src/panfrost/midgard/mir.c

   1 /*
   2  * Copyright (C) 2019 Alyssa Rosenzweig <alyssa@rosenzweig.io>
   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 FROM,
  20  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  21  * SOFTWARE.
  22  */
  23
  24 #include "compiler.h"
  25 #include "midgard_ops.h"
  26
  27 void mir_rewrite_index_src_single(midgard_instruction *ins, unsigned old, unsigned new)
  28 {
  29         for (unsigned i = 0; i < ARRAY_SIZE(ins->src); ++i) {
  30                 if (ins->src[i] == old)
  31                         ins->src[i] = new;
  32         }
  33 }
  34
  35 void mir_rewrite_index_dst_single(midgard_instruction *ins, unsigned old, unsigned new)
  36 {
  37         if (ins->dest == old)
  38                 ins->dest = new;
  39 }
  40
  41 static midgard_vector_alu_src
  42 mir_get_alu_src(midgard_instruction *ins, unsigned idx)
  43 {
  44         unsigned b = (idx == 0) ? ins->alu.src1 : ins->alu.src2;
  45         return vector_alu_from_unsigned(b);
  46 }
  47
  48 unsigned
  49 mir_get_swizzle(midgard_instruction *ins, unsigned idx)
  50 {
  51         if (ins->type == TAG_ALU_4) {
  52                 if (idx == 2 || ins->compact_branch)
  53                         return ins->cond_swizzle;
  54
  55                 return (mir_get_alu_src(ins, idx)).swizzle;
  56         } else if (ins->type == TAG_LOAD_STORE_4) {
  57                 /* Main swizzle of a load is on the destination */
  58                 if (!OP_IS_STORE(ins->load_store.op))
  59                         idx++;
  60
  61                 switch (idx) {
  62                 case 0:
  63                         return ins->load_store.swizzle;
  64                 case 1:
  65                 case 2: {
  66                         uint8_t raw =
  67                                 (idx == 2) ? ins->load_store.arg_2 : ins->load_store.arg_1;
  68
  69                         /* TODO: Integrate component count with properties */
  70                         unsigned components = 1;
  71                         switch (ins->load_store.op) {
  72                         case midgard_op_ld_int4:
  73                                 components = (idx == 0) ? 2 : 1;
  74                                 break;
  75                         case midgard_op_st_int4:
  76                                 components = (idx == 1) ? 2 : 1;
  77                                 break;
  78                         case midgard_op_ld_cubemap_coords:
  79                                 components = 3;
  80                                 break;
  81                         case midgard_op_ldst_perspective_division_z:
  82                                 components = 3;
  83                                 break;
  84                         case midgard_op_ldst_perspective_division_w:
  85                                 components = 4;
  86                                 break;
  87                         default:
  88                                 components = 1;
  89                                 break;
  90                         }
  91
  92                         return component_to_swizzle(midgard_ldst_select(raw).component, components);
  93                 }
  94                 default:
  95                         unreachable("Unknown load/store source");
  96                 }
  97         } else if (ins->type == TAG_TEXTURE_4) {
  98                 switch (idx) {
  99                 case 0:
 100                         return ins->texture.in_reg_swizzle;
 101                 case 1:
 102                         /* Swizzle on bias doesn't make sense */
 103                         return 0;
 104                 default:
 105                         unreachable("Unknown texture source");
 106                 }
 107         } else {
 108                 unreachable("Unknown type");
 109         }
 110 }
 111
 112 void
 113 mir_set_swizzle(midgard_instruction *ins, unsigned idx, unsigned new)
 114 {
 115         if (ins->type == TAG_ALU_4) {
 116                 if (idx == 2 || ins->compact_branch) {
 117                         ins->cond_swizzle = new;
 118                         return;
 119                 }
 120
 121                 unsigned b = (idx == 0) ? ins->alu.src1 : ins->alu.src2;
 122
 123                 midgard_vector_alu_src s =
 124                         vector_alu_from_unsigned(b);
