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 #include "brw_fs.h"
25 #include "brw_cfg.h"
26
27 /** @file brw_fs_cse.cpp
28 *
29 * Support for local common subexpression elimination.
30 *
31 * See Muchnick's Advanced Compiler Design and Implementation, section
32 * 13.1 (p378).
33 */
34
35 using namespace brw;
36
37 namespace {
38 struct aeb_entry : public exec_node {
39 /** The instruction that generates the expression value. */
40 fs_inst *generator;
41
42 /** The temporary where the value is stored. */
43 fs_reg tmp;
44 };
45 }
46
47 static bool
is_expression(const fs_visitor * v,const fs_inst * const inst)48 is_expression(const fs_visitor *v, const fs_inst *const inst)
49 {
50 switch (inst->opcode) {
51 case BRW_OPCODE_MOV:
52 case BRW_OPCODE_SEL:
53 case BRW_OPCODE_NOT:
54 case BRW_OPCODE_AND:
55 case BRW_OPCODE_OR:
56 case BRW_OPCODE_XOR:
57 case BRW_OPCODE_SHR:
58 case BRW_OPCODE_SHL:
59 case BRW_OPCODE_ASR:
60 case BRW_OPCODE_CMP:
61 case BRW_OPCODE_CMPN:
62 case BRW_OPCODE_ADD:
63 case BRW_OPCODE_MUL:
64 case SHADER_OPCODE_MULH:
65 case BRW_OPCODE_FRC:
66 case BRW_OPCODE_RNDU:
67 case BRW_OPCODE_RNDD:
68 case BRW_OPCODE_RNDE:
69 case BRW_OPCODE_RNDZ:
70 case BRW_OPCODE_LINE:
71 case BRW_OPCODE_PLN:
72 case BRW_OPCODE_MAD:
73 case BRW_OPCODE_LRP:
74 case FS_OPCODE_FB_READ_LOGICAL:
75 case FS_OPCODE_UNIFORM_PULL_CONSTANT_LOAD:
76 case FS_OPCODE_VARYING_PULL_CONSTANT_LOAD_LOGICAL:
77 case FS_OPCODE_VARYING_PULL_CONSTANT_LOAD_GEN7:
78 case FS_OPCODE_CINTERP:
79 case FS_OPCODE_LINTERP:
80 case SHADER_OPCODE_FIND_LIVE_CHANNEL:
81 case SHADER_OPCODE_BROADCAST:
82 case SHADER_OPCODE_MOV_INDIRECT:
83 case SHADER_OPCODE_TEX_LOGICAL:
84 case SHADER_OPCODE_TXD_LOGICAL:
85 case SHADER_OPCODE_TXF_LOGICAL:
86 case SHADER_OPCODE_TXL_LOGICAL:
87 case SHADER_OPCODE_TXS_LOGICAL:
88 case FS_OPCODE_TXB_LOGICAL:
89 case SHADER_OPCODE_TXF_CMS_LOGICAL:
90 case SHADER_OPCODE_TXF_CMS_W_LOGICAL:
91 case SHADER_OPCODE_TXF_UMS_LOGICAL:
92 case SHADER_OPCODE_TXF_MCS_LOGICAL:
93 case SHADER_OPCODE_LOD_LOGICAL:
94 case SHADER_OPCODE_TG4_LOGICAL:
95 case SHADER_OPCODE_TG4_OFFSET_LOGICAL:
96 case FS_OPCODE_PACK:
97 return true;
98 case SHADER_OPCODE_RCP:
99 case SHADER_OPCODE_RSQ:
100 case SHADER_OPCODE_SQRT:
101 case SHADER_OPCODE_EXP2:
102 case SHADER_OPCODE_LOG2:
103 case SHADER_OPCODE_POW:
104 case SHADER_OPCODE_INT_QUOTIENT:
105 case SHADER_OPCODE_INT_REMAINDER:
106 case SHADER_OPCODE_SIN:
107 case SHADER_OPCODE_COS:
108 return inst->mlen < 2;
109 case SHADER_OPCODE_LOAD_PAYLOAD:
110 return !inst->is_copy_payload(v->alloc);
111 default:
112 return inst->is_send_from_grf() && !inst->has_side_effects() &&
113 !inst->is_volatile();
114 }
115 }
116
117 static bool
operands_match(const fs_inst * a,const fs_inst * b,bool * negate)118 operands_match(const fs_inst *a, const fs_inst *b, bool *negate)
119 {
120 fs_reg *xs = a->src;
121 fs_reg *ys = b->src;
122
123 if (a->opcode == BRW_OPCODE_MAD) {
124 return xs[0].equals(ys[0]) &&
125 ((xs[1].equals(ys[1]) && xs[2].equals(ys[2])) ||
126 (xs[2].equals(ys[1]) && xs[1].equals(ys[2])));
127 } else if (a->opcode == BRW_OPCODE_MUL && a->dst.type == BRW_REGISTER_TYPE_F) {
128 bool xs0_negate = xs[0].negate;
129 bool xs1_negate = xs[1].file == IMM ? xs[1].f < 0.0f
130 : xs[1].negate;
131 bool ys0_negate = ys[0].negate;
132 bool ys1_negate = ys[1].file == IMM ? ys[1].f < 0.0f
133 : ys[1].negate;
134 float xs1_imm = xs[1].f;
135 float ys1_imm = ys[1].f;
136
137 xs[0].negate = false;
138 xs[1].negate = false;
139 ys[0].negate = false;
140 ys[1].negate = false;
