373aa2d5e3f35e9c0dc224e711e8d75ee4abc8ab
[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 /** @file brw_fs_live_variables.cpp
34 *
35 * Support for computing at the basic block level which variables
36 * (virtual GRFs in our case) are live at entry and exit.
37 *
38 * See Muchnik's Advanced Compiler Design and Implementation, section
39 * 14.1 (p444).
40 */
41
42 /**
43 * Sets up the use[] and def[] bitsets.
44 *
45 * The basic-block-level live variable analysis needs to know which
46 * variables get used before they're completely defined, and which
47 * variables are completely defined before they're used.
48 */
49 void
50 fs_live_variables::setup_def_use()
51 {
52 int ip = 0;
53
54 for (int b = 0; b < cfg->num_blocks; b++) {
55 bblock_t *block = cfg->blocks[b];
56
57 assert(ip == block->start_ip);
58 if (b > 0)
59 assert(cfg->blocks[b - 1]->end_ip == ip - 1);
60
61 for (fs_inst *inst = (fs_inst *)block->start;
62 inst != block->end->next;
63 inst = (fs_inst *)inst->next) {
64
65 /* Set use[] for this instruction */
66 for (unsigned int i = 0; i < 3; i++) {
67 if (inst->src[i].file == GRF) {
68 int reg = inst->src[i].reg;
69
70 if (!BITSET_TEST(bd[b].def, reg))
71 BITSET_SET(bd[b].use, reg);
72 }
73 }
74
75 /* Check for unconditional writes to whole registers. These
76 * are the things that screen off preceding definitions of a
77 * variable, and thus qualify for being in def[].
78 */
79 if (inst->dst.file == GRF &&
80 inst->regs_written == v->virtual_grf_sizes[inst->dst.reg] &&
81 !inst->predicate &&
82 !inst->force_uncompressed &&
83 !inst->force_sechalf) {
84 int reg = inst->dst.reg;
85 if (!BITSET_TEST(bd[b].use, reg))
86 BITSET_SET(bd[b].def, reg);
87 }
88
89 ip++;
90 }
91 }
92 }
93
94 /**
95 * The algorithm incrementally sets bits in liveout and livein,
96 * propagating it through control flow. It will eventually terminate
97 * because it only ever adds bits, and stops when no bits are added in
98 * a pass.
99 */
100 void
101 fs_live_variables::compute_live_variables()
102 {
103 bool cont = true;
104
105 while (cont) {
106 cont = false;
107
108 for (int b = 0; b < cfg->num_blocks; b++) {
109 /* Update livein */
110 for (int i = 0; i < bitset_words; i++) {
111 BITSET_WORD new_livein = (bd[b].use[i] |
112 (bd[b].liveout[i] & ~bd[b].def[i]));
113 if (new_livein & ~bd[b].livein[i]) {
114 bd[b].livein[i] |= new_livein;
115 cont = true;
116 }
117 }
118
119 /* Update liveout */
120 foreach_list(block_node, &cfg->blocks[b]->children) {
121 bblock_link *link = (bblock_link *)block_node;
122 bblock_t *block = link->block;
123
124 for (int i = 0; i < bitset_words; i++) {
125 BITSET_WORD new_liveout = (bd[block->block_num].livein[i] &
126 ~bd[b].liveout[i]);
127 if (new_liveout) {
128 bd[b].liveout[i] |= new_liveout;
129 cont = true;
130 }
131 }
132 }
133 }
134 }
135 }
136
137 fs_live_variables::fs_live_variables(fs_visitor *v, cfg_t *cfg)
138 : v(v), cfg(cfg)
139 {
140 mem_ctx = ralloc_context(cfg->mem_ctx);
141
142 num_vars = v->virtual_grf_count;
143 bd = rzalloc_array(mem_ctx, struct block_data, cfg->num_blocks);
144
145 bitset_words = (ALIGN(v->virtual_grf_count, BITSET_WORDBITS) /
146 BITSET_WORDBITS);
147 for (int i = 0; i < cfg->num_blocks; i++) {
148 bd[i].def = rzalloc_array(mem_ctx, BITSET_WORD, bitset_words);
149 bd[i].use = rzalloc_array(mem_ctx, BITSET_WORD, bitset_words);
150 bd[i].livein = rzalloc_array(mem_ctx, BITSET_WORD, bitset_words);
151 bd[i].liveout = rzalloc_array(mem_ctx, BITSET_WORD, bitset_words);
152 }
153
154 setup_def_use();
155 compute_live_variables();
156 }
157
158 fs_live_variables::~fs_live_variables()
159 {
160 ralloc_free(mem_ctx);
161 }
162
163 #define MAX_INSTRUCTION (1 << 30)
164
165 void
166 fs_visitor::calculate_live_intervals()
167 {
168 int num_vars = this->virtual_grf_count;
169
170 if (this->live_intervals_valid)
171 return;
172
173 int *def = ralloc_array(mem_ctx, int, num_vars);
174 int *use = ralloc_array(mem_ctx, int, num_vars);
175 ralloc_free(this->virtual_grf_def);
176 ralloc_free(this->virtual_grf_use);
177 this->virtual_grf_def = def;
178 this->virtual_grf_use = use;
179
180 for (int i = 0; i < num_vars; i++) {
181 def[i] = MAX_INSTRUCTION;
182 use[i] = -1;
183 }
184
185 /* Start by setting up the intervals with no knowledge of control
186 * flow.
