&
240 intr->intrinsic != nir_intrinsic_interp_deref_at_offset &&
241 intr->intrinsic != nir_intrinsic_interp_deref_at_vertex)
242 continue;
243
244 nir_deref_instr *deref = nir_src_as_deref(intr->src[0]);
245 if (!nir_deref_mode_is(deref, mode))
246 continue;
247
248 nir_variable *var = nir_deref_instr_get_variable(deref);
249
250 nir_deref_path path;
251 nir_deref_path_init(&path, deref, NULL);
252
253 int loc = var->data.location * 4 + var->data.location_frac;
254 if (deref_has_indirect(&b, var, &path))
255 BITSET_SET(indirects, loc);
256
257 nir_deref_path_finish(&path);
258 }
259 }
260 }
261 }
262
263 static void
lower_io_arrays_to_elements(nir_shader * shader,nir_variable_mode mask,BITSET_WORD * indirects,struct hash_table * varyings,bool after_cross_stage_opts)264 lower_io_arrays_to_elements(nir_shader *shader, nir_variable_mode mask,
265 BITSET_WORD *indirects,
266 struct hash_table *varyings,
267 bool after_cross_stage_opts)
268 {
269 nir_foreach_function_impl(impl, shader) {
270 nir_builder b = nir_builder_create(impl);
271
272 nir_foreach_block(block, impl) {
273 nir_foreach_instr_safe(instr, block) {
274 if (instr->type != nir_instr_type_intrinsic)
275 continue;
276
277 nir_intrinsic_instr *intr = nir_instr_as_intrinsic(instr);
278
279 if (intr->intrinsic != nir_intrinsic_load_deref &&
280 intr->intrinsic != nir_intrinsic_store_deref &&
281 intr->intrinsic != nir_intrinsic_interp_deref_at_centroid &&
282 intr->intrinsic != nir_intrinsic_interp_deref_at_sample &&
283 intr->intrinsic != nir_intrinsic_interp_deref_at_offset &&
284 intr->intrinsic != nir_intrinsic_interp_deref_at_vertex)
285 continue;
286
287 nir_deref_instr *deref = nir_src_as_deref(intr->src[0]);
288 if (!nir_deref_mode_is_one_of(deref, mask))
289 continue;
290
291 nir_variable *var = nir_deref_instr_get_variable(deref);
292
293 /* Drivers assume compact arrays are, in fact, arrays. */
294 if (var->data.compact)
295 continue;
296
297 /* Per-view variables are expected to remain arrays. */
298 if (var->data.per_view)
299 continue;
300
301 /* Skip indirects */
302 int loc = var->data.location * 4 + var->data.location_frac;
303 if (BITSET_TEST(indirects, loc))
304 continue;
305
306 nir_variable_mode mode = var->data.mode;
307
308 const struct glsl_type *type = var->type;
309 if (nir_is_arrayed_io(var, b.shader->info.stage)) {
310 assert(glsl_type_is_array(type));
311 type = glsl_get_array_element(type);
312 }
313
314 /* Skip types we cannot split.
315 *
316 * TODO: Add support for struct splitting.
317 */
318 if ((!glsl_type_is_array(type) && !glsl_type_is_matrix(type)) ||
319 glsl_type_is_struct_or_ifc(glsl_without_array(type)))
320 continue;
321
322 /* Skip builtins */
323 if (!after_cross_stage_opts &&
324 var->data.location < VARYING_SLOT_VAR0 &&
325 var->data.location >= 0)
326 continue;
327
328 /* Don't bother splitting if we can't opt away any unused
329 * elements.
330 */
331 if (!after_cross_stage_opts && var->data.always_active_io)
332 continue;
333
334 switch (intr->intrinsic) {
335 case nir_intrinsic_interp_deref_at_centroid:
336 case nir_intrinsic_interp_deref_at_sample:
337 case nir_intrinsic_interp_deref_at_offset:
338 case nir_intrinsic_interp_deref_at_vertex:
339 case nir_intrinsic_load_deref:
340 case nir_intrinsic_store_deref:
341 if ((mask & nir_var_shader_in && mode == nir_var_shader_in) ||
342 (mask & nir_var_shader_out && mode == nir_var_shader_out))
343 lower_array(&b, intr, var, varyings);
344 break;
345 default:
346 break;
347 }
348 }
349 }
350 }
351 }
352
353 bool
nir_lower_io_arrays_to_elements_no_indirects(nir_shader * shader,bool outputs_only)354 nir_lower_io_arrays_to_elements_no_indirects(nir_shader *shader,
355 bool outputs_only)
356 {
357 struct hash_table *split_inputs = _mesa_pointer_hash_table_create(NULL);
358 struct hash_table *split_outputs = _mesa_pointer_hash_table_create(NULL);
359
360 BITSET_DECLARE(indirects, 4 * VARYING_SLOT_TESS_MAX) = { 0 };
361
362 lower_io_arrays_to_elements(shader, nir_var_shader_out,
363 indirects, split_outputs, true);
364
365 if (!outputs_only) {
366 lower_io_arrays_to_elements(shader, nir_var_shader_in,
367 indirects, split_inputs, true);
368
369 /* Remove old input from the shaders inputs list */
370 hash_table_foreach(split_inputs, entry) {
371 nir_variable *var = (nir_variable *)entry->key;
372 exec_node_remove(&var->node);
373
374 free(entry->data);
375 }
376 }
377
378 /* Remove old output from the shaders outputs list */
379 hash_table_foreach(split_outputs, entry) {
380 nir_variable *var = (nir_variable *)entry->key;
381 exec_node_remove(&var->node);
382
383 free(entry->data);
384 }
385
386 _mesa_hash_table_destroy(split_inputs, NULL);
387 _mesa_hash_table_destroy(split_outputs, NULL);
388
389 nir_remove_dead_derefs(shader);
390 return true;
391 }
392
393 void
nir_lower_io_arrays_to_elements(nir_shader * producer,nir_shader * consumer)394 nir_lower_io_arrays_to_elements(nir_shader *producer, nir_shader *consumer)
395 {
396 struct hash_table *split_inputs = _mesa_pointer_hash_table_create(NULL);
397 struct hash_table *split_outputs = _mesa_pointer_hash_table_create(NULL);
398
399 BITSET_DECLARE(indirects, 4 * VARYING_SLOT_TESS_MAX) = { 0 };
400
401 create_indirects_mask(producer, indirects, nir_var_shader_out);
402 create_indirects_mask(consumer, indirects, nir_var_shader_in);
403
404 lower_io_arrays_to_elements(producer, nir_var_shader_out,
405 indirects, split_outputs, false);
406
407 lower_io_arrays_to_elements(consumer, nir_var_shader_in,
408 indirects, split_inputs, false);
409
410 /* Remove old input from the shaders inputs list */
411 hash_table_foreach(split_inputs, entry) {
412 nir_variable *var = (nir_variable *)entry->key;
413 exec_node_remove(&var->node);
414
415 free(entry->data);
416 }
417
418 /* Remove old output from the shaders outputs list */
419 hash_table_foreach(split_outputs, entry) {
420 nir_variable *var = (nir_variable *)entry->key;
421 exec_node_remove(&var->node);
422
423 free(entry->data);
424 }
425
426 _mesa_hash_table_destroy(split_inputs, NULL);
427 _mesa_hash_table_destroy(split_outputs, NULL);
428
429 nir_remove_dead_derefs(producer);
430 nir_remove_dead_derefs(consumer);
431 }
432