1 /* -*- mode: C; c-file-style: "k&r"; tab-width 4; indent-tabs-mode: t; -*- */
4 * Copyright (C) 2014 Rob Clark <robclark@freedesktop.org>
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:
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
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
26 * Rob Clark <robclark@freedesktop.org>
29 #include "util/u_math.h"
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:
42 * // for each src register:
43 * foreach (src in instr->regs[1..n])
44 * d = max(d, delayslots(src->instr, n) + depth(src->instr));
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.
52 /* generally don't count false dependencies, since this can just be
53 * something like a barrier, or SSBO store. The exception is array
54 * dependencies if the assigner is an array write and the consumer
55 * reads the same array.
58 ignore_dep(struct ir3_instruction
*assigner
,
59 struct ir3_instruction
*consumer
, unsigned n
)
61 if (!__is_false_dep(consumer
, n
))
64 if (assigner
->barrier_class
& IR3_BARRIER_ARRAY_W
) {
65 struct ir3_register
*dst
= assigner
->regs
[0];
66 struct ir3_register
*src
;
68 debug_assert(dst
->flags
& IR3_REG_ARRAY
);
70 foreach_src(src
, consumer
) {
71 if ((src
->flags
& IR3_REG_ARRAY
) &&
72 (dst
->array
.id
== src
->array
.id
)) {
81 /* calculate required # of delay slots between the instruction that
82 * assigns a value and the one that consumes
84 int ir3_delayslots(struct ir3_instruction
*assigner
,
85 struct ir3_instruction
*consumer
, unsigned n
)
87 if (ignore_dep(assigner
, consumer
, n
))
90 /* worst case is cat1-3 (alu) -> cat4/5 needing 6 cycles, normal
91 * alu -> alu needs 3 cycles, cat4 -> alu and texture fetch
92 * handled with sync bits
95 if (is_meta(assigner
))
98 if (writes_addr(assigner
))
101 /* handled via sync flags: */
102 if (is_sfu(assigner
) || is_tex(assigner
) || is_mem(assigner
))
105 /* assigner must be alu: */
106 if (is_flow(consumer
) || is_sfu(consumer
) || is_tex(consumer
) ||
109 } else if ((is_mad(consumer
->opc
) || is_madsh(consumer
->opc
)) &&
111 /* special case, 3rd src to cat3 not required on first cycle */
119 ir3_insert_by_depth(struct ir3_instruction
*instr
, struct list_head
*list
)
121 /* remove from existing spot in list: */
122 list_delinit(&instr
->node
);
124 /* find where to re-insert instruction: */
125 list_for_each_entry (struct ir3_instruction
, pos
, list
, node
) {
126 if (pos
->depth
> instr
->depth
) {
127 list_add(&instr
->node
, &pos
->node
);
131 /* if we get here, we didn't find an insertion spot: */
132 list_addtail(&instr
->node
, list
);
136 ir3_instr_depth(struct ir3_instruction
*instr
, unsigned boost
, bool falsedep
)
138 struct ir3_instruction
*src
;
140 /* don't mark falsedep's as used, but otherwise process them normally: */
142 instr
->flags
&= ~IR3_INSTR_UNUSED
;
144 if (ir3_instr_check_mark(instr
))
149 foreach_ssa_src_n(src
, i
, instr
) {
152 /* visit child to compute it's depth: */
153 ir3_instr_depth(src
, boost
, __is_false_dep(instr
, i
));
155 /* for array writes, no need to delay on previous write: */
159 sd
= ir3_delayslots(src
, instr
, i
) + src
->depth
;
162 instr
->depth
= MAX2(instr
->depth
, sd
);
168 ir3_insert_by_depth(instr
, &instr
->block
->instr_list
);
172 remove_unused_by_block(struct ir3_block
*block
)
174 bool progress
= false;
175 list_for_each_entry_safe (struct ir3_instruction
, instr
, &block
->instr_list
, node
) {
176 if (instr
->opc
== OPC_END
)
178 if (instr
->flags
& IR3_INSTR_UNUSED
) {
179 list_delinit(&instr
->node
);
187 compute_depth_and_remove_unused(struct ir3
*ir
)
190 bool progress
= false;
194 /* initially mark everything as unused, we'll clear the flag as we
195 * visit the instructions:
197 list_for_each_entry (struct ir3_block
, block
, &ir
->block_list
, node
) {
198 list_for_each_entry (struct ir3_instruction
, instr
, &block
->instr_list
, node
) {
199 instr
->flags
|= IR3_INSTR_UNUSED
;
203 for (i
= 0; i
< ir
->noutputs
; i
++)
205 ir3_instr_depth(ir
->outputs
[i
], 0, false);
207 list_for_each_entry (struct ir3_block
, block
, &ir
->block_list
, node
) {
208 for (i
= 0; i
< block
->keeps_count
; i
++)
209 ir3_instr_depth(block
->keeps
[i
], 0, false);
211 /* We also need to account for if-condition: */
212 if (block
->condition
)
213 ir3_instr_depth(block
->condition
, 6, false);
216 /* mark un-used instructions: */
217 list_for_each_entry (struct ir3_block
, block
, &ir
->block_list
, node
) {
218 progress
|= remove_unused_by_block(block
);
221 /* note that we can end up with unused indirects, but we should
222 * not end up with unused predicates.
224 for (i
= 0; i
< ir
->indirects_count
; i
++) {
225 struct ir3_instruction
*instr
= ir
->indirects
[i
];
226 if (instr
&& (instr
->flags
& IR3_INSTR_UNUSED
))
227 ir
->indirects
[i
] = NULL
;
230 /* cleanup unused inputs: */
231 for (i
= 0; i
< ir
->ninputs
; i
++) {
232 struct ir3_instruction
*in
= ir
->inputs
[i
];
233 if (in
&& (in
->flags
& IR3_INSTR_UNUSED
))
234 ir
->inputs
[i
] = NULL
;
241 ir3_depth(struct ir3
*ir
)
245 progress
= compute_depth_and_remove_unused(ir
);