1 /*
2 * Copyright © 2015 Intel Corporation
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 /*
25 * Implements a pass that lowers output and/or input variables to a
26 * temporary plus an output variable with a single copy at each exit
27 * point of the shader and/or an input variable with a single copy
28 * at the entrance point of the shader. This way the output variable
29 * is only ever written once and/or input is only read once, and there
30 * are no indirect outut/input accesses.
31 */
32
33 #include "nir.h"
34 #include "nir_builder.h"
35 #include "nir_deref.h"
36
37 struct lower_io_state {
38 nir_shader *shader;
39 nir_function_impl *entrypoint;
40 struct exec_list old_outputs;
41 struct exec_list old_inputs;
42 struct exec_list new_outputs;
43 struct exec_list new_inputs;
44
45 /* map from temporary to new input */
46 struct hash_table *input_map;
47 };
48
49 static void
emit_copies(nir_builder * b,struct exec_list * dest_vars,struct exec_list * src_vars)50 emit_copies(nir_builder *b, struct exec_list *dest_vars,
51 struct exec_list *src_vars)
52 {
53 assert(exec_list_length(dest_vars) == exec_list_length(src_vars));
54
55 foreach_two_lists(dest_node, dest_vars, src_node, src_vars) {
56 nir_variable *dest = exec_node_data(nir_variable, dest_node, node);
57 nir_variable *src = exec_node_data(nir_variable, src_node, node);
58
59 /* No need to copy the contents of a non-fb_fetch_output output variable
60 * to the temporary allocated for it, since its initial value is
61 * undefined.
62 */
63 if (src->data.mode == nir_var_shader_out &&
64 !src->data.fb_fetch_output)
65 continue;
66
67 /* Can't copy the contents of the temporary back to a read-only
68 * interface variable. The value of the temporary won't have been
69 * modified by the shader anyway.
70 */
71 if (dest->data.read_only)
72 continue;
73
74 nir_copy_var(b, dest, src);
75 }
76 }
77
78 static void
emit_output_copies_impl(struct lower_io_state * state,nir_function_impl * impl)79 emit_output_copies_impl(struct lower_io_state *state, nir_function_impl *impl)
80 {
81 nir_builder b;
82 nir_builder_init(&b, impl);
83
84 if (state->shader->info.stage == MESA_SHADER_GEOMETRY) {
85 /* For geometry shaders, we have to emit the output copies right
86 * before each EmitVertex call.
87 */
88 nir_foreach_block(block, impl) {
89 nir_foreach_instr(instr, block) {
90 if (instr->type != nir_instr_type_intrinsic)
91 continue;
92
93 nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr);
94 if (intrin->intrinsic == nir_intrinsic_emit_vertex ||
95 intrin->intrinsic == nir_intrinsic_emit_vertex_with_counter) {
96 b.cursor = nir_before_instr(&intrin->instr);
97 emit_copies(&b, &state->new_outputs, &state->old_outputs);
98 }
99 }
100 }
101 } else if (impl == state->entrypoint) {
102 b.cursor = nir_before_block(nir_start_block(impl));
103 emit_copies(&b, &state->old_outputs, &state->new_outputs);
104
105 /* For all other shader types, we need to do the copies right before
106 * the jumps to the end block.
107 */
108 set_foreach(impl->end_block->predecessors, block_entry) {
109 struct nir_block *block = (void *)block_entry->key;
110 b.cursor = nir_after_block_before_jump(block);
111 emit_copies(&b, &state->new_outputs, &state->old_outputs);
112 }
113 }
114 }
115
116 /* For fragment shader inputs, when we lower to temporaries we'll invalidate
117 * interpolateAt*() because now they'll be pointing to the temporary instead
118 * of the actual variable. Since the caller presumably doesn't support
119 * indirect indexing of inputs, we'll need to lower something like:
120 *
121 * in vec4 foo[3];
122 *
123 * ... = interpolateAtCentroid(foo[i]);
124 *
125 * to a sequence of interpolations that store to our temporary, then a
126 * load at the end:
127 *
128 * in vec4 foo[3];
129 * vec4 foo_tmp[3];
130 *
131 * foo_tmp[0] = interpolateAtCentroid(foo[0]);
132 * foo_tmp[1] = interpolateAtCentroid(foo[1]);
133 * ... = foo_tmp[i];
134 */
135
136 /*
137 * Recursively emit the interpolation instructions. Here old_interp_deref
138 * refers to foo[i], temp_deref is foo_tmp[0/1], and new_interp_deref is
139 * foo[0/1].
