3ef67731926bffca5bd16d6b57f7afd3ff975f65
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>
30 #include "util/u_math.h"
40 * Instruction Scheduling:
42 * Using the depth sorted list from depth pass, attempt to recursively
43 * schedule deepest unscheduled path. The first instruction that cannot
44 * be scheduled, returns the required delay slots it needs, at which
45 * point we return back up to the top and attempt to schedule by next
46 * highest depth. After a sufficient number of instructions have been
47 * scheduled, return back to beginning of list and start again. If you
48 * reach the end of depth sorted list without being able to insert any
49 * instruction, insert nop's. Repeat until no more unscheduled
52 * There are a few special cases that need to be handled, since sched
53 * is currently independent of register allocation. Usages of address
54 * register (a0.x) or predicate register (p0.x) must be serialized. Ie.
55 * if you have two pairs of instructions that write the same special
56 * register and then read it, then those pairs cannot be interleaved.
57 * To solve this, when we are in such a scheduling "critical section",
58 * and we encounter a conflicting write to a special register, we try
59 * to schedule any remaining instructions that use that value first.
62 struct ir3_sched_ctx
{
63 struct ir3_instruction
*scheduled
; /* last scheduled instr */
64 struct ir3_instruction
*addr
; /* current a0.x user, if any */
65 struct ir3_instruction
*pred
; /* current p0.x user, if any */
69 static struct ir3_instruction
*
70 deepest(struct ir3_instruction
**srcs
, unsigned nsrcs
)
72 struct ir3_instruction
*d
= NULL
;
73 unsigned i
= 0, id
= 0;
75 while ((i
< nsrcs
) && !(d
= srcs
[id
= i
]))
81 for (; i
< nsrcs
; i
++)
82 if (srcs
[i
] && (srcs
[i
]->depth
> d
->depth
))
90 static unsigned distance(struct ir3_sched_ctx
*ctx
,
91 struct ir3_instruction
*instr
, unsigned maxd
)
93 struct ir3_instruction
*n
= ctx
->scheduled
;
95 while (n
&& (n
!= instr
) && (d
< maxd
)) {
96 if (is_alu(n
) || is_flow(n
))
103 /* TODO maybe we want double linked list? */
104 static struct ir3_instruction
* prev(struct ir3_instruction
*instr
)
106 struct ir3_instruction
*p
= instr
->block
->head
;
107 while (p
&& (p
->next
!= instr
))
112 static void schedule(struct ir3_sched_ctx
*ctx
,
113 struct ir3_instruction
*instr
, bool remove
)
115 struct ir3_block
*block
= instr
->block
;
117 /* maybe there is a better way to handle this than just stuffing
118 * a nop.. ideally we'd know about this constraint in the
119 * scheduling and depth calculation..
121 if (ctx
->scheduled
&& is_sfu(ctx
->scheduled
) && is_sfu(instr
))
122 schedule(ctx
, ir3_instr_create(block
, 0, OPC_NOP
), false);
124 /* remove from depth list:
127 struct ir3_instruction
*p
= prev(instr
);
129 /* NOTE: this can happen for inputs which are not
130 * read.. in that case there is no need to schedule
131 * the input, so just bail:
133 if (instr
!= (p
? p
->next
: block
->head
))
137 p
->next
= instr
->next
;
139 block
->head
= instr
->next
;
142 if (writes_addr(instr
)) {
143 assert(ctx
->addr
== NULL
);
147 if (writes_pred(instr
)) {
148 assert(ctx
->pred
== NULL
);
152 instr
->flags
|= IR3_INSTR_MARK
;
154 instr
->next
= ctx
->scheduled
;
155 ctx
->scheduled
= instr
;
161 * Delay-slot calculation. Follows fanin/fanout.
