nir_opt_memcpy.c revision 7ec681f3 
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
26static bool
27opt_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
72static bool
73type_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
125static bool
126try_lower_memcpy(nir_builder *b, nir_intrinsic_instr *cpy)
127{
128   nir_deref_instr *dst = nir_src_as_deref(cpy->src[0]);
129   nir_deref_instr *src = nir_src_as_deref(cpy->src[1]);
130
131   /* A self-copy can always be eliminated */
132   if (dst == src) {
133      nir_instr_remove(&cpy->instr);
134      return true;
135   }
136
137   if (!nir_src_is_const(cpy->src[2]))
138      return false;
139
140   uint64_t size = nir_src_as_uint(cpy->src[2]);
141   if (size == 0) {
142      nir_instr_remove(&cpy->instr);
143      return true;
144   }
145
146   if (glsl_type_is_vector_or_scalar(src->type) &&
147       glsl_type_is_vector_or_scalar(dst->type) &&
148       glsl_get_explicit_size(dst->type, false) == size &&
149       glsl_get_explicit_size(src->type, false) == size) {
150      b->cursor = nir_instr_remove(&cpy->instr);
151      nir_ssa_def *data =
152         nir_load_deref_with_access(b, src, nir_intrinsic_src_access(cpy));
153      data = nir_bitcast_vector(b, data, glsl_get_bit_size(dst->type));
154      assert(data->num_components == glsl_get_vector_elements(dst->type));
155      nir_store_deref_with_access(b, dst, data, ~0 /* write mask */,
156                                  nir_intrinsic_dst_access(cpy));
157      return true;
158   }
159
160   unsigned type_size;
161   if (dst->type == src->type &&
162       type_is_tightly_packed(dst->type, &type_size) &&
163       type_size == size) {
164      b->cursor = nir_instr_remove(&cpy->instr);
165      nir_copy_deref_with_access(b, dst, src,
166                                 nir_intrinsic_dst_access(cpy),
167                                 nir_intrinsic_src_access(cpy));
168      return true;
169   }
170
171   return false;
172}
173
174static bool
175opt_memcpy_impl(nir_function_impl *impl)
176{
177   bool progress = false;
178
179   nir_builder b;
180   nir_builder_init(&b, impl);
181
182   nir_foreach_block(block, impl) {
183      nir_foreach_instr_safe(instr, block) {
184         if (instr->type != nir_instr_type_intrinsic)
185            continue;
186
187         nir_intrinsic_instr *cpy = nir_instr_as_intrinsic(instr);
188         if (cpy->intrinsic != nir_intrinsic_memcpy_deref)
189            continue;
190
191         while (opt_memcpy_deref_cast(cpy, &cpy->src[0]))
192            progress = true;
193         while (opt_memcpy_deref_cast(cpy, &cpy->src[1]))
194            progress = true;
195
196         if (try_lower_memcpy(&b, cpy)) {
197            progress = true;
198            continue;
199         }
200      }
201   }
202
203   if (progress) {
204      nir_metadata_preserve(impl, nir_metadata_block_index |
205                                  nir_metadata_dominance);
206   } else {
207      nir_metadata_preserve(impl, nir_metadata_all);
208   }
209
210   return progress;
211}
212
213bool
214nir_opt_memcpy(nir_shader *shader)
215{
216   bool progress = false;
217
218   nir_foreach_function(function, shader) {
219      if (function->impl && opt_memcpy_impl(function->impl))
220         progress = true;
221   }
222
223   return progress;
224}
225