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_ssa_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_ssa_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 = nir_ssa_for_src(b, (*p)->arr.index, 1);
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) ?
67 glsl_get_aoa_size((*p)->type) : 1;
68
69 num_elements *= glsl_type_is_matrix(glsl_without_array((*p)->type)) ?
70 glsl_get_matrix_columns(glsl_without_array((*p)->type)) : 1;
71
72 *element_index += num_elements * index;
73 } else if ((*p)->deref_type == nir_deref_type_struct) {
74 /* TODO: we could also add struct splitting support to this pass */
75 break;
76 }
77 }
78
79 nir_deref_path_finish(&path);
80
81 return offset;
82 }
83
84 static nir_variable **
get_array_elements(struct hash_table * ht,nir_variable * var,gl_shader_stage stage)85 get_array_elements(struct hash_table *ht, nir_variable *var,
86 gl_shader_stage stage)
87 {
88 nir_variable **elements;
89 struct hash_entry *entry = _mesa_hash_table_search(ht, var);
90 if (!entry) {
91 const struct glsl_type *type = var->type;
92 if (nir_is_arrayed_io(var, stage)) {
93 assert(glsl_type_is_array(type));
94 type = glsl_get_array_element(type);
95 }
96
97 unsigned num_elements = glsl_type_is_array(type) ?
98 glsl_get_aoa_size(type) : 1;
99
100 num_elements *= glsl_type_is_matrix(glsl_without_array(type)) ?
101 glsl_get_matrix_columns(glsl_without_array(type)) : 1;
102
103 elements = (nir_variable **) calloc(num_elements, sizeof(nir_variable *));
104 _mesa_hash_table_insert(ht, var, elements);
105 } else {
106 elements = (nir_variable **) entry->data;
107 }
108
109 return elements;
110 }
111
112 static void
lower_array(nir_builder * b,nir_intrinsic_instr * intr,nir_variable * var,struct hash_table * varyings)113 lower_array(nir_builder *b, nir_intrinsic_instr *intr, nir_variable *var,
114 struct hash_table *varyings)
115 {
116 b->cursor = nir_before_instr(&intr->instr);
117
118 if (nir_deref_instr_is_known_out_of_bounds(nir_src_as_deref(intr->src[0]))) {
119 /* See Section 5.11 (Out-of-Bounds Accesses) of the GLSL 4.60 */
120 if (intr->intrinsic != nir_intrinsic_store_deref) {
121 nir_ssa_def *zero = nir_imm_zero(b, intr->dest.ssa.num_components,
122 intr->dest.ssa.bit_size);
123 nir_ssa_def_rewrite_uses(&intr->dest.ssa,
124 zero);
125 }
126 nir_instr_remove(&intr->instr);
127 return;
128 }
129
130 nir_variable **elements =
131 get_array_elements(varyings, var, b->shader->info.stage);
132
133 nir_ssa_def *array_index = NULL;
134 unsigned elements_index = 0;
135 unsigned xfb_offset = 0;
136 unsigned io_offset = get_io_offset(b, nir_src_as_deref(intr->src[0]),
137 var, &elements_index, &xfb_offset,
138 &array_index);
139
140 nir_variable *element = elements[elements_index];
141 if (!element) {
142 element = nir_variable_clone(var, b->shader);
143 element->data.location = var->data.location + io_offset;
144
145 if (var->data.explicit_offset)
146 element->data.offset = var->data.offset + xfb_offset;
147
148 const struct glsl_type *type = glsl_without_array(element->type);
149
150 /* This pass also splits matrices so we need give them a new type. */
151 if (glsl_type_is_matrix(type))
152 type = glsl_get_column_type(type);
153
154 if (nir_is_arrayed_io(var, b->shader->info.stage)) {
155 type = glsl_array_type(type, glsl_get_length(element->type),
156 glsl_get_explicit_stride(element->type));
157 }
158
159 element->type = type;
160 elements[elements_index] = element;
161
162 nir_shader_add_variable(b->shader, element);
163 }
164
165 nir_deref_instr *element_deref = nir_build_deref_var(b, element);
166
167 if (nir_is_arrayed_io(var, b->shader->info.stage)) {
168 assert(array_index);
169 element_deref = nir_build_deref_array(b, element_deref, array_index);
170 }
171
172 nir_intrinsic_instr *element_intr =
173 nir_intrinsic_instr_create(b->shader, intr->intrinsic);
174 element_intr->num_components = intr->num_components;
175 element_intr->src[0] = nir_src_for_ssa(&element_deref->dest.ssa);
176
177 if (intr->intrinsic != nir_intrinsic_store_deref) {
178 nir_ssa_dest_init(&element_intr->instr, &element_intr->dest,
179 intr->num_components, intr->dest.ssa.bit_size, NULL);
180
181 if (intr->intrinsic == nir_intrinsic_interp_deref_at_offset ||
182 intr->intrinsic == nir_intrinsic_interp_deref_at_sample ||
183 intr->intrinsic == nir_intrinsic_interp_deref_at_vertex) {
184 nir_src_copy(&element_intr->src[1], &intr->src[1]);
185 }
186
187 nir_ssa_def_rewrite_uses(&intr->dest.ssa,
188 &element_intr->dest.ssa);
189 } else {
190 nir_intrinsic_set_write_mask(element_intr,
191 nir_intrinsic_write_mask(intr));
192 nir_src_copy(&element_intr->src[1], &intr->src[1]);
193 }
194
195 nir_builder_instr_insert(b, &element_intr->instr);
196
197 /* Remove the old load intrinsic */
198 nir_instr_remove(&intr->instr);
199 }
200
201 static bool
deref_has_indirect(nir_builder * b,nir_variable * var,nir_deref_path * path)202 deref_has_indirect(nir_builder *b, nir_variable *var, nir_deref_path *path)
203 {
204 assert(path->path[0]->deref_type == nir_deref_type_var);
205 nir_deref_instr **p = &path->path[1];
206
207 if (nir_is_arrayed_io(var, b->shader->info.stage)) {
208 p++;
209 }
210
211 for (; *p; p++) {
212 if ((*p)->deref_type != nir_deref_type_array)
213 continue;
214
215 if (!nir_src_is_const((*p)->arr.index))
216 return true;
217 }
218
219 return false;
220 }
221
222 /* Creates a mask of locations that contains arrays that are indexed via
223 * indirect indexing.
