1 /*
2 * Copyright © 2016 Red Hat
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 FROM,
20 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
21 * SOFTWARE.
22 */
23
24 /* Lowering pass that lowers accesses to built-in uniform variables.
25 * Built-in uniforms are not necessarily packed the same way that
26 * normal uniform structs are, for example:
27 *
28 * struct gl_FogParameters {
29 * vec4 color;
30 * float density;
31 * float start;
32 * float end;
33 * float scale;
34 * };
35 *
36 * is packed into vec4[2], whereas the same struct would be packed
37 * (by gallium), as vec4[5] if it where not built-in. Because of
38 * this, we need to replace (for example) access like:
39 *
40 * vec1 ssa_1 = intrinsic load_var () (gl_Fog.start) ()
41 *
42 * with:
43 *
44 * vec4 ssa_2 = intrinsic load_var () (fog.params) ()
45 * vec1 ssa_1 = ssa_2.y
46 *
47 * with appropriate substitutions in the uniform variables list:
48 *
49 * decl_var uniform INTERP_MODE_NONE gl_FogParameters gl_Fog (0, 0)
50 *
51 * would become:
52 *
53 * decl_var uniform INTERP_MODE_NONE vec4 state.fog.color (0, 0)
54 * decl_var uniform INTERP_MODE_NONE vec4 state.fog.params (0, 1)
55 *
56 * See in particular 'struct gl_builtin_uniform_element'.
57 */
58
59 #include "compiler/nir/nir.h"
60 #include "compiler/nir/nir_builder.h"
61 #include "compiler/nir/nir_deref.h"
62 #include "st_nir.h"
63 #include "compiler/glsl/ir.h"
64 #include "uniforms.h"
65 #include "program/prog_instruction.h"
66
67 typedef struct {
68 nir_shader *shader;
69 nir_builder builder;
70 void *mem_ctx;
71 } lower_builtin_state;
72
73 static const struct gl_builtin_uniform_element *
get_element(const struct gl_builtin_uniform_desc * desc,nir_deref_path * path)74 get_element(const struct gl_builtin_uniform_desc *desc, nir_deref_path *path)
75 {
76 int idx = 1;
77
78 assert(path->path[0]->deref_type == nir_deref_type_var);
79
80 if ((desc->num_elements == 1) && (desc->elements[0].field == NULL))
81 return NULL;
82
83 /* we handle arrays in get_variable(): */
84 if (path->path[idx]->deref_type == nir_deref_type_array)
85 idx++;
86
87 /* don't need to deal w/ non-struct or array of non-struct: */
88 if (!path->path[idx])
89 return NULL;
90
91 if (path->path[idx]->deref_type != nir_deref_type_struct)
92 return NULL;
93
94 assert(path->path[idx]->strct.index < desc->num_elements);
95
96 return &desc->elements[path->path[idx]->strct.index ];
97 }
98
99 static nir_variable *
get_variable(lower_builtin_state * state,nir_deref_path * path,const struct gl_builtin_uniform_element * element)100 get_variable(lower_builtin_state *state, nir_deref_path *path,
101 const struct gl_builtin_uniform_element *element)
102 {
103 nir_shader *shader = state->shader;
104 gl_state_index16 tokens[STATE_LENGTH];
105 int idx = 1;
106
107 memcpy(tokens, element->tokens, sizeof(tokens));
108
109 if (path->path[idx]->deref_type == nir_deref_type_array) {
110 /* we need to fixup the array index slot: */
111 switch (tokens[0]) {
112 case STATE_MODELVIEW_MATRIX:
113 case STATE_MODELVIEW_MATRIX_INVERSE:
114 case STATE_MODELVIEW_MATRIX_TRANSPOSE:
115 case STATE_MODELVIEW_MATRIX_INVTRANS:
116 case STATE_PROJECTION_MATRIX:
117 case STATE_PROJECTION_MATRIX_INVERSE:
118 case STATE_PROJECTION_MATRIX_TRANSPOSE:
119 case STATE_PROJECTION_MATRIX_INVTRANS:
120 case STATE_MVP_MATRIX:
121 case STATE_MVP_MATRIX_INVERSE:
122 case STATE_MVP_MATRIX_TRANSPOSE:
123 case STATE_MVP_MATRIX_INVTRANS:
124 case STATE_TEXTURE_MATRIX:
125 case STATE_TEXTURE_MATRIX_INVERSE:
126 case STATE_TEXTURE_MATRIX_TRANSPOSE:
127 case STATE_TEXTURE_MATRIX_INVTRANS:
128 case STATE_PROGRAM_MATRIX:
129 case STATE_PROGRAM_MATRIX_INVERSE:
130 case STATE_PROGRAM_MATRIX_TRANSPOSE:
131 case STATE_PROGRAM_MATRIX_INVTRANS:
132 case STATE_LIGHT:
133 case STATE_LIGHTPROD:
134 case STATE_TEXGEN:
135 case STATE_TEXENV_COLOR:
136 case STATE_CLIPPLANE:
137 tokens[1] = nir_src_as_uint(path->path[idx]->arr.index);
138 break;
139 }
140 }
141
