100e511a56c84b02b47e4bfa5a41ab9117dee3c0
[mesa.git] / src / mesa / drivers / dri / i965 / brw_vec4_cse.cpp
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
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 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 VEC4_OPCODE_UNPACK_UNIFORM:
75 return true;
76 case SHADER_OPCODE_RCP:
77 case SHADER_OPCODE_RSQ:
78 case SHADER_OPCODE_SQRT:
79 case SHADER_OPCODE_EXP2:
80 case SHADER_OPCODE_LOG2:
81 case SHADER_OPCODE_POW:
82 case SHADER_OPCODE_INT_QUOTIENT:
83 case SHADER_OPCODE_INT_REMAINDER:
84 case SHADER_OPCODE_SIN:
85 case SHADER_OPCODE_COS:
86 return inst->mlen == 0;
87 default:
88 return false;
89 }
90 }
91
92 static bool
93 operands_match(const vec4_instruction *a, const vec4_instruction *b)
94 {
95 const src_reg *xs = a->src;
96 const src_reg *ys = b->src;
97
98 if (a->opcode == BRW_OPCODE_MAD) {
99 return xs[0].equals(ys[0]) &&
100 ((xs[1].equals(ys[1]) && xs[2].equals(ys[2])) ||
101 (xs[2].equals(ys[1]) && xs[1].equals(ys[2])));
102 } else if (!a->is_commutative()) {
103 return xs[0].equals(ys[0]) && xs[1].equals(ys[1]) && xs[2].equals(ys[2]);
104 } else {
105 return (xs[0].equals(ys[0]) && xs[1].equals(ys[1])) ||
106 (xs[1].equals(ys[0]) && xs[0].equals(ys[1]));
107 }
108 }
109
110 static bool
111 instructions_match(vec4_instruction *a, vec4_instruction *b)
112 {
113 return a->opcode == b->opcode &&
114 a->saturate == b->saturate &&
115 a->conditional_mod == b->conditional_mod &&
116 a->dst.type == b->dst.type &&
117 a->dst.writemask == b->dst.writemask &&
118 a->force_writemask_all == b->force_writemask_all &&
119 a->regs_written == b->regs_written &&
120 operands_match(a, b);
121 }
122
123 bool
124 vec4_visitor::opt_cse_local(bblock_t *block)
125 {
126 bool progress = false;
127 exec_list aeb;
128
129 void *cse_ctx = ralloc_context(NULL);
130
131 int ip = block->start_ip;
132 foreach_inst_in_block (vec4_instruction, inst, block) {
133 /* Skip some cases. */
134 if (is_expression(inst) && !inst->predicate && inst->mlen == 0 &&
135 (inst->dst.file != HW_REG || inst->dst.is_null()))
136 {
137 bool found = false;
138
139 foreach_in_list_use_after(aeb_entry, entry, &aeb) {
140 /* Match current instruction's expression against those in AEB. */
141 if (!(entry->generator->dst.is_null() && !inst->dst.is_null()) &&
142 instructions_match(inst, entry->generator)) {
143 found = true;
144 progress = true;
145 break;
146 }
147 }
148
149 if (!found) {
150 if (inst->opcode != BRW_OPCODE_MOV ||
151 (inst->opcode == BRW_OPCODE_MOV &&
152 inst->src[0].file == IMM &&
153 inst->src[0].type == BRW_REGISTER_TYPE_VF)) {
154 /* Our first sighting of this expression. Create an entry. */
155 aeb_entry *entry = ralloc(cse_ctx, aeb_entry);
156 entry->tmp = src_reg(); /* file will be BAD_FILE */
157 entry->generator = inst;
158 aeb.push_tail(entry);
159 }
160 } else {
161 /* This is at least our second sighting of this expression.
162 * If we don't have a temporary already, make one.
163 */
164 bool no_existing_temp = entry->tmp.file == BAD_FILE;
165 if (no_existing_temp && !entry->generator->dst.is_null()) {
166 entry->tmp = retype(src_reg(GRF, alloc.allocate(
167 entry->generator->regs_written),
168 NULL), inst->dst.type);
169
170 for (unsigned i = 0; i < entry->generator->regs_written; ++i) {
171 vec4_instruction *copy = MOV(offset(entry->generator->dst, i),
172 offset(entry->tmp, i));
173 copy->force_writemask_all =
174 entry->generator->force_writemask_all;
175 entry->generator->insert_after(block, copy);
176 }
177
178 entry->generator->dst = dst_reg(entry->tmp);
179 }
180
181 /* dest <- temp */
182 if (!inst->dst.is_null()) {
183 assert(inst->dst.type == entry->tmp.type);
184
185 for (unsigned i = 0; i < inst->regs_written; ++i) {
186 vec4_instruction *copy = MOV(offset(inst->dst, i),
187 offset(entry->tmp, i));
188 copy->force_writemask_all = inst->force_writemask_all;
189 inst->insert_before(block, copy);
190 }
191 }
192
193 /* Set our iterator so that next time through the loop inst->next
194 * will get the instruction in the basic block after the one we've
195 * removed.
196 */
197 vec4_instruction *prev = (vec4_instruction *)inst->prev;
198
199 inst->remove(block);
200 inst = prev;
201 }
202 }
203
204 foreach_in_list_safe(aeb_entry, entry, &aeb) {
205 /* Kill all AEB entries that write a different value to or read from
206 * the flag register if we just wrote it.
207 */
208 if (inst->writes_flag()) {
209 if (entry->generator->reads_flag() ||
210 (entry->generator->writes_flag() &&
211 !instructions_match(inst, entry->generator))) {
212 entry->remove();
213 ralloc_free(entry);
214 continue;
215 }
216 }
217
218 for (int i = 0; i < 3; i++) {
219 src_reg *src = &entry->generator->src[i];
220
221 /* Kill all AEB entries that use the destination we just
222 * overwrote.
223 */
224 if (inst->dst.file == entry->generator->src[i].file &&
225 inst->dst.reg == entry->generator->src[i].reg) {
226 entry->remove();
227 ralloc_free(entry);
228 break;
229 }
230
231 /* Kill any AEB entries using registers that don't get reused any
232 * more -- a sure sign they'll fail operands_match().
233 */
234 if (src->file == GRF) {
235 if (var_range_end(var_from_reg(alloc, *src), 4) < ip) {
236 entry->remove();
237 ralloc_free(entry);
238 break;
239 }
240 }
241 }
242 }
243
244 ip++;
245 }
246
247 ralloc_free(cse_ctx);
248
249 return progress;
250 }
251
252 bool
253 vec4_visitor::opt_cse()
254 {
255 bool progress = false;
256
257 calculate_live_intervals();
258
259 foreach_block (block, cfg) {
260 progress = opt_cse_local(block) || progress;
261 }
262
263 if (progress)
264 invalidate_live_intervals();
265
266 return progress;
267 }