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 bool
nir_lower_to_source_mods_block(nir_block * block)37 nir_lower_to_source_mods_block(nir_block *block)
38 {
39 bool progress = false;
40
41 nir_foreach_instr(instr, block) {
42 if (instr->type != nir_instr_type_alu)
43 continue;
44
45 nir_alu_instr *alu = nir_instr_as_alu(instr);
46
47 for (unsigned i = 0; i < nir_op_infos[alu->op].num_inputs; i++) {
48 if (!alu->src[i].src.is_ssa)
49 continue;
50
51 if (alu->src[i].src.ssa->parent_instr->type != nir_instr_type_alu)
52 continue;
53
54 nir_alu_instr *parent = nir_instr_as_alu(alu->src[i].src.ssa->parent_instr);
55
56 if (parent->dest.saturate)
57 continue;
58
59 switch (nir_alu_type_get_base_type(nir_op_infos[alu->op].input_types[i])) {
60 case nir_type_float:
61 if (parent->op != nir_op_fmov)
62 continue;
63 break;
64 case nir_type_int:
65 if (parent->op != nir_op_imov)
66 continue;
67 break;
68 default:
69 continue;
70 }
71
72 /* We can only do a rewrite if the source we are copying is SSA.
73 * Otherwise, moving the read might invalidly reorder reads/writes
74 * on a register.
75 */
76 if (!parent->src[0].src.is_ssa)
77 continue;
78
79 nir_instr_rewrite_src(instr, &alu->src[i].src, parent->src[0].src);
80 if (alu->src[i].abs) {
81 /* abs trumps both neg and abs, do nothing */
82 } else {
83 alu->src[i].negate = (alu->src[i].negate != parent->src[0].negate);
84 alu->src[i].abs |= parent->src[0].abs;
85 }
86
87 for (int j = 0; j < 4; ++j) {
88 if (!nir_alu_instr_channel_used(alu, i, j))
89 continue;
90 alu->src[i].swizzle[j] = parent->src[0].swizzle[alu->src[i].swizzle[j]];
91 }
92
93 if (list_empty(&parent->dest.dest.ssa.uses) &&
94 list_empty(&parent->dest.dest.ssa.if_uses))
95 nir_instr_remove(&parent->instr);
96
97 progress = true;
98 }
99
100 switch (alu->op) {
101 case nir_op_fsat:
102 alu->op = nir_op_fmov;
103 alu->dest.saturate = true;
104 break;
105 case nir_op_ineg:
106 alu->op = nir_op_imov;
107 alu->src[0].negate = !alu->src[0].negate;
108 break;
109 case nir_op_fneg:
110 alu->op = nir_op_fmov;
111 alu->src[0].negate = !alu->src[0].negate;
112 break;
113 case nir_op_iabs:
114 alu->op = nir_op_imov;
115 alu->src[0].abs = true;
116 alu->src[0].negate = false;
117 break;
118 case nir_op_fabs:
119 alu->op = nir_op_fmov;
120 alu->src[0].abs = true;
121 alu->src[0].negate = false;
122 break;
123 default:
124 break;
125 }
126
127 /* We've covered sources. Now we're going to try and saturate the
128 * destination if we can.
129 */
130
131 if (!alu->dest.dest.is_ssa)
132 continue;
133
134 /* We can only saturate float destinations */
135 if (nir_alu_type_get_base_type(nir_op_infos[alu->op].output_type) !=
136 nir_type_float)
137 continue;
138
139 if (!list_empty(&alu->dest.dest.ssa.if_uses))
140 continue;
141
142 bool all_children_are_sat = true;
143 nir_foreach_use(child_src, &alu->dest.dest.ssa) {
144 assert(child_src->is_ssa);
145 nir_instr *child = child_src->parent_instr;
146 if (child->type != nir_instr_type_alu) {
147 all_children_are_sat = false;
148 continue;
149 }
150
151 nir_alu_instr *child_alu = nir_instr_as_alu(child);
152 if (child_alu->src[0].negate || child_alu->src[0].abs) {
153 all_children_are_sat = false;
154 continue;
155 }
156
157 if (child_alu->op != nir_op_fsat &&
158 !(child_alu->op == nir_op_fmov && child_alu->dest.saturate)) {
159 all_children_are_sat = false;
160 continue;
161 }
162 }
163
164 if (!all_children_are_sat)
165 continue;
166
167 alu->dest.saturate = true;
168 progress = true;
169
170 nir_foreach_use(child_src, &alu->dest.dest.ssa) {
171 assert(child_src->is_ssa);
172 nir_alu_instr *child_alu = nir_instr_as_alu(child_src->parent_instr);
173
174 child_alu->op = nir_op_fmov;
175 child_alu->dest.saturate = false;
176 /* We could propagate the dest of our instruction to the
177 * destinations of the uses here. However, one quick round of
178 * copy propagation will clean that all up and then we don't have
179 * the complexity.
180 */
181 }
182 }
183
184 return progress;
185 }
186
187 static bool
nir_lower_to_source_mods_impl(nir_function_impl * impl)188 nir_lower_to_source_mods_impl(nir_function_impl *impl)
189 {
190 bool progress = false;
191
192 nir_foreach_block(block, impl) {
193 progress |= nir_lower_to_source_mods_block(block);
194 }
195
196 if (progress)
197 nir_metadata_preserve(impl, nir_metadata_block_index |
198 nir_metadata_dominance);
199
200 return progress;
201 }
202
203 bool
nir_lower_to_source_mods(nir_shader * shader)204 nir_lower_to_source_mods(nir_shader *shader)
205 {
206 bool progress = false;
207
208 nir_foreach_function(function, shader) {
209 if (function->impl) {
210 progress |= nir_lower_to_source_mods_impl(function->impl);
211 }
212 }
213
214 return progress;
215 }
216