src/gallium/drivers/vc4/vc4_register_allocate.c

   1 /*
   2  * Copyright © 2014 Broadcom
   3  *
   4  * Permission is hereby granted, free of charge, to any person obtaining a
   5  * copy of this software and associated documentation files (the "Software"),
   6  * to deal in the Software without restriction, including without limitation
   7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   8  * and/or sell copies of the Software, and to permit persons to whom the
   9  * Software is furnished to do so, subject to the following conditions:
  10  *
  11  * The above copyright notice and this permission notice (including the next
  12  * paragraph) shall be included in all copies or substantial portions of the
  13  * Software.
  14  *
  15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  16  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  17  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  18  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  19  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
  20  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
  21  * IN THE SOFTWARE.
  22  */
  23
  24 #include "util/ralloc.h"
  25 #include "util/register_allocate.h"
  26 #include "vc4_context.h"
  27 #include "vc4_qir.h"
  28 #include "vc4_qpu.h"
  29
  30 #define QPU_R(file, index) { QPU_MUX_##file, index }
  31
  32 static const struct qpu_reg vc4_regs[] = {
  33         { QPU_MUX_R0, 0},
  34         { QPU_MUX_R1, 0},
  35         { QPU_MUX_R2, 0},
  36         { QPU_MUX_R3, 0},
  37         { QPU_MUX_R4, 0},
  38         QPU_R(A, 0),
  39         QPU_R(B, 0),
  40         QPU_R(A, 1),
  41         QPU_R(B, 1),
  42         QPU_R(A, 2),
  43         QPU_R(B, 2),
  44         QPU_R(A, 3),
  45         QPU_R(B, 3),
  46         QPU_R(A, 4),
  47         QPU_R(B, 4),
  48         QPU_R(A, 5),
  49         QPU_R(B, 5),
  50         QPU_R(A, 6),
  51         QPU_R(B, 6),
  52         QPU_R(A, 7),
  53         QPU_R(B, 7),
  54         QPU_R(A, 8),
  55         QPU_R(B, 8),
  56         QPU_R(A, 9),
  57         QPU_R(B, 9),
  58         QPU_R(A, 10),
  59         QPU_R(B, 10),
  60         QPU_R(A, 11),
  61         QPU_R(B, 11),
  62         QPU_R(A, 12),
  63         QPU_R(B, 12),
  64         QPU_R(A, 13),
  65         QPU_R(B, 13),
  66         QPU_R(A, 14),
  67         QPU_R(B, 14),
  68         QPU_R(A, 15),
  69         QPU_R(B, 15),
  70         QPU_R(A, 16),
  71         QPU_R(B, 16),
  72         QPU_R(A, 17),
  73         QPU_R(B, 17),
  74         QPU_R(A, 18),
  75         QPU_R(B, 18),
  76         QPU_R(A, 19),
  77         QPU_R(B, 19),
  78         QPU_R(A, 20),
  79         QPU_R(B, 20),
  80         QPU_R(A, 21),
  81         QPU_R(B, 21),
  82         QPU_R(A, 22),
  83         QPU_R(B, 22),
  84         QPU_R(A, 23),
  85         QPU_R(B, 23),
  86         QPU_R(A, 24),
  87         QPU_R(B, 24),
  88         QPU_R(A, 25),
  89         QPU_R(B, 25),
  90         QPU_R(A, 26),
  91         QPU_R(B, 26),
  92         QPU_R(A, 27),
  93         QPU_R(B, 27),
  94         QPU_R(A, 28),
  95         QPU_R(B, 28),
  96         QPU_R(A, 29),
  97         QPU_R(B, 29),
  98         QPU_R(A, 30),
  99         QPU_R(B, 30),
 100         QPU_R(A, 31),
 101         QPU_R(B, 31),
 102 };
 103 #define ACC_INDEX     0
 104 #define AB_INDEX      (ACC_INDEX + 5)
 105
 106 static void
 107 vc4_alloc_reg_set(struct vc4_context *vc4)
 108 {
 109         assert(vc4_regs[AB_INDEX].addr == 0);
 110         assert(vc4_regs[AB_INDEX + 1].addr == 0);
 111         STATIC_ASSERT(ARRAY_SIZE(vc4_regs) == AB_INDEX + 64);
 112
 113         if (vc4->regs)
 114                 return;
 115
 116         vc4->regs = ra_alloc_reg_set(vc4, ARRAY_SIZE(vc4_regs), true);
 117
 118         /* The physical regfiles split us into two classes, with [0] being the
 119          * whole space and [1] being the bottom half (for threaded fragment
 120          * shaders).
