summaryrefslogtreecommitdiffstats
path: root/src/mesa/drivers/dri/i965/brw_vec4_reg_allocate.cpp
blob: afc326612a2812f3ff0b667a356788e6fcdab48a (plain)
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
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
/*
 * Copyright © 2011 Intel Corporation
 *
 * 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.
 */

#include "util/register_allocate.h"
#include "brw_vec4.h"
#include "brw_cfg.h"

using namespace brw;

namespace brw {

static void
assign(unsigned int *reg_hw_locations, backend_reg *reg)
{
   if (reg->file == VGRF) {
      reg->nr = reg_hw_locations[reg->nr] + reg->reg_offset;
      reg->reg_offset = 0;
   }
}

bool
vec4_visitor::reg_allocate_trivial()
{
   unsigned int hw_reg_mapping[this->alloc.count];
   bool virtual_grf_used[this->alloc.count];
   int next;

   /* Calculate which virtual GRFs are actually in use after whatever
    * optimization passes have occurred.
    */
   for (unsigned i = 0; i < this->alloc.count; i++) {
      virtual_grf_used[i] = false;
   }

   foreach_block_and_inst(block, vec4_instruction, inst, cfg) {
      if (inst->dst.file == VGRF)
         virtual_grf_used[inst->dst.nr] = true;

      for (unsigned i = 0; i < 3; i++) {
	 if (inst->src[i].file == VGRF)
            virtual_grf_used[inst->src[i].nr] = true;
      }
   }

   hw_reg_mapping[0] = this->first_non_payload_grf;
   next = hw_reg_mapping[0] + this->alloc.sizes[0];
   for (unsigned i = 1; i < this->alloc.count; i++) {
      if (virtual_grf_used[i]) {
	 hw_reg_mapping[i] = next;
	 next += this->alloc.sizes[i];
      }
   }
   prog_data->total_grf = next;

   foreach_block_and_inst(block, vec4_instruction, inst, cfg) {
      assign(hw_reg_mapping, &inst->dst);
      assign(hw_reg_mapping, &inst->src[0]);
      assign(hw_reg_mapping, &inst->src[1]);
      assign(hw_reg_mapping, &inst->src[2]);
   }

   if (prog_data->total_grf > max_grf) {
      fail("Ran out of regs on trivial allocator (%d/%d)\n",
	   prog_data->total_grf, max_grf);
      return false;
   }

   return true;
}

extern "C" void
brw_vec4_alloc_reg_set(struct brw_compiler *compiler)
{
   int base_reg_count =
      compiler->devinfo->gen >= 7 ? GEN7_MRF_HACK_START : BRW_MAX_GRF;

   /* After running split_virtual_grfs(), almost all VGRFs will be of size 1.
    * SEND-from-GRF sources cannot be split, so we also need classes for each
    * potential message length.
    */
   const int class_count = MAX_VGRF_SIZE;
   int class_sizes[MAX_VGRF_SIZE];

   for (int i = 0; i < class_count; i++)
      class_sizes[i] = i + 1;

   /* Compute the total number of registers across all classes. */
   int ra_reg_count = 0;
   for (int i = 0; i < class_count; i++) {
      ra_reg_count += base_reg_count - (class_sizes[i] - 1);
   }

   ralloc_free(compiler->vec4_reg_set.ra_reg_to_grf);
   compiler->vec4_reg_set.ra_reg_to_grf = ralloc_array(compiler, uint8_t, ra_reg_count);
   ralloc_free(compiler->vec4_reg_set.regs);
   compiler->vec4_reg_set.regs = ra_alloc_reg_set(compiler, ra_reg_count, false);
   if (compiler->devinfo->gen >= 6)
      ra_set_allocate_round_robin(compiler->vec4_reg_set.regs);
   ralloc_free(compiler->vec4_reg_set.classes);
   compiler->vec4_reg_set.classes = ralloc_array(compiler, int, class_count);

