1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
|
/* -*- mode: C; c-file-style: "k&r"; tab-width 4; indent-tabs-mode: t; -*- */
/*
* Copyright (C) 2014 Rob Clark <robclark@freedesktop.org>
*
* Permission is hereby granted, free of charge, to any person obtaining a
* copy of this software and associated documentation files (the "Software"),
* to deal in the Software without restriction, including without limitation
* the rights to use, copy, modify, merge, publish, distribute, sublicense,
* and/or sell copies of the Software, and to permit persons to whom the
* Software is furnished to do so, subject to the following conditions:
*
* The above copyright notice and this permission notice (including the next
* paragraph) shall be included in all copies or substantial portions of the
* Software.
*
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
* THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
* LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
* OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
* SOFTWARE.
*
* Authors:
* Rob Clark <robclark@freedesktop.org>
*/
#include "util/u_math.h"
#include "ir3.h"
enum {
SCHEDULED = -1,
DELAYED = -2,
};
/*
* Instruction Scheduling:
*
* Using the depth sorted list from depth pass, attempt to recursively
* schedule deepest unscheduled path. The first instruction that cannot
* be scheduled, returns the required delay slots it needs, at which
* point we return back up to the top and attempt to schedule by next
* highest depth. After a sufficient number of instructions have been
* scheduled, return back to beginning of list and start again. If you
* reach the end of depth sorted list without being able to insert any
* instruction, insert nop's. Repeat until no more unscheduled
* instructions.
*
* There are a few special cases that need to be handled, since sched
* is currently independent of register allocation. Usages of address
* register (a0.x) or predicate register (p0.x) must be serialized. Ie.
* if you have two pairs of instructions that write the same special
* register and then read it, then those pairs cannot be interleaved.
* To solve this, when we are in such a scheduling "critical section",
* and we encounter a conflicting write to a special register, we try
* to schedule any remaining instructions that use that value first.
*/
struct ir3_sched_ctx {
struct ir3_instruction *scheduled; /* last scheduled instr */
struct ir3_instruction *addr; /* current a0.x user, if any */
struct ir3_instruction *pred; /* current p0.x user, if any */
unsigned cnt;
};
static struct ir3_instruction *
deepest(struct ir3_instruction **srcs, unsigned nsrcs)
{
struct ir3_instruction *d = NULL;
unsigned i = 0, id = 0;
while ((i < nsrcs) && !(d = srcs[id = i]))
i++;
if (!d)
return NULL;
for (; i < nsrcs; i++)
if (srcs[i] && (srcs[i]->depth > d->depth))
d = srcs[id = i];
srcs[id] = NULL;
return d;
}
static unsigned distance(struct ir3_sched_ctx *ctx,
struct ir3_instruction *instr, unsigned maxd)
{
struct ir3_instruction *n = ctx->scheduled;
unsigned d = 0;
while (n && (n != instr) && (d < maxd)) {
if (is_alu(n) || is_flow(n))
d++;
n = n->next;
}
return d;
}
/* TODO maybe we want double linked list? */
static struct ir3_instruction * prev(struct ir3_instruction *instr)
{
struct ir3_instruction *p = instr->block->head;
while (p && (p->next != instr))
p = p->next;
return p;
}
static void schedule(struct ir3_sched_ctx *ctx,
struct ir3_instruction *instr, bool remove)
{
struct ir3_block *block = instr->block;
/* maybe there is a better way to handle this than just stuffing
* a nop.. ideally we'd know about this constraint in the
* scheduling and depth calculation..
*/
if (ctx->scheduled && is_sfu(ctx->scheduled) && is_sfu(instr))
schedule(ctx, ir3_instr_create(block, 0, OPC_NOP), false);
/* remove from depth list:
*/
if (remove) {
struct ir3_instruction *p = prev(instr);
/* NOTE: this can happen for inputs which are not
* read.. in that case there is no need to schedule
* the input, so just bail:
*/
if (instr != (p ? p->next : block->head))
return;
if (p)
p->next = instr->next;
else
block->head = instr->next;
}
if (writes_addr(instr)) {
assert(ctx->addr == NULL);
ctx->addr = instr;
}
if (writes_pred(instr)) {
assert(ctx->pred == NULL);
ctx->pred = instr;
}
instr->flags |= IR3_INSTR_MARK;
instr->next = ctx->scheduled;
ctx->scheduled = instr;
ctx->cnt++;
}
/*
* Delay-slot calculation. Follows fanin/fanout.
