1 /*
2 * Copyright © 2012, 2013, 2014 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_vec4.h"
25 #include "brw_vec4_live_variables.h"
26 #include "brw_cfg.h"
27
28 using namespace brw;
29
30 /** @file brw_vec4_cse.cpp
31 *
32 * Support for local common subexpression elimination.
33 *
34 * See Muchnick's Advanced Compiler Design and Implementation, section
35 * 13.1 (p378).
36 */
37
38 namespace {
39 struct aeb_entry : public exec_node {
40 /** The instruction that generates the expression value. */
41 vec4_instruction *generator;
42
43 /** The temporary where the value is stored. */
44 src_reg tmp;
45 };
46 }
47
48 static bool
is_expression(const vec4_instruction * const inst)49 is_expression(const vec4_instruction *const inst)
50 {
51 switch (inst->opcode) {
52 case BRW_OPCODE_MOV:
53 case BRW_OPCODE_SEL:
54 case BRW_OPCODE_NOT:
55 case BRW_OPCODE_AND:
56 case BRW_OPCODE_OR:
57 case BRW_OPCODE_XOR:
58 case BRW_OPCODE_SHR:
59 case BRW_OPCODE_SHL:
60 case BRW_OPCODE_ASR:
61 case BRW_OPCODE_CMP:
62 case BRW_OPCODE_CMPN:
63 case BRW_OPCODE_ADD:
64 case BRW_OPCODE_MUL:
65 case SHADER_OPCODE_MULH:
66 case BRW_OPCODE_FRC:
67 case BRW_OPCODE_RNDU:
68 case BRW_OPCODE_RNDD:
69 case BRW_OPCODE_RNDE:
70 case BRW_OPCODE_RNDZ:
71 case BRW_OPCODE_LINE:
72 case BRW_OPCODE_PLN:
73 case BRW_OPCODE_MAD:
74 case BRW_OPCODE_LRP:
75 case VEC4_OPCODE_UNPACK_UNIFORM:
76 case SHADER_OPCODE_FIND_LIVE_CHANNEL:
77 case SHADER_OPCODE_BROADCAST:
78 case TCS_OPCODE_SET_INPUT_URB_OFFSETS:
79 case TCS_OPCODE_SET_OUTPUT_URB_OFFSETS:
80 return true;
81 case SHADER_OPCODE_RCP:
82 case SHADER_OPCODE_RSQ:
83 case SHADER_OPCODE_SQRT:
84 case SHADER_OPCODE_EXP2:
85 case SHADER_OPCODE_LOG2:
86 case SHADER_OPCODE_POW:
87 case SHADER_OPCODE_INT_QUOTIENT:
88 case SHADER_OPCODE_INT_REMAINDER:
89 case SHADER_OPCODE_SIN:
90 case SHADER_OPCODE_COS:
91 return inst->mlen == 0;
92 default:
93 return false;
94 }
95 }
96
97 static bool
operands_match(const vec4_instruction * a,const vec4_instruction * b)98 operands_match(const vec4_instruction *a, const vec4_instruction *b)
99 {
100 const src_reg *xs = a->src;
101 const src_reg *ys = b->src;
102
103 if (a->opcode == BRW_OPCODE_MAD) {
104 return xs[0].equals(ys[0]) &&
105 ((xs[1].equals(ys[1]) && xs[2].equals(ys[2])) ||
106 (xs[2].equals(ys[1]) && xs[1].equals(ys[2])));
107 } else if (a->opcode == BRW_OPCODE_MOV &&
108 xs[0].file == IMM &&
109 xs[0].type == BRW_REGISTER_TYPE_VF) {
110 src_reg tmp_x = xs[0];
111 src_reg tmp_y = ys[0];
112
113 /* Smash out the values that are not part of the writemask. Otherwise
114 * the equals operator will fail due to mismatches in unused components.