   /* Now, add the registers to their classes, and add the conflicts
    * between them and the base GRF registers (and also each other).
    */
   int reg = 0;
   unsigned *q_values[MAX_VGRF_SIZE];
   for (int i = 0; i < class_count; i++) {
      int class_reg_count = base_reg_count - (class_sizes[i] - 1);
      compiler->vec4_reg_set.classes[i] = ra_alloc_reg_class(compiler->vec4_reg_set.regs);

      q_values[i] = new unsigned[MAX_VGRF_SIZE];

      for (int j = 0; j < class_reg_count; j++) {
	 ra_class_add_reg(compiler->vec4_reg_set.regs, compiler->vec4_reg_set.classes[i], reg);

	 compiler->vec4_reg_set.ra_reg_to_grf[reg] = j;

	 for (int base_reg = j;
	      base_reg < j + class_sizes[i];
	      base_reg++) {
	    ra_add_reg_conflict(compiler->vec4_reg_set.regs, base_reg, reg);
	 }

	 reg++;
      }

      for (int j = 0; j < class_count; j++) {
         /* Calculate the q values manually because the algorithm used by
          * ra_set_finalize() to do it has higher complexity affecting the
          * start-up time of some applications.  q(i, j) is just the maximum
          * number of registers from class i a register from class j can
          * conflict with.
          */
         q_values[i][j] = class_sizes[i] + class_sizes[j] - 1;
      }
   }
   assert(reg == ra_reg_count);

   for (int reg = 0; reg < base_reg_count; reg++)
      ra_make_reg_conflicts_transitive(compiler->vec4_reg_set.regs, reg);

   ra_set_finalize(compiler->vec4_reg_set.regs, q_values);

   for (int i = 0; i < MAX_VGRF_SIZE; i++)
      delete[] q_values[i];
}

void
vec4_visitor::setup_payload_interference(struct ra_graph *g,
                                         int first_payload_node,
                                         int reg_node_count)
{
   int payload_node_count = this->first_non_payload_grf;

   for (int i = 0; i < payload_node_count; i++) {
      /* Mark each payload reg node as being allocated to its physical register.
       *
       * The alternative would be to have per-physical register classes, which
       * would just be silly.
       */
      ra_set_node_reg(g, first_payload_node + i, i);

      /* For now, just mark each payload node as interfering with every other
       * node to be allocated.
       */
      for (int j = 0; j < reg_node_count; j++) {
         ra_add_node_interference(g, first_payload_node + i, j);
      }
   }
}

bool
vec4_visitor::reg_allocate()
{
   unsigned int hw_reg_mapping[alloc.count];
   int payload_reg_count = this->first_non_payload_grf;

   /* Using the trivial allocator can be useful in debugging undefined
    * register access as a result of broken optimization passes.
    */
   if (0)
      return reg_allocate_trivial();

   calculate_live_intervals();

   int node_count = alloc.count;
   int first_payload_node = node_count;
   node_count += payload_reg_count;
   struct ra_graph *g =
      ra_alloc_interference_graph(compiler->vec4_reg_set.regs, node_count);

   for (unsigned i = 0; i < alloc.count; i++) {
      int size = this->alloc.sizes[i];
      assert(size >= 1 && size <= MAX_VGRF_SIZE);
      ra_set_node_class(g, i, compiler->vec4_reg_set.classes[size - 1]);

      for (unsigned j = 0; j < i; j++) {
	 if (virtual_grf_interferes(i, j)) {
	    ra_add_node_interference(g, i, j);
	 }
      }
   }

   /* Certain instructions can't safely use the same register for their
    * sources and destination.  Add interference.
    */
   foreach_block_and_inst(block, vec4_instruction, inst, cfg) {
      if (inst->dst.file == VGRF && inst->has_source_and_destination_hazard()) {
         for (unsigned i = 0; i < 3; i++) {
            if (inst->src[i].file == VGRF) {
               ra_add_node_interference(g, inst->dst.nr, inst->src[i].nr);
            }
         }
      }
   }

   setup_payload_interference(g, first_payload_node, node_count);

   if (!ra_allocate(g)) {
      /* Failed to allocate registers.  Spill a reg, and the caller will
       * loop back into here to try again.
       */
      int reg = choose_spill_reg(g);
      if (this->no_spills) {
         fail("Failure to register allocate.  Reduce number of live "
              "values to avoid this.");
      } else if (reg == -1) {
         fail("no register to spill\n");
      } else {
         spill_reg(reg);
      }
      ralloc_free(g);
      return false;
   }

