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