i965/fs_reg: Allocate double the number of vgrfs in SIMD16 mode
[mesa.git] / src / mesa / drivers / dri / i965 / brw_fs_register_coalesce.cpp
1 /*
2 * Copyright © 2012 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 /** @file brw_fs_register_coalesce.cpp
25 *
26 * Implements register coalescing: Checks if the two registers involved in a
27 * raw move don't interfere, in which case they can both be stored in the same
28 * place and the MOV removed.
29 *
30 * To do this, all uses of the source of the MOV in the shader are replaced
31 * with the destination of the MOV. For example:
32 *
33 * add vgrf3:F, vgrf1:F, vgrf2:F
34 * mov vgrf4:F, vgrf3:F
35 * mul vgrf5:F, vgrf5:F, vgrf4:F
36 *
37 * becomes
38 *
39 * add vgrf4:F, vgrf1:F, vgrf2:F
40 * mul vgrf5:F, vgrf5:F, vgrf4:F
41 */
42
43 #include "brw_fs.h"
44 #include "brw_cfg.h"
45 #include "brw_fs_live_variables.h"
46
47 static bool
48 is_nop_mov(const fs_inst *inst)
49 {
50 if (inst->opcode == SHADER_OPCODE_LOAD_PAYLOAD) {
51 fs_reg dst = inst->dst;
52 for (int i = 0; i < inst->sources; i++) {
53 dst.reg_offset = i;
54 if (!dst.equals(inst->src[i])) {
55 return false;
56 }
57 }
58 return true;
59 } else if (inst->opcode == BRW_OPCODE_MOV) {
60 return inst->dst.equals(inst->src[0]);
61 }
62
63 return false;
64 }
65
66 static bool
67 is_copy_payload(const fs_visitor *v, const fs_inst *inst)
68 {
69 if (v->virtual_grf_sizes[inst->src[0].reg] != inst->regs_written)
70 return false;
71
72 fs_reg reg = inst->src[0];
73
74 for (int i = 0; i < inst->sources; i++)
75 if (!inst->src[i].equals(offset(reg, i)))
76 return false;
77
78 return true;
79 }
80
81 static bool
82 is_coalesce_candidate(const fs_visitor *v, const fs_inst *inst)
83 {
84 if ((inst->opcode != BRW_OPCODE_MOV &&
85 inst->opcode != SHADER_OPCODE_LOAD_PAYLOAD) ||
86 inst->is_partial_write() ||
87 inst->saturate ||
88 inst->src[0].file != GRF ||
89 inst->src[0].negate ||
90 inst->src[0].abs ||
91 !inst->src[0].is_contiguous() ||
92 inst->dst.file != GRF ||
93 inst->dst.type != inst->src[0].type) {
94 return false;
95 }
96
97 if (v->virtual_grf_sizes[inst->src[0].reg] >
98 v->virtual_grf_sizes[inst->dst.reg])
99 return false;
100
101 if (inst->opcode == SHADER_OPCODE_LOAD_PAYLOAD) {
102 if (!is_copy_payload(v, inst)) {
103 return false;
104 }
105 }
106
107 return true;
108 }
109
110 static bool
111 can_coalesce_vars(brw::fs_live_variables *live_intervals,
112 const cfg_t *cfg, const fs_inst *inst,
113 int var_to, int var_from)
114 {
115 if (!live_intervals->vars_interfere(var_from, var_to))
116 return true;
117
118 int start_to = live_intervals->start[var_to];
119 int end_to = live_intervals->end[var_to];
120 int start_from = live_intervals->start[var_from];
121 int end_from = live_intervals->end[var_from];
122
123 /* Variables interfere and one line range isn't a subset of the other. */
124 if ((end_to > end_from && start_from < start_to) ||
125 (end_from > end_to && start_to < start_from))
126 return false;
127
128 int start_ip = MIN2(start_to, start_from);
129 int scan_ip = -1;
130
131 foreach_block_and_inst(block, fs_inst, scan_inst, cfg) {
132 scan_ip++;
133
134 if (scan_ip < start_ip)
135 continue;
136
137 if (scan_inst->is_control_flow())
138 return false;
139
140 if (scan_ip <= live_intervals->start[var_to])
141 continue;
142
143 if (scan_ip > live_intervals->end[var_to])
144 return true;
145
146 if (scan_inst->dst.equals(inst->dst) ||
147 scan_inst->dst.equals(inst->src[0]))
148 return false;
149 }
150
151 return true;
152 }
153