115 */
116 const unsigned ab_writemask = a->dst.writemask & b->dst.writemask;
117 const uint32_t mask = ((ab_writemask & WRITEMASK_X) ? 0x000000ff : 0) |
118 ((ab_writemask & WRITEMASK_Y) ? 0x0000ff00 : 0) |
119 ((ab_writemask & WRITEMASK_Z) ? 0x00ff0000 : 0) |
120 ((ab_writemask & WRITEMASK_W) ? 0xff000000 : 0);
121
122 tmp_x.ud &= mask;
123 tmp_y.ud &= mask;
124
125 return tmp_x.equals(tmp_y);
126 } else if (!a->is_commutative()) {
127 return xs[0].equals(ys[0]) && xs[1].equals(ys[1]) && xs[2].equals(ys[2]);
128 } else {
129 return (xs[0].equals(ys[0]) && xs[1].equals(ys[1])) ||
130 (xs[1].equals(ys[0]) && xs[0].equals(ys[1]));
131 }
132 }
133
134 /**
135 * Checks if instructions match, exactly for sources, but loosely for
136 * destination writemasks.
137 *
138 * \param 'a' is the generating expression from the AEB entry.
139 * \param 'b' is the second occurrence of the expression that we're
140 * considering eliminating.
141 */
142 static bool
instructions_match(vec4_instruction * a,vec4_instruction * b)143 instructions_match(vec4_instruction *a, vec4_instruction *b)
144 {
145 return a->opcode == b->opcode &&
146 a->saturate == b->saturate &&
147 a->predicate == b->predicate &&
148 a->predicate_inverse == b->predicate_inverse &&
149 a->conditional_mod == b->conditional_mod &&
150 a->flag_subreg == b->flag_subreg &&
151 a->dst.type == b->dst.type &&
152 a->offset == b->offset &&
153 a->mlen == b->mlen &&
154 a->base_mrf == b->base_mrf &&
155 a->header_size == b->header_size &&
156 a->shadow_compare == b->shadow_compare &&
157 ((a->dst.writemask & b->dst.writemask) == a->dst.writemask) &&
158 a->force_writemask_all == b->force_writemask_all &&
159 a->size_written == b->size_written &&
160 a->exec_size == b->exec_size &&
161 a->group == b->group &&
162 operands_match(a, b);
163 }
164
165 bool
opt_cse_local(bblock_t * block,const vec4_live_variables & live)166 vec4_visitor::opt_cse_local(bblock_t *block, const vec4_live_variables &live)
167 {
168 bool progress = false;
169 exec_list aeb;
170
171 void *cse_ctx = ralloc_context(NULL);
172
173 int ip = block->start_ip;
174 foreach_inst_in_block (vec4_instruction, inst, block) {
175 /* Skip some cases. */
176 if (is_expression(inst) && !inst->predicate && inst->mlen == 0 &&
177 ((inst->dst.file != ARF && inst->dst.file != FIXED_GRF) ||
178 inst->dst.is_null()))
179 {
180 bool found = false;
181
182 foreach_in_list_use_after(aeb_entry, entry, &aeb) {
183 /* Match current instruction's expression against those in AEB. */
184 if (!(entry->generator->dst.is_null() && !inst->dst.is_null()) &&
185 instructions_match(inst, entry->generator)) {
186 found = true;
187 progress = true;
188 break;
189 }
190 }
191
192 if (!found) {
193 if (inst->opcode != BRW_OPCODE_MOV ||
194 (inst->opcode == BRW_OPCODE_MOV &&
195 inst->src[0].file == IMM &&
196 inst->src[0].type == BRW_REGISTER_TYPE_VF)) {
197 /* Our first sighting of this expression. Create an entry. */
198 aeb_entry *entry = ralloc(cse_ctx, aeb_entry);
199 entry->tmp = src_reg(); /* file will be BAD_FILE */
200 entry->generator = inst;
201 aeb.push_tail(entry);
202 }
203 } else {
204 /* This is at least our second sighting of this expression.
205 * If we don't have a temporary already, make one.