   /* Get the chosen virtual registers for each node, and map virtual
    * regs in the register classes back down to real hardware reg
    * numbers.
    */
   prog_data->total_grf = payload_reg_count;
   for (unsigned i = 0; i < alloc.count; i++) {
      int reg = ra_get_node_reg(g, i);

      hw_reg_mapping[i] = compiler->vec4_reg_set.ra_reg_to_grf[reg];
      prog_data->total_grf = MAX2(prog_data->total_grf,
				  hw_reg_mapping[i] + alloc.sizes[i]);
   }

   foreach_block_and_inst(block, vec4_instruction, inst, cfg) {
      assign(hw_reg_mapping, &inst->dst);
      assign(hw_reg_mapping, &inst->src[0]);
      assign(hw_reg_mapping, &inst->src[1]);
      assign(hw_reg_mapping, &inst->src[2]);
   }

   ralloc_free(g);

   return true;
}

/**
 * When we decide to spill a register, instead of blindly spilling every use,
 * save unspills when the spill register is used (read) in consecutive
 * instructions. This can potentially save a bunch of unspills that would
 * have very little impact in register allocation anyway.
 *
 * Notice that we need to account for this behavior when spilling a register
 * and when evaluating spilling costs. This function is designed so it can
 * be called from both places and avoid repeating the logic.
 *
 *  - When we call this function from spill_reg(), we pass in scratch_reg the
 *    actual unspill/spill register that we want to reuse in the current
 *    instruction.
 *
 *  - When we call this from evaluate_spill_costs(), we pass the register for
 *    which we are evaluating spilling costs.
 *
 * In either case, we check if the previous instructions read scratch_reg until
 * we find one that writes to it with a compatible mask or does not read/write
 * scratch_reg at all.
 */
static bool
can_use_scratch_for_source(const vec4_instruction *inst, unsigned i,
                           unsigned scratch_reg)
{
   assert(inst->src[i].file == VGRF);
   bool prev_inst_read_scratch_reg = false;

   /* See if any previous source in the same instructions reads scratch_reg */
   for (unsigned n = 0; n < i; n++) {
      if (inst->src[n].file == VGRF && inst->src[n].nr == scratch_reg)
         prev_inst_read_scratch_reg = true;
   }

   /* Now check if previous instructions read/write scratch_reg */
   for (vec4_instruction *prev_inst = (vec4_instruction *) inst->prev;
        !prev_inst->is_head_sentinel();
        prev_inst = (vec4_instruction *) prev_inst->prev) {

      /* If the previous instruction writes to scratch_reg then we can reuse
       * it if the write is not conditional and the channels we write are
       * compatible with our read mask
       */
      if (prev_inst->dst.file == VGRF && prev_inst->dst.nr == scratch_reg) {
         return (!prev_inst->predicate || prev_inst->opcode == BRW_OPCODE_SEL) &&
                (brw_mask_for_swizzle(inst->src[i].swizzle) &
                 ~prev_inst->dst.writemask) == 0;
      }

      /* Skip scratch read/writes so that instructions generated by spilling
       * other registers (that won't read/write scratch_reg) do not stop us from
       * reusing scratch_reg for this instruction.
       */
      if (prev_inst->opcode == SHADER_OPCODE_GEN4_SCRATCH_WRITE ||
          prev_inst->opcode == SHADER_OPCODE_GEN4_SCRATCH_READ)
         continue;

      /* If the previous instruction does not write to scratch_reg, then check
       * if it reads it
       */
      int n;
      for (n = 0; n < 3; n++) {
         if (prev_inst->src[n].file == VGRF &&
             prev_inst->src[n].nr == scratch_reg) {
            prev_inst_read_scratch_reg = true;
            break;
         }
      }
      if (n == 3) {
         /* The previous instruction does not read scratch_reg. At this point,
          * if no previous instruction has read scratch_reg it means that we
          * will need to unspill it here and we can't reuse it (so we return
          * false). Otherwise, if we found at least one consecutive instruction
          * that read scratch_reg, then we know that we got here from
          * evaluate_spill_costs (since for the spill_reg path any block of
          * consecutive instructions using scratch_reg must start with a write
          * to that register, so we would've exited the loop in the check for
          * the write that we have at the start of this loop), and in that case
          * it means that we found the point at which the scratch_reg would be
          * unspilled. Since we always unspill a full vec4, it means that we
          * have all the channels available and we can just return true to
          * signal that we can reuse the register in the current instruction
          * too.
          */
         return prev_inst_read_scratch_reg;
      }
   }