140 */
141
142 static void
emit_interp(nir_builder * b,nir_deref_instr ** old_interp_deref,nir_deref_instr * temp_deref,nir_deref_instr * new_interp_deref,nir_intrinsic_instr * interp)143 emit_interp(nir_builder *b, nir_deref_instr **old_interp_deref,
144 nir_deref_instr *temp_deref, nir_deref_instr *new_interp_deref,
145 nir_intrinsic_instr *interp)
146 {
147 while (*old_interp_deref) {
148 switch ((*old_interp_deref)->deref_type) {
149 case nir_deref_type_struct:
150 temp_deref =
151 nir_build_deref_struct(b, temp_deref,
152 (*old_interp_deref)->strct.index);
153 new_interp_deref =
154 nir_build_deref_struct(b, new_interp_deref,
155 (*old_interp_deref)->strct.index);
156 break;
157 case nir_deref_type_array:
158 if (nir_src_is_const((*old_interp_deref)->arr.index)) {
159 temp_deref =
160 nir_build_deref_array(b, temp_deref,
161 (*old_interp_deref)->arr.index.ssa);
162 new_interp_deref =
163 nir_build_deref_array(b, new_interp_deref,
164 (*old_interp_deref)->arr.index.ssa);
165 break;
166 } else {
167 /* We have an indirect deref, so we have to emit interpolations
168 * for every index. Recurse in case we have an array of arrays.
169 */
170 unsigned length = glsl_get_length(temp_deref->type);
171 for (unsigned i = 0; i < length; i++) {
172 nir_deref_instr *new_temp =
173 nir_build_deref_array_imm(b, temp_deref, i);
174 nir_deref_instr *new_interp =
175 nir_build_deref_array_imm(b, new_interp_deref, i);
176
177 emit_interp(b, old_interp_deref + 1, new_temp, new_interp,
178 interp);
179 }
180
181 return;
182 }
183
184 case nir_deref_type_var:
185 case nir_deref_type_array_wildcard:
186 case nir_deref_type_ptr_as_array:
187 case nir_deref_type_cast:
188 unreachable("bad deref type");
189 }
190
191 old_interp_deref++;
192 }
193
194 /* Now that we've constructed a fully-qualified deref with all the indirect
195 * derefs replaced with direct ones, it's time to actually emit the new
196 * interpolation instruction.
197 */
198
199 nir_intrinsic_instr *new_interp =
200 nir_intrinsic_instr_create(b->shader, interp->intrinsic);
201
202 new_interp->src[0] = nir_src_for_ssa(&new_interp_deref->dest.ssa);
203 if (interp->intrinsic == nir_intrinsic_interp_deref_at_sample ||
204 interp->intrinsic == nir_intrinsic_interp_deref_at_offset ||
205 interp->intrinsic == nir_intrinsic_interp_deref_at_vertex) {
206 new_interp->src[1] = interp->src[1];
207 }
208
209 new_interp->num_components = interp->num_components;
210 nir_ssa_dest_init(&new_interp->instr, &new_interp->dest,
211 interp->dest.ssa.num_components,
212 interp->dest.ssa.bit_size, NULL);
213
214 nir_builder_instr_insert(b, &new_interp->instr);
215 nir_store_deref(b, temp_deref, &new_interp->dest.ssa,
216 (1 << interp->dest.ssa.num_components) - 1);
217 }
218
219 static void
fixup_interpolation_instr(struct lower_io_state * state,nir_intrinsic_instr * interp,nir_builder * b)220 fixup_interpolation_instr(struct lower_io_state *state,
221 nir_intrinsic_instr *interp, nir_builder *b)
222 {
223 nir_deref_path interp_path;
224 nir_deref_path_init(&interp_path, nir_src_as_deref(interp->src[0]), NULL);
225
226 b->cursor = nir_before_instr(&interp->instr);
227
228 /* The original interpolation instruction should contain a deref path
229 * starting with the original variable, which is now the temporary.
230 */
231 nir_deref_instr *temp_root = interp_path.path[0];
232
233 /* Fish out the newly-created input variable. */
234 assert(temp_root->deref_type == nir_deref_type_var);
235 struct hash_entry *entry = _mesa_hash_table_search(state->input_map,
236 temp_root->var);
237 assert(entry);
238 nir_variable *input = entry->data;
239 nir_deref_instr *input_root = nir_build_deref_var(b, input);
240
241 /* Emit the interpolation instructions. */
242 emit_interp(b, interp_path.path + 1, temp_root, input_root, interp);
243
244 /* Now the temporary contains the interpolation results, and we can just
245 * load from it. We can reuse the original deref, since it points to the
246 * correct part of the temporary.