 125
 126                 s.swizzle = new;
 127                 unsigned pack = vector_alu_srco_unsigned(s);
 128
 129                 if (idx == 0)
 130                         ins->alu.src1 = pack;
 131                 else
 132                         ins->alu.src2 = pack;
 133         } else if (ins->type == TAG_LOAD_STORE_4) {
 134                 /* Main swizzle of a load is on the destination */
 135                 if (!OP_IS_STORE(ins->load_store.op))
 136                         idx++;
 137
 138                 switch (idx) {
 139                 case 0:
 140                         ins->load_store.swizzle = new;
 141                         break;
 142                 case 1:
 143                 case 2: {
 144                         uint8_t raw =
 145                                 (idx == 2) ? ins->load_store.arg_2 : ins->load_store.arg_1;
 146
 147                         midgard_ldst_register_select sel
 148                                 = midgard_ldst_select(raw);
 149                         sel.component = swizzle_to_component(new);
 150                         uint8_t packed = midgard_ldst_pack(sel);
 151
 152                         if (idx == 2)
 153                                 ins->load_store.arg_2 = packed;
 154                         else
 155                                 ins->load_store.arg_1 = packed;
 156
 157                         break;
 158                 }
 159                 default:
 160                         assert(new == 0);
 161                         break;
 162                 }
 163         } else if (ins->type == TAG_TEXTURE_4) {
 164                 switch (idx) {
 165                 case 0:
 166                         ins->texture.in_reg_swizzle = new;
 167                         break;
 168                 default:
 169                         assert(new == 0);
 170                         break;
 171                 }
 172         } else {
 173                 unreachable("Unknown type");
 174         }
 175 }
 176
 177 static void
 178 mir_rewrite_index_src_single_swizzle(midgard_instruction *ins, unsigned old, unsigned new, unsigned swizzle)
 179 {
 180         for (unsigned i = 0; i < ARRAY_SIZE(ins->src); ++i) {
 181                 if (ins->src[i] != old) continue;
 182
 183                 ins->src[i] = new;
 184
 185                 mir_set_swizzle(ins, i,
 186                         pan_compose_swizzle(mir_get_swizzle(ins, i), swizzle));
 187         }
 188 }
 189
 190 void
 191 mir_rewrite_index_src(compiler_context *ctx, unsigned old, unsigned new)
 192 {
 193         mir_foreach_instr_global(ctx, ins) {
 194                 mir_rewrite_index_src_single(ins, old, new);
 195         }
 196 }
 197
 198 void
 199 mir_rewrite_index_src_swizzle(compiler_context *ctx, unsigned old, unsigned new, unsigned swizzle)
 200 {
 201         mir_foreach_instr_global(ctx, ins) {
 202                 mir_rewrite_index_src_single_swizzle(ins, old, new, swizzle);
 203         }
 204 }
 205
 206 void
 207 mir_rewrite_index_dst(compiler_context *ctx, unsigned old, unsigned new)
 208 {
 209         mir_foreach_instr_global(ctx, ins) {
 210                 mir_rewrite_index_dst_single(ins, old, new);
 211         }
 212 }
 213
 214 void
 215 mir_rewrite_index(compiler_context *ctx, unsigned old, unsigned new)
 216 {
 217         mir_rewrite_index_src(ctx, old, new);
 218         mir_rewrite_index_dst(ctx, old, new);
 219 }
 220
 221 unsigned
 222 mir_use_count(compiler_context *ctx, unsigned value)
 223 {
 224         unsigned used_count = 0;
 225
 226         mir_foreach_instr_global(ctx, ins) {
 227                 if (mir_has_arg(ins, value))
 228                         ++used_count;
 229         }
 230
 231         return used_count;
 232 }
 233
 234 /* Checks if a value is used only once (or totally dead), which is an important
 235  * heuristic to figure out if certain optimizations are Worth It (TM) */
 236
 237 bool
 238 mir_single_use(compiler_context *ctx, unsigned value)
 239 {
 240         /* We can replicate constants in places so who cares */
 241         if (value == SSA_FIXED_REGISTER(REGISTER_CONSTANT))
 242                 return true;
 243
 244         return mir_use_count(ctx, value) <= 1;
 245 }
 246
 247 static bool
 248 mir_nontrivial_raw_mod(midgard_vector_alu_src src, bool is_int)
 249 {
 250         if (is_int)