 121          */
 122         for (int i = 0; i < 2; i++) {
 123                 vc4->reg_class_any[i] = ra_alloc_reg_class(vc4->regs);
 124                 vc4->reg_class_a_or_b[i] = ra_alloc_reg_class(vc4->regs);
 125                 vc4->reg_class_a_or_b_or_acc[i] = ra_alloc_reg_class(vc4->regs);
 126                 vc4->reg_class_r4_or_a[i] = ra_alloc_reg_class(vc4->regs);
 127                 vc4->reg_class_a[i] = ra_alloc_reg_class(vc4->regs);
 128         }
 129         vc4->reg_class_r0_r3 = ra_alloc_reg_class(vc4->regs);
 130
 131         /* r0-r3 */
 132         for (uint32_t i = ACC_INDEX; i < ACC_INDEX + 4; i++) {
 133                 ra_class_add_reg(vc4->regs, vc4->reg_class_r0_r3, i);
 134                 ra_class_add_reg(vc4->regs, vc4->reg_class_a_or_b_or_acc[0], i);
 135                 ra_class_add_reg(vc4->regs, vc4->reg_class_a_or_b_or_acc[1], i);
 136         }
 137
 138         /* R4 gets a special class because it can't be written as a general
 139          * purpose register. (it's TMU_NOSWAP as a write address).
 140          */
 141         for (int i = 0; i < 2; i++) {
 142                 ra_class_add_reg(vc4->regs, vc4->reg_class_r4_or_a[i],
 143                                  ACC_INDEX + 4);
 144                 ra_class_add_reg(vc4->regs, vc4->reg_class_any[i],
 145                                  ACC_INDEX + 4);
 146         }
 147
 148         /* A/B */
 149         for (uint32_t i = AB_INDEX; i < AB_INDEX + 64; i ++) {
 150                 /* Reserve ra14/rb14 for spilling fixup_raddr_conflict() in
 151                  * vc4_qpu_emit.c
 152                  */
 153                 if (vc4_regs[i].addr == 14)
 154                         continue;
 155
 156                 ra_class_add_reg(vc4->regs, vc4->reg_class_any[0], i);
 157                 ra_class_add_reg(vc4->regs, vc4->reg_class_a_or_b[0], i);
 158                 ra_class_add_reg(vc4->regs, vc4->reg_class_a_or_b_or_acc[0], i);
 159
 160                 if (vc4_regs[i].addr < 16) {
 161                         ra_class_add_reg(vc4->regs, vc4->reg_class_any[1], i);
 162                         ra_class_add_reg(vc4->regs, vc4->reg_class_a_or_b[1], i);
 163                         ra_class_add_reg(vc4->regs, vc4->reg_class_a_or_b_or_acc[1], i);
 164                 }
 165
 166
 167                 /* A only */
 168                 if (((i - AB_INDEX) & 1) == 0) {
 169                         ra_class_add_reg(vc4->regs, vc4->reg_class_a[0], i);
 170                         ra_class_add_reg(vc4->regs, vc4->reg_class_r4_or_a[0], i);
 171
 172                         if (vc4_regs[i].addr < 16) {
 173                                 ra_class_add_reg(vc4->regs,
 174                                                  vc4->reg_class_a[1], i);
 175                                 ra_class_add_reg(vc4->regs,
 176                                                  vc4->reg_class_r4_or_a[1], i);
 177                         }
 178                 }
 179         }
 180
 181         ra_set_finalize(vc4->regs, NULL);
 182 }
 183
 184 struct node_to_temp_map {
 185         uint32_t temp;
 186         uint32_t priority;
 187 };
 188
 189 static int
 190 node_to_temp_priority(const void *in_a, const void *in_b)
 191 {
 192         const struct node_to_temp_map *a = in_a;
 193         const struct node_to_temp_map *b = in_b;
 194
 195         return a->priority - b->priority;
 196 }
 197
 198 #define CLASS_BIT_A                     (1 << 0)
 199 #define CLASS_BIT_B                     (1 << 1)
 200 #define CLASS_BIT_R4                    (1 << 2)
 201 #define CLASS_BIT_R0_R3                 (1 << 4)
 202
 203 /**
 204  * Returns a mapping from QFILE_TEMP indices to struct qpu_regs.
