1 /*
2 * Copyright © 2018 Red Hat
3 * Copyright © 2019 Valve Corporation
4 *
5 * Permission is hereby granted, free of charge, to any person obtaining a
6 * copy of this software and associated documentation files (the "Software"),
7 * to deal in the Software without restriction, including without limitation
8 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
9 * and/or sell copies of the Software, and to permit persons to whom the
10 * Software is furnished to do so, subject to the following conditions:
11 *
12 * The above copyright notice and this permission notice (including the next
13 * paragraph) shall be included in all copies or substantial portions of the
14 * Software.
15 *
16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
19 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
20 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
21 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
22 * IN THE SOFTWARE.
23 *
24 * Authors:
25 * Rob Clark (robdclark@gmail.com>
26 * Daniel Schürmann (daniel.schuermann@campus.tu-berlin.de)
27 * Rhys Perry (pendingchaos02@gmail.com)
28 *
29 */
30
31 #include "nir.h"
32
33
34 /*
35 * A simple pass that moves some instructions into the least common
36 * anscestor of consuming instructions.
37 */
38
39 bool
nir_can_move_instr(nir_instr * instr,nir_move_options options)40 nir_can_move_instr(nir_instr *instr, nir_move_options options)
41 {
42 switch (instr->type) {
43 case nir_instr_type_load_const:
44 case nir_instr_type_ssa_undef: {
45 return options & nir_move_const_undef;
46 }
47 case nir_instr_type_alu: {
48 if (nir_op_is_vec(nir_instr_as_alu(instr)->op) ||
49 nir_instr_as_alu(instr)->op == nir_op_b2i32)
50 return options & nir_move_copies;
51 if (nir_alu_instr_is_comparison(nir_instr_as_alu(instr)))
52 return options & nir_move_comparisons;
53 return false;
54 }
55 case nir_instr_type_intrinsic: {
56 nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr);
57 switch (intrin->intrinsic) {
58 case nir_intrinsic_load_ubo:
59 case nir_intrinsic_load_ubo_vec4:
60 return options & nir_move_load_ubo;
61 case nir_intrinsic_load_ssbo:
62 return (options & nir_move_load_ssbo) && nir_intrinsic_can_reorder(intrin);
63 case nir_intrinsic_load_input:
64 case nir_intrinsic_load_interpolated_input:
65 case nir_intrinsic_load_per_vertex_input:
66 return options & nir_move_load_input;
67 case nir_intrinsic_load_uniform:
68 return options & nir_move_load_uniform;
69 default:
70 return false;
71 }
72 }
73 default:
74 return false;
75 }
76 }
77
78 static nir_loop *
get_innermost_loop(nir_cf_node * node)79 get_innermost_loop(nir_cf_node *node)
80 {
81 for (; node != NULL; node = node->parent) {
82 if (node->type == nir_cf_node_loop)
83 return (nir_loop*)node;
84 }
85 return NULL;
86 }
87
88 static bool
loop_contains_block(nir_loop * loop,nir_block * block)89 loop_contains_block(nir_loop *loop, nir_block *block)
90 {
91 nir_block *before = nir_cf_node_as_block(nir_cf_node_prev(&loop->cf_node));
92 nir_block *after = nir_cf_node_as_block(nir_cf_node_next(&loop->cf_node));
93
94 return block->index > before->index && block->index < after->index;
95 }
96
97 /* Given the LCA of all uses and the definition, find a block on the path
98 * between them in the dominance tree that is outside of as many loops as
99 * possible. If "sink_out_of_loops" is false, then we disallow sinking the
100 * definition outside of the loop it's defined in (if any).