   return prev_inst_read_scratch_reg;
}

void
vec4_visitor::evaluate_spill_costs(float *spill_costs, bool *no_spill)
{
   float loop_scale = 1.0;

   for (unsigned i = 0; i < this->alloc.count; i++) {
      spill_costs[i] = 0.0;
      no_spill[i] = alloc.sizes[i] != 1;
   }

   /* Calculate costs for spilling nodes.  Call it a cost of 1 per
    * spill/unspill we'll have to do, and guess that the insides of
    * loops run 10 times.
    */
   foreach_block_and_inst(block, vec4_instruction, inst, cfg) {
      for (unsigned int i = 0; i < 3; i++) {
         if (inst->src[i].file == VGRF) {
            /* We will only unspill src[i] it it wasn't unspilled for the
             * previous instruction, in which case we'll just reuse the scratch
             * reg for this instruction.
             */
            if (!can_use_scratch_for_source(inst, i, inst->src[i].nr)) {
               spill_costs[inst->src[i].nr] += loop_scale;
               if (inst->src[i].reladdr)
                  no_spill[inst->src[i].nr] = true;
            }
         }
      }

      if (inst->dst.file == VGRF) {
         spill_costs[inst->dst.nr] += loop_scale;
         if (inst->dst.reladdr)
            no_spill[inst->dst.nr] = true;
      }

      switch (inst->opcode) {

      case BRW_OPCODE_DO:
         loop_scale *= 10;
         break;

      case BRW_OPCODE_WHILE:
         loop_scale /= 10;
         break;

      case SHADER_OPCODE_GEN4_SCRATCH_READ:
      case SHADER_OPCODE_GEN4_SCRATCH_WRITE:
         for (int i = 0; i < 3; i++) {
            if (inst->src[i].file == VGRF)
               no_spill[inst->src[i].nr] = true;
         }
         if (inst->dst.file == VGRF)
            no_spill[inst->dst.nr] = true;
         break;

      default:
         break;
      }
   }
}

int
vec4_visitor::choose_spill_reg(struct ra_graph *g)
{
   float spill_costs[this->alloc.count];
   bool no_spill[this->alloc.count];

   evaluate_spill_costs(spill_costs, no_spill);

   for (unsigned i = 0; i < this->alloc.count; i++) {
      if (!no_spill[i])
         ra_set_node_spill_cost(g, i, spill_costs[i]);
   }

   return ra_get_best_spill_node(g);
}

void
vec4_visitor::spill_reg(int spill_reg_nr)
{
   assert(alloc.sizes[spill_reg_nr] == 1);
   unsigned int spill_offset = last_scratch++;

   /* Generate spill/unspill instructions for the objects being spilled. */
   int scratch_reg = -1;
   foreach_block_and_inst(block, vec4_instruction, inst, cfg) {
      for (unsigned int i = 0; i < 3; i++) {
         if (inst->src[i].file == VGRF && inst->src[i].nr == spill_reg_nr) {
            if (scratch_reg == -1 ||
                !can_use_scratch_for_source(inst, i, scratch_reg)) {
               /* We need to unspill anyway so make sure we read the full vec4
                * in any case. This way, the cached register can be reused
                * for consecutive instructions that read different channels of
                * the same vec4.
                */
               scratch_reg = alloc.allocate(1);
               src_reg temp = inst->src[i];
               temp.nr = scratch_reg;
               temp.swizzle = BRW_SWIZZLE_XYZW;
               emit_scratch_read(block, inst,
                                 dst_reg(temp), inst->src[i], spill_offset);
            }
            assert(scratch_reg != -1);
            inst->src[i].nr = scratch_reg;
         }
      }

      if (inst->dst.file == VGRF && inst->dst.nr == spill_reg_nr) {
         emit_scratch_write(block, inst, spill_offset);
         scratch_reg = inst->dst.nr;
      }
   }

   invalidate_live_intervals();
}

} /* namespace brw */