1 /*
2 * Copyright © 2020 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 #include "nir_builder.h"
25
26 static bool
opt_memcpy_deref_cast(nir_intrinsic_instr * cpy,nir_src * deref_src)27 opt_memcpy_deref_cast(nir_intrinsic_instr *cpy, nir_src *deref_src)
28 {
29 assert(cpy->intrinsic == nir_intrinsic_memcpy_deref);
30
31 nir_deref_instr *cast = nir_src_as_deref(*deref_src);
32 if (cast == NULL || cast->deref_type != nir_deref_type_cast)
33 return false;
34
35 /* We always have to replace the source with a deref, not a bare uint
36 * pointer. If it's the first deref in the chain, bail.
37 */
38 nir_deref_instr *parent = nir_src_as_deref(cast->parent);
39 if (parent == NULL)
40 return false;
41
42 /* If it has useful alignment information, we want to keep that */
43 if (cast->cast.align_mul > 0)
44 return false;
45
46 /* Casts to uint8 or int8 never do us any good; get rid of them */
47 if (cast->type == glsl_int8_t_type() ||
48 cast->type == glsl_uint8_t_type()) {
49 nir_src_rewrite(deref_src, &parent->def);
50 return true;
51 }
52
53 int64_t parent_type_size = glsl_get_explicit_size(parent->type, false);
54 if (parent_type_size < 0)
55 return false;
56
57 if (!nir_src_is_const(cpy->src[2]))
58 return false;
59
60 /* We don't want to get rid of the cast if the resulting type would be
61 * smaller than the amount of data we're copying.
62 */
63 if (nir_src_as_uint(cpy->src[2]) < (uint64_t)parent_type_size)
64 return false;
65
66 nir_src_rewrite(deref_src, &parent->def);
67 return true;
68 }
69
70 static bool
type_is_tightly_packed(const struct glsl_type * type,unsigned * size_out)71 type_is_tightly_packed(const struct glsl_type *type, unsigned *size_out)
72 {
73 unsigned size = 0;
74 if (glsl_type_is_struct_or_ifc(type)) {
75 unsigned num_fields = glsl_get_length(type);
76 for (unsigned i = 0; i < num_fields; i++) {
77 const struct glsl_struct_field *field =
78 glsl_get_struct_field_data(type, i);
79
80 if (field->offset < 0 || field->offset != size)
81 return false;
82
83 unsigned field_size;
84 if (!type_is_tightly_packed(field->type, &field_size))
85 return false;
86
87 size = field->offset + field_size;
88 }
89 } else if (glsl_type_is_array_or_matrix(type)) {
90 if (glsl_type_is_unsized_array(type))
91 return false;
92
93 unsigned stride = glsl_get_explicit_stride(type);
94 if (stride == 0)
95 return false;
96
97 const struct glsl_type *elem_type = glsl_get_array_element(type);
98
99 unsigned elem_size;
100 if (!type_is_tightly_packed(elem_type, &elem_size))
101 return false;
102
103 if (elem_size != stride)
104 return false;
105
106 size = stride * glsl_get_length(type);
107 } else {
108 assert(glsl_type_is_vector_or_scalar(type));
109 if (glsl_get_explicit_stride(type) > 0)
110 return false;
111
112 if (glsl_type_is_boolean(type))
113 return false;
114
115 size = glsl_get_explicit_size(type, false);
116 }
117
118 if (size_out)
119 *size_out = size;
120 return true;
121 }
122
123 static bool
try_lower_memcpy(nir_builder * b,nir_intrinsic_instr * cpy,struct set * complex_vars)124 try_lower_memcpy(nir_builder *b, nir_intrinsic_instr *cpy,
125 struct set *complex_vars)
126 {
127 nir_deref_instr *dst = nir_src_as_deref(cpy->src[0]);
128 nir_deref_instr *src = nir_src_as_deref(cpy->src[1]);
129
130 /* A self-copy can always be eliminated */
131 if (dst == src) {
132 nir_instr_remove(&cpy->instr);
133 return true;
134 }
135
136 if (!nir_src_is_const(cpy->src[2]))
137 return false;
138
139 uint64_t size = nir_src_as_uint(cpy->src[2]);
140 if (size == 0) {
141 nir_instr_remove(&cpy->instr);
142 return true;
143 }
144
145 if (glsl_type_is_vector_or_scalar(src->type) &&
146 glsl_type_is_vector_or_scalar(dst->type) &&
147 glsl_get_explicit_size(dst->type, false) == size &&
148 glsl_get_explicit_size(src->type, false) == size) {
149 b->cursor = nir_instr_remove(&cpy->instr);
150 nir_def *data =
151 nir_load_deref_with_access(b, src, nir_intrinsic_src_access(cpy));
152 data = nir_bitcast_vector(b, data, glsl_get_bit_size(dst->type));
153 assert(data->num_components == glsl_get_vector_elements(dst->type));
154 nir_store_deref_with_access(b, dst, data, ~0 /* write mask */,
155 nir_intrinsic_dst_access(cpy));
156 return true;
157 }
158
159 unsigned type_size;
160 if (dst->type == src->type &&
161 type_is_tightly_packed(dst->type, &type_size) &&
162 type_size == size) {
163 b->cursor = nir_instr_remove(&cpy->instr);
164 nir_copy_deref_with_access(b, dst, src,
165 nir_intrinsic_dst_access(cpy),
166 nir_intrinsic_src_access(cpy));
167 return true;
168 }
169
170 /* If one of the two types is tightly packed and happens to equal the
171 * memcpy size, then we can get the memcpy by casting to that type and
172 * doing a deref copy.