 251                 return src.mod == midgard_int_shift;
 252         else
 253                 return src.mod;
 254 }
 255
 256 bool
 257 mir_nontrivial_mod(midgard_vector_alu_src src, bool is_int, unsigned mask)
 258 {
 259         if (mir_nontrivial_raw_mod(src, is_int)) return true;
 260
 261         /* size-conversion */
 262         if (src.half) return true;
 263
 264         /* swizzle */
 265         for (unsigned c = 0; c < 4; ++c) {
 266                 if (!(mask & (1 << c))) continue;
 267                 if (((src.swizzle >> (2*c)) & 3) != c) return true;
 268         }
 269
 270         return false;
 271 }
 272
 273 bool
 274 mir_nontrivial_source2_mod(midgard_instruction *ins)
 275 {
 276         bool is_int = midgard_is_integer_op(ins->alu.op);
 277
 278         midgard_vector_alu_src src2 =
 279                 vector_alu_from_unsigned(ins->alu.src2);
 280
 281         return mir_nontrivial_mod(src2, is_int, ins->mask);
 282 }
 283
 284 bool
 285 mir_nontrivial_source2_mod_simple(midgard_instruction *ins)
 286 {
 287         bool is_int = midgard_is_integer_op(ins->alu.op);
 288
 289         midgard_vector_alu_src src2 =
 290                 vector_alu_from_unsigned(ins->alu.src2);
 291
 292         return mir_nontrivial_raw_mod(src2, is_int) || src2.half;
 293 }
 294
 295 bool
 296 mir_nontrivial_outmod(midgard_instruction *ins)
 297 {
 298         bool is_int = midgard_is_integer_op(ins->alu.op);
 299         unsigned mod = ins->alu.outmod;
 300
 301         /* Pseudo-outmod */
 302         if (ins->invert)
 303                 return true;
 304
 305         /* Type conversion is a sort of outmod */
 306         if (ins->alu.dest_override != midgard_dest_override_none)
 307                 return true;
 308
 309         if (is_int)
 310                 return mod != midgard_outmod_int_wrap;
 311         else
 312                 return mod != midgard_outmod_none;
 313 }
 314
 315 /* Checks if an index will be used as a special register -- basically, if we're
 316  * used as the input to a non-ALU op */
 317
 318 bool
 319 mir_special_index(compiler_context *ctx, unsigned idx)
 320 {
 321         mir_foreach_instr_global(ctx, ins) {
 322                 bool is_ldst = ins->type == TAG_LOAD_STORE_4;
 323                 bool is_tex = ins->type == TAG_TEXTURE_4;
 324                 bool is_writeout = ins->compact_branch && ins->writeout;
 325
 326                 if (!(is_ldst || is_tex || is_writeout))
 327                         continue;
 328
 329                 if (mir_has_arg(ins, idx))
 330                         return true;
 331         }
 332
 333         return false;
 334 }
 335
 336 /* Is a node written before a given instruction? */
 337
 338 bool
 339 mir_is_written_before(compiler_context *ctx, midgard_instruction *ins, unsigned node)
 340 {
 341         if (node >= SSA_FIXED_MINIMUM)
 342                 return true;
 343
 344         mir_foreach_instr_global(ctx, q) {
 345                 if (q == ins)
 346                         break;
 347
 348                 if (q->dest == node)
 349                         return true;
 350         }
 351
 352         return false;
 353 }
 354
 355 /* Grabs the type size. */
 356
 357 midgard_reg_mode
 358 mir_typesize(midgard_instruction *ins)
 359 {
 360         if (ins->compact_branch)
 361                 return midgard_reg_mode_32;
 362
 363         /* TODO: Type sizes for texture */
 364         if (ins->type == TAG_TEXTURE_4)
 365                 return midgard_reg_mode_32;
 366
 367         if (ins->type == TAG_LOAD_STORE_4)
 368                 return GET_LDST_SIZE(load_store_opcode_props[ins->load_store.op].props);
 369
 370         if (ins->type == TAG_ALU_4) {
 371                 midgard_reg_mode mode = ins->alu.reg_mode;
 372
 373                 /* If we have an override, step down by half */
 374                 if (ins->alu.dest_override != midgard_dest_override_none) {
 375                         assert(mode > midgard_reg_mode_8);
 376                         mode--;
 377                 }
 378
 379                 return mode;
 380         }
 381
 382         unreachable("Invalid instruction type");
 383 }
 384
 385 /* Grabs the size of a source */
 386
 387 static midgard_reg_mode
 388 mir_srcsize(midgard_instruction *ins, unsigned i)
 389 {
 390         /* TODO: 16-bit textures/ldst */
 391         if (ins->type == TAG_TEXTURE_4 || ins->type == TAG_LOAD_STORE_4)