164 static unsigned delay_calc2(struct ir3_sched_ctx
*ctx
,
165 struct ir3_instruction
*assigner
,
166 struct ir3_instruction
*consumer
, unsigned srcn
)
170 if (is_meta(assigner
)) {
172 for (i
= 1; i
< assigner
->regs_count
; i
++) {
173 struct ir3_register
*reg
= assigner
->regs
[i
];
174 if (reg
->flags
& IR3_REG_SSA
) {
175 unsigned d
= delay_calc2(ctx
, reg
->instr
,
177 delay
= MAX2(delay
, d
);
181 delay
= ir3_delayslots(assigner
, consumer
, srcn
);
182 delay
-= distance(ctx
, assigner
, delay
);
188 static unsigned delay_calc(struct ir3_sched_ctx
*ctx
,
189 struct ir3_instruction
*instr
)
191 unsigned i
, delay
= 0;
193 for (i
= 1; i
< instr
->regs_count
; i
++) {
194 struct ir3_register
*reg
= instr
->regs
[i
];
195 if (reg
->flags
& IR3_REG_SSA
) {
196 unsigned d
= delay_calc2(ctx
, reg
->instr
,
198 delay
= MAX2(delay
, d
);
205 /* A negative return value signals that an instruction has been newly
206 * scheduled, return back up to the top of the stack (to block_sched())
208 static int trysched(struct ir3_sched_ctx
*ctx
,
209 struct ir3_instruction
*instr
)
211 struct ir3_instruction
*srcs
[ARRAY_SIZE(instr
->regs
) - 1];
212 struct ir3_instruction
*src
;
213 unsigned i
, delay
, nsrcs
= 0;
215 /* if already scheduled: */
216 if (instr
->flags
& IR3_INSTR_MARK
)
219 /* figure out our src's: */
220 for (i
= 1; i
< instr
->regs_count
; i
++) {
221 struct ir3_register
*reg
= instr
->regs
[i
];
222 if (reg
->flags
& IR3_REG_SSA
)
223 srcs
[nsrcs
++] = reg
->instr
;
226 /* for each src register in sorted order:
229 while ((src
= deepest(srcs
, nsrcs
))) {
230 delay
= trysched(ctx
, src
);
235 /* all our dependents are scheduled, figure out if
236 * we have enough delay slots to schedule ourself:
238 delay
= delay_calc(ctx
, instr
);
242 /* if this is a write to address/predicate register, and that
243 * register is currently in use, we need to defer until it is
246 if (writes_addr(instr
) && ctx
->addr
) {
247 assert(ctx
->addr
!= instr
);
250 if (writes_pred(instr
) && ctx
->pred
) {
251 assert(ctx
->pred
!= instr
);
255 schedule(ctx
, instr
, true);
259 static struct ir3_instruction
* reverse(struct ir3_instruction
*instr
)
261 struct ir3_instruction
*reversed
= NULL
;
263 struct ir3_instruction
*next
= instr
->next
;
264 instr
->next
= reversed
;
271 static bool uses_current_addr(struct ir3_sched_ctx
*ctx
,
272 struct ir3_instruction
*instr
)
275 for (i
= 1; i
< instr
->regs_count
; i
++) {
276 struct ir3_register
*reg
= instr
->regs
[i
];
277 if (reg
->flags
& IR3_REG_SSA
) {
278 if (is_addr(reg
->instr
)) {
279 struct ir3_instruction
*addr
;
280 addr
= reg
->instr
->regs
[1]->instr
; /* the mova */
281 if (ctx
->addr
== addr
)
289 static bool uses_current_pred(struct ir3_sched_ctx
*ctx
,
290 struct ir3_instruction
*instr
)
293 for (i
= 1; i
< instr
->regs_count
; i
++) {
294 struct ir3_register
*reg
= instr
->regs
[i
];
295 if ((reg
->flags
& IR3_REG_SSA
) && (ctx
->pred
== reg
->instr
))
301 /* when we encounter an instruction that writes to the address register
302 * when it is in use, we delay that instruction and try to schedule all
303 * other instructions using the current address register:
305 static int block_sched_undelayed(struct ir3_sched_ctx
*ctx
,
306 struct ir3_block
*block
)
308 struct ir3_instruction
*instr
= block
->head
;
309 bool addr_in_use
= false;
310 bool pred_in_use
= false;
314 struct ir3_instruction
*next
= instr
->next
;
315 bool addr
= uses_current_addr(ctx
, instr
);
316 bool pred
= uses_current_pred(ctx
, instr
);
319 int ret
= trysched(ctx
, instr
);
320 if (ret
== SCHEDULED
)
323 cnt
= MIN2(cnt
, ret
);
342 static void block_sched(struct ir3_sched_ctx
*ctx
, struct ir3_block
*block
)
344 struct ir3_instruction
*instr
;
346 /* schedule all the shader input's (meta-instr) first so that
347 * the RA step sees that the input registers contain a value
348 * from the start of the shader:
350 if (!block
->parent
) {
352 for (i
= 0; i
< block
->ninputs
; i
++) {
353 struct ir3_instruction
*in
= block
->inputs
[i
];
355 schedule(ctx
, in
, true);
359 while ((instr
= block
->head
)) {
360 /* NOTE: always grab next *before* trysched(), in case the
361 * instruction is actually scheduled (and therefore moved
362 * from depth list into scheduled list)
364 struct ir3_instruction
*next
= instr
->next
;
365 int cnt
= trysched(ctx
, instr
);
368 cnt
= block_sched_undelayed(ctx
, block
);
370 /* -1 is signal to return up stack, but to us means same as 0: */
375 /* if deepest remaining instruction cannot be scheduled, try
376 * the increasingly more shallow instructions until needed
377 * number of delay slots is filled:
379 while (instr
&& (cnt
> ctx
->cnt
)) {
381 trysched(ctx
, instr
);
385 /* and if we run out of instructions that can be scheduled,
386 * then it is time for nop's:
388 while (cnt
> ctx
->cnt
)
389 schedule(ctx
, ir3_instr_create(block
, 0, OPC_NOP
), false);
392 /* at this point, scheduled list is in reverse order, so fix that: */
393 block
->head
= reverse(ctx
->scheduled
);
396 void ir3_block_sched(struct ir3_block
*block
)
398 struct ir3_sched_ctx ctx
= {0};
399 ir3_clear_mark(block
->shader
);
400 block_sched(&ctx
, block
);