224 */
225 static void
create_indirects_mask(nir_shader * shader,BITSET_WORD * indirects,nir_variable_mode mode)226 create_indirects_mask(nir_shader *shader,
227 BITSET_WORD *indirects, nir_variable_mode mode)
228 {
229 nir_foreach_function(function, shader) {
230 if (function->impl) {
231 nir_builder b;
232 nir_builder_init(&b, function->impl);
233
234 nir_foreach_block(block, function->impl) {
235 nir_foreach_instr_safe(instr, block) {
236
237 if (instr->type != nir_instr_type_intrinsic)
238 continue;
239
240 nir_intrinsic_instr *intr = nir_instr_as_intrinsic(instr);
241
242 if (intr->intrinsic != nir_intrinsic_load_deref &&
243 intr->intrinsic != nir_intrinsic_store_deref &&
244 intr->intrinsic != nir_intrinsic_interp_deref_at_centroid &&
245 intr->intrinsic != nir_intrinsic_interp_deref_at_sample &&
246 intr->intrinsic != nir_intrinsic_interp_deref_at_offset &&
247 intr->intrinsic != nir_intrinsic_interp_deref_at_vertex)
248 continue;
249
250 nir_deref_instr *deref = nir_src_as_deref(intr->src[0]);
251 if (!nir_deref_mode_is(deref, mode))
252 continue;
253
254 nir_variable *var = nir_deref_instr_get_variable(deref);
255
256 nir_deref_path path;
257 nir_deref_path_init(&path, deref, NULL);
258
259 int loc = var->data.location * 4 + var->data.location_frac;
260 if (deref_has_indirect(&b, var, &path))
261 BITSET_SET(indirects, loc);
262
263 nir_deref_path_finish(&path);
264 }
265 }
266 }
267 }
268 }
269
270 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)271 lower_io_arrays_to_elements(nir_shader *shader, nir_variable_mode mask,
272 BITSET_WORD *indirects,
273 struct hash_table *varyings,
274 bool after_cross_stage_opts)
275 {
276 nir_foreach_function(function, shader) {
277 if (function->impl) {
278 nir_builder b;
279 nir_builder_init(&b, function->impl);
280
281 nir_foreach_block(block, function->impl) {
282 nir_foreach_instr_safe(instr, block) {
283 if (instr->type != nir_instr_type_intrinsic)
284 continue;
285
286 nir_intrinsic_instr *intr = nir_instr_as_intrinsic(instr);
287
288 if (intr->intrinsic != nir_intrinsic_load_deref &&
289 intr->intrinsic != nir_intrinsic_store_deref &&
290 intr->intrinsic != nir_intrinsic_interp_deref_at_centroid &&
291 intr->intrinsic != nir_intrinsic_interp_deref_at_sample &&
292 intr->intrinsic != nir_intrinsic_interp_deref_at_offset &&
293 intr->intrinsic != nir_intrinsic_interp_deref_at_vertex)
294 continue;
295
296 nir_deref_instr *deref = nir_src_as_deref(intr->src[0]);
297 if (!nir_deref_mode_is_one_of(deref, mask))
298 continue;
299
300 nir_variable *var = nir_deref_instr_get_variable(deref);
301
302 /* Drivers assume compact arrays are, in fact, arrays. */
303 if (var->data.compact)
304 continue;
305
306 /* Per-view variables are expected to remain arrays. */
307 if (var->data.per_view)
308 continue;
309
310 /* Skip indirects */
311 int loc = var->data.location * 4 + var->data.location_frac;
312 if (BITSET_TEST(indirects, loc))
313 continue;
314
315 nir_variable_mode mode = var->data.mode;
316
317 const struct glsl_type *type = var->type;
318 if (nir_is_arrayed_io(var, b.shader->info.stage)) {
319 assert(glsl_type_is_array(type));
320 type = glsl_get_array_element(type);
321 }
322
323 /* Skip types we cannot split.