187 */
188 int ip = 0;
189 foreach_list(node, &this->instructions) {
190 fs_inst *inst = (fs_inst *)node;
191
192 for (unsigned int i = 0; i < 3; i++) {
193 if (inst->src[i].file == GRF) {
194 int reg = inst->src[i].reg;
195
196 use[reg] = ip;
197
198 /* In most cases, a register can be written over safely by the
199 * same instruction that is its last use. For a single
200 * instruction, the sources are dereferenced before writing of the
201 * destination starts (naturally). This gets more complicated for
202 * simd16, because the instruction:
203 *
204 * mov(16) g4<1>F g4<8,8,1>F g6<8,8,1>F
205 *
206 * is actually decoded in hardware as:
207 *
208 * mov(8) g4<1>F g4<8,8,1>F g6<8,8,1>F
209 * mov(8) g5<1>F g5<8,8,1>F g7<8,8,1>F
210 *
211 * Which is safe. However, if we have uniform accesses
212 * happening, we get into trouble:
213 *
214 * mov(8) g4<1>F g4<0,1,0>F g6<8,8,1>F
215 * mov(8) g5<1>F g4<0,1,0>F g7<8,8,1>F
216 *
217 * Now our destination for the first instruction overwrote the
218 * second instruction's src0, and we get garbage for those 8
219 * pixels. There's a similar issue for the pre-gen6
220 * pixel_x/pixel_y, which are registers of 16-bit values and thus
221 * would get stomped by the first decode as well.
222 */
223 if (dispatch_width == 16 && (inst->src[i].smear ||
224 (this->pixel_x.reg == reg ||
225 this->pixel_y.reg == reg))) {
226 use[reg]++;
227 }
228 }
229 }
230
231 if (inst->dst.file == GRF) {
232 int reg = inst->dst.reg;
233
234 def[reg] = MIN2(def[reg], ip);
235 }
236
237 ip++;
238 }
239
240 /* Now, extend those intervals using our analysis of control flow. */
241 cfg_t cfg(this);
242 fs_live_variables livevars(this, &cfg);
243
244 for (int b = 0; b < cfg.num_blocks; b++) {
245 for (int i = 0; i < num_vars; i++) {
246 if (BITSET_TEST(livevars.bd[b].livein, i)) {
247 def[i] = MIN2(def[i], cfg.blocks[b]->start_ip);
248 use[i] = MAX2(use[i], cfg.blocks[b]->start_ip);
249 }
250
251 if (BITSET_TEST(livevars.bd[b].liveout, i)) {
252 def[i] = MIN2(def[i], cfg.blocks[b]->end_ip);
253 use[i] = MAX2(use[i], cfg.blocks[b]->end_ip);
254 }
255 }
256 }
257
258 this->live_intervals_valid = true;
259
260 /* Note in the non-control-flow code above, that we only take def[] as the
261 * first store, and use[] as the last use. We use this in dead code
262 * elimination, to determine when a store never gets used. However, we
263 * also use these arrays to answer the virtual_grf_interferes() question
264 * (live interval analysis), which is used for register coalescing and
265 * register allocation.
266 *
267 * So, there's a conflict over what the array should mean: if use[]
268 * considers a def after the last use, then the dead code elimination pass
269 * never does anything (and it's an important pass!). But if we don't
270 * include dead code, then virtual_grf_interferes() lies and we'll do
271 * horrible things like coalesce the register that is dead-code-written
272 * into another register that was live across the dead write (causing the
273 * use of the second register to take the dead write's source value instead
274 * of the coalesced MOV's source value).
275 *
276 * To resolve the conflict, immediately after calculating live intervals,
277 * detect dead code, nuke it, and if we changed anything, calculate again
278 * before returning to the caller. Now we happen to produce def[] and
279 * use[] arrays that will work for virtual_grf_interferes().
280 */
281 if (dead_code_eliminate())
282 calculate_live_intervals();
283 }
284
285 bool
286 fs_visitor::virtual_grf_interferes(int a, int b)
287 {
288 int a_def = this->virtual_grf_def[a], a_use = this->virtual_grf_use[a];
289 int b_def = this->virtual_grf_def[b], b_use = this->virtual_grf_use[b];
290
291 /* If there's dead code (def but not use), it would break our test
292 * unless we consider it used.
293 */
294 if ((a_use == -1 && a_def != MAX_INSTRUCTION) ||
295 (b_use == -1 && b_def != MAX_INSTRUCTION)) {
296 return true;
297 }
298
299 int start = MAX2(a_def, b_def);
300 int end = MIN2(a_use, b_use);
301
302 return start < end;
303 }