1 /*
2 * Copyright © 2014 Broadcom
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 /**
25 * @file v3d_opt_dead_code.c
26 *
27 * This is a simple dead code eliminator for SSA values in VIR.
28 *
29 * It walks all the instructions finding what temps are used, then walks again
30 * to remove instructions writing unused temps.
31 *
32 * This is an inefficient implementation if you have long chains of
33 * instructions where the entire chain is dead, but we expect those to have
34 * been eliminated at the NIR level, and here we're just cleaning up small
35 * problems produced by NIR->VIR.
36 */
37
38 #include "v3d_compiler.h"
39
40 static bool debug;
41
42 static void
dce(struct v3d_compile * c,struct qinst * inst)43 dce(struct v3d_compile *c, struct qinst *inst)
44 {
45 if (debug) {
46 fprintf(stderr, "Removing: ");
47 vir_dump_inst(c, inst);
48 fprintf(stderr, "\n");
49 }
50 assert(inst->qpu.flags.apf == V3D_QPU_PF_NONE);
51 assert(inst->qpu.flags.mpf == V3D_QPU_PF_NONE);
52 assert(inst->qpu.flags.auf == V3D_QPU_UF_NONE);
53 assert(inst->qpu.flags.muf == V3D_QPU_UF_NONE);
54 vir_remove_instruction(c, inst);
55 }
56
57 static bool
has_nonremovable_reads(struct v3d_compile * c,struct qinst * inst)58 has_nonremovable_reads(struct v3d_compile *c, struct qinst *inst)
59 {
60 for (int i = 0; i < vir_get_nsrc(inst); i++) {
61 if (inst->src[i].file == QFILE_VPM) {
62 /* Instance ID, Vertex ID: Should have been removed at
63 * the NIR level
64 */
65 if (inst->src[i].index == ~0)
66 return true;
67
68 uint32_t attr = inst->src[i].index / 4;
69 uint32_t offset = inst->src[i].index % 4;
70
71 if (c->vattr_sizes[attr] != offset)
72 return true;
73
74 /* Can't get rid of the last VPM read, or the
75 * simulator (at least) throws an error.
76 */
77 uint32_t total_size = 0;
78 for (uint32_t i = 0; i < ARRAY_SIZE(c->vattr_sizes); i++)
79 total_size += c->vattr_sizes[i];
80 if (total_size == 1)
81 return true;
82 }
83 }
84
85 return false;
86 }
87
88 bool
vir_opt_dead_code(struct v3d_compile * c)89 vir_opt_dead_code(struct v3d_compile *c)
90 {
91 bool progress = false;
92 bool *used = calloc(c->num_temps, sizeof(bool));
93
94 vir_for_each_inst_inorder(inst, c) {
95 for (int i = 0; i < vir_get_nsrc(inst); i++) {
96 if (inst->src[i].file == QFILE_TEMP)
97 used[inst->src[i].index] = true;
98 }
99 }
100
101 vir_for_each_block(block, c) {
102 vir_for_each_inst_safe(inst, block) {
103 if (inst->dst.file != QFILE_NULL &&
104 !(inst->dst.file == QFILE_TEMP &&
105 !used[inst->dst.index])) {
106 continue;
107 }
108
109 if (vir_has_side_effects(c, inst))
110 continue;
111
112 if (inst->qpu.flags.apf != V3D_QPU_PF_NONE ||
113 inst->qpu.flags.mpf != V3D_QPU_PF_NONE ||
114 inst->qpu.flags.auf != V3D_QPU_UF_NONE ||
115 inst->qpu.flags.muf != V3D_QPU_UF_NONE ||
116 has_nonremovable_reads(c, inst)) {
117 /* If we can't remove the instruction, but we
118 * don't need its destination value, just
119 * remove the destination. The register
120 * allocator would trivially color it and it
121 * wouldn't cause any register pressure, but
122 * it's nicer to read the VIR code without
123 * unused destination regs.
124 */
125 if (inst->dst.file == QFILE_TEMP) {
126 if (debug) {
127 fprintf(stderr,
128 "Removing dst from: ");
129 vir_dump_inst(c, inst);
130 fprintf(stderr, "\n");
131 }
132 c->defs[inst->dst.index] = NULL;
133 inst->dst.file = QFILE_NULL;
134 progress = true;
135 }
136 continue;
137 }
138
139 for (int i = 0; i < vir_get_nsrc(inst); i++) {
140 if (inst->src[i].file != QFILE_VPM)
141 continue;
142 uint32_t attr = inst->src[i].index / 4;
143 uint32_t offset = (inst->src[i].index % 4);
144
145 if (c->vattr_sizes[attr] == offset) {
146 c->num_inputs--;
147 c->vattr_sizes[attr]--;
148 }
149 }
150
151 dce(c, inst);
152 progress = true;
153 continue;
154 }
155 }
156
157 free(used);
158
159 return progress;
160 }
161