i965/gen8: Add instruction compaction tables.
[mesa.git] / src / mesa / drivers / dri / i965 / brw_fs_live_variables.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 * Authors:
24 * Eric Anholt <eric@anholt.net>
25 *
26 */
27
28 #include "brw_cfg.h"
29 #include "brw_fs_live_variables.h"
30
31 using namespace brw;
32
33 #define MAX_INSTRUCTION (1 << 30)
34
35 /** @file brw_fs_live_variables.cpp
36 *
37 * Support for calculating liveness information about virtual GRFs.
38 *
39 * This produces a live interval for each whole virtual GRF. We could
40 * choose to expose per-component live intervals for VGRFs of size > 1,
41 * but we currently do not. It is easier for the consumers of this
42 * information to work with whole VGRFs.
43 *
44 * However, we internally track use/def information at the per-component
45 * (reg_offset) level for greater accuracy. Large VGRFs may be accessed
46 * piecemeal over many (possibly non-adjacent) instructions. In this case,
47 * examining a single instruction is insufficient to decide whether a whole
48 * VGRF is ultimately used or defined. Tracking individual components
49 * allows us to easily assemble this information.
50 *
51 * See Muchnick's Advanced Compiler Design and Implementation, section
52 * 14.1 (p444).
53 */
54
55 void
56 fs_live_variables::setup_one_read(bblock_t *block, fs_inst *inst,
57 int ip, fs_reg reg)
58 {
59 int var = var_from_vgrf[reg.reg] + reg.reg_offset;
60 assert(var < num_vars);
61
62 /* In most cases, a register can be written over safely by the
63 * same instruction that is its last use. For a single
64 * instruction, the sources are dereferenced before writing of the
65 * destination starts (naturally). This gets more complicated for
66 * simd16, because the instruction:
67 *
68 * add(16) g4<1>F g4<8,8,1>F g6<8,8,1>F
69 *
70 * is actually decoded in hardware as:
71 *
72 * add(8) g4<1>F g4<8,8,1>F g6<8,8,1>F
73 * add(8) g5<1>F g5<8,8,1>F g7<8,8,1>F
74 *
75 * Which is safe. However, if we have uniform accesses
76 * happening, we get into trouble:
77 *
78 * add(8) g4<1>F g4<0,1,0>F g6<8,8,1>F
79 * add(8) g5<1>F g4<0,1,0>F g7<8,8,1>F
80 *
81 * Now our destination for the first instruction overwrote the
82 * second instruction's src0, and we get garbage for those 8
83 * pixels. There's a similar issue for the pre-gen6
84 * pixel_x/pixel_y, which are registers of 16-bit values and thus
85 * would get stomped by the first decode as well.
86 */
87 int end_ip = ip;
88 if (v->dispatch_width == 16 && (reg.stride == 0 ||
89 reg.type == BRW_REGISTER_TYPE_UW ||
90 reg.type == BRW_REGISTER_TYPE_W ||
91 reg.type == BRW_REGISTER_TYPE_UB ||
92 reg.type == BRW_REGISTER_TYPE_B)) {
93 end_ip++;
94 }
95
96 start[var] = MIN2(start[var], ip);
97 end[var] = MAX2(end[var], end_ip);
98
99 /* The use[] bitset marks when the block makes use of a variable (VGRF
100 * channel) without having completely defined that variable within the
101 * block.
102 */
103 if (!BITSET_TEST(bd[block->block_num].def, var))
104 BITSET_SET(bd[block->block_num].use, var);
105 }
106
107 void
108 fs_live_variables::setup_one_write(bblock_t *block, fs_inst *inst,
109 int ip, fs_reg reg)
110 {
111 int var = var_from_vgrf[reg.reg] + reg.reg_offset;
112 assert(var < num_vars);
113
114 start[var] = MIN2(start[var], ip);
115 end[var] = MAX2(end[var], ip);
116
117 /* The def[] bitset marks when an initialization in a block completely
118 * screens off previous updates of that variable (VGRF channel).
