nir/sink: Rewrite loop handling logic
[mesa.git] / src / compiler / nir / nir_opt_sink.c
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
40 nir_can_move_instr(nir_instr *instr, nir_move_options options)
41 {
42 if ((options & nir_move_const_undef) && instr->type == nir_instr_type_load_const) {
43 return true;
44 }
45
46 if (instr->type == nir_instr_type_intrinsic) {
47 nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr);
48 if ((options & nir_move_load_ubo) && intrin->intrinsic == nir_intrinsic_load_ubo)
49 return true;
50
51 if ((options & nir_move_load_input) &&
52 (intrin->intrinsic == nir_intrinsic_load_interpolated_input ||
53 intrin->intrinsic == nir_intrinsic_load_input))
54 return true;
55 }
56
57 if ((options & nir_move_const_undef) && instr->type == nir_instr_type_ssa_undef) {
58 return true;
59 }
60
61 if ((options & nir_move_comparisons) && instr->type == nir_instr_type_alu &&
62 nir_alu_instr_is_comparison(nir_instr_as_alu(instr))) {
63 return true;
64 }
65
66 return false;
67 }
68
69 static nir_loop *
70 get_innermost_loop(nir_cf_node *node)
71 {
72 for (; node != NULL; node = node->parent) {
73 if (node->type == nir_cf_node_loop)
74 return (nir_loop*)node;
75 }
76 return NULL;
77 }
78
79 static bool
80 loop_contains_block(nir_loop *loop, nir_block *block)
81 {
82 nir_block *before = nir_cf_node_as_block(nir_cf_node_prev(&loop->cf_node));
83 nir_block *after = nir_cf_node_as_block(nir_cf_node_next(&loop->cf_node));
84
85 return block->index > before->index && block->index < after->index;
86 }
87
88 /* Given the LCA of all uses and the definition, find a block on the path
89 * between them in the dominance tree that is outside of as many loops as
90 * possible.
91 */
92
93 static nir_block *
94 adjust_block_for_loops(nir_block *use_block, nir_block *def_block)
95 {
96 for (nir_block *cur_block = use_block; cur_block != def_block->imm_dom;
97 cur_block = cur_block->imm_dom) {
98 nir_cf_node *next = nir_cf_node_next(&cur_block->cf_node);
99 if (next && next->type == nir_cf_node_loop) {
100 nir_loop *following_loop = nir_cf_node_as_loop(next);
101 if (loop_contains_block(following_loop, use_block)) {
102 use_block = cur_block;
103 continue;
104 }
105 }
106 }
107
108 return use_block;
109 }
110
111 /* iterate a ssa def's use's and try to find a more optimal block to
112 * move it to, using the dominance tree. In short, if all of the uses
113 * are contained in a single block, the load will be moved there,
114 * otherwise it will be move to the least common ancestor block of all
115 * the uses
116 */
117 static nir_block *
118 get_preferred_block(nir_ssa_def *def, bool sink_into_loops)
119 {
120 nir_block *lca = NULL;
121
122 nir_foreach_use(use, def) {
123 nir_instr *instr = use->parent_instr;
124 nir_block *use_block = instr->block;
125
126 /*
127 * Kind of an ugly special-case, but phi instructions
128 * need to appear first in the block, so by definition
129 * we can't move an instruction into a block where it is
130 * consumed by a phi instruction. We could conceivably
131 * move it into a dominator block.
132 */
133 if (instr->type == nir_instr_type_phi) {
134 nir_phi_instr *phi = nir_instr_as_phi(instr);
135 nir_block *phi_lca = NULL;
136 nir_foreach_phi_src(src, phi) {
137 if (&src->src == use)
138 phi_lca = nir_dominance_lca(phi_lca, src->pred);
139 }
140 use_block = phi_lca;
141 }
142
143 lca = nir_dominance_lca(lca, use_block);
144 }
145
146 nir_foreach_if_use(use, def) {
147 nir_block *use_block =
148 nir_cf_node_as_block(nir_cf_node_prev(&use->parent_if->cf_node));
149
150 lca = nir_dominance_lca(lca, use_block);
151 }
152
153 /* If we're moving a load_ubo or load_interpolated_input, we don't want to
154 * sink it down into loops, which may result in accessing memory or shared
155 * functions multiple times. Sink it just above the start of the loop
156 * where it's used. For load_consts, undefs, and comparisons, we expect
157 * the driver to be able to emit them as simple ALU ops, so sinking as far
158 * in as we can go is probably worth it for register pressure.
159 */
160 if (!sink_into_loops) {
161 lca = adjust_block_for_loops(lca, def->parent_instr->block);
162 assert(nir_block_dominates(def->parent_instr->block, lca));
163 }
164
165
166 return lca;
167 }
168
169 /* insert before first non-phi instruction: */
170 static void
171 insert_after_phi(nir_instr *instr, nir_block *block)
172 {
173 nir_foreach_instr(instr2, block) {
174 if (instr2->type == nir_instr_type_phi)
175 continue;
176
177 exec_node_insert_node_before(&instr2->node,
178 &instr->node);
179
180 return;
181 }
182
183 /* if haven't inserted it, push to tail (ie. empty block or possibly
184 * a block only containing phi's?)
185 */
186 exec_list_push_tail(&block->instr_list, &instr->node);
187 }
188
189 bool
190 nir_opt_sink(nir_shader *shader, nir_move_options options)
191 {
192 bool progress = false;
193
194 nir_foreach_function(function, shader) {
195 if (!function->impl)
196 continue;
197
198 nir_metadata_require(function->impl,
199 nir_metadata_block_index | nir_metadata_dominance);
200
201 nir_foreach_block_reverse(block, function->impl) {
202 nir_foreach_instr_reverse_safe(instr, block) {
203 if (!nir_can_move_instr(instr, options))
204 continue;
205
206 nir_ssa_def *def = nir_instr_ssa_def(instr);
207 nir_block *use_block =
208 get_preferred_block(def, instr->type != nir_instr_type_intrinsic);
209
210 if (!use_block || use_block == instr->block)
211 continue;
212
213 exec_node_remove(&instr->node);
214
215 insert_after_phi(instr, use_block);
216
217 instr->block = use_block;
218
219 progress = true;
220 }
221 }
222
223 nir_metadata_preserve(function->impl,
224 nir_metadata_block_index | nir_metadata_dominance);
225 }
226
227 return progress;
228 }