1 /*
2 * Copyright © 2014 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 * Authors:
24 * Jason Ekstrand (jason@jlekstrand.net)
25 *
26 */
27
28 #include "nir.h"
29
30 /*
31 * This pass lowers the neg, abs, and sat operations to source modifiers on
32 * ALU operations to make things nicer for the backend. It's just much
33 * easier to not have them when we're doing optimizations.
34 */
35
36 static void
alu_src_consume_abs(nir_alu_src * src)37 alu_src_consume_abs(nir_alu_src *src)
38 {
39 src->abs = true;
40 }
41
42 static void
alu_src_consume_negate(nir_alu_src * src)43 alu_src_consume_negate(nir_alu_src *src)
44 {
45 /* If abs is set on the source, the negate goes away */
46 if (!src->abs)
47 src->negate = !src->negate;
48 }
49
50 static bool
nir_lower_to_source_mods_block(nir_block * block,nir_lower_to_source_mods_flags options)51 nir_lower_to_source_mods_block(nir_block *block,
52 nir_lower_to_source_mods_flags options)
53 {
54 bool progress = false;
55
56 nir_foreach_instr(instr, block) {
57 if (instr->type != nir_instr_type_alu)
58 continue;
59
60 nir_alu_instr *alu = nir_instr_as_alu(instr);
61
62 bool lower_abs = (nir_op_infos[alu->op].num_inputs < 3) ||
63 (options & nir_lower_triop_abs);
64
65 for (unsigned i = 0; i < nir_op_infos[alu->op].num_inputs; i++) {
66 if (!alu->src[i].src.is_ssa)
67 continue;
68
69 if (alu->src[i].src.ssa->parent_instr->type != nir_instr_type_alu)
70 continue;
71
72 nir_alu_instr *parent = nir_instr_as_alu(alu->src[i].src.ssa->parent_instr);
73
74 if (parent->dest.saturate)
75 continue;
76
77 switch (nir_alu_type_get_base_type(nir_op_infos[alu->op].input_types[i])) {
78 case nir_type_float:
79 if (!(options & nir_lower_float_source_mods))
80 continue;
81 if (!(parent->op == nir_op_fabs && (options & nir_lower_fabs_source_mods)) &&
82 !(parent->op == nir_op_fneg && (options & nir_lower_fneg_source_mods))) {
83 continue;
84 }
85 break;
86 case nir_type_int:
87 if (!(options & nir_lower_int_source_mods))
88 continue;
89 if (parent->op != nir_op_iabs && parent->op != nir_op_ineg)
90 continue;
91 break;
92 default:
93 continue;
94 }
95
96 if (nir_src_bit_size(alu->src[i].src) == 64 &&
97 !(options & nir_lower_64bit_source_mods)) {
98 continue;
99 }
100
101 /* We can only do a rewrite if the source we are copying is SSA.
102 * Otherwise, moving the read might invalidly reorder reads/writes
103 * on a register.
104 */
105 if (!parent->src[0].src.is_ssa)
106 continue;
107
108 if (!lower_abs && (parent->op == nir_op_fabs ||
109 parent->op == nir_op_iabs ||
110 parent->src[0].abs))
111 continue;
112
113 nir_instr_rewrite_src(instr, &alu->src[i].src, parent->src[0].src);
114
115 /* Apply any modifiers that come from the parent opcode */
116 if (parent->op == nir_op_fneg || parent->op == nir_op_ineg)
117 alu_src_consume_negate(&alu->src[i]);
118 if (parent->op == nir_op_fabs || parent->op == nir_op_iabs)
119 alu_src_consume_abs(&alu->src[i]);
120
121 /* Apply modifiers from the parent source */
122 if (parent->src[0].negate)
123 alu_src_consume_negate(&alu->src[i]);
124 if (parent->src[0].abs)
125 alu_src_consume_abs(&alu->src[i]);
126
127 for (int j = 0; j < 4; ++j) {
128 if (!nir_alu_instr_channel_used(alu, i, j))
129 continue;
130 alu->src[i].swizzle[j] = parent->src[0].swizzle[alu->src[i].swizzle[j]];
131 }
132
133 if (nir_ssa_def_is_unused(&parent->dest.dest.ssa))
134 nir_instr_remove(&parent->instr);
135
136 progress = true;
137 }
138
139 /* We've covered sources. Now we're going to try and saturate the
140 * destination if we can.
141 */
142
143 if (!alu->dest.dest.is_ssa)
144 continue;
145
146 if (nir_dest_bit_size(alu->dest.dest) == 64 &&
147 !(options & nir_lower_64bit_source_mods)) {
148 continue;
149 }
150
151 /* We can only saturate float destinations */
152 if (nir_alu_type_get_base_type(nir_op_infos[alu->op].output_type) !=
153 nir_type_float)
154 continue;
155
156 if (!(options & nir_lower_float_source_mods))
157 continue;
158
159 if (!list_is_empty(&alu->dest.dest.ssa.if_uses))
160 continue;
161
162 bool all_children_are_sat = true;
163 nir_foreach_use(child_src, &alu->dest.dest.ssa) {
164 assert(child_src->is_ssa);
165 nir_instr *child = child_src->parent_instr;
166 if (child->type != nir_instr_type_alu) {
167 all_children_are_sat = false;
168 continue;
169 }
170
171 nir_alu_instr *child_alu = nir_instr_as_alu(child);
172 if (child_alu->src[0].negate || child_alu->src[0].abs) {
173 all_children_are_sat = false;
174 continue;
175 }
176
177 if (child_alu->op != nir_op_fsat) {
178 all_children_are_sat = false;
179 continue;
180 }
181 }
182
183 if (!all_children_are_sat)
184 continue;
185
186 alu->dest.saturate = true;
187 progress = true;
188
189 nir_foreach_use(child_src, &alu->dest.dest.ssa) {
190 assert(child_src->is_ssa);
191 nir_alu_instr *child_alu = nir_instr_as_alu(child_src->parent_instr);
192
193 child_alu->op = nir_op_mov;
194 child_alu->dest.saturate = false;
195 /* We could propagate the dest of our instruction to the
196 * destinations of the uses here. However, one quick round of
197 * copy propagation will clean that all up and then we don't have
198 * the complexity.
199 */
200 }
201 }
202
203 return progress;
204 }
205
206 static bool
nir_lower_to_source_mods_impl(nir_function_impl * impl,nir_lower_to_source_mods_flags options)207 nir_lower_to_source_mods_impl(nir_function_impl *impl,
208 nir_lower_to_source_mods_flags options)
209 {
210 bool progress = false;
211
212 nir_foreach_block(block, impl) {
213 progress |= nir_lower_to_source_mods_block(block, options);
214 }
215
216 if (progress)
217 nir_metadata_preserve(impl, nir_metadata_block_index |
218 nir_metadata_dominance);
219
220 return progress;
221 }
222
223 bool
nir_lower_to_source_mods(nir_shader * shader,nir_lower_to_source_mods_flags options)224 nir_lower_to_source_mods(nir_shader *shader,
225 nir_lower_to_source_mods_flags options)
226 {
227 bool progress = false;
228
229 nir_foreach_function(function, shader) {
230 if (function->impl) {
231 progress |= nir_lower_to_source_mods_impl(function->impl, options);
232 }
233 }
234
235 return progress;
236 }
237