119 */
120 if (inst->dst.file == GRF && !inst->is_partial_write()) {
121 if (!BITSET_TEST(bd[block->block_num].use, var))
122 BITSET_SET(bd[block->block_num].def, var);
123 }
124 }
125
126 /**
127 * Sets up the use[] and def[] bitsets.
128 *
129 * The basic-block-level live variable analysis needs to know which
130 * variables get used before they're completely defined, and which
131 * variables are completely defined before they're used.
132 *
133 * These are tracked at the per-component level, rather than whole VGRFs.
134 */
135 void
136 fs_live_variables::setup_def_use()
137 {
138 int ip = 0;
139
140 for (int b = 0; b < cfg->num_blocks; b++) {
141 bblock_t *block = cfg->blocks[b];
142
143 assert(ip == block->start_ip);
144 if (b > 0)
145 assert(cfg->blocks[b - 1]->end_ip == ip - 1);
146
147 foreach_inst_in_block(fs_inst, inst, block) {
148 /* Set use[] for this instruction */
149 for (unsigned int i = 0; i < inst->sources; i++) {
150 fs_reg reg = inst->src[i];
151
152 if (reg.file != GRF)
153 continue;
154
155 for (int j = 0; j < inst->regs_read(v, i); j++) {
156 setup_one_read(block, inst, ip, reg);
157 reg.reg_offset++;
158 }
159 }
160
161 /* Set def[] for this instruction */
162 if (inst->dst.file == GRF) {
163 fs_reg reg = inst->dst;
164 for (int j = 0; j < inst->regs_written; j++) {
165 setup_one_write(block, inst, ip, reg);
166 reg.reg_offset++;
167 }
168 }
169
170 ip++;
171 }
172 }
173 }
174
175 /**
176 * The algorithm incrementally sets bits in liveout and livein,
177 * propagating it through control flow. It will eventually terminate
178 * because it only ever adds bits, and stops when no bits are added in
179 * a pass.
180 */
181 void
182 fs_live_variables::compute_live_variables()
183 {
184 bool cont = true;
185
186 while (cont) {
187 cont = false;
188
189 for (int b = 0; b < cfg->num_blocks; b++) {
190 /* Update livein */
191 for (int i = 0; i < bitset_words; i++) {
192 BITSET_WORD new_livein = (bd[b].use[i] |
193 (bd[b].liveout[i] & ~bd[b].def[i]));
194 if (new_livein & ~bd[b].livein[i]) {
195 bd[b].livein[i] |= new_livein;
196 cont = true;
197 }
198 }
199
200 /* Update liveout */
201 foreach_list_typed(bblock_link, link, link, &cfg->blocks[b]->children) {
202 bblock_t *block = link->block;
203
204 for (int i = 0; i < bitset_words; i++) {
205 BITSET_WORD new_liveout = (bd[block->block_num].livein[i] &
206 ~bd[b].liveout[i]);
207 if (new_liveout) {
208 bd[b].liveout[i] |= new_liveout;
209 cont = true;
210 }
211 }
212 }
213 }
214 }
215 }
216
217 /**
218 * Extend the start/end ranges for each variable to account for the
219 * new information calculated from control flow.
