1 /*
2 * Copyright © 2014 Rob Clark <[email protected]>
3 * SPDX-License-Identifier: MIT
4 *
5 * Authors:
6 * Rob Clark <[email protected]>
7 */
8
9 #include "util/u_math.h"
10
11 #include "ir3.h"
12 #include "ir3_shader.h"
13
14 /*
15 * Dead code elimination:
16 */
17
18 static void
mark_array_use(struct ir3_instruction * instr,struct ir3_register * reg)19 mark_array_use(struct ir3_instruction *instr, struct ir3_register *reg)
20 {
21 if (reg->flags & IR3_REG_ARRAY) {
22 struct ir3_array *arr =
23 ir3_lookup_array(instr->block->shader, reg->array.id);
24 arr->unused = false;
25 }
26 }
27
28 static void
instr_dce(struct ir3_instruction * instr,bool falsedep)29 instr_dce(struct ir3_instruction *instr, bool falsedep)
30 {
31 /* don't mark falsedep's as used, but otherwise process them normally: */
32 if (!falsedep)
33 instr->flags &= ~IR3_INSTR_UNUSED;
34
35 if (ir3_instr_check_mark(instr))
36 return;
37
38 foreach_dst (dst, instr) {
39 if (is_dest_gpr(dst))
40 mark_array_use(instr, dst);
41 }
42
43 foreach_src (reg, instr)
44 mark_array_use(instr, reg); /* src */
45
46 foreach_ssa_src_n (src, i, instr) {
47 instr_dce(src, __is_false_dep(instr, i));
48 }
49 }
50
51 static bool
remove_unused_by_block(struct ir3_block * block)52 remove_unused_by_block(struct ir3_block *block)
53 {
54 bool progress = false;
55 foreach_instr_safe (instr, &block->instr_list) {
56 if (instr->opc == OPC_END || instr->opc == OPC_CHSH ||
57 instr->opc == OPC_CHMASK || instr->opc == OPC_LOCK ||
58 instr->opc == OPC_UNLOCK)
59 continue;
60 if (instr->flags & IR3_INSTR_UNUSED) {
61 if (instr->opc == OPC_META_SPLIT) {
62 struct ir3_instruction *src = ssa(instr->srcs[0]);
63 /* tex (cat5) instructions have a writemask, so we can
64 * mask off unused components. Other instructions do not.
65 */
66 if (src && is_tex_or_prefetch(src) && (src->dsts[0]->wrmask > 1)) {
67 src->dsts[0]->wrmask &= ~(1 << instr->split.off);
68 }
69 }
70
71 /* prune false-deps, etc: */
72 foreach_ssa_use (use, instr)
73 foreach_ssa_srcp_n (srcp, n, use)
74 if (*srcp == instr)
75 *srcp = NULL;
76
77 ir3_instr_remove(instr);
78 progress = true;
79 }
80 }
81 return progress;
82 }
83
84 static bool
find_and_remove_unused(struct ir3 * ir,struct ir3_shader_variant * so)85 find_and_remove_unused(struct ir3 *ir, struct ir3_shader_variant *so)
86 {
87 unsigned i;
88 bool progress = false;
89
90 ir3_clear_mark(ir);
91
92 /* initially mark everything as unused, we'll clear the flag as we
93 * visit the instructions:
94 */
95 foreach_block (block, &ir->block_list) {
96 foreach_instr (instr, &block->instr_list) {
97 if (instr->opc == OPC_META_INPUT) {
98 /* Without GS header geometry shader is never invoked. */
99 if (instr->input.sysval == SYSTEM_VALUE_GS_HEADER_IR3)
100 continue;
101 }
102
103 instr->flags |= IR3_INSTR_UNUSED;
104 }
105 }
106
107 foreach_array (arr, &ir->array_list)
108 arr->unused = true;
109
110 foreach_block (block, &ir->block_list) {
111 for (i = 0; i < block->keeps_count; i++)
112 instr_dce(block->keeps[i], false);
113
114 /* We also need to account for if-condition: */
115 struct ir3_instruction *terminator = ir3_block_get_terminator(block);
116 if (terminator) {
117 instr_dce(terminator, false);
118 }
119 }
120
121 /* remove un-used instructions: */
122 foreach_block (block, &ir->block_list) {
123 progress |= remove_unused_by_block(block);
124 }
125
126 /* remove un-used arrays: */
127 foreach_array_safe (arr, &ir->array_list) {
128 if (arr->unused)
129 list_delinit(&arr->node);
130 }
131
132 /* fixup wrmask of split instructions to account for adjusted tex
133 * wrmask's:
134 */
135 foreach_block (block, &ir->block_list) {
136 foreach_instr (instr, &block->instr_list) {
137 if (instr->opc != OPC_META_SPLIT)
138 continue;
139
140 struct ir3_instruction *src = ssa(instr->srcs[0]);
141 if (!is_tex_or_prefetch(src))
142 continue;
143
144 instr->srcs[0]->wrmask = src->dsts[0]->wrmask;
145 }
146 }
147
148 for (i = 0; i < ir->a0_users_count; i++) {
149 struct ir3_instruction *instr = ir->a0_users[i];
150 if (instr && (instr->flags & IR3_INSTR_UNUSED))
151 ir->a0_users[i] = NULL;
152 }
153
154 for (i = 0; i < ir->a1_users_count; i++) {
155 struct ir3_instruction *instr = ir->a1_users[i];
156 if (instr && (instr->flags & IR3_INSTR_UNUSED))
157 ir->a1_users[i] = NULL;
158 }
159
160 /* cleanup unused inputs: */
161 foreach_input_n (in, n, ir)
162 if (in->flags & IR3_INSTR_UNUSED)
163 ir->inputs[n] = NULL;
164
165 return progress;
166 }
167
168 bool
ir3_dce(struct ir3 * ir,struct ir3_shader_variant * so)169 ir3_dce(struct ir3 *ir, struct ir3_shader_variant *so)
170 {
171 void *mem_ctx = ralloc_context(NULL);
172 bool progress, made_progress = false;
173
174 ir3_find_ssa_uses(ir, mem_ctx, true);
175
176 do {
177 progress = find_and_remove_unused(ir, so);
178 made_progress |= progress;
179 } while (progress);
180
181 ralloc_free(mem_ctx);
182
183 return made_progress;
184 }
185