src/panfrost/midgard/disassemble.c

   1 /* Author(s):
   2  *   Connor Abbott
   3  *   Alyssa Rosenzweig
   4  *
   5  * Copyright (c) 2013 Connor Abbott (connor@abbott.cx)
   6  * Copyright (c) 2018 Alyssa Rosenzweig (alyssa@rosenzweig.io)
   7  *
   8  * Permission is hereby granted, free of charge, to any person obtaining a copy
   9  * of this software and associated documentation files (the "Software"), to deal
  10  * in the Software without restriction, including without limitation the rights
  11  * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
  12  * copies of the Software, and to permit persons to whom the Software is
  13  * furnished to do so, subject to the following conditions:
  14  *
  15  * The above copyright notice and this permission notice shall be included in
  16  * all copies or substantial portions of the Software.
  17  *
  18  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  19  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  20  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
  21  * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  22  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
  23  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
  24  * THE SOFTWARE.
  25  */
  26
  27 #include <stdio.h>
  28 #include <stdint.h>
  29 #include <assert.h>
  30 #include <inttypes.h>
  31 #include <ctype.h>
  32 #include <string.h>
  33 #include "midgard.h"
  34 #include "midgard-parse.h"
  35 #include "midgard_ops.h"
  36 #include "disassemble.h"
  37 #include "helpers.h"
  38 #include "util/half_float.h"
  39 #include "util/u_math.h"
  40
  41 #define DEFINE_CASE(define, str) case define: { printf(str); break; }
  42
  43 static bool is_instruction_int = false;
  44
  45 /* Stats */
  46
  47 static struct midgard_disasm_stats midg_stats;
  48
  49 /* Prints a short form of the tag for branching, the minimum needed to be
  50  * legible and unambiguous */
  51
  52 static void
  53 print_tag_short(unsigned tag)
  54 {
  55         switch (midgard_word_types[tag]) {
  56         case midgard_word_type_texture:
  57                 printf("tex/%X", tag);
  58                 break;
  59
  60         case midgard_word_type_load_store:
  61                 printf("ldst");
  62                 break;
  63
  64         case midgard_word_type_alu:
  65                 printf("alu%u/%X", midgard_word_size[tag], tag);
  66                 break;
  67
  68         default:
  69                 printf("%s%X", (tag > 0) ? "" : "unk", tag);
  70                 break;
  71         }
  72 }
  73
  74 static void
  75 print_alu_opcode(midgard_alu_op op)
  76 {
  77         bool int_op = false;
  78
  79         if (alu_opcode_props[op].name) {
  80                 printf("%s", alu_opcode_props[op].name);
  81
  82                 int_op = midgard_is_integer_op(op);
  83         } else
  84                 printf("alu_op_%02X", op);
  85
  86         /* For constant analysis */
  87         is_instruction_int = int_op;
  88 }
  89
  90 static void
  91 print_ld_st_opcode(midgard_load_store_op op)
  92 {
  93         if (load_store_opcode_names[op])
  94                 printf("%s", load_store_opcode_names[op]);
  95         else
  96                 printf("ldst_op_%02X", op);
  97 }
  98
  99 static bool is_embedded_constant_half = false;
 100 static bool is_embedded_constant_int = false;
 101
 102 static char
 103 prefix_for_bits(unsigned bits)
 104 {
 105         switch (bits) {
 106         case 8:
 107                 return 'q';
 108         case 16:
 109                 return 'h';
 110         case 64:
 111                 return 'd';
 112         default:
 113                 return 0;
 114         }
 115 }
 116
 117 /* For static analysis to ensure all registers are written at least once before
 118  * use along the source code path (TODO: does this break done for complex CF?)
