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_instr_rewrite_src(&cpy->instr, deref_src,
50 nir_src_for_ssa(&parent->dest.ssa));
51 return true;
52 }
53
54 int64_t parent_type_size = glsl_get_explicit_size(parent->type, false);
55 if (parent_type_size < 0)
56 return false;
57
58 if (!nir_src_is_const(cpy->src[2]))
59 return false;
60
61 /* We don't want to get rid of the cast if the resulting type would be
62 * smaller than the amount of data we're copying.
63 */
64 if (nir_src_as_uint(cpy->src[2]) < (uint64_t)parent_type_size)
65 return false;
66
67 nir_instr_rewrite_src(&cpy->instr, deref_src,
68 nir_src_for_ssa(&parent->dest.ssa));
69 return true;
70 }
71
72 static bool
type_is_tightly_packed(const struct glsl_type * type,unsigned * size_out)73 type_is_tightly_packed(const struct glsl_type *type, unsigned *size_out)
74 {
75 unsigned size = 0;
76 if (glsl_type_is_struct_or_ifc(type)) {
77 unsigned num_fields = glsl_get_length(type);
78 for (unsigned i = 0; i < num_fields; i++) {
79 const struct glsl_struct_field *field =
80 glsl_get_struct_field_data(type, i);
81
82 if (field->offset < 0 || field->offset != size)
83 return false;
84
85 unsigned field_size;
86 if (!type_is_tightly_packed(field->type, &field_size))
87 return false;
88
89 size = field->offset + field_size;
90 }
91 } else if (glsl_type_is_array_or_matrix(type)) {
92 if (glsl_type_is_unsized_array(type))
93 return false;
94
95 unsigned stride = glsl_get_explicit_stride(type);
96 if (stride == 0)
97 return false;
98
99 const struct glsl_type *elem_type = glsl_get_array_element(type);
100
101 unsigned elem_size;
102 if (!type_is_tightly_packed(elem_type, &elem_size))
103 return false;
104
105 if (elem_size != stride)
106 return false;
107
108 size = stride * glsl_get_length(type);
109 } else {
110 assert(glsl_type_is_vector_or_scalar(type));
111 if (glsl_get_explicit_stride(type) > 0)
112 return false;
113
114 if (glsl_type_is_boolean(type))
115 return false;
116
117 size = glsl_get_explicit_size(type, false);
118 }
119
120 if (size_out)
121 *size_out = size;
122 return true;
123 }
124
125 static bool
try_lower_memcpy(nir_builder * b,nir_intrinsic_instr * cpy,struct set * complex_vars)126 try_lower_memcpy(nir_builder *b, nir_intrinsic_instr *cpy,
127 struct set *complex_vars)
128 {
129 nir_deref_instr *dst = nir_src_as_deref(cpy->src[0]);
130 nir_deref_instr *src = nir_src_as_deref(cpy->src[1]);
131
132 /* A self-copy can always be eliminated */
133 if (dst == src) {
134 nir_instr_remove(&cpy->instr);
135 return true;
136 }
137
138 if (!nir_src_is_const(cpy->src[2]))
139 return false;
140
141 uint64_t size = nir_src_as_uint(cpy->src[2]);
142 if (size == 0) {
143 nir_instr_remove(&cpy->instr);
144 return true;
145 }
146
147 if (glsl_type_is_vector_or_scalar(src->type) &&
148 glsl_type_is_vector_or_scalar(dst->type) &&
149 glsl_get_explicit_size(dst->type, false) == size &&
150 glsl_get_explicit_size(src->type, false) == size) {
151 b->cursor = nir_instr_remove(&cpy->instr);
152 nir_ssa_def *data =
153 nir_load_deref_with_access(b, src, nir_intrinsic_src_access(cpy));
154 data = nir_bitcast_vector(b, data, glsl_get_bit_size(dst->type));
155 assert(data->num_components == glsl_get_vector_elements(dst->type));
156 nir_store_deref_with_access(b, dst, data, ~0 /* write mask */,
157 nir_intrinsic_dst_access(cpy));
158 return true;
159 }
160
161 unsigned type_size;
162 if (dst->type == src->type &&
163 type_is_tightly_packed(dst->type, &type_size) &&
164 type_size == size) {
165 b->cursor = nir_instr_remove(&cpy->instr);
166 nir_copy_deref_with_access(b, dst, src,
167 nir_intrinsic_dst_access(cpy),
168 nir_intrinsic_src_access(cpy));
169 return true;
170 }
171
172 /* If one of the two types is tightly packed and happens to equal the
173 * memcpy size, then we can get the memcpy by casting to that type and
174 * doing a deref copy.