141 xs[1].f = fabsf(xs[1].f);
142 ys[1].f = fabsf(ys[1].f);
143
144 bool ret = (xs[0].equals(ys[0]) && xs[1].equals(ys[1])) ||
145 (xs[1].equals(ys[0]) && xs[0].equals(ys[1]));
146
147 xs[0].negate = xs0_negate;
148 xs[1].negate = xs[1].file == IMM ? false : xs1_negate;
149 ys[0].negate = ys0_negate;
150 ys[1].negate = ys[1].file == IMM ? false : ys1_negate;
151 xs[1].f = xs1_imm;
152 ys[1].f = ys1_imm;
153
154 *negate = (xs0_negate != xs1_negate) != (ys0_negate != ys1_negate);
155 if (*negate && (a->saturate || b->saturate))
156 return false;
157 return ret;
158 } else if (!a->is_commutative()) {
159 bool match = true;
160 for (int i = 0; i < a->sources; i++) {
161 if (!xs[i].equals(ys[i])) {
162 match = false;
163 break;
164 }
165 }
166 return match;
167 } else {
168 return (xs[0].equals(ys[0]) && xs[1].equals(ys[1])) ||
169 (xs[1].equals(ys[0]) && xs[0].equals(ys[1]));
170 }
171 }
172
173 static bool
instructions_match(fs_inst * a,fs_inst * b,bool * negate)174 instructions_match(fs_inst *a, fs_inst *b, bool *negate)
175 {
176 return a->opcode == b->opcode &&
177 a->force_writemask_all == b->force_writemask_all &&
178 a->exec_size == b->exec_size &&
179 a->group == b->group &&
180 a->saturate == b->saturate &&
181 a->predicate == b->predicate &&
182 a->predicate_inverse == b->predicate_inverse &&
183 a->conditional_mod == b->conditional_mod &&
184 a->flag_subreg == b->flag_subreg &&
185 a->dst.type == b->dst.type &&
186 a->offset == b->offset &&
187 a->mlen == b->mlen &&
188 a->size_written == b->size_written &&
189 a->base_mrf == b->base_mrf &&
190 a->eot == b->eot &&
191 a->header_size == b->header_size &&
192 a->shadow_compare == b->shadow_compare &&
193 a->pi_noperspective == b->pi_noperspective &&
194 a->target == b->target &&
195 a->sources == b->sources &&
196 operands_match(a, b, negate);
197 }
198
199 static void
create_copy_instr(const fs_builder & bld,fs_inst * inst,fs_reg src,bool negate)200 create_copy_instr(const fs_builder &bld, fs_inst *inst, fs_reg src, bool negate)
201 {
202 unsigned written = regs_written(inst);
203 unsigned dst_width =
204 DIV_ROUND_UP(inst->dst.component_size(inst->exec_size), REG_SIZE);
205 fs_inst *copy;
206
207 if (inst->opcode == SHADER_OPCODE_LOAD_PAYLOAD ||
208 written != dst_width) {
209 fs_reg *payload;
210 int sources, header_size;
211 if (inst->opcode == SHADER_OPCODE_LOAD_PAYLOAD) {
212 sources = inst->sources;
213 header_size = inst->header_size;
214 } else {
215 assert(written % dst_width == 0);
216 sources = written / dst_width;
217 header_size = 0;
218 }
219
220 assert(src.file == VGRF);
221 payload = ralloc_array(bld.shader->mem_ctx, fs_reg, sources);
222 for (int i = 0; i < header_size; i++) {
223 payload[i] = src;
224 src.offset += REG_SIZE;
225 }
226 for (int i = header_size; i < sources; i++) {
227 payload[i] = src;
228 src = offset(src, bld, 1);
229 }
230 copy = bld.LOAD_PAYLOAD(inst->dst, payload, sources, header_size);
231 } else {
232 copy = bld.MOV(inst->dst, src);
233 copy->group = inst->group;
234 copy->force_writemask_all = inst->force_writemask_all;
235 copy->src[0].negate = negate;
236 }
237 assert(regs_written(copy) == written);
238 }
239
240 bool
opt_cse_local(bblock_t * block)241 fs_visitor::opt_cse_local(bblock_t *block)
242 {
243 bool progress = false;
244 exec_list aeb;
245
246 void *cse_ctx = ralloc_context(NULL);
247
248 int ip = block->start_ip;
249 foreach_inst_in_block(fs_inst, inst, block) {
250 /* Skip some cases. */
251 if (is_expression(this, inst) && !inst->is_partial_write() &&