142 char *name = _mesa_program_state_string(tokens);
143
144 nir_foreach_uniform_variable(var, shader) {
145 if (strcmp(var->name, name) == 0) {
146 free(name);
147 return var;
148 }
149 }
150
151 /* variable doesn't exist yet, so create it: */
152 nir_variable *var =
153 nir_variable_create(shader, nir_var_uniform, glsl_vec4_type(), name);
154
155 var->num_state_slots = 1;
156 var->state_slots = rzalloc_array(var, nir_state_slot, 1);
157 memcpy(var->state_slots[0].tokens, tokens,
158 sizeof(var->state_slots[0].tokens));
159
160 free(name);
161
162 return var;
163 }
164
165 static bool
lower_builtin_block(lower_builtin_state * state,nir_block * block)166 lower_builtin_block(lower_builtin_state *state, nir_block *block)
167 {
168 nir_builder *b = &state->builder;
169 bool progress = false;
170
171 nir_foreach_instr_safe(instr, block) {
172 if (instr->type != nir_instr_type_intrinsic)
173 continue;
174
175 nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr);
176
177 if (intrin->intrinsic != nir_intrinsic_load_deref)
178 continue;
179
180 nir_variable *var =
181 nir_deref_instr_get_variable(nir_src_as_deref(intrin->src[0]));
182 if (var->data.mode != nir_var_uniform)
183 continue;
184
185 /* built-in's will always start with "gl_" */
186 if (strncmp(var->name, "gl_", 3) != 0)
187 continue;
188
189 const struct gl_builtin_uniform_desc *desc =
190 _mesa_glsl_get_builtin_uniform_desc(var->name);
191
192 /* if no descriptor, it isn't something we need to handle specially: */
193 if (!desc)
194 continue;
195
196 nir_deref_path path;
197 nir_deref_path_init(&path, nir_src_as_deref(intrin->src[0]), NULL);
198
199 const struct gl_builtin_uniform_element *element = get_element(desc, &path);
200
201 /* matrix elements (array_deref) do not need special handling: */
202 if (!element) {
203 nir_deref_path_finish(&path);
204 continue;
205 }
206
207 /* remove existing var from uniform list: */
208 exec_node_remove(&var->node);
209 /* the _self_link() ensures we can remove multiple times, rather than
210 * trying to keep track of what we have already removed:
211 */
212 exec_node_self_link(&var->node);
213
214 nir_variable *new_var = get_variable(state, &path, element);
215 nir_deref_path_finish(&path);
216
217 b->cursor = nir_before_instr(instr);
218
219 nir_ssa_def *def = nir_load_var(b, new_var);
220
221 /* swizzle the result: */
222 unsigned swiz[NIR_MAX_VEC_COMPONENTS] = {0};
223 for (unsigned i = 0; i < 4; i++) {
224 swiz[i] = GET_SWZ(element->swizzle, i);
225 assert(swiz[i] <= SWIZZLE_W);
226 }
227 def = nir_swizzle(b, def, swiz, intrin->num_components);
228
229 /* and rewrite uses of original instruction: */
230 assert(intrin->dest.is_ssa);
231 nir_ssa_def_rewrite_uses(&intrin->dest.ssa, def);
232
233 /* at this point intrin should be unused. We need to remove it
234 * (rather than waiting for DCE pass) to avoid dangling reference
235 * to remove'd var. And we have to remove the original uniform
236 * var since we don't want it to get uniform space allocated.
237 */
238 nir_instr_remove(&intrin->instr);
239
240 progress = true;
241 }
242
243 return progress;
244 }
245
246 static void
lower_builtin_impl(lower_builtin_state * state,nir_function_impl * impl)247 lower_builtin_impl(lower_builtin_state *state, nir_function_impl *impl)
248 {
249 nir_builder_init(&state->builder, impl);
250 state->mem_ctx = ralloc_parent(impl);
251
252 bool progress = false;
253 nir_foreach_block(block, impl) {
254 progress |= lower_builtin_block(state, block);
255 }
256
257 if (progress)
258 nir_remove_dead_derefs_impl(impl);
259
260 nir_metadata_preserve(impl, nir_metadata_block_index |
261 nir_metadata_dominance);
262 }
263
264 void
st_nir_lower_builtin(nir_shader * shader)265 st_nir_lower_builtin(nir_shader *shader)
266 {
267 lower_builtin_state state;
268 state.shader = shader;
269 nir_foreach_function(function, shader) {
270 if (function->impl)
271 lower_builtin_impl(&state, function->impl);
272 }
273 }
274