1/*
2 * Copyright © 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/** @file brw_vec4_dead_code_eliminate.cpp
29 *
30 * Dataflow-aware dead code elimination.
31 *
32 * Walks the instruction list from the bottom, removing instructions that
33 * have results that both aren't used in later blocks and haven't been read
34 * yet in the tail end of this block.
35 */
36
37using namespace brw;
38
39bool
40vec4_visitor::dead_code_eliminate()
41{
42   bool progress = false;
43
44   calculate_live_intervals();
45
46   int num_vars = live_intervals->num_vars;
47   BITSET_WORD *live = rzalloc_array(NULL, BITSET_WORD, BITSET_WORDS(num_vars));
48   BITSET_WORD *flag_live = rzalloc_array(NULL, BITSET_WORD, 1);
49
50   foreach_block_reverse_safe(block, cfg) {
51      memcpy(live, live_intervals->block_data[block->num].liveout,
52             sizeof(BITSET_WORD) * BITSET_WORDS(num_vars));
53      memcpy(flag_live, live_intervals->block_data[block->num].flag_liveout,
54             sizeof(BITSET_WORD));
55
56      foreach_inst_in_block_reverse_safe(vec4_instruction, inst, block) {
57         if ((inst->dst.file == VGRF && !inst->has_side_effects()) ||
58             (inst->dst.is_null() && inst->writes_flag())){
59            bool result_live[4] = { false };
60            if (inst->dst.file == VGRF) {
61               for (unsigned i = 0; i < DIV_ROUND_UP(inst->size_written, 16); i++) {
62                  for (int c = 0; c < 4; c++) {
63                     const unsigned v = var_from_reg(alloc, inst->dst, c, i);
64                     result_live[c] |= BITSET_TEST(live, v);
65                  }
66               }
67            } else {
68               for (unsigned c = 0; c < 4; c++)
69                  result_live[c] = BITSET_TEST(flag_live, c);
70            }
71
72            /* If the instruction can't do writemasking, then it's all or
73             * nothing.
74             */
75            if (!inst->can_do_writemask(devinfo)) {
76               bool result = result_live[0] | result_live[1] |
77                             result_live[2] | result_live[3];
78               result_live[0] = result;
79               result_live[1] = result;
80               result_live[2] = result;
81               result_live[3] = result;
82            }
83
84            if (inst->writes_flag()) {
85               /* Independently calculate the usage of the flag components and
86                * the destination value components.
87                */
88               uint8_t flag_mask = inst->dst.writemask;
89               uint8_t dest_mask = inst->dst.writemask;
90
91               for (int c = 0; c < 4; c++) {
92                  if (!result_live[c] && dest_mask & (1 << c))
93                     dest_mask &= ~(1 << c);
94
95                  if (!BITSET_TEST(flag_live, c))
96                     flag_mask &= ~(1 << c);
97               }
98
99               if (inst->dst.writemask != (flag_mask | dest_mask)) {
100                  progress = true;
101                  inst->dst.writemask = flag_mask | dest_mask;
102               }
103
104               /* If none of the destination components are read, replace the
105                * destination register with the NULL register.
106                */
107               if (dest_mask == 0) {
108                  progress = true;
109                  inst->dst = dst_reg(retype(brw_null_reg(), inst->dst.type));
110               }
111            } else {
112               for (int c = 0; c < 4; c++) {
113                  if (!result_live[c] && inst->dst.writemask & (1 << c)) {
114                     inst->dst.writemask &= ~(1 << c);
115                     progress = true;
116
117                     if (inst->dst.writemask == 0) {
118                        if (inst->writes_accumulator) {
119                           inst->dst = dst_reg(retype(brw_null_reg(), inst->dst.type));
120                        } else {
121                           inst->opcode = BRW_OPCODE_NOP;
122                           break;
123                        }
124                     }
125                  }
126               }
127            }
128         }
129
130         if (inst->dst.is_null() && inst->writes_flag()) {
131            bool combined_live = false;
132            for (unsigned c = 0; c < 4; c++)
133               combined_live |= BITSET_TEST(flag_live, c);
134
135            if (!combined_live) {
136               inst->opcode = BRW_OPCODE_NOP;
137               progress = true;
138            }
139         }
140
141         if (inst->dst.file == VGRF && !inst->predicate &&
142             !inst->is_align1_partial_write()) {
143            for (unsigned i = 0; i < DIV_ROUND_UP(inst->size_written, 16); i++) {
144               for (int c = 0; c < 4; c++) {
145                  if (inst->dst.writemask & (1 << c)) {
146                     const unsigned v = var_from_reg(alloc, inst->dst, c, i);
147                     BITSET_CLEAR(live, v);
148                  }
149               }
150            }
151         }
152
153         if (inst->writes_flag() && !inst->predicate && inst->exec_size == 8) {
154            for (unsigned c = 0; c < 4; c++)
155               BITSET_CLEAR(flag_live, c);
156         }
157
158         if (inst->opcode == BRW_OPCODE_NOP) {
159            inst->remove(block);
160            continue;
161         }
162
163         for (int i = 0; i < 3; i++) {
164            if (inst->src[i].file == VGRF) {
165               for (unsigned j = 0; j < DIV_ROUND_UP(inst->size_read(i), 16); j++) {
166                  for (int c = 0; c < 4; c++) {
167                     const unsigned v = var_from_reg(alloc, inst->src[i], c, j);
168                     BITSET_SET(live, v);
169                  }
170               }
171            }
172         }
173
174         for (unsigned c = 0; c < 4; c++) {
175            if (inst->reads_flag(c)) {
176               BITSET_SET(flag_live, c);
177            }
178         }
179      }
180   }
181
182   ralloc_free(live);
183   ralloc_free(flag_live);
184
185   if (progress)
186      invalidate_live_intervals();
187
188   return progress;
189}
190