247 */
248 nir_ssa_def *load = nir_load_deref(b, nir_src_as_deref(interp->src[0]));
249 nir_ssa_def_rewrite_uses(&interp->dest.ssa, nir_src_for_ssa(load));
250 nir_instr_remove(&interp->instr);
251
252 nir_deref_path_finish(&interp_path);
253 }
254
255 static void
fixup_interpolation(struct lower_io_state * state,nir_function_impl * impl,nir_builder * b)256 fixup_interpolation(struct lower_io_state *state, nir_function_impl *impl,
257 nir_builder *b)
258 {
259 nir_foreach_block(block, impl) {
260 nir_foreach_instr_safe(instr, block) {
261 if (instr->type != nir_instr_type_intrinsic)
262 continue;
263
264 nir_intrinsic_instr *interp = nir_instr_as_intrinsic(instr);
265
266 if (interp->intrinsic == nir_intrinsic_interp_deref_at_centroid ||
267 interp->intrinsic == nir_intrinsic_interp_deref_at_sample ||
268 interp->intrinsic == nir_intrinsic_interp_deref_at_offset ||
269 interp->intrinsic == nir_intrinsic_interp_deref_at_vertex) {
270 fixup_interpolation_instr(state, interp, b);
271 }
272 }
273 }
274 }
275
276 static void
emit_input_copies_impl(struct lower_io_state * state,nir_function_impl * impl)277 emit_input_copies_impl(struct lower_io_state *state, nir_function_impl *impl)
278 {
279 if (impl == state->entrypoint) {
280 nir_builder b;
281 nir_builder_init(&b, impl);
282 b.cursor = nir_before_block(nir_start_block(impl));
283 emit_copies(&b, &state->old_inputs, &state->new_inputs);
284 if (state->shader->info.stage == MESA_SHADER_FRAGMENT)
285 fixup_interpolation(state, impl, &b);
286 }
287 }
288
289 static nir_variable *
create_shadow_temp(struct lower_io_state * state,nir_variable * var)290 create_shadow_temp(struct lower_io_state *state, nir_variable *var)
291 {
292 nir_variable *nvar = ralloc(state->shader, nir_variable);
293 memcpy(nvar, var, sizeof *nvar);
294 nvar->data.cannot_coalesce = true;
295
296 /* The original is now the temporary */
297 nir_variable *temp = var;
298
299 /* Reparent the name to the new variable */
300 ralloc_steal(nvar, nvar->name);
301
302 assert(nvar->constant_initializer == NULL && nvar->pointer_initializer == NULL);
303
304 /* Give the original a new name with @<mode>-temp appended */
305 const char *mode = (temp->data.mode == nir_var_shader_in) ? "in" : "out";
306 temp->name = ralloc_asprintf(var, "%s@%s-temp", mode, nvar->name);
307 temp->data.mode = nir_var_shader_temp;
308 temp->data.read_only = false;
309 temp->data.fb_fetch_output = false;
310 temp->data.compact = false;
311
312 return nvar;
313 }
314
315 static void
move_variables_to_list(nir_shader * shader,nir_variable_mode mode,struct exec_list * dst_list)316 move_variables_to_list(nir_shader *shader, nir_variable_mode mode,
317 struct exec_list *dst_list)
318 {
319 nir_foreach_variable_with_modes_safe(var, shader, mode) {
320 exec_node_remove(&var->node);
321 exec_list_push_tail(dst_list, &var->node);
322 }
323 }
324
325 void
nir_lower_io_to_temporaries(nir_shader * shader,nir_function_impl * entrypoint,bool outputs,bool inputs)326 nir_lower_io_to_temporaries(nir_shader *shader, nir_function_impl *entrypoint,
327 bool outputs, bool inputs)
328 {
329 struct lower_io_state state;
330
331 if (shader->info.stage == MESA_SHADER_TESS_CTRL)
332 return;
333
334 state.shader = shader;
335 state.entrypoint = entrypoint;
336 state.input_map = _mesa_pointer_hash_table_create(NULL);
337
338 exec_list_make_empty(&state.old_inputs);
339 if (inputs)
340 move_variables_to_list(shader, nir_var_shader_in, &state.old_inputs);
341
342 exec_list_make_empty(&state.old_outputs);
343 if (outputs)
344 move_variables_to_list(shader, nir_var_shader_out, &state.old_outputs);
345
346 exec_list_make_empty(&state.new_inputs);
347 exec_list_make_empty(&state.new_outputs);
348
349 /* Walk over all of the outputs turn each output into a temporary and
350 * make a new variable for the actual output.
351 */
352 nir_foreach_variable_in_list(var, &state.old_outputs) {
353 nir_variable *output = create_shadow_temp(&state, var);
354 exec_list_push_tail(&state.new_outputs, &output->node);
355 }
356
357 /* and same for inputs: */
358 nir_foreach_variable_in_list(var, &state.old_inputs) {
359 nir_variable *input = create_shadow_temp(&state, var);
360 exec_list_push_tail(&state.new_inputs, &input->node);
361 _mesa_hash_table_insert(state.input_map, var, input);
362 }
363
364 nir_foreach_function(function, shader) {
365 if (function->impl == NULL)
366 continue;
367
368 if (inputs)
369 emit_input_copies_impl(&state, function->impl);
370
371 if (outputs)
372 emit_output_copies_impl(&state, function->impl);
373
374 nir_metadata_preserve(function->impl, nir_metadata_block_index |
375 nir_metadata_dominance);
376 }
377
378 exec_list_append(&shader->variables, &state.old_inputs);
379 exec_list_append(&shader->variables, &state.old_outputs);
380 exec_list_append(&shader->variables, &state.new_inputs);
381 exec_list_append(&shader->variables, &state.new_outputs);
382
383 nir_fixup_deref_modes(shader);
384
385 _mesa_hash_table_destroy(state.input_map, NULL);
386 }
387