175 *
176 * However, if we blindly apply this logic, we may end up with extra casts
177 * where we don't want them. The whole point of converting memcpy to
178 * copy_deref is in the hopes that nir_opt_copy_prop_vars or
179 * nir_lower_vars_to_ssa will get rid of the copy and those passes don't
180 * handle casts well. Heuristically, only do this optimization if the
181 * tightly packed type is on a deref with nir_var_function_temp so we stick
182 * the cast on the other mode.
183 */
184 if (dst->modes == nir_var_function_temp &&
185 type_is_tightly_packed(dst->type, &type_size) &&
186 type_size == size) {
187 b->cursor = nir_instr_remove(&cpy->instr);
188 src = nir_build_deref_cast(b, &src->dest.ssa,
189 src->modes, dst->type, 0);
190 nir_copy_deref_with_access(b, dst, src,
191 nir_intrinsic_dst_access(cpy),
192 nir_intrinsic_src_access(cpy));
193 return true;
194 }
195
196 /* If we can get at the variable AND the only complex use of that variable
197 * is as a memcpy destination, then we don't have to care about any empty
198 * space in the variable. In particular, we know that the variable is never
199 * cast to any other type and it's never used as a memcpy source so nothing
200 * can see any padding bytes. This holds even if some other memcpy only
201 * writes to part of the variable.
202 */
203 if (dst->deref_type == nir_deref_type_var &&
204 dst->modes == nir_var_function_temp &&
205 _mesa_set_search(complex_vars, dst->var) == NULL &&
206 glsl_get_explicit_size(dst->type, false) <= size) {
207 b->cursor = nir_instr_remove(&cpy->instr);
208 src = nir_build_deref_cast(b, &src->dest.ssa,
209 src->modes, dst->type, 0);
210 nir_copy_deref_with_access(b, dst, src,
211 nir_intrinsic_dst_access(cpy),
212 nir_intrinsic_src_access(cpy));
213 return true;
214 }
215
216 if (src->modes == nir_var_function_temp &&
217 type_is_tightly_packed(src->type, &type_size) &&
218 type_size == size) {
219 b->cursor = nir_instr_remove(&cpy->instr);
220 dst = nir_build_deref_cast(b, &dst->dest.ssa,
221 dst->modes, src->type, 0);
222 nir_copy_deref_with_access(b, dst, src,
223 nir_intrinsic_dst_access(cpy),
224 nir_intrinsic_src_access(cpy));
225 return true;
226 }
227
228 return false;
229 }
230
231 static bool
opt_memcpy_impl(nir_function_impl * impl)232 opt_memcpy_impl(nir_function_impl *impl)
233 {
234 bool progress = false;
235
236 nir_builder b;
237 nir_builder_init(&b, impl);
238
239 struct set *complex_vars = _mesa_pointer_set_create(NULL);
240
241 nir_foreach_block(block, impl) {
242 nir_foreach_instr(instr, block) {
243 if (instr->type != nir_instr_type_deref)
244 continue;
245
246 nir_deref_instr *deref = nir_instr_as_deref(instr);
247 if (deref->deref_type != nir_deref_type_var)
248 continue;
249
250 nir_deref_instr_has_complex_use_options opts =
251 nir_deref_instr_has_complex_use_allow_memcpy_dst;
252 if (nir_deref_instr_has_complex_use(deref, opts))
253 _mesa_set_add(complex_vars, deref->var);
254 }
255 }
256
257 nir_foreach_block(block, impl) {
258 nir_foreach_instr_safe(instr, block) {
259 if (instr->type != nir_instr_type_intrinsic)
260 continue;
261
262 nir_intrinsic_instr *cpy = nir_instr_as_intrinsic(instr);
263 if (cpy->intrinsic != nir_intrinsic_memcpy_deref)
264 continue;
265
266 while (opt_memcpy_deref_cast(cpy, &cpy->src[0]))
267 progress = true;
268 while (opt_memcpy_deref_cast(cpy, &cpy->src[1]))
269 progress = true;
270
271 if (try_lower_memcpy(&b, cpy, complex_vars)) {
272 progress = true;
273 continue;
274 }
275 }
276 }
277
278 _mesa_set_destroy(complex_vars, NULL);
279
280 if (progress) {
281 nir_metadata_preserve(impl, nir_metadata_block_index |
282 nir_metadata_dominance);
283 } else {
284 nir_metadata_preserve(impl, nir_metadata_all);
285 }
286
287 return progress;
288 }
289
290 bool
nir_opt_memcpy(nir_shader * shader)291 nir_opt_memcpy(nir_shader *shader)
292 {
293 bool progress = false;
294
295 nir_foreach_function(function, shader) {
296 if (function->impl && opt_memcpy_impl(function->impl))
297 progress = true;
298 }
299
300 return progress;
301 }
302