 392                 return midgard_reg_mode_32;
 393
 394         /* TODO: 16-bit branches */
 395         if (ins->compact_branch)
 396                 return midgard_reg_mode_32;
 397
 398         if (i >= 2) {
 399                 /* TODO: 16-bit conditions, ffma */
 400                 assert(i == 2);
 401                 return midgard_reg_mode_32;
 402         }
 403
 404         /* Default to type of the instruction */
 405
 406         midgard_reg_mode mode = ins->alu.reg_mode;
 407
 408         /* If we have a half modifier, step down by half */
 409
 410         if ((mir_get_alu_src(ins, i)).half) {
 411                 assert(mode > midgard_reg_mode_8);
 412                 mode--;
 413         }
 414
 415         return mode;
 416 }
 417
 418 /* Converts per-component mask to a byte mask */
 419
 420 static uint16_t
 421 mir_to_bytemask(midgard_reg_mode mode, unsigned mask)
 422 {
 423         switch (mode) {
 424         case midgard_reg_mode_8:
 425                 return mask;
 426
 427         case midgard_reg_mode_16: {
 428                 unsigned space =
 429                         ((mask & 0x1) << (0 - 0)) |
 430                         ((mask & 0x2) << (2 - 1)) |
 431                         ((mask & 0x4) << (4 - 2)) |
 432                         ((mask & 0x8) << (6 - 3)) |
 433                         ((mask & 0x10) << (8 - 4)) |
 434                         ((mask & 0x20) << (10 - 5)) |
 435                         ((mask & 0x40) << (12 - 6)) |
 436                         ((mask & 0x80) << (14 - 7));
 437
 438                 return space | (space << 1);
 439         }
 440
 441         case midgard_reg_mode_32: {
 442                 unsigned space =
 443                         ((mask & 0x1) << (0 - 0)) |
 444                         ((mask & 0x2) << (4 - 1)) |
 445                         ((mask & 0x4) << (8 - 2)) |
 446                         ((mask & 0x8) << (12 - 3));
 447
 448                 return space | (space << 1) | (space << 2) | (space << 3);
 449         }
 450
 451         case midgard_reg_mode_64: {
 452                 unsigned A = (mask & 0x1) ? 0xFF : 0x00;
 453                 unsigned B = (mask & 0x2) ? 0xFF : 0x00;
 454                 return A | (B << 8);
 455         }
 456
 457         default:
 458                 unreachable("Invalid register mode");
 459         }
 460 }
 461
 462 /* ...and the inverse */
 463
 464 static unsigned
 465 mir_bytes_for_mode(midgard_reg_mode mode)
 466 {
 467         switch (mode) {
 468         case midgard_reg_mode_8:
 469                 return 1;
 470         case midgard_reg_mode_16:
 471                 return 2;
 472         case midgard_reg_mode_32:
 473                 return 4;
 474         case midgard_reg_mode_64:
 475                 return 8;
 476         default:
 477                 unreachable("Invalid register mode");
 478         }
 479 }
 480
 481 uint16_t
 482 mir_from_bytemask(uint16_t bytemask, midgard_reg_mode mode)
 483 {
 484         unsigned value = 0;
 485         unsigned count = mir_bytes_for_mode(mode);
 486
 487         for (unsigned c = 0, d = 0; c < 16; c += count, ++d) {
 488                 bool a = (bytemask & (1 << c)) != 0;
 489
 490                 for (unsigned q = c; q < count; ++q)
 491                         assert(((bytemask & (1 << q)) != 0) == a);
 492
 493                 value |= (a << d);
 494         }
 495
 496         return value;
 497 }
 498
 499 /* Rounds down a bytemask to fit a given component count. Iterate each
 500  * component, and check if all bytes in the component are masked on */
 501
 502 uint16_t
 503 mir_round_bytemask_down(uint16_t mask, midgard_reg_mode mode)
 504 {
 505         unsigned bytes = mir_bytes_for_mode(mode);
 506         unsigned maxmask = mask_of(bytes);
 507         unsigned channels = 16 / bytes;
 508
 509         for (unsigned c = 0; c < channels; ++c) {
 510                 /* Get bytes in component */
 511                 unsigned submask = (mask >> c * channels) & maxmask;
 512
 513                 if (submask != maxmask)
 514                         mask &= ~(maxmask << (c * channels));
 515         }
 516
 517         return mask;
 518 }
 519
 520 /* Grabs the per-byte mask of an instruction (as opposed to per-component) */
 521
 522 uint16_t
 523 mir_bytemask(midgard_instruction *ins)
 524 {