206 */
207 bool no_existing_temp = entry->tmp.file == BAD_FILE;
208 if (no_existing_temp && !entry->generator->dst.is_null()) {
209 entry->tmp = retype(src_reg(VGRF, alloc.allocate(
210 regs_written(entry->generator)),
211 NULL), inst->dst.type);
212
213 const unsigned width = entry->generator->exec_size;
214 unsigned component_size = width * type_sz(entry->tmp.type);
215 unsigned num_copy_movs =
216 DIV_ROUND_UP(entry->generator->size_written, component_size);
217 for (unsigned i = 0; i < num_copy_movs; ++i) {
218 vec4_instruction *copy =
219 MOV(offset(entry->generator->dst, width, i),
220 offset(entry->tmp, width, i));
221 copy->exec_size = width;
222 copy->group = entry->generator->group;
223 copy->force_writemask_all =
224 entry->generator->force_writemask_all;
225 entry->generator->insert_after(block, copy);
226 }
227
228 entry->generator->dst = dst_reg(entry->tmp);
229 }
230
231 /* dest <- temp */
232 if (!inst->dst.is_null()) {
233 assert(inst->dst.type == entry->tmp.type);
234 const unsigned width = inst->exec_size;
235 unsigned component_size = width * type_sz(inst->dst.type);
236 unsigned num_copy_movs =
237 DIV_ROUND_UP(inst->size_written, component_size);
238 for (unsigned i = 0; i < num_copy_movs; ++i) {
239 vec4_instruction *copy =
240 MOV(offset(inst->dst, width, i),
241 offset(entry->tmp, width, i));
242 copy->exec_size = inst->exec_size;
243 copy->group = inst->group;
244 copy->force_writemask_all = inst->force_writemask_all;
245 inst->insert_before(block, copy);
246 }
247 }
248
249 /* Set our iterator so that next time through the loop inst->next
250 * will get the instruction in the basic block after the one we've
251 * removed.
252 */
253 vec4_instruction *prev = (vec4_instruction *)inst->prev;
254
255 inst->remove(block);
256 inst = prev;
257 }
258 }
259
260 foreach_in_list_safe(aeb_entry, entry, &aeb) {
261 /* Kill all AEB entries that write a different value to or read from
262 * the flag register if we just wrote it.
263 */
264 if (inst->writes_flag()) {
265 if (entry->generator->reads_flag() ||
266 (entry->generator->writes_flag() &&
267 !instructions_match(inst, entry->generator))) {
268 entry->remove();
269 ralloc_free(entry);
270 continue;
271 }
272 }
273
274 for (int i = 0; i < 3; i++) {
275 src_reg *src = &entry->generator->src[i];
276
277 /* Kill all AEB entries that use the destination we just
278 * overwrote.
279 */
280 if (inst->dst.file == entry->generator->src[i].file &&
281 inst->dst.nr == entry->generator->src[i].nr) {
282 entry->remove();
283 ralloc_free(entry);
284 break;
285 }
286
287 /* Kill any AEB entries using registers that don't get reused any
288 * more -- a sure sign they'll fail operands_match().
289 */
290 if (src->file == VGRF) {
291 if (live.var_range_end(var_from_reg(alloc, dst_reg(*src)), 8) < ip) {
292 entry->remove();
293 ralloc_free(entry);
294 break;
295 }
296 }
297 }
298 }
299
300 ip++;
301 }
302
303 ralloc_free(cse_ctx);
304
305 return progress;
306 }
307
308 bool
opt_cse()309 vec4_visitor::opt_cse()
310 {
311 bool progress = false;
312 const vec4_live_variables &live = live_analysis.require();
313
314 foreach_block (block, cfg) {
315 progress = opt_cse_local(block, live) || progress;
316 }
317
318 if (progress)
319 invalidate_analysis(DEPENDENCY_INSTRUCTIONS | DEPENDENCY_VARIABLES);
320
321 return progress;
322 }
323