*/
static unsigned delay_calc2(struct ir3_sched_ctx *ctx,
struct ir3_instruction *assigner,
struct ir3_instruction *consumer, unsigned srcn)
{
unsigned delay = 0;
if (is_meta(assigner)) {
unsigned i;
for (i = 1; i < assigner->regs_count; i++) {
struct ir3_register *reg = assigner->regs[i];
if (reg->flags & IR3_REG_SSA) {
unsigned d = delay_calc2(ctx, reg->instr,
consumer, srcn);
delay = MAX2(delay, d);
}
}
} else {
delay = ir3_delayslots(assigner, consumer, srcn);
delay -= distance(ctx, assigner, delay);
}
return delay;
}
static unsigned delay_calc(struct ir3_sched_ctx *ctx,
struct ir3_instruction *instr)
{
unsigned i, delay = 0;
for (i = 1; i < instr->regs_count; i++) {
struct ir3_register *reg = instr->regs[i];
if (reg->flags & IR3_REG_SSA) {
unsigned d = delay_calc2(ctx, reg->instr,
instr, i - 1);
delay = MAX2(delay, d);
}
}
return delay;
}
/* A negative return value signals that an instruction has been newly
* scheduled, return back up to the top of the stack (to block_sched())
*/
static int trysched(struct ir3_sched_ctx *ctx,
struct ir3_instruction *instr)
{
struct ir3_instruction *srcs[ARRAY_SIZE(instr->regs) - 1];
struct ir3_instruction *src;
unsigned i, delay, nsrcs = 0;
/* if already scheduled: */
if (instr->flags & IR3_INSTR_MARK)
return 0;
/* figure out our src's: */
for (i = 1; i < instr->regs_count; i++) {
struct ir3_register *reg = instr->regs[i];
if (reg->flags & IR3_REG_SSA)
srcs[nsrcs++] = reg->instr;
}
/* for each src register in sorted order:
*/
delay = 0;
while ((src = deepest(srcs, nsrcs))) {
delay = trysched(ctx, src);
if (delay)
return delay;
}
/* all our dependents are scheduled, figure out if
* we have enough delay slots to schedule ourself:
*/
delay = delay_calc(ctx, instr);
if (delay)
return delay;
/* if this is a write to address/predicate register, and that
* register is currently in use, we need to defer until it is
* free:
*/
if (writes_addr(instr) && ctx->addr) {
assert(ctx->addr != instr);
return DELAYED;
}
if (writes_pred(instr) && ctx->pred) {
assert(ctx->pred != instr);
return DELAYED;
}
schedule(ctx, instr, true);
return SCHEDULED;
}
static struct ir3_instruction * reverse(struct ir3_instruction *instr)
{
struct ir3_instruction *reversed = NULL;
while (instr) {
struct ir3_instruction *next = instr->next;
instr->next = reversed;
reversed = instr;
instr = next;
}
return reversed;
}
static bool uses_current_addr(struct ir3_sched_ctx *ctx,
struct ir3_instruction *instr)
{
unsigned i;
for (i = 1; i < instr->regs_count; i++) {
struct ir3_register *reg = instr->regs[i];
if (reg->flags & IR3_REG_SSA) {
if (is_addr(reg->instr)) {
struct ir3_instruction *addr;
addr = reg->instr->regs[1]->instr; /* the mova */
if (ctx->addr == addr)
return true;
}
}
}
return false;
}
static bool uses_current_pred(struct ir3_sched_ctx *ctx,
struct ir3_instruction *instr)
{
unsigned i;
for (i = 1; i < instr->regs_count; i++) {
struct ir3_register *reg = instr->regs[i];
if ((reg->flags & IR3_REG_SSA) && (ctx->pred == reg->instr))
return true;
}
return false;
}
/* when we encounter an instruction that writes to the address register
* when it is in use, we delay that instruction and try to schedule all
* other instructions using the current address register:
*/
static int block_sched_undelayed(struct ir3_sched_ctx *ctx,
struct ir3_block *block)
{
struct ir3_instruction *instr = block->head;
bool addr_in_use = false;
bool pred_in_use = false;
unsigned cnt = ~0;
while (instr) {
struct ir3_instruction *next = instr->next;
bool addr = uses_current_addr(ctx, instr);
bool pred = uses_current_pred(ctx, instr);
if (addr || pred) {
int ret = trysched(ctx, instr);
if (ret == SCHEDULED)
cnt = 0;
else if (ret > 0)
cnt = MIN2(cnt, ret);
if (addr)
addr_in_use = true;
if (pred)
pred_in_use = true;
}
instr = next;
}
if (!addr_in_use)
ctx->addr = NULL;
if (!pred_in_use)
ctx->pred = NULL;
return cnt;
}
static void block_sched(struct ir3_sched_ctx *ctx, struct ir3_block *block)
{
struct ir3_instruction *instr;
/* schedule all the shader input's (meta-instr) first so that
* the RA step sees that the input registers contain a value
* from the start of the shader:
*/
if (!block->parent) {
unsigned i;
for (i = 0; i < block->ninputs; i++) {
struct ir3_instruction *in = block->inputs[i];
if (in)
schedule(ctx, in, true);
}
}
while ((instr = block->head)) {
/* NOTE: always grab next *before* trysched(), in case the
* instruction is actually scheduled (and therefore moved
* from depth list into scheduled list)
*/
struct ir3_instruction *next = instr->next;
int cnt = trysched(ctx, instr);
if (cnt == DELAYED)
cnt = block_sched_undelayed(ctx, block);
/* -1 is signal to return up stack, but to us means same as 0: */
cnt = MAX2(0, cnt);
cnt += ctx->cnt;
instr = next;
/* if deepest remaining instruction cannot be scheduled, try
* the increasingly more shallow instructions until needed
* number of delay slots is filled:
*/
while (instr && (cnt > ctx->cnt)) {
next = instr->next;
trysched(ctx, instr);
instr = next;
}
/* and if we run out of instructions that can be scheduled,
* then it is time for nop's:
*/
while (cnt > ctx->cnt)
schedule(ctx, ir3_instr_create(block, 0, OPC_NOP), false);
}
/* at this point, scheduled list is in reverse order, so fix that: */
block->head = reverse(ctx->scheduled);
}
void ir3_block_sched(struct ir3_block *block)
{
struct ir3_sched_ctx ctx = {0};
ir3_clear_mark(block->shader);
block_sched(&ctx, block);
}
|