 525         return mir_to_bytemask(mir_typesize(ins), ins->mask);
 526 }
 527
 528 /* Creates a mask of the components of a node read by an instruction, by
 529  * analyzing the swizzle with respect to the instruction's mask. E.g.:
 530  *
 531  *  fadd r0.xz, r1.yyyy, r2.zwyx
 532  *
 533  * will return a mask of Z/Y for r2
 534  */
 535
 536 static uint16_t
 537 mir_bytemask_of_read_components_single(unsigned swizzle, unsigned inmask, midgard_reg_mode mode)
 538 {
 539         unsigned cmask = 0;
 540
 541         for (unsigned c = 0; c < 4; ++c) {
 542                 if (!(inmask & (1 << c))) continue;
 543
 544                 unsigned comp = (swizzle >> (2*c)) & 3;
 545                 cmask |= (1 << comp);
 546         }
 547
 548         return mir_to_bytemask(mode, cmask);
 549 }
 550
 551 static unsigned
 552 mir_source_count(midgard_instruction *ins)
 553 {
 554         if (ins->type == TAG_ALU_4) {
 555                 /* ALU is always binary, except csel */
 556                 return OP_IS_CSEL(ins->alu.op) ? 3 : 2;
 557         } else if (ins->type == TAG_LOAD_STORE_4) {
 558                 bool load = !OP_IS_STORE(ins->load_store.op);
 559                 return (load ? 2 : 3);
 560         } else if (ins->type == TAG_TEXTURE_4) {
 561                 /* Coords, bias.. TODO: Offsets? */
 562                 return 2;
 563         } else {
 564                 unreachable("Invalid instruction type");
 565         }
 566 }
 567
 568 uint16_t
 569 mir_bytemask_of_read_components(midgard_instruction *ins, unsigned node)
 570 {
 571         uint16_t mask = 0;
 572
 573         for (unsigned i = 0; i < mir_source_count(ins); ++i) {
 574                 if (ins->src[i] != node) continue;
 575
 576                 /* Branch writeout uses all components */
 577                 if (ins->compact_branch && ins->writeout && (i == 0))
 578                         return 0xFFFF;
 579
 580                 /* Conditional branches read one 32-bit component = 4 bytes (TODO: multi branch??) */
 581                 if (ins->compact_branch && !ins->prepacked_branch && ins->branch.conditional && (i == 0))
 582                         return 0xF;
 583
 584                 /* ALU ops act componentwise so we need to pay attention to
 585                  * their mask. Texture/ldst does not so we don't clamp source
 586                  * readmasks based on the writemask */
 587                 unsigned qmask = (ins->type == TAG_ALU_4) ? ins->mask : ~0;
 588
 589                 /* Handle dot products and things */
 590                 if (ins->type == TAG_ALU_4 && !ins->compact_branch) {
 591                         unsigned props = alu_opcode_props[ins->alu.op].props;
 592
 593                         unsigned channel_override = GET_CHANNEL_COUNT(props);
 594
 595                         if (channel_override)
 596                                 qmask = mask_of(channel_override);
 597                 }
 598
 599                 unsigned swizzle = mir_get_swizzle(ins, i);
 600                 mask |= mir_bytemask_of_read_components_single(swizzle, qmask, mir_srcsize(ins, i));
 601         }
 602
 603         return mask;
 604 }
 605
 606 unsigned
 607 mir_ubo_shift(midgard_load_store_op op)
 608 {
 609         switch (op) {
 610         case midgard_op_ld_ubo_char:
 611                 return 0;
 612         case midgard_op_ld_ubo_char2:
 613                 return 1;
 614         case midgard_op_ld_ubo_char4:
 615                 return 2;
 616         case midgard_op_ld_ubo_short4:
 617                 return 3;
 618         case midgard_op_ld_ubo_int4:
 619                 return 4;
 620         default:
 621                 unreachable("Invalid op");
 622         }
 623 }
 624
 625 /* Register allocation occurs after instruction scheduling, which is fine until
 626  * we start needing to spill registers and therefore insert instructions into
 627  * an already-scheduled program. We don't have to be terribly efficient about
 628  * this, since spilling is already slow. So just semantically we need to insert
 629  * the instruction into a new bundle before/after the bundle of the instruction
 630  * in question */
 631
 632 static midgard_bundle
 633 mir_bundle_for_op(compiler_context *ctx, midgard_instruction ins)
 634 {
 635         midgard_instruction *u = mir_upload_ins(ctx, ins);
 636
 637         midgard_bundle bundle = {
