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
24 #include "nir.h"
25 #include "nir_builder.h"
26
27 static void
28 emit_load_store(nir_builder *b, nir_intrinsic_instr *orig_instr,
29 nir_deref_var *deref, nir_deref *tail,
30 nir_ssa_def **dest, nir_ssa_def *src);
31
32 static void
emit_indirect_load_store(nir_builder * b,nir_intrinsic_instr * orig_instr,nir_deref_var * deref,nir_deref * arr_parent,int start,int end,nir_ssa_def ** dest,nir_ssa_def * src)33 emit_indirect_load_store(nir_builder *b, nir_intrinsic_instr *orig_instr,
34 nir_deref_var *deref, nir_deref *arr_parent,
35 int start, int end,
36 nir_ssa_def **dest, nir_ssa_def *src)
37 {
38 nir_deref_array *arr = nir_deref_as_array(arr_parent->child);
39 assert(arr->deref_array_type == nir_deref_array_type_indirect);
40 assert(arr->indirect.is_ssa);
41
42 assert(start < end);
43 if (start == end - 1) {
44 /* Base case. Just emit the load/store op */
45 nir_deref_array direct = *arr;
46 direct.deref_array_type = nir_deref_array_type_direct;
47 direct.base_offset += start;
48 direct.indirect = NIR_SRC_INIT;
49
50 arr_parent->child = &direct.deref;
51 emit_load_store(b, orig_instr, deref, &direct.deref, dest, src);
52 arr_parent->child = &arr->deref;
53 } else {
54 int mid = start + (end - start) / 2;
55
56 nir_ssa_def *then_dest, *else_dest;
57
58 nir_push_if(b, nir_ilt(b, arr->indirect.ssa, nir_imm_int(b, mid)));
59 emit_indirect_load_store(b, orig_instr, deref, arr_parent,
60 start, mid, &then_dest, src);
61 nir_push_else(b, NULL);
62 emit_indirect_load_store(b, orig_instr, deref, arr_parent,
63 mid, end, &else_dest, src);
64 nir_pop_if(b, NULL);
65
66 if (src == NULL)
67 *dest = nir_if_phi(b, then_dest, else_dest);
68 }
69 }
70
71 static void
emit_load_store(nir_builder * b,nir_intrinsic_instr * orig_instr,nir_deref_var * deref,nir_deref * tail,nir_ssa_def ** dest,nir_ssa_def * src)72 emit_load_store(nir_builder *b, nir_intrinsic_instr *orig_instr,
73 nir_deref_var *deref, nir_deref *tail,
74 nir_ssa_def **dest, nir_ssa_def *src)
75 {
76 for (; tail->child; tail = tail->child) {
77 if (tail->child->deref_type != nir_deref_type_array)
78 continue;
79
80 nir_deref_array *arr = nir_deref_as_array(tail->child);
81 if (arr->deref_array_type != nir_deref_array_type_indirect)
82 continue;
83
84 int length = glsl_get_length(tail->type);
85
86 emit_indirect_load_store(b, orig_instr, deref, tail, -arr->base_offset,
87 length - arr->base_offset, dest, src);
88 return;
89 }
90
91 assert(tail && tail->child == NULL);
92
93 /* We reached the end of the deref chain. Emit the instruction */
94
95 if (src == NULL) {
96 /* This is a load instruction */
97 nir_intrinsic_instr *load =
98 nir_intrinsic_instr_create(b->shader, orig_instr->intrinsic);
99 load->num_components = orig_instr->num_components;
100 load->variables[0] = nir_deref_var_clone(deref, load);
101
102 /* Copy over any sources. This is needed for interp_var_at */
103 for (unsigned i = 0;
104 i < nir_intrinsic_infos[orig_instr->intrinsic].num_srcs; i++)
105 nir_src_copy(&load->src[i], &orig_instr->src[i], load);
106
107 unsigned bit_size = orig_instr->dest.ssa.bit_size;
108 nir_ssa_dest_init(&load->instr, &load->dest,
109 load->num_components, bit_size, NULL);
110 nir_builder_instr_insert(b, &load->instr);
111 *dest = &load->dest.ssa;
112 } else {
113 /* This is a store instruction */
114 nir_intrinsic_instr *store =
115 nir_intrinsic_instr_create(b->shader, nir_intrinsic_store_var);
