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         vc4->reg_class_any = ra_alloc_reg_class(vc4->regs);
 119         vc4->reg_class_a_or_b_or_acc = ra_alloc_reg_class(vc4->regs);
 120         vc4->reg_class_r4_or_a = ra_alloc_reg_class(vc4->regs);
 121         vc4->reg_class_a = ra_alloc_reg_class(vc4->regs);
 122         vc4->reg_class_r0_r3 = ra_alloc_reg_class(vc4->regs);
 123
 124         /* r0-r3 */
 125         for (uint32_t i = ACC_INDEX; i < ACC_INDEX + 4; i++) {
 126                 ra_class_add_reg(vc4->regs, vc4->reg_class_r0_r3, i);
 127                 ra_class_add_reg(vc4->regs, vc4->reg_class_a_or_b_or_acc, i);
 128         }
 129
 130         /* R4 gets a special class because it can't be written as a general
 131          * purpose register. (it's TMU_NOSWAP as a write address).
 132          */
 133         ra_class_add_reg(vc4->regs, vc4->reg_class_r4_or_a, ACC_INDEX + 4);
 134
 135         /* A/B */
 136         for (uint32_t i = AB_INDEX; i < AB_INDEX + 64; i ++) {
 137                 /* Reserve ra31/rb31 for spilling fixup_raddr_conflict() in
 138                  * vc4_qpu_emit.c
 139                  */
 140                 if (vc4_regs[i].addr == 31)
 141                         continue;
 142
 143                 ra_class_add_reg(vc4->regs, vc4->reg_class_any, i);
 144                 ra_class_add_reg(vc4->regs, vc4->reg_class_a_or_b_or_acc, i);
 145
 146                 /* A only */
 147                 if (((i - AB_INDEX) & 1) == 0) {
 148                         ra_class_add_reg(vc4->regs, vc4->reg_class_a, i);
 149                         ra_class_add_reg(vc4->regs, vc4->reg_class_r4_or_a, i);
 150                 }
 151         }
 152
 153         ra_set_finalize(vc4->regs, NULL);
 154 }
 155
 156 struct node_to_temp_map {
 157         uint32_t temp;
 158         uint32_t priority;
 159 };
 160
 161 static int
 162 node_to_temp_priority(const void *in_a, const void *in_b)
 163 {
 164         const struct node_to_temp_map *a = in_a;
 165         const struct node_to_temp_map *b = in_b;
 166
 167         return a->priority - b->priority;
 168 }
 169
 170 #define CLASS_BIT_A                     (1 << 0)
 171 #define CLASS_BIT_B                     (1 << 1)
 172 #define CLASS_BIT_R4                    (1 << 2)
 173 #define CLASS_BIT_R0_R3                 (1 << 4)
 174
 175 /**
 176  * Returns a mapping from QFILE_TEMP indices to struct qpu_regs.
 177  *
 178  * The return value should be freed by the caller.
 179  */
 180 struct qpu_reg *
 181 vc4_register_allocate(struct vc4_context *vc4, struct vc4_compile *c)
 182 {
 183         struct node_to_temp_map map[c->num_temps];
 184         uint32_t temp_to_node[c->num_temps];
 185         uint8_t class_bits[c->num_temps];
 186         struct qpu_reg *temp_registers = calloc(c->num_temps,
 187                                                 sizeof(*temp_registers));
 188
 189         /* If things aren't ever written (undefined values), just read from
 190          * r0.
 191          */
 192         for (uint32_t i = 0; i < c->num_temps; i++)
 193                 temp_registers[i] = qpu_rn(0);
 194
 195         vc4_alloc_reg_set(vc4);
 196
 197         struct ra_graph *g = ra_alloc_interference_graph(vc4->regs,
 198                                                          c->num_temps);
 199
 200         /* Compute the live ranges so we can figure out interference. */
 201         qir_calculate_live_intervals(c);
 202
 203         for (uint32_t i = 0; i < c->num_temps; i++) {
 204                 map[i].temp = i;
 205                 map[i].priority = c->temp_end[i] - c->temp_start[i];
 206         }
 207         qsort(map, c->num_temps, sizeof(map[0]), node_to_temp_priority);
 208         for (uint32_t i = 0; i < c->num_temps; i++) {
 209                 temp_to_node[map[i].temp] = i;
 210         }
 211
 212         /* Figure out our register classes and preallocated registers.  We
 213          * start with any temp being able to be in any file, then instructions
 214          * incrementally remove bits that the temp definitely can't be in.
 215          */
 216         memset(class_bits,
 217                CLASS_BIT_A | CLASS_BIT_B | CLASS_BIT_R4 | CLASS_BIT_R0_R3,
 218                sizeof(class_bits));
 219
 220         int ip = 0;
 221         qir_for_each_inst_inorder(inst, c) {
 222                 if (qir_writes_r4(inst)) {
 223                         /* This instruction writes r4 (and optionally moves
 224                          * its result to a temp), so nothing else can be
 225                          * stored in r4 across it.