 638                 .tag = ins.type,
 639                 .instruction_count = 1,
 640                 .instructions = { u },
 641         };
 642
 643         if (bundle.tag == TAG_ALU_4) {
 644                 assert(OP_IS_MOVE(u->alu.op));
 645                 u->unit = UNIT_VMUL;
 646
 647                 size_t bytes_emitted = sizeof(uint32_t) + sizeof(midgard_reg_info) + sizeof(midgard_vector_alu);
 648                 bundle.padding = ~(bytes_emitted - 1) & 0xF;
 649                 bundle.control = ins.type | u->unit;
 650         }
 651
 652         return bundle;
 653 }
 654
 655 static unsigned
 656 mir_bundle_idx_for_ins(midgard_instruction *tag, midgard_block *block)
 657 {
 658         midgard_bundle *bundles =
 659                 (midgard_bundle *) block->bundles.data;
 660
 661         size_t count = (block->bundles.size / sizeof(midgard_bundle));
 662
 663         for (unsigned i = 0; i < count; ++i) {
 664                 for (unsigned j = 0; j < bundles[i].instruction_count; ++j) {
 665                         if (bundles[i].instructions[j] == tag)
 666                                 return i;
 667                 }
 668         }
 669
 670         mir_print_instruction(tag);
 671         unreachable("Instruction not scheduled in block");
 672 }
 673
 674 void
 675 mir_insert_instruction_before_scheduled(
 676         compiler_context *ctx,
 677         midgard_block *block,
 678         midgard_instruction *tag,
 679         midgard_instruction ins)
 680 {
 681         unsigned before = mir_bundle_idx_for_ins(tag, block);
 682         size_t count = util_dynarray_num_elements(&block->bundles, midgard_bundle);
 683         UNUSED void *unused = util_dynarray_grow(&block->bundles, midgard_bundle, 1);
 684
 685         midgard_bundle *bundles = (midgard_bundle *) block->bundles.data;
 686         memmove(bundles + before + 1, bundles + before, (count - before) * sizeof(midgard_bundle));
 687         midgard_bundle *before_bundle = bundles + before + 1;
 688
 689         midgard_bundle new = mir_bundle_for_op(ctx, ins);
 690         memcpy(bundles + before, &new, sizeof(new));
 691
 692         list_addtail(&new.instructions[0]->link, &before_bundle->instructions[0]->link);
 693 }
 694
 695 void
 696 mir_insert_instruction_after_scheduled(
 697         compiler_context *ctx,
 698         midgard_block *block,
 699         midgard_instruction *tag,
 700         midgard_instruction ins)
 701 {
 702         /* We need to grow the bundles array to add our new bundle */
 703         size_t count = util_dynarray_num_elements(&block->bundles, midgard_bundle);
 704         UNUSED void *unused = util_dynarray_grow(&block->bundles, midgard_bundle, 1);
 705
 706         /* Find the bundle that we want to insert after */
 707         unsigned after = mir_bundle_idx_for_ins(tag, block);
 708
 709         /* All the bundles after that one, we move ahead by one */
 710         midgard_bundle *bundles = (midgard_bundle *) block->bundles.data;
 711         memmove(bundles + after + 2, bundles + after + 1, (count - after - 1) * sizeof(midgard_bundle));
 712         midgard_bundle *after_bundle = bundles + after;
 713
 714         midgard_bundle new = mir_bundle_for_op(ctx, ins);
 715         memcpy(bundles + after + 1, &new, sizeof(new));
 716         list_add(&new.instructions[0]->link, &after_bundle->instructions[after_bundle->instruction_count - 1]->link);
 717 }
 718
 719 /* Flip the first-two arguments of a (binary) op. Currently ALU
 720  * only, no known uses for ldst/tex */
 721
 722 void
 723 mir_flip(midgard_instruction *ins)
 724 {
 725         unsigned temp = ins->src[0];
 726         ins->src[0] = ins->src[1];
 727         ins->src[1] = temp;
 728
 729         assert(ins->type == TAG_ALU_4);
 730
 731         temp = ins->alu.src1;
 732         ins->alu.src1 = ins->alu.src2;
 733         ins->alu.src2 = temp;
 734 }
 735
 736 /* Before squashing, calculate ctx->temp_count just by observing the MIR */
 737
 738 void
 739 mir_compute_temp_count(compiler_context *ctx)
 740 {
 741         if (ctx->temp_count)
 742                 return;
 743
 744         unsigned max_dest = 0;
 745
 746         mir_foreach_instr_global(ctx, ins) {
 747                 if (ins->dest < SSA_FIXED_MINIMUM)
 748                         max_dest = MAX2(max_dest, ins->dest + 1);
 749         }
 750
 751         ctx->temp_count = max_dest;
 752 }