173 *
174 * However, if we blindly apply this logic, we may end up with extra casts
175 * where we don't want them. The whole point of converting memcpy to
176 * copy_deref is in the hopes that nir_opt_copy_prop_vars or
177 * nir_lower_vars_to_ssa will get rid of the copy and those passes don't
178 * handle casts well. Heuristically, only do this optimization if the
179 * tightly packed type is on a deref with nir_var_function_temp so we stick
180 * the cast on the other mode.
181 */
182 if (dst->modes == nir_var_function_temp &&
183 type_is_tightly_packed(dst->type, &type_size) &&
184 type_size == size) {
185 b->cursor = nir_instr_remove(&cpy->instr);
186 src = nir_build_deref_cast(b, &src->def,
187 src->modes, dst->type, 0);
188 nir_copy_deref_with_access(b, dst, src,
189 nir_intrinsic_dst_access(cpy),
190 nir_intrinsic_src_access(cpy));
191 return true;
192 }
193
194 /* If we can get at the variable AND the only complex use of that variable
195 * is as a memcpy destination, then we don't have to care about any empty
196 * space in the variable. In particular, we know that the variable is never
197 * cast to any other type and it's never used as a memcpy source so nothing
198 * can see any padding bytes. This holds even if some other memcpy only
199 * writes to part of the variable.
200 */
201 if (dst->deref_type == nir_deref_type_var &&
202 dst->modes == nir_var_function_temp &&
203 _mesa_set_search(complex_vars, dst->var) == NULL &&
204 glsl_get_explicit_size(dst->type, false) <= size) {
205 b->cursor = nir_instr_remove(&cpy->instr);
206 src = nir_build_deref_cast(b, &src->def,
207 src->modes, dst->type, 0);
208 nir_copy_deref_with_access(b, dst, src,
209 nir_intrinsic_dst_access(cpy),
210 nir_intrinsic_src_access(cpy));
211 return true;
212 }
213
214 if (src->modes == nir_var_function_temp &&
215 type_is_tightly_packed(src->type, &type_size) &&
216 type_size == size) {
217 b->cursor = nir_instr_remove(&cpy->instr);
218 dst = nir_build_deref_cast(b, &dst->def,
219 dst->modes, src->type, 0);
220 nir_copy_deref_with_access(b, dst, src,
221 nir_intrinsic_dst_access(cpy),
222 nir_intrinsic_src_access(cpy));
223 return true;
224 }
225
226 return false;
227 }
228
229 static bool
opt_memcpy_impl(nir_function_impl * impl)230 opt_memcpy_impl(nir_function_impl *impl)
231 {
232 bool progress = false;
233
234 nir_builder b = nir_builder_create(impl);
235
236 struct set *complex_vars = _mesa_pointer_set_create(NULL);
237
238 nir_foreach_block(block, impl) {
239 nir_foreach_instr(instr, block) {
240 if (instr->type != nir_instr_type_deref)
241 continue;
242
243 nir_deref_instr *deref = nir_instr_as_deref(instr);
244 if (deref->deref_type != nir_deref_type_var)
245 continue;
246
247 nir_deref_instr_has_complex_use_options opts =
248 nir_deref_instr_has_complex_use_allow_memcpy_dst;
249 if (nir_deref_instr_has_complex_use(deref, opts))
250 _mesa_set_add(complex_vars, deref->var);
251 }
252 }
253
254 nir_foreach_block(block, impl) {
255 nir_foreach_instr_safe(instr, block) {
256 if (instr->type != nir_instr_type_intrinsic)
257 continue;
258
259 nir_intrinsic_instr *cpy = nir_instr_as_intrinsic(instr);
260 if (cpy->intrinsic != nir_intrinsic_memcpy_deref)
261 continue;
262
263 while (opt_memcpy_deref_cast(cpy, &cpy->src[0]))
264 progress = true;
265 while (opt_memcpy_deref_cast(cpy, &cpy->src[1]))
266 progress = true;
267
268 if (try_lower_memcpy(&b, cpy, complex_vars)) {
269 progress = true;
270 continue;
271 }
272 }
273 }
274
275 _mesa_set_destroy(complex_vars, NULL);
276
277 if (progress) {
278 nir_metadata_preserve(impl, nir_metadata_control_flow);
279 } else {
280 nir_metadata_preserve(impl, nir_metadata_all);
281 }
282
283 return progress;
284 }
285
286 bool
nir_opt_memcpy(nir_shader * shader)287 nir_opt_memcpy(nir_shader *shader)
288 {
289 bool progress = false;
290
291 nir_foreach_function_impl(impl, shader) {
292 if (opt_memcpy_impl(impl))
293 progress = true;
294 }
295
296 return progress;
297 }
298