 205  *
 206  * The return value should be freed by the caller.
 207  */
 208 struct qpu_reg *
 209 vc4_register_allocate(struct vc4_context *vc4, struct vc4_compile *c)
 210 {
 211         struct node_to_temp_map map[c->num_temps];
 212         uint32_t temp_to_node[c->num_temps];
 213         uint8_t class_bits[c->num_temps];
 214         struct qpu_reg *temp_registers = calloc(c->num_temps,
 215                                                 sizeof(*temp_registers));
 216
 217         /* If things aren't ever written (undefined values), just read from
 218          * r0.
 219          */
 220         for (uint32_t i = 0; i < c->num_temps; i++)
 221                 temp_registers[i] = qpu_rn(0);
 222
 223         vc4_alloc_reg_set(vc4);
 224
 225         struct ra_graph *g = ra_alloc_interference_graph(vc4->regs,
 226                                                          c->num_temps);
 227
 228         /* Compute the live ranges so we can figure out interference. */
 229         qir_calculate_live_intervals(c);
 230
 231         for (uint32_t i = 0; i < c->num_temps; i++) {
 232                 map[i].temp = i;
 233                 map[i].priority = c->temp_end[i] - c->temp_start[i];
 234         }
 235         qsort(map, c->num_temps, sizeof(map[0]), node_to_temp_priority);
 236         for (uint32_t i = 0; i < c->num_temps; i++) {
 237                 temp_to_node[map[i].temp] = i;
 238         }
 239
 240         /* Figure out our register classes and preallocated registers.  We
 241          * start with any temp being able to be in any file, then instructions
 242          * incrementally remove bits that the temp definitely can't be in.
 243          */
 244         memset(class_bits,
 245                CLASS_BIT_A | CLASS_BIT_B | CLASS_BIT_R4 | CLASS_BIT_R0_R3,
 246                sizeof(class_bits));
 247
 248         int ip = 0;
 249         qir_for_each_inst_inorder(inst, c) {
 250                 if (qir_writes_r4(inst)) {
 251                         /* This instruction writes r4 (and optionally moves
 252                          * its result to a temp), so nothing else can be
 253                          * stored in r4 across it.
 254                          */
 255                         for (int i = 0; i < c->num_temps; i++) {
 256                                 if (c->temp_start[i] < ip && c->temp_end[i] > ip)
 257                                         class_bits[i] &= ~CLASS_BIT_R4;
 258                         }
 259
 260                         /* If we're doing a conditional write of something
 261                          * writing R4 (math, tex results), then make sure that
 262                          * we store in a temp so that we actually
 263                          * conditionally move the result.
 264                          */
 265                         if (inst->cond != QPU_COND_ALWAYS)
 266                                 class_bits[inst->dst.index] &= ~CLASS_BIT_R4;
 267                 } else {
 268                         /* R4 can't be written as a general purpose
 269                          * register. (it's TMU_NOSWAP as a write address).
 270                          */
 271                         if (inst->dst.file == QFILE_TEMP)
 272                                 class_bits[inst->dst.index] &= ~CLASS_BIT_R4;
 273                 }
 274
 275                 switch (inst->op) {
 276                 case QOP_FRAG_Z:
 277                         ra_set_node_reg(g, temp_to_node[inst->dst.index],
 278                                         AB_INDEX + QPU_R_FRAG_PAYLOAD_ZW * 2 + 1);
 279                         break;
 280
 281                 case QOP_FRAG_W:
 282                         ra_set_node_reg(g, temp_to_node[inst->dst.index],
 283                                         AB_INDEX + QPU_R_FRAG_PAYLOAD_ZW * 2);
 284                         break;
 285
 286                 case QOP_ROT_MUL:
 287                         assert(inst->src[0].file == QFILE_TEMP);
 288                         class_bits[inst->src[0].index] &= CLASS_BIT_R0_R3;
 289                         break;
 290
 291                 case QOP_THRSW:
 292                         /* All accumulators are invalidated across a thread
 293                          * switch.
 294                          */
 295                         for (int i = 0; i < c->num_temps; i++) {
 296                                 if (c->temp_start[i] < ip && c->temp_end[i] > ip)
 297                                         class_bits[i] &= ~(CLASS_BIT_R0_R3 |
 298                                                            CLASS_BIT_R4);
 299                         }
 300                         break;
 301
 302                 default:
 303                         break;
 304                 }
 305
 306                 if (inst->dst.pack && !qir_is_mul(inst)) {
 307                         /* The non-MUL pack flags require an A-file dst
 308                          * register.