101 */
102
103 static nir_block *
adjust_block_for_loops(nir_block * use_block,nir_block * def_block,bool sink_out_of_loops)104 adjust_block_for_loops(nir_block *use_block, nir_block *def_block,
105 bool sink_out_of_loops)
106 {
107 nir_loop *def_loop = NULL;
108 if (!sink_out_of_loops)
109 def_loop = get_innermost_loop(&def_block->cf_node);
110
111 for (nir_block *cur_block = use_block; cur_block != def_block->imm_dom;
112 cur_block = cur_block->imm_dom) {
113 if (!sink_out_of_loops && def_loop &&
114 !loop_contains_block(def_loop, use_block)) {
115 use_block = cur_block;
116 continue;
117 }
118
119 nir_cf_node *next = nir_cf_node_next(&cur_block->cf_node);
120 if (next && next->type == nir_cf_node_loop) {
121 nir_loop *following_loop = nir_cf_node_as_loop(next);
122 if (loop_contains_block(following_loop, use_block)) {
123 use_block = cur_block;
124 continue;
125 }
126 }
127 }
128
129 return use_block;
130 }
131
132 /* iterate a ssa def's use's and try to find a more optimal block to
133 * move it to, using the dominance tree. In short, if all of the uses
134 * are contained in a single block, the load will be moved there,
135 * otherwise it will be move to the least common ancestor block of all
136 * the uses
137 */
138 static nir_block *
get_preferred_block(nir_ssa_def * def,bool sink_out_of_loops)139 get_preferred_block(nir_ssa_def *def, bool sink_out_of_loops)
140 {
141 nir_block *lca = NULL;
142
143 nir_foreach_use(use, def) {
144 nir_instr *instr = use->parent_instr;
145 nir_block *use_block = instr->block;
146
147 /*
148 * Kind of an ugly special-case, but phi instructions
149 * need to appear first in the block, so by definition
150 * we can't move an instruction into a block where it is
151 * consumed by a phi instruction. We could conceivably
152 * move it into a dominator block.
153 */
154 if (instr->type == nir_instr_type_phi) {
155 nir_phi_instr *phi = nir_instr_as_phi(instr);
156 nir_block *phi_lca = NULL;
157 nir_foreach_phi_src(src, phi) {
158 if (&src->src == use)
159 phi_lca = nir_dominance_lca(phi_lca, src->pred);
160 }
161 use_block = phi_lca;
162 }
163
164 lca = nir_dominance_lca(lca, use_block);
165 }
166
167 nir_foreach_if_use(use, def) {
168 nir_block *use_block =
169 nir_cf_node_as_block(nir_cf_node_prev(&use->parent_if->cf_node));
170
171 lca = nir_dominance_lca(lca, use_block);
172 }
173
174 /* return in case, we didn't find a reachable user */
175 if (!lca)
176 return NULL;
177
178 /* We don't sink any instructions into loops to avoid repeated executions
179 * This might occasionally increase register pressure, but seems overall
180 * the better choice.
181 */
182 lca = adjust_block_for_loops(lca, def->parent_instr->block,
183 sink_out_of_loops);
184 assert(nir_block_dominates(def->parent_instr->block, lca));
185
186 return lca;
187 }
188
189 static bool
can_sink_out_of_loop(nir_intrinsic_instr * intrin)190 can_sink_out_of_loop(nir_intrinsic_instr *intrin)
191 {
192 /* Don't sink buffer loads out of loops because that can make its
193 * resource divergent and break code like that which is generated
194 * by nir_lower_non_uniform_access.
195 */
196 return intrin->intrinsic != nir_intrinsic_load_ubo &&
197 intrin->intrinsic != nir_intrinsic_load_ssbo;
198 }
199
200 bool
nir_opt_sink(nir_shader * shader,nir_move_options options)201 nir_opt_sink(nir_shader *shader, nir_move_options options)
202 {
203 bool progress = false;
204
205 nir_foreach_function(function, shader) {
206 if (!function->impl)
207 continue;
208
209 nir_metadata_require(function->impl,
210 nir_metadata_block_index | nir_metadata_dominance);
211
212 nir_foreach_block_reverse(block, function->impl) {
213 nir_foreach_instr_reverse_safe(instr, block) {
214 if (!nir_can_move_instr(instr, options))
215 continue;
216
217 nir_ssa_def *def = nir_instr_ssa_def(instr);
218
219 bool sink_out_of_loops =
220 instr->type != nir_instr_type_intrinsic ||
221 can_sink_out_of_loop(nir_instr_as_intrinsic(instr));
222 nir_block *use_block =
223 get_preferred_block(def, sink_out_of_loops);
224
225 if (!use_block || use_block == instr->block)
226 continue;
227
228 nir_instr_remove(instr);
229 nir_instr_insert(nir_after_phis(use_block), instr);
230
231 progress = true;
232 }
233 }
234
235 nir_metadata_preserve(function->impl,
236 nir_metadata_block_index | nir_metadata_dominance);
237 }
238
239 return progress;
240 }
241