252 ((inst->dst.file != ARF && inst->dst.file != FIXED_GRF) ||
253 inst->dst.is_null()))
254 {
255 bool found = false;
256 bool negate = false;
257
258 foreach_in_list_use_after(aeb_entry, entry, &aeb) {
259 /* Match current instruction's expression against those in AEB. */
260 if (!(entry->generator->dst.is_null() && !inst->dst.is_null()) &&
261 instructions_match(inst, entry->generator, &negate)) {
262 found = true;
263 progress = true;
264 break;
265 }
266 }
267
268 if (!found) {
269 if (inst->opcode != BRW_OPCODE_MOV ||
270 (inst->opcode == BRW_OPCODE_MOV &&
271 inst->src[0].file == IMM &&
272 inst->src[0].type == BRW_REGISTER_TYPE_VF)) {
273 /* Our first sighting of this expression. Create an entry. */
274 aeb_entry *entry = ralloc(cse_ctx, aeb_entry);
275 entry->tmp = reg_undef;
276 entry->generator = inst;
277 aeb.push_tail(entry);
278 }
279 } else {
280 /* This is at least our second sighting of this expression.
281 * If we don't have a temporary already, make one.
282 */
283 bool no_existing_temp = entry->tmp.file == BAD_FILE;
284 if (no_existing_temp && !entry->generator->dst.is_null()) {
285 const fs_builder ibld = fs_builder(this, block, entry->generator)
286 .at(block, entry->generator->next);
287 int written = regs_written(entry->generator);
288
289 entry->tmp = fs_reg(VGRF, alloc.allocate(written),
290 entry->generator->dst.type);
291
292 create_copy_instr(ibld, entry->generator, entry->tmp, false);
293
294 entry->generator->dst = entry->tmp;
295 }
296
297 /* dest <- temp */
298 if (!inst->dst.is_null()) {
299 assert(inst->size_written == entry->generator->size_written);
300 assert(inst->dst.type == entry->tmp.type);
301 const fs_builder ibld(this, block, inst);
302
303 create_copy_instr(ibld, inst, entry->tmp, negate);
304 }
305
306 /* Set our iterator so that next time through the loop inst->next
307 * will get the instruction in the basic block after the one we've
308 * removed.
309 */
310 fs_inst *prev = (fs_inst *)inst->prev;
311
312 inst->remove(block);
313 inst = prev;
314 }
315 }
316
317 foreach_in_list_safe(aeb_entry, entry, &aeb) {
318 /* Kill all AEB entries that write a different value to or read from
319 * the flag register if we just wrote it.
320 */
321 if (inst->flags_written()) {
322 bool negate; /* dummy */
323 if (entry->generator->flags_read(devinfo) ||
324 (entry->generator->flags_written() &&
325 !instructions_match(inst, entry->generator, &negate))) {
326 entry->remove();
327 ralloc_free(entry);
328 continue;
329 }
330 }
331
332 for (int i = 0; i < entry->generator->sources; i++) {
333 fs_reg *src_reg = &entry->generator->src[i];
334
335 /* Kill all AEB entries that use the destination we just
336 * overwrote.
337 */
338 if (regions_overlap(inst->dst, inst->size_written,
339 entry->generator->src[i],
340 entry->generator->size_read(i))) {
341 entry->remove();
342 ralloc_free(entry);
343 break;
344 }
345
346 /* Kill any AEB entries using registers that don't get reused any
347 * more -- a sure sign they'll fail operands_match().
348 */
349 if (src_reg->file == VGRF && virtual_grf_end[src_reg->nr] < ip) {
350 entry->remove();
351 ralloc_free(entry);
352 break;
353 }
354 }
355 }
356
357 ip++;
358 }
359
360 ralloc_free(cse_ctx);
361
362 return progress;
363 }
364
365 bool
opt_cse()366 fs_visitor::opt_cse()
367 {
368 bool progress = false;
369
370 calculate_live_intervals();
371
372 foreach_block (block, cfg) {
373 progress = opt_cse_local(block) || progress;
374 }
375
376 if (progress)
377 invalidate_live_intervals();
378
379 return progress;
380 }
381