1 /*
2 * Copyright © 2013 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 "brw_fs.h"
25 #include "brw_fs_live_variables.h"
26 #include "brw_cfg.h"
27
28 using namespace brw;
29
30 /** @file
31 *
32 * Implements a pass that propagates the SAT modifier from a MOV.SAT into the
33 * instruction that produced the source of the MOV.SAT, thereby allowing the
34 * MOV's src and dst to be coalesced and the MOV removed.
35 *
36 * For instance,
37 *
38 * ADD tmp, src0, src1
39 * MOV.SAT dst, tmp
40 *
41 * would be transformed into
42 *
43 * ADD.SAT tmp, src0, src1
44 * MOV dst, tmp
45 */
46
47 static bool
propagate_sat(fs_inst * inst,fs_inst * scan_inst)48 propagate_sat(fs_inst *inst, fs_inst *scan_inst)
49 {
50 if (scan_inst->dst.type != inst->dst.type) {
51 scan_inst->dst.type = inst->dst.type;
52 for (int i = 0; i < scan_inst->sources; i++) {
53 scan_inst->src[i].type = inst->dst.type;
54 }
55 }
56
57 if (inst->src[0].negate) {
58 if (scan_inst->opcode == BRW_OPCODE_MUL) {
59 scan_inst->src[0].negate = !scan_inst->src[0].negate;
60 inst->src[0].negate = false;
61 } else if (scan_inst->opcode == BRW_OPCODE_MAD) {
62 for (int i = 0; i < 2; i++) {
63 if (scan_inst->src[i].file == IMM) {
64 brw_reg_negate_immediate(&scan_inst->src[i]);
65 } else {
66 scan_inst->src[i].negate = !scan_inst->src[i].negate;
67 }
68 }
69 inst->src[0].negate = false;
70 } else if (scan_inst->opcode == BRW_OPCODE_ADD) {
71 if (scan_inst->src[1].file == IMM) {
72 if (!brw_reg_negate_immediate(&scan_inst->src[1])) {
73 return false;
74 }
75 } else {
76 scan_inst->src[1].negate = !scan_inst->src[1].negate;
77 }
78 scan_inst->src[0].negate = !scan_inst->src[0].negate;
79 inst->src[0].negate = false;
80 } else {
81 return false;
82 }
83 }
84
85 scan_inst->saturate = true;
86 inst->saturate = false;
87 return true;
88 }
89
90 static bool
opt_saturate_propagation_local(fs_visitor & s,bblock_t * block)91 opt_saturate_propagation_local(fs_visitor &s, bblock_t *block)
92 {
93 bool progress = false;
94 int ip = block->end_ip + 1;
95
96 foreach_inst_in_block_reverse(fs_inst, inst, block) {
97 ip--;
98
99 if (inst->opcode != BRW_OPCODE_MOV ||
100 !inst->saturate ||
101 inst->dst.file != VGRF ||
102 inst->dst.type != inst->src[0].type ||
103 inst->src[0].file != VGRF ||
104 inst->src[0].abs)
105 continue;
106
107 const brw::def_analysis &defs = s.def_analysis.require();
108 fs_inst *def = defs.get(inst->src[0]);
109
110 if (def != NULL) {
111 if (def->exec_size != inst->exec_size)
112 continue;
113
114 if (def->dst.type != inst->dst.type && !def->can_change_types())
115 continue;
116
117 if (def->flags_written(s.devinfo) != 0)
118 continue;
119
120 if (def->saturate) {
121 inst->saturate = false;
122 progress = true;
123 continue;
124 } else if (defs.get_use_count(def->dst) == 1 &&
125 def->can_do_saturate() &&
126 propagate_sat(inst, def)) {
127 progress = true;
128 continue;
129 }
130
131 /* If the def is in a different block the liveness based pass will
132 * not be able to make progress, so skip it.
133 */
134 if (block != defs.get_block(inst->src[0]))
135 continue;
136 }
137
138 const fs_live_variables &live = s.live_analysis.require();
139 int src_var = live.var_from_reg(inst->src[0]);
140 int src_end_ip = live.end[src_var];
141
142 bool interfered = false;
143 foreach_inst_in_block_reverse_starting_from(fs_inst, scan_inst, inst) {
144 if (scan_inst->exec_size == inst->exec_size &&
145 regions_overlap(scan_inst->dst, scan_inst->size_written,
146 inst->src[0], inst->size_read(0))) {
147 if (scan_inst->is_partial_write() ||
148 (scan_inst->dst.type != inst->dst.type &&
149 !scan_inst->can_change_types()))
150 break;
151
152 if (scan_inst->flags_written(s.devinfo) != 0)
153 break;
154
155 if (scan_inst->saturate) {
156 inst->saturate = false;
157 progress = true;
158 } else if (src_end_ip == ip || inst->dst.equals(inst->src[0])) {
159 if (scan_inst->can_do_saturate() &&
160 propagate_sat(inst, scan_inst)) {
161 progress = true;
162 }
163 }
164 break;
165 }
166 for (int i = 0; i < scan_inst->sources; i++) {
167 if (scan_inst->src[i].file == VGRF &&
168 scan_inst->src[i].nr == inst->src[0].nr &&
169 regions_overlap(
170 scan_inst->src[i], scan_inst->size_read(i),
171 inst->src[0], inst->size_read(0))) {
172 if (scan_inst->opcode != BRW_OPCODE_MOV ||
173 !scan_inst->saturate ||
174 scan_inst->src[0].abs ||
175 scan_inst->src[0].negate ||
176 scan_inst->src[0].abs != inst->src[0].abs ||
177 scan_inst->src[0].negate != inst->src[0].negate) {
178 interfered = true;
179 break;
180 }
181 }
182 }
183
184 if (interfered)
185 break;
186 }
187 }
188
189 return progress;
190 }
191
192 bool
brw_fs_opt_saturate_propagation(fs_visitor & s)193 brw_fs_opt_saturate_propagation(fs_visitor &s)
194 {
195 bool progress = false;
196
197 foreach_block (block, s.cfg) {
198 progress = opt_saturate_propagation_local(s, block) || progress;
199 }
200
201 /* Live intervals are still valid. */
202
203 return progress;
204 }
205