324 *
325 * TODO: Add support for struct splitting.
326 */
327 if ((!glsl_type_is_array(type) && !glsl_type_is_matrix(type))||
328 glsl_type_is_struct_or_ifc(glsl_without_array(type)))
329 continue;
330
331 /* Skip builtins */
332 if (!after_cross_stage_opts &&
333 var->data.location < VARYING_SLOT_VAR0 &&
334 var->data.location >= 0)
335 continue;
336
337 /* Don't bother splitting if we can't opt away any unused
338 * elements.
339 */
340 if (!after_cross_stage_opts && var->data.always_active_io)
341 continue;
342
343 switch (intr->intrinsic) {
344 case nir_intrinsic_interp_deref_at_centroid:
345 case nir_intrinsic_interp_deref_at_sample:
346 case nir_intrinsic_interp_deref_at_offset:
347 case nir_intrinsic_interp_deref_at_vertex:
348 case nir_intrinsic_load_deref:
349 case nir_intrinsic_store_deref:
350 if ((mask & nir_var_shader_in && mode == nir_var_shader_in) ||
351 (mask & nir_var_shader_out && mode == nir_var_shader_out))
352 lower_array(&b, intr, var, varyings);
353 break;
354 default:
355 break;
356 }
357 }
358 }
359 }
360 }
361 }
362
363 void
nir_lower_io_arrays_to_elements_no_indirects(nir_shader * shader,bool outputs_only)364 nir_lower_io_arrays_to_elements_no_indirects(nir_shader *shader,
365 bool outputs_only)
366 {
367 struct hash_table *split_inputs = _mesa_pointer_hash_table_create(NULL);
368 struct hash_table *split_outputs = _mesa_pointer_hash_table_create(NULL);
369
370 BITSET_DECLARE(indirects, 4 * VARYING_SLOT_TESS_MAX) = {0};
371
372 lower_io_arrays_to_elements(shader, nir_var_shader_out,
373 indirects, split_outputs, true);
374
375 if (!outputs_only) {
376 lower_io_arrays_to_elements(shader, nir_var_shader_in,
377 indirects, split_inputs, true);
378
379 /* Remove old input from the shaders inputs list */
380 hash_table_foreach(split_inputs, entry) {
381 nir_variable *var = (nir_variable *) entry->key;
382 exec_node_remove(&var->node);
383
384 free(entry->data);
385 }
386 }
387
388 /* Remove old output from the shaders outputs list */
389 hash_table_foreach(split_outputs, entry) {
390 nir_variable *var = (nir_variable *) entry->key;
391 exec_node_remove(&var->node);
392
393 free(entry->data);
394 }
395
396 _mesa_hash_table_destroy(split_inputs, NULL);
397 _mesa_hash_table_destroy(split_outputs, NULL);
398
399 nir_remove_dead_derefs(shader);
400 }
401
402 void
nir_lower_io_arrays_to_elements(nir_shader * producer,nir_shader * consumer)403 nir_lower_io_arrays_to_elements(nir_shader *producer, nir_shader *consumer)
404 {
405 struct hash_table *split_inputs = _mesa_pointer_hash_table_create(NULL);
406 struct hash_table *split_outputs = _mesa_pointer_hash_table_create(NULL);
407
408 BITSET_DECLARE(indirects, 4 * VARYING_SLOT_TESS_MAX) = {0};
409
410 create_indirects_mask(producer, indirects, nir_var_shader_out);
411 create_indirects_mask(consumer, indirects, nir_var_shader_in);
412
413 lower_io_arrays_to_elements(producer, nir_var_shader_out,
414 indirects, split_outputs, false);
415
416 lower_io_arrays_to_elements(consumer, nir_var_shader_in,
417 indirects, split_inputs, false);
418
419 /* Remove old input from the shaders inputs list */
420 hash_table_foreach(split_inputs, entry) {
421 nir_variable *var = (nir_variable *) entry->key;
422 exec_node_remove(&var->node);
423
424 free(entry->data);
425 }
426
427 /* Remove old output from the shaders outputs list */
428 hash_table_foreach(split_outputs, entry) {
429 nir_variable *var = (nir_variable *) entry->key;
430 exec_node_remove(&var->node);
431
432 free(entry->data);
433 }
434
435 _mesa_hash_table_destroy(split_inputs, NULL);
436 _mesa_hash_table_destroy(split_outputs, NULL);
437
438 nir_remove_dead_derefs(producer);
439 nir_remove_dead_derefs(consumer);
440 }
441