1 /*
2 * Copyright © 2016 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 * Authors:
24 * Jason Ekstrand (jason@jlekstrand.net)
25 *
26 */
27
28 /*
29 * This lowering pass converts references to variables with loads/stores to
30 * scratch space based on a few configurable parameters.
31 */
32
33 #include "nir.h"
34 #include "nir_builder.h"
35 #include "nir_deref.h"
36
37 static void
lower_load_store(nir_builder * b,nir_intrinsic_instr * intrin,glsl_type_size_align_func size_align)38 lower_load_store(nir_builder *b,
39 nir_intrinsic_instr *intrin,
40 glsl_type_size_align_func size_align)
41 {
42 b->cursor = nir_before_instr(&intrin->instr);
43
44 nir_deref_instr *deref = nir_src_as_deref(intrin->src[0]);
45 nir_variable *var = nir_deref_instr_get_variable(deref);
46
47 nir_ssa_def *offset =
48 nir_iadd_imm(b, nir_build_deref_offset(b, deref, size_align),
49 var->data.location);
50
51 unsigned align, UNUSED size;
52 size_align(deref->type, &size, &align);
53
54 if (intrin->intrinsic == nir_intrinsic_load_deref) {
55 unsigned bit_size = intrin->dest.ssa.bit_size;
56 nir_ssa_def *value = nir_load_scratch(
57 b, intrin->num_components, bit_size == 1 ? 32 : bit_size, offset, .align_mul=align);
58 if (bit_size == 1)
59 value = nir_b2b1(b, value);
60
61 nir_ssa_def_rewrite_uses(&intrin->dest.ssa, value);
62 } else {
63 assert(intrin->intrinsic == nir_intrinsic_store_deref);
64
65 assert(intrin->src[1].is_ssa);
66 nir_ssa_def *value = intrin->src[1].ssa;
67 if (value->bit_size == 1)
68 value = nir_b2b32(b, value);
69
70 nir_store_scratch(b, value, offset, .align_mul=align,
71 .write_mask=nir_intrinsic_write_mask(intrin));
72 }
73
74 nir_instr_remove(&intrin->instr);
75 nir_deref_instr_remove_if_unused(deref);
76 }
77
only_used_for_load_store(nir_deref_instr * deref)78 static bool only_used_for_load_store(nir_deref_instr *deref)
79 {
80 nir_foreach_use(src, &deref->dest.ssa) {
81 if (!src->parent_instr)
82 return false;
83 if (src->parent_instr->type == nir_instr_type_deref) {
84 if (!only_used_for_load_store(nir_instr_as_deref(src->parent_instr)))
85 return false;
86 } else if (src->parent_instr->type != nir_instr_type_intrinsic) {
87 return false;
88 } else {
89 nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(src->parent_instr);
90 if (intrin->intrinsic != nir_intrinsic_load_deref &&
91 intrin->intrinsic != nir_intrinsic_store_deref)
92 return false;
93 }
94 }
95 return true;
96 }
97
98 bool
nir_lower_vars_to_scratch(nir_shader * shader,nir_variable_mode modes,int size_threshold,glsl_type_size_align_func size_align)99 nir_lower_vars_to_scratch(nir_shader *shader,
100 nir_variable_mode modes,
101 int size_threshold,
102 glsl_type_size_align_func size_align)
103 {
104 struct set *set = _mesa_pointer_set_create(NULL);
105
106 /* First, we walk the instructions and flag any variables we want to lower
107 * by removing them from their respective list and setting the mode to 0.
108 */
109 nir_foreach_function(function, shader) {
110 nir_foreach_block(block, function->impl) {
111 nir_foreach_instr(instr, block) {
112 if (instr->type != nir_instr_type_intrinsic)
113 continue;
114
115 nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr);
116 if (intrin->intrinsic != nir_intrinsic_load_deref &&
117 intrin->intrinsic != nir_intrinsic_store_deref)
118 continue;
119
120 nir_deref_instr *deref = nir_src_as_deref(intrin->src[0]);
121 if (!nir_deref_mode_is_one_of(deref, modes))
122 continue;
123
124 if (!nir_deref_instr_has_indirect(nir_src_as_deref(intrin->src[0])))
125 continue;
126
127 nir_variable *var = nir_deref_instr_get_variable(deref);
128 if (!var)
129 continue;
130
131 /* We set var->mode to 0 to indicate that a variable will be moved
132 * to scratch. Don't assign a scratch location twice.
133 */
134 if (var->data.mode == 0)
135 continue;
136
137 unsigned var_size, var_align;
138 size_align(var->type, &var_size, &var_align);
139 if (var_size <= size_threshold)
140 continue;
141
142 _mesa_set_add(set, var);
143 }
144 }
145 }
146
147 if (set->entries == 0) {
148 _mesa_set_destroy(set, NULL);
149 return false;
150 }
151
152 nir_foreach_function(function, shader) {
153 nir_foreach_block(block, function->impl) {
154 nir_foreach_instr(instr, block) {
155 if (instr->type != nir_instr_type_deref)
156 continue;
157
158 nir_deref_instr *deref = nir_instr_as_deref(instr);
159 if (deref->deref_type != nir_deref_type_var)
160 continue;
161
162 struct set_entry *entry = _mesa_set_search(set, deref->var);
163 if (!entry)
164 continue;
165
166 if (!only_used_for_load_store(deref))
167 _mesa_set_remove(set, entry);
168 }
169 }
170 }
171
172 set_foreach(set, entry) {
173 nir_variable* var = (void*)entry->key;
174
175 /* Remove it from its list */
176 exec_node_remove(&var->node);
177 /* Invalid mode used to flag "moving to scratch" */
178 var->data.mode = 0;
179
180 /* We don't allocate space here as iteration in this loop is
181 * non-deterministic due to the nir_variable pointers. */
182 var->data.location = INT_MAX;
183 }
184
185 bool progress = false;
186 nir_foreach_function(function, shader) {
187 if (!function->impl)
188 continue;
189
190 nir_builder build;
191 nir_builder_init(&build, function->impl);
192
193 bool impl_progress = false;
194 nir_foreach_block(block, function->impl) {
195 nir_foreach_instr_safe(instr, block) {
196 if (instr->type != nir_instr_type_intrinsic)
197 continue;
198
199 nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr);
200 if (intrin->intrinsic != nir_intrinsic_load_deref &&
201 intrin->intrinsic != nir_intrinsic_store_deref)
202 continue;
203
204 nir_variable *var = nir_intrinsic_get_var(intrin, 0);
205 /* Variables flagged for lowering above have mode == 0 */
206 if (!var || var->data.mode)
207 continue;
208
209 if (var->data.location == INT_MAX) {
210 unsigned var_size, var_align;
211 size_align(var->type, &var_size, &var_align);
212
213 var->data.location = ALIGN_POT(shader->scratch_size, var_align);
214 shader->scratch_size = var->data.location + var_size;
215 }
216
217 lower_load_store(&build, intrin, size_align);
218 impl_progress = true;
219 }
220 }
221
222 if (impl_progress) {
223 progress = true;
224 nir_metadata_preserve(function->impl, nir_metadata_block_index |
225 nir_metadata_dominance);
226 } else {
227 nir_metadata_preserve(function->impl, nir_metadata_all);
228 }
229 }
230
231 _mesa_set_destroy(set, NULL);
232
233 return progress;
234 }
235