1 /* -*- mode: C; c-file-style: "k&r"; tab-width 4; indent-tabs-mode: t; -*- */
2
3 /*
4 * Copyright (C) 2014 Rob Clark <robclark@freedesktop.org>
5 *
6 * Permission is hereby granted, free of charge, to any person obtaining a
7 * copy of this software and associated documentation files (the "Software"),
8 * to deal in the Software without restriction, including without limitation
9 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
10 * and/or sell copies of the Software, and to permit persons to whom the
11 * Software is furnished to do so, subject to the following conditions:
12 *
13 * The above copyright notice and this permission notice (including the next
14 * paragraph) shall be included in all copies or substantial portions of the
15 * Software.
16 *
17 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
18 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
19 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
20 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
21 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
22 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
23 * SOFTWARE.
24 *
25 * Authors:
26 * Rob Clark <robclark@freedesktop.org>
27 */
28
29 #include "util/u_math.h"
30
31 #include "ir3.h"
32
33 /*
34 * Instruction Depth:
35 *
36 * Calculates weighted instruction depth, ie. the sum of # of needed
37 * instructions plus delay slots back to original input (ie INPUT or
38 * CONST). That is to say, an instructions depth is:
39 *
40 * depth(instr) {
41 * d = 0;
42 * // for each src register:
43 * foreach (src in instr->regs[1..n])
44 * d = max(d, delayslots(src->instr, n) + depth(src->instr));
45 * return d + 1;
46 * }
47 *
48 * After an instruction's depth is calculated, it is inserted into the
49 * blocks depth sorted list, which is used by the scheduling pass.
50 */
51
52 /* calculate required # of delay slots between the instruction that
53 * assigns a value and the one that consumes
54 */
ir3_delayslots(struct ir3_instruction * assigner,struct ir3_instruction * consumer,unsigned n)55 int ir3_delayslots(struct ir3_instruction *assigner,
56 struct ir3_instruction *consumer, unsigned n)
57 {
58 /* don't count false-dependencies: */
59 if (__is_false_dep(consumer, n))
60 return 0;
61
62 /* worst case is cat1-3 (alu) -> cat4/5 needing 6 cycles, normal
63 * alu -> alu needs 3 cycles, cat4 -> alu and texture fetch
64 * handled with sync bits
65 */
66
67 if (is_meta(assigner))
68 return 0;
69
70 if (writes_addr(assigner))
71 return 6;
72
73 /* handled via sync flags: */
74 if (is_sfu(assigner) || is_tex(assigner) || is_mem(assigner))
75 return 0;
76
77 /* assigner must be alu: */
78 if (is_flow(consumer) || is_sfu(consumer) || is_tex(consumer) ||
79 is_mem(consumer)) {
80 return 6;
81 } else if ((is_mad(consumer->opc) || is_madsh(consumer->opc)) &&
82 (n == 3)) {
83 /* special case, 3rd src to cat3 not required on first cycle */
84 return 1;
85 } else {
86 return 3;
87 }
88 }
89
90 void
ir3_insert_by_depth(struct ir3_instruction * instr,struct list_head * list)91 ir3_insert_by_depth(struct ir3_instruction *instr, struct list_head *list)
92 {
93 /* remove from existing spot in list: */
94 list_delinit(&instr->node);
95
96 /* find where to re-insert instruction: */
97 list_for_each_entry (struct ir3_instruction, pos, list, node) {
98 if (pos->depth > instr->depth) {
99 list_add(&instr->node, &pos->node);
100 return;
101 }
102 }
103 /* if we get here, we didn't find an insertion spot: */
104 list_addtail(&instr->node, list);
105 }
106
107 static void
ir3_instr_depth(struct ir3_instruction * instr)108 ir3_instr_depth(struct ir3_instruction *instr)
109 {
110 struct ir3_instruction *src;
111
112 /* if we've already visited this instruction, bail now: */
113 if (ir3_instr_check_mark(instr))
114 return;
115
116 instr->depth = 0;
117
118 foreach_ssa_src_n(src, i, instr) {
119 unsigned sd;
120
121 /* visit child to compute it's depth: */
122 ir3_instr_depth(src);
123
124 /* for array writes, no need to delay on previous write: */
125 if (i == 0)
126 continue;
127
128 sd = ir3_delayslots(src, instr, i) + src->depth;
129
130 instr->depth = MAX2(instr->depth, sd);
131 }
132
133 if (!is_meta(instr))
134 instr->depth++;
135
136 ir3_insert_by_depth(instr, &instr->block->instr_list);
137 }
138
139 static void
remove_unused_by_block(struct ir3_block * block)140 remove_unused_by_block(struct ir3_block *block)
141 {
142 list_for_each_entry_safe (struct ir3_instruction, instr, &block->instr_list, node) {
143 if (!ir3_instr_check_mark(instr)) {
144 if (instr->opc == OPC_END)
145 continue;
146 /* mark it, in case it is input, so we can
147 * remove unused inputs:
148 */
149 instr->flags |= IR3_INSTR_UNUSED;
150 /* and remove from instruction list: */
151 list_delinit(&instr->node);
152 }
153 }
154 }
155
156 void
ir3_depth(struct ir3 * ir)157 ir3_depth(struct ir3 *ir)
158 {
159 unsigned i;
160
161 ir3_clear_mark(ir);
162 for (i = 0; i < ir->noutputs; i++)
163 if (ir->outputs[i])
164 ir3_instr_depth(ir->outputs[i]);
165
166 list_for_each_entry (struct ir3_block, block, &ir->block_list, node) {
167 for (i = 0; i < block->keeps_count; i++)
168 ir3_instr_depth(block->keeps[i]);
169
170 /* We also need to account for if-condition: */
171 if (block->condition)
172 ir3_instr_depth(block->condition);
173 }
174
175 /* mark un-used instructions: */
176 list_for_each_entry (struct ir3_block, block, &ir->block_list, node) {
177 remove_unused_by_block(block);
178 }
179
180 /* note that we can end up with unused indirects, but we should
181 * not end up with unused predicates.
182 */
183 for (i = 0; i < ir->indirects_count; i++) {
184 struct ir3_instruction *instr = ir->indirects[i];
185 if (instr->flags & IR3_INSTR_UNUSED)
186 ir->indirects[i] = NULL;
187 }
188
189 /* cleanup unused inputs: */
190 for (i = 0; i < ir->ninputs; i++) {
191 struct ir3_instruction *in = ir->inputs[i];
192 if (in && (in->flags & IR3_INSTR_UNUSED))
193 ir->inputs[i] = NULL;
194 }
195 }
196