 226                          */
 227                         for (int i = 0; i < c->num_temps; i++) {
 228                                 if (c->temp_start[i] < ip && c->temp_end[i] > ip)
 229                                         class_bits[i] &= ~CLASS_BIT_R4;
 230                         }
 231                 } else {
 232                         /* R4 can't be written as a general purpose
 233                          * register. (it's TMU_NOSWAP as a write address).
 234                          */
 235                         if (inst->dst.file == QFILE_TEMP)
 236                                 class_bits[inst->dst.index] &= ~CLASS_BIT_R4;
 237                 }
 238
 239                 switch (inst->op) {
 240                 case QOP_FRAG_Z:
 241                         ra_set_node_reg(g, temp_to_node[inst->dst.index],
 242                                         AB_INDEX + QPU_R_FRAG_PAYLOAD_ZW * 2 + 1);
 243                         break;
 244
 245                 case QOP_FRAG_W:
 246                         ra_set_node_reg(g, temp_to_node[inst->dst.index],
 247                                         AB_INDEX + QPU_R_FRAG_PAYLOAD_ZW * 2);
 248                         break;
 249
 250                 case QOP_ROT_MUL:
 251                         assert(inst->src[0].file == QFILE_TEMP);
 252                         class_bits[inst->src[0].index] &= CLASS_BIT_R0_R3;
 253                         break;
 254
 255                 default:
 256                         break;
 257                 }
 258
 259                 if (inst->dst.pack && !qir_is_mul(inst)) {
 260                         /* The non-MUL pack flags require an A-file dst
 261                          * register.
 262                          */
 263                         class_bits[inst->dst.index] &= CLASS_BIT_A;
 264                 }
 265
 266                 /* Apply restrictions for src unpacks.  The integer unpacks
 267                  * can only be done from regfile A, while float unpacks can be
 268                  * either A or R4.
 269                  */
 270                 for (int i = 0; i < qir_get_op_nsrc(inst->op); i++) {
 271                         if (inst->src[i].file == QFILE_TEMP &&
 272                             inst->src[i].pack) {
 273                                 if (qir_is_float_input(inst)) {
 274                                         class_bits[inst->src[i].index] &=
 275                                                 CLASS_BIT_A | CLASS_BIT_R4;
 276                                 } else {
 277                                         class_bits[inst->src[i].index] &=
 278                                                 CLASS_BIT_A;
 279                                 }
 280                         }
 281                 }
 282
 283                 ip++;
 284         }
 285
 286         for (uint32_t i = 0; i < c->num_temps; i++) {
 287                 int node = temp_to_node[i];
 288
 289                 switch (class_bits[i]) {
 290                 case CLASS_BIT_A | CLASS_BIT_B | CLASS_BIT_R4 | CLASS_BIT_R0_R3:
 291                         ra_set_node_class(g, node, vc4->reg_class_any);
 292                         break;
 293                 case CLASS_BIT_A | CLASS_BIT_B | CLASS_BIT_R0_R3:
 294                         ra_set_node_class(g, node, vc4->reg_class_a_or_b_or_acc);
 295                         break;
 296                 case CLASS_BIT_A | CLASS_BIT_R4:
 297                         ra_set_node_class(g, node, vc4->reg_class_r4_or_a);
 298                         break;
 299                 case CLASS_BIT_A:
 300                         ra_set_node_class(g, node, vc4->reg_class_a);
 301                         break;
 302                 case CLASS_BIT_R0_R3:
 303                         ra_set_node_class(g, node, vc4->reg_class_r0_r3);
 304                         break;
 305                 default:
 306                         fprintf(stderr, "temp %d: bad class bits: 0x%x\n",
 307                                 i, class_bits[i]);
 308                         abort();
 309                         break;
 310                 }
 311         }
 312
 313         for (uint32_t i = 0; i < c->num_temps; i++) {
 314                 for (uint32_t j = i + 1; j < c->num_temps; j++) {
 315                         if (!(c->temp_start[i] >= c->temp_end[j] ||
 316                               c->temp_start[j] >= c->temp_end[i])) {
 317                                 ra_add_node_interference(g,
 318                                                          temp_to_node[i],
 319                                                          temp_to_node[j]);
 320                         }
 321                 }
 322         }
 323
 324         bool ok = ra_allocate(g);
 325         if (!ok) {
 326                 fprintf(stderr, "Failed to register allocate:\n");
 327                 qir_dump(c);
 328                 c->failed = true;
 329                 return NULL;
 330         }
 331
 332         for (uint32_t i = 0; i < c->num_temps; i++) {
 333                 temp_registers[i] = vc4_regs[ra_get_node_reg(g, temp_to_node[i])];
 334
 335                 /* If the value's never used, just write to the NOP register
 336                  * for clarity in debug output.
 337                  */
 338                 if (c->temp_start[i] == c->temp_end[i])
 339                         temp_registers[i] = qpu_ra(QPU_W_NOP);
 340         }
 341
 342         ralloc_free(g);
 343
 344         return temp_registers;
 345 }