 119  */
 120
 121 uint16_t midg_ever_written = 0;
 122
 123 static void
 124 print_reg(unsigned reg, unsigned bits)
 125 {
 126         /* Perform basic static analysis for expanding constants correctly */
 127
 128         if (reg == 26) {
 129                 is_embedded_constant_int = is_instruction_int;
 130                 is_embedded_constant_half = (bits < 32);
 131         }
 132
 133         unsigned uniform_reg = 23 - reg;
 134         bool is_uniform = false;
 135
 136         /* For r8-r15, it could be a work or uniform. We distinguish based on
 137          * the fact work registers are ALWAYS written before use, but uniform
 138          * registers are NEVER written before use. */
 139
 140         if ((reg >= 8 && reg < 16) && !(midg_ever_written & (1 << reg)))
 141                 is_uniform = true;
 142
 143         /* r16-r23 are always uniform */
 144
 145         if (reg >= 16 && reg <= 23)
 146                 is_uniform = true;
 147
 148         /* Update the uniform count appropriately */
 149
 150         if (is_uniform)
 151                 midg_stats.uniform_count =
 152                         MAX2(uniform_reg + 1, midg_stats.uniform_count);
 153
 154         char prefix = prefix_for_bits(bits);
 155
 156         if (prefix)
 157                 putchar(prefix);
 158
 159         printf("r%u", reg);
 160 }
 161
 162 static char *outmod_names_float[4] = {
 163         "",
 164         ".pos",
 165         ".unk2",
 166         ".sat"
 167 };
 168
 169 static char *outmod_names_int[4] = {
 170         ".isat",
 171         ".usat",
 172         "",
 173         ".hi"
 174 };
 175
 176 static char *srcmod_names_int[4] = {
 177         "sext(",
 178         "zext(",
 179         "",
 180         "("
 181 };
 182
 183 static void
 184 print_outmod(unsigned outmod, bool is_int)
 185 {
 186         printf("%s", is_int ? outmod_names_int[outmod] :
 187                outmod_names_float[outmod]);
 188 }
 189
 190 static void
 191 print_quad_word(uint32_t *words, unsigned tabs)
 192 {
 193         unsigned i;
 194
 195         for (i = 0; i < 4; i++)
 196                 printf("0x%08X%s ", words[i], i == 3 ? "" : ",");
 197
 198         printf("\n");
 199 }
 200
 201 static const char components[16] = "xyzwefghijklmnop";
 202
 203 /* Helper to print 4 chars of a swizzle */
 204 static void
 205 print_swizzle_helper(unsigned swizzle, bool upper)
 206 {
 207         for (unsigned i = 0; i < 4; ++i) {
 208                 unsigned c = (swizzle >> (i * 2)) & 3;
 209                 c += upper*4;
 210                 printf("%c", components[c]);
 211         }
 212 }
 213
 214 /* Helper to print 8 chars of a swizzle, duplicating over */
 215 static void
 216 print_swizzle_helper_8(unsigned swizzle, bool upper)
 217 {
 218         for (unsigned i = 0; i < 4; ++i) {
 219                 unsigned c = (swizzle >> (i * 2)) & 3;
 220                 c *= 2;
 221                 c += upper*8;
 222                 printf("%c%c", components[c], components[c+1]);
 223         }
 224 }
 225
 226 static void
 227 print_swizzle_vec16(unsigned swizzle, bool rep_high, bool rep_low,
 228                     midgard_dest_override override)
 229 {
 230         printf(".");
 231
 232         if (override == midgard_dest_override_upper) {
 233                 if (rep_high)
 234                         printf(" /* rep_high */ ");
 235                 if (rep_low)
 236                         printf(" /* rep_low */ ");
 237
 238                 if (!rep_high && rep_low)
 239                         print_swizzle_helper_8(swizzle, true);
 240                 else
 241                         print_swizzle_helper_8(swizzle, false);
 242         } else {
 243                 print_swizzle_helper_8(swizzle, rep_high & 1);
 244                 print_swizzle_helper_8(swizzle, !(rep_low & 1));
 245         }
 246 }
 247
 248 static void
 249 print_swizzle_vec8(unsigned swizzle, bool rep_high, bool rep_low)
 250 {
 251         printf(".");
 252
 253         print_swizzle_helper(swizzle, rep_high & 1);
 254         print_swizzle_helper(swizzle, !(rep_low & 1));
 255 }
 256
 257 static void
 258 print_swizzle_vec4(unsigned swizzle, bool rep_high, bool rep_low)
 259 {
 260         if (rep_high)
 261                 printf(" /* rep_high */ ");
 262         if (rep_low)
 263                 printf(" /* rep_low */ ");
 264
 265         if (swizzle == 0xE4) return; /* xyzw */
 266
 267         printf(".");
 268         print_swizzle_helper(swizzle, 0);
 269 }
 270 static void
 271 print_swizzle_vec2(unsigned swizzle, bool rep_high, bool rep_low)
 272 {
 273         if (rep_high)
 274                 printf(" /* rep_high */ ");
 275         if (rep_low)
 276                 printf(" /* rep_low */ ");
 277
 278         if (swizzle == 0xE4) return; /* XY */
 279
 280         printf(".");
 281
 282         for (unsigned i = 0; i < 4; i += 2) {
 283                 unsigned a = (swizzle >> (i * 2)) & 3;
 284                 unsigned b = (swizzle >> ((i+1) * 2)) & 3;
 285
 286                 /* Normally we're adjacent, but if there's an issue, don't make
 287                  * it ambiguous */
 288
 289                 if (a & 0x1)
 290                         printf("[%c%c]", components[a], components[b]);
 291                 else if (a == b)
 292                         printf("%c", components[a >> 1]);
 293                 else if (b == (a + 1))
 294                         printf("%c", "XY"[a >> 1]);
 295                 else
 296                         printf("[%c%c]", components[a], components[b]);
 297         }
 298 }
 299
 300 static int
 301 bits_for_mode(midgard_reg_mode mode)