154 bool
155 fs_visitor::register_coalesce()
156 {
157 bool progress = false;
158
159 calculate_live_intervals();
160
161 int src_size = 0;
162 int channels_remaining = 0;
163 int reg_from = -1, reg_to = -1;
164 int reg_to_offset[MAX_SAMPLER_MESSAGE_SIZE];
165 fs_inst *mov[MAX_SAMPLER_MESSAGE_SIZE];
166 int var_to[MAX_SAMPLER_MESSAGE_SIZE];
167 int var_from[MAX_SAMPLER_MESSAGE_SIZE];
168
169 foreach_block_and_inst(block, fs_inst, inst, cfg) {
170 if (!is_coalesce_candidate(this, inst))
171 continue;
172
173 if (is_nop_mov(inst)) {
174 inst->opcode = BRW_OPCODE_NOP;
175 progress = true;
176 continue;
177 }
178
179 if (reg_from != inst->src[0].reg) {
180 reg_from = inst->src[0].reg;
181
182 src_size = virtual_grf_sizes[inst->src[0].reg];
183 assert(src_size <= MAX_SAMPLER_MESSAGE_SIZE);
184
185 assert(inst->src[0].width % 8 == 0);
186 channels_remaining = src_size;
187 memset(mov, 0, sizeof(mov));
188
189 reg_to = inst->dst.reg;
190 }
191
192 if (reg_to != inst->dst.reg)
193 continue;
194
195 if (inst->opcode == SHADER_OPCODE_LOAD_PAYLOAD) {
196 for (int i = 0; i < src_size; i++) {
197 reg_to_offset[i] = i;
198 }
199 mov[0] = inst;
200 channels_remaining -= inst->regs_written;
201 } else {
202 const int offset = inst->src[0].reg_offset;
203 reg_to_offset[offset] = inst->dst.reg_offset;
204 if (inst->src[0].width == 16)
205 reg_to_offset[offset + 1] = inst->dst.reg_offset + 1;
206 mov[offset] = inst;
207 channels_remaining -= inst->regs_written;
208 }
209
210 if (channels_remaining)
211 continue;
212
213 bool can_coalesce = true;
214 for (int i = 0; i < src_size; i++) {
215 var_to[i] = live_intervals->var_from_vgrf[reg_to] + reg_to_offset[i];
216 var_from[i] = live_intervals->var_from_vgrf[reg_from] + i;
217
218 if (!can_coalesce_vars(live_intervals, cfg, inst,
219 var_to[i], var_from[i])) {
220 can_coalesce = false;
221 reg_from = -1;
222 break;
223 }
224 }
225
226 if (!can_coalesce)
227 continue;
228
229 progress = true;
230 bool was_load_payload = inst->opcode == SHADER_OPCODE_LOAD_PAYLOAD;
231
232 for (int i = 0; i < src_size; i++) {
233 if (mov[i]) {
234 mov[i]->opcode = BRW_OPCODE_NOP;
235 mov[i]->conditional_mod = BRW_CONDITIONAL_NONE;
236 mov[i]->dst = reg_undef;
237 for (int j = 0; j < mov[i]->sources; j++) {
238 mov[i]->src[j] = reg_undef;
239 }
240 }
241 }
242
243 foreach_block_and_inst(block, fs_inst, scan_inst, cfg) {
244 for (int i = 0; i < src_size; i++) {
245 if (mov[i] || was_load_payload) {
246 if (scan_inst->dst.file == GRF &&
247 scan_inst->dst.reg == reg_from &&
248 scan_inst->dst.reg_offset == i) {
249 scan_inst->dst.reg = reg_to;
250 scan_inst->dst.reg_offset = reg_to_offset[i];
251 }
252 for (int j = 0; j < scan_inst->sources; j++) {
253 if (scan_inst->src[j].file == GRF &&
254 scan_inst->src[j].reg == reg_from &&
255 scan_inst->src[j].reg_offset == i) {
256 scan_inst->src[j].reg = reg_to;
257 scan_inst->src[j].reg_offset = reg_to_offset[i];
258 }
259 }
260 }
261 }
262 }
263
264 for (int i = 0; i < src_size; i++) {
265 live_intervals->start[var_to[i]] =
266 MIN2(live_intervals->start[var_to[i]],
267 live_intervals->start[var_from[i]]);
268 live_intervals->end[var_to[i]] =
269 MAX2(live_intervals->end[var_to[i]],
270 live_intervals->end[var_from[i]]);
271 }
272 reg_from = -1;
273 }
274
275 if (progress) {
276 foreach_block_and_inst_safe (block, backend_instruction, inst, cfg) {
277 if (inst->opcode == BRW_OPCODE_NOP) {
278 inst->remove(block);
279 }
280 }
281
282 invalidate_live_intervals();
283 }
284
285 return progress;
286 }