220 */
221 void
222 fs_live_variables::compute_start_end()
223 {
224 for (int b = 0; b < cfg->num_blocks; b++) {
225 for (int i = 0; i < num_vars; i++) {
226 if (BITSET_TEST(bd[b].livein, i)) {
227 start[i] = MIN2(start[i], cfg->blocks[b]->start_ip);
228 end[i] = MAX2(end[i], cfg->blocks[b]->start_ip);
229 }
230
231 if (BITSET_TEST(bd[b].liveout, i)) {
232 start[i] = MIN2(start[i], cfg->blocks[b]->end_ip);
233 end[i] = MAX2(end[i], cfg->blocks[b]->end_ip);
234 }
235
236 }
237 }
238 }
239
240 int
241 fs_live_variables::var_from_reg(fs_reg *reg)
242 {
243 return var_from_vgrf[reg->reg] + reg->reg_offset;
244 }
245
246 fs_live_variables::fs_live_variables(fs_visitor *v, const cfg_t *cfg)
247 : v(v), cfg(cfg)
248 {
249 mem_ctx = ralloc_context(NULL);
250
251 num_vgrfs = v->virtual_grf_count;
252 num_vars = 0;
253 var_from_vgrf = rzalloc_array(mem_ctx, int, num_vgrfs);
254 for (int i = 0; i < num_vgrfs; i++) {
255 var_from_vgrf[i] = num_vars;
256 num_vars += v->virtual_grf_sizes[i];
257 }
258
259 vgrf_from_var = rzalloc_array(mem_ctx, int, num_vars);
260 for (int i = 0; i < num_vgrfs; i++) {
261 for (int j = 0; j < v->virtual_grf_sizes[i]; j++) {
262 vgrf_from_var[var_from_vgrf[i] + j] = i;
263 }
264 }
265
266 start = ralloc_array(mem_ctx, int, num_vars);
267 end = rzalloc_array(mem_ctx, int, num_vars);
268 for (int i = 0; i < num_vars; i++) {
269 start[i] = MAX_INSTRUCTION;
270 end[i] = -1;
271 }
272
273 bd = rzalloc_array(mem_ctx, struct block_data, cfg->num_blocks);
274
275 bitset_words = BITSET_WORDS(num_vars);
276 for (int i = 0; i < cfg->num_blocks; i++) {
277 bd[i].def = rzalloc_array(mem_ctx, BITSET_WORD, bitset_words);
278 bd[i].use = rzalloc_array(mem_ctx, BITSET_WORD, bitset_words);
279 bd[i].livein = rzalloc_array(mem_ctx, BITSET_WORD, bitset_words);
280 bd[i].liveout = rzalloc_array(mem_ctx, BITSET_WORD, bitset_words);
281 }
282
283 setup_def_use();
284 compute_live_variables();
285 compute_start_end();
286 }
287
288 fs_live_variables::~fs_live_variables()
289 {
290 ralloc_free(mem_ctx);
291 }
292
293 void
294 fs_visitor::invalidate_live_intervals()
295 {
296 ralloc_free(live_intervals);
297 live_intervals = NULL;
298
299 invalidate_cfg();
300 }
301
302 /**
303 * Compute the live intervals for each virtual GRF.
304 *
305 * This uses the per-component use/def data, but combines it to produce
306 * information about whole VGRFs.
307 */
308 void
309 fs_visitor::calculate_live_intervals()
310 {
311 if (this->live_intervals)
312 return;
313
314 int num_vgrfs = this->virtual_grf_count;
315 ralloc_free(this->virtual_grf_start);
316 ralloc_free(this->virtual_grf_end);
317 virtual_grf_start = ralloc_array(mem_ctx, int, num_vgrfs);
318 virtual_grf_end = ralloc_array(mem_ctx, int, num_vgrfs);
319
320 for (int i = 0; i < num_vgrfs; i++) {
321 virtual_grf_start[i] = MAX_INSTRUCTION;
322 virtual_grf_end[i] = -1;
323 }
324
325 calculate_cfg();
326 this->live_intervals = new(mem_ctx) fs_live_variables(this, cfg);
327
328 /* Merge the per-component live ranges to whole VGRF live ranges. */
329 for (int i = 0; i < live_intervals->num_vars; i++) {
330 int vgrf = live_intervals->vgrf_from_var[i];
331 virtual_grf_start[vgrf] = MIN2(virtual_grf_start[vgrf],
332 live_intervals->start[i]);
333 virtual_grf_end[vgrf] = MAX2(virtual_grf_end[vgrf],
334 live_intervals->end[i]);
335 }
336 }
337
338 bool
339 fs_live_variables::vars_interfere(int a, int b)
340 {
341 return !(end[b] <= start[a] ||
342 end[a] <= start[b]);
343 }
344
345 bool
346 fs_visitor::virtual_grf_interferes(int a, int b)
347 {
348 return !(virtual_grf_end[a] <= virtual_grf_start[b] ||
349 virtual_grf_end[b] <= virtual_grf_start[a]);
350 }