freedreno/ir3: Add unit tests for derivatives disasm.
[mesa.git] / src / freedreno / ir3 / ir3_dce.c
1 /*
2 * Copyright (C) 2014 Rob Clark <robclark@freedesktop.org>
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 FROM,
20 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
21 * SOFTWARE.
22 *
23 * Authors:
24 * Rob Clark <robclark@freedesktop.org>
25 */
26
27 #include "util/u_math.h"
28
29 #include "ir3.h"
30 #include "ir3_shader.h"
31
32 /*
33 * Dead code elimination:
34 */
35
36 static void
37 mark_array_use(struct ir3_instruction *instr, struct ir3_register *reg)
38 {
39 if (reg->flags & IR3_REG_ARRAY) {
40 struct ir3_array *arr =
41 ir3_lookup_array(instr->block->shader, reg->array.id);
42 arr->unused = false;
43 }
44 }
45
46 static void
47 instr_dce(struct ir3_instruction *instr, bool falsedep)
48 {
49 /* don't mark falsedep's as used, but otherwise process them normally: */
50 if (!falsedep)
51 instr->flags &= ~IR3_INSTR_UNUSED;
52
53 if (ir3_instr_check_mark(instr))
54 return;
55
56 if (writes_gpr(instr))
57 mark_array_use(instr, instr->regs[0]); /* dst */
58
59 foreach_src (reg, instr)
60 mark_array_use(instr, reg); /* src */
61
62 foreach_ssa_src_n (src, i, instr) {
63 instr_dce(src, __is_false_dep(instr, i));
64 }
65 }
66
67 static bool
68 remove_unused_by_block(struct ir3_block *block)
69 {
70 bool progress = false;
71 foreach_instr_safe (instr, &block->instr_list) {
72 if (instr->opc == OPC_END || instr->opc == OPC_CHSH || instr->opc == OPC_CHMASK)
73 continue;
74 if (instr->flags & IR3_INSTR_UNUSED) {
75 if (instr->opc == OPC_META_SPLIT) {
76 struct ir3_instruction *src = ssa(instr->regs[1]);
77 /* tex (cat5) instructions have a writemask, so we can
78 * mask off unused components. Other instructions do not.
79 */
80 if (src && is_tex_or_prefetch(src) && (src->regs[0]->wrmask > 1)) {
81 src->regs[0]->wrmask &= ~(1 << instr->split.off);
82
83 /* prune no-longer needed right-neighbors. We could
84 * probably do the same for left-neighbors (ie. tex
85 * fetch that only need .yw components), but that
86 * makes RA a bit more confusing than it already is
87 */
88 struct ir3_instruction *n = instr;
89 while (n && n->cp.right)
90 n = n->cp.right;
91 while (n->flags & IR3_INSTR_UNUSED) {
92 n = n->cp.left;
93 if (!n)
94 break;
95 n->cp.right = NULL;
96 }
97 }
98 }
99
100 /* prune false-deps, etc: */
101 foreach_ssa_use (use, instr)
102 foreach_ssa_srcp_n (srcp, n, use)
103 if (*srcp == instr)
104 *srcp = NULL;
105
106 list_delinit(&instr->node);
107 progress = true;
108 }
109 }
110 return progress;
111 }
112
113 static bool
114 find_and_remove_unused(struct ir3 *ir, struct ir3_shader_variant *so)
115 {
116 unsigned i;
117 bool progress = false;
118
119 ir3_clear_mark(ir);
120
121 /* initially mark everything as unused, we'll clear the flag as we
122 * visit the instructions:
123 */
124 foreach_block (block, &ir->block_list) {
125 foreach_instr (instr, &block->instr_list) {
126 /* special case, if pre-fs texture fetch used, we cannot
127 * eliminate the barycentric i/j input
128 */
129 if (so->num_sampler_prefetch &&
130 (instr->opc == OPC_META_INPUT) &&
131 (instr->input.sysval == SYSTEM_VALUE_BARYCENTRIC_PERSP_PIXEL))
132 continue;
133 instr->flags |= IR3_INSTR_UNUSED;
134 }
135 }
136
137 foreach_array (arr, &ir->array_list)
138 arr->unused = true;
139
140 foreach_output (out, ir)
141 instr_dce(out, false);
142
143 foreach_block (block, &ir->block_list) {
144 for (i = 0; i < block->keeps_count; i++)
145 instr_dce(block->keeps[i], false);
146
147 /* We also need to account for if-condition: */
148 if (block->condition)
149 instr_dce(block->condition, false);
150 }
151
152 /* remove un-used instructions: */
153 foreach_block (block, &ir->block_list) {
154 progress |= remove_unused_by_block(block);
155 }
156
157 /* remove un-used arrays: */
158 foreach_array_safe (arr, &ir->array_list) {
159 if (arr->unused)
160 list_delinit(&arr->node);
161 }
162
163 /* fixup wrmask of split instructions to account for adjusted tex
164 * wrmask's:
165 */
166 foreach_block (block, &ir->block_list) {
167 foreach_instr (instr, &block->instr_list) {
168 if (instr->opc != OPC_META_SPLIT)
169 continue;
170
171 struct ir3_instruction *src = ssa(instr->regs[1]);
172 if (!is_tex_or_prefetch(src))
173 continue;
174
175 instr->regs[1]->wrmask = src->regs[0]->wrmask;
176 }
177 }
178
179 /* note that we can end up with unused indirects, but we should
180 * not end up with unused predicates.
181 */
182 for (i = 0; i < ir->a0_users_count; i++) {
183 struct ir3_instruction *instr = ir->a0_users[i];
184 if (instr && (instr->flags & IR3_INSTR_UNUSED))
185 ir->a0_users[i] = NULL;
186 }
187
188 for (i = 0; i < ir->a1_users_count; i++) {
189 struct ir3_instruction *instr = ir->a1_users[i];
190 if (instr && (instr->flags & IR3_INSTR_UNUSED))
191 ir->a1_users[i] = NULL;
192 }
193
194 /* cleanup unused inputs: */
195 foreach_input_n (in, n, ir)
196 if (in->flags & IR3_INSTR_UNUSED)
197 ir->inputs[n] = NULL;
198
199 return progress;
200 }
201
202 bool
203 ir3_dce(struct ir3 *ir, struct ir3_shader_variant *so)
204 {
205 void *mem_ctx = ralloc_context(NULL);
206 bool progress, made_progress = false;
207
208 ir3_find_ssa_uses(ir, mem_ctx, true);
209
210 do {
211 progress = find_and_remove_unused(ir, so);
212 made_progress |= progress;
213 } while (progress);
214
215 ralloc_free(mem_ctx);
216
217 return made_progress;
218 }