 302 {
 303         switch (mode) {
 304         case midgard_reg_mode_8:
 305                 return 8;
 306         case midgard_reg_mode_16:
 307                 return 16;
 308         case midgard_reg_mode_32:
 309                 return 32;
 310         case midgard_reg_mode_64:
 311                 return 64;
 312         default:
 313                 unreachable("Invalid reg mode");
 314                 return 0;
 315         }
 316 }
 317
 318 static int
 319 bits_for_mode_halved(midgard_reg_mode mode, bool half)
 320 {
 321         unsigned bits = bits_for_mode(mode);
 322
 323         if (half)
 324                 bits >>= 1;
 325
 326         return bits;
 327 }
 328
 329 static void
 330 print_vector_src(unsigned src_binary,
 331                  midgard_reg_mode mode, unsigned reg,
 332                  midgard_dest_override override, bool is_int)
 333 {
 334         midgard_vector_alu_src *src = (midgard_vector_alu_src *)&src_binary;
 335
 336         /* Modifiers change meaning depending on the op's context */
 337
 338         midgard_int_mod int_mod = src->mod;
 339
 340         if (is_int) {
 341                 printf("%s", srcmod_names_int[int_mod]);
 342         } else {
 343                 if (src->mod & MIDGARD_FLOAT_MOD_NEG)
 344                         printf("-");
 345
 346                 if (src->mod & MIDGARD_FLOAT_MOD_ABS)
 347                         printf("abs(");
 348         }
 349
 350         //register
 351         unsigned bits = bits_for_mode_halved(mode, src->half);
 352         print_reg(reg, bits);
 353
 354         //swizzle
 355         if (bits == 16)
 356                 print_swizzle_vec8(src->swizzle, src->rep_high, src->rep_low);
 357         else if (bits == 8)
 358                 print_swizzle_vec16(src->swizzle, src->rep_high, src->rep_low, override);
 359         else if (bits == 32)
 360                 print_swizzle_vec4(src->swizzle, src->rep_high, src->rep_low);
 361         else if (bits == 64)
 362                 print_swizzle_vec2(src->swizzle, src->rep_high, src->rep_low);
 363
 364         /* Since we wrapped with a function-looking thing */
 365
 366         if (is_int && int_mod == midgard_int_shift)
 367                 printf(") << %u", bits);
 368         else if ((is_int && (int_mod != midgard_int_normal))
 369                  || (!is_int && src->mod & MIDGARD_FLOAT_MOD_ABS))
 370                 printf(")");
 371 }
 372
 373 static uint16_t
 374 decode_vector_imm(unsigned src2_reg, unsigned imm)
 375 {
 376         uint16_t ret;
 377         ret = src2_reg << 11;
 378         ret |= (imm & 0x7) << 8;
 379         ret |= (imm >> 3) & 0xFF;
 380         return ret;
 381 }
 382
 383 static void
 384 print_immediate(uint16_t imm)
 385 {
 386         if (is_instruction_int)
 387                 printf("#%u", imm);
 388         else
 389                 printf("#%g", _mesa_half_to_float(imm));
 390 }
 391
 392 static void
 393 update_dest(unsigned reg)
 394 {
 395         /* We should record writes as marking this as a work register. Store
 396          * the max register in work_count; we'll add one at the end */
 397
 398         if (reg < 16) {
 399                 midg_stats.work_count = MAX2(reg, midg_stats.work_count);
 400                 midg_ever_written |= (1 << reg);
 401         }
 402 }
 403
 404 static void
 405 print_dest(unsigned reg, midgard_reg_mode mode, midgard_dest_override override)
 406 {
 407         /* Depending on the mode and override, we determine the type of
 408          * destination addressed. Absent an override, we address just the
 409          * type of the operation itself */
 410
 411         unsigned bits = bits_for_mode(mode);
 412
 413         if (override != midgard_dest_override_none)
 414                 bits /= 2;
 415
 416         update_dest(reg);
 417         print_reg(reg, bits);
 418 }
 419
 420 static void
 421 print_mask_vec16(uint8_t mask, midgard_dest_override override)
 422 {
 423         printf(".");
 424
 425         for (unsigned i = 0; i < 8; i++) {
 426                 if (mask & (1 << i))
 427                         printf("%c%c",
 428                                components[i*2 + 0],
 429                                components[i*2 + 1]);
 430         }
 431 }
 432
 433 /* For 16-bit+ masks, we read off from the 8-bit mask field. For 16-bit (vec8),
 434  * it's just one bit per channel, easy peasy. For 32-bit (vec4), it's one bit
 435  * per channel with one duplicate bit in the middle. For 64-bit (vec2), it's
 436  * one-bit per channel with _3_ duplicate bits in the middle. Basically, just
 437  * subdividing the 128-bit word in 16-bit increments. For 64-bit, we uppercase
 438  * the mask to make it obvious what happened */
 439
 440 static void
 441 print_mask(uint8_t mask, unsigned bits, midgard_dest_override override)
 442 {
 443         if (bits == 8) {
 444                 print_mask_vec16(mask, override);
 445                 return;
 446         }
 447
 448         /* Skip 'complete' masks */
 449
 450         if (override == midgard_dest_override_none) {
 451                 if (bits >= 32 && mask == 0xFF) return;
 452
 453                 if (bits == 16) {
 454                         if (mask == 0x0F)
 455                                 return;
 456                         else if (mask == 0xF0) {
 457                                 printf("'");
 458                                 return;
 459                         }
 460                 }
 461         }
 462
 463         printf(".");
 464
 465         unsigned skip = (bits / 16);
 466         bool uppercase = bits > 32;
 467         bool tripped = false;
 