 309                          */
 310                         class_bits[inst->dst.index] &= CLASS_BIT_A;
 311                 }
 312
 313                 /* Apply restrictions for src unpacks.  The integer unpacks
 314                  * can only be done from regfile A, while float unpacks can be
 315                  * either A or R4.
 316                  */
 317                 for (int i = 0; i < qir_get_nsrc(inst); i++) {
 318                         if (inst->src[i].file == QFILE_TEMP &&
 319                             inst->src[i].pack) {
 320                                 if (qir_is_float_input(inst)) {
 321                                         class_bits[inst->src[i].index] &=
 322                                                 CLASS_BIT_A | CLASS_BIT_R4;
 323                                 } else {
 324                                         class_bits[inst->src[i].index] &=
 325                                                 CLASS_BIT_A;
 326                                 }
 327                         }
 328                 }
 329
 330                 ip++;
 331         }
 332
 333         for (uint32_t i = 0; i < c->num_temps; i++) {
 334                 int node = temp_to_node[i];
 335
 336                 switch (class_bits[i]) {
 337                 case CLASS_BIT_A | CLASS_BIT_B | CLASS_BIT_R4 | CLASS_BIT_R0_R3:
 338                         ra_set_node_class(g, node,
 339                                           vc4->reg_class_any[c->fs_threaded]);
 340                         break;
 341                 case CLASS_BIT_A | CLASS_BIT_B:
 342                         ra_set_node_class(g, node,
 343                                           vc4->reg_class_a_or_b[c->fs_threaded]);
 344                         break;
 345                 case CLASS_BIT_A | CLASS_BIT_B | CLASS_BIT_R0_R3:
 346                         ra_set_node_class(g, node,
 347                                           vc4->reg_class_a_or_b_or_acc[c->fs_threaded]);
 348                         break;
 349                 case CLASS_BIT_A | CLASS_BIT_R4:
 350                         ra_set_node_class(g, node,
 351                                           vc4->reg_class_r4_or_a[c->fs_threaded]);
 352                         break;
 353                 case CLASS_BIT_A:
 354                         ra_set_node_class(g, node,
 355                                           vc4->reg_class_a[c->fs_threaded]);
 356                         break;
 357                 case CLASS_BIT_R0_R3:
 358                         ra_set_node_class(g, node, vc4->reg_class_r0_r3);
 359                         break;
 360
 361                 default:
 362                         /* DDX/DDY used across thread switched might get us
 363                          * here.
 364                          */
 365                         if (c->fs_threaded) {
 366                                 c->failed = true;
 367                                 free(temp_registers);
 368                                 return NULL;
 369                         }
 370
 371                         fprintf(stderr, "temp %d: bad class bits: 0x%x\n",
 372                                 i, class_bits[i]);
 373                         abort();
 374                         break;
 375                 }
 376         }
 377
 378         for (uint32_t i = 0; i < c->num_temps; i++) {
 379                 for (uint32_t j = i + 1; j < c->num_temps; j++) {
 380                         if (!(c->temp_start[i] >= c->temp_end[j] ||
 381                               c->temp_start[j] >= c->temp_end[i])) {
 382                                 ra_add_node_interference(g,
 383                                                          temp_to_node[i],
 384                                                          temp_to_node[j]);
 385                         }
 386                 }
 387         }
 388
 389         bool ok = ra_allocate(g);
 390         if (!ok) {
 391                 if (!c->fs_threaded) {
 392                         fprintf(stderr, "Failed to register allocate:\n");
 393                         qir_dump(c);
 394                 }
 395
 396                 c->failed = true;
 397                 free(temp_registers);
 398                 return NULL;
 399         }
 400
 401         for (uint32_t i = 0; i < c->num_temps; i++) {
 402                 temp_registers[i] = vc4_regs[ra_get_node_reg(g, temp_to_node[i])];
 403
 404                 /* If the value's never used, just write to the NOP register
 405                  * for clarity in debug output.
 406                  */
 407                 if (c->temp_start[i] == c->temp_end[i])
 408                         temp_registers[i] = qpu_ra(QPU_W_NOP);
 409         }
 410
 411         ralloc_free(g);
 412
 413         return temp_registers;
 414 }