116 store->num_components = orig_instr->num_components;
117 nir_intrinsic_set_write_mask(store, nir_intrinsic_write_mask(orig_instr));
118 store->variables[0] = nir_deref_var_clone(deref, store);
119 store->src[0] = nir_src_for_ssa(src);
120 nir_builder_instr_insert(b, &store->instr);
121 }
122 }
123
124 static bool
deref_has_indirect(nir_deref_var * deref)125 deref_has_indirect(nir_deref_var *deref)
126 {
127 for (nir_deref *tail = deref->deref.child; tail; tail = tail->child) {
128 if (tail->deref_type != nir_deref_type_array)
129 continue;
130
131 nir_deref_array *arr = nir_deref_as_array(tail);
132 if (arr->deref_array_type == nir_deref_array_type_indirect)
133 return true;
134 }
135
136 return false;
137 }
138
139 static bool
lower_indirect_block(nir_block * block,nir_builder * b,nir_variable_mode modes)140 lower_indirect_block(nir_block *block, nir_builder *b,
141 nir_variable_mode modes)
142 {
143 bool progress = false;
144
145 nir_foreach_instr_safe(instr, block) {
146 if (instr->type != nir_instr_type_intrinsic)
147 continue;
148
149 nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr);
150 if (intrin->intrinsic != nir_intrinsic_load_var &&
151 intrin->intrinsic != nir_intrinsic_interp_var_at_centroid &&
152 intrin->intrinsic != nir_intrinsic_interp_var_at_sample &&
153 intrin->intrinsic != nir_intrinsic_interp_var_at_offset &&
154 intrin->intrinsic != nir_intrinsic_store_var)
155 continue;
156
157 if (!deref_has_indirect(intrin->variables[0]))
158 continue;
159
160 /* Only lower variables whose mode is in the mask, or compact
161 * array variables. (We can't handle indirects on tightly packed
162 * scalar arrays, so we need to lower them regardless.)
163 */
164 if (!(modes & intrin->variables[0]->var->data.mode) &&
165 !intrin->variables[0]->var->data.compact)
166 continue;
167
168 b->cursor = nir_before_instr(&intrin->instr);
169
170 if (intrin->intrinsic != nir_intrinsic_store_var) {
171 nir_ssa_def *result;
172 emit_load_store(b, intrin, intrin->variables[0],
173 &intrin->variables[0]->deref, &result, NULL);
174 nir_ssa_def_rewrite_uses(&intrin->dest.ssa, nir_src_for_ssa(result));
175 } else {
176 assert(intrin->src[0].is_ssa);
177 emit_load_store(b, intrin, intrin->variables[0],
178 &intrin->variables[0]->deref, NULL, intrin->src[0].ssa);
179 }
180 nir_instr_remove(&intrin->instr);
181 progress = true;
182 }
183
184 return progress;
185 }
186
187 static bool
lower_indirects_impl(nir_function_impl * impl,nir_variable_mode modes)188 lower_indirects_impl(nir_function_impl *impl, nir_variable_mode modes)
189 {
190 nir_builder builder;
191 nir_builder_init(&builder, impl);
192 bool progress = false;
193
194 nir_foreach_block_safe(block, impl) {
195 progress |= lower_indirect_block(block, &builder, modes);
196 }
197
198 if (progress)
199 nir_metadata_preserve(impl, nir_metadata_none);
200
201 return progress;
202 }
203
204 /** Lowers indirect variable loads/stores to direct loads/stores.
205 *
206 * The pass works by replacing any indirect load or store with an if-ladder
207 * that does a binary search on the array index.
208 */
209 bool
nir_lower_indirect_derefs(nir_shader * shader,nir_variable_mode modes)210 nir_lower_indirect_derefs(nir_shader *shader, nir_variable_mode modes)
211 {
212 bool progress = false;
213
214 nir_foreach_function(function, shader) {
215 if (function->impl)
216 progress = lower_indirects_impl(function->impl, modes) || progress;
217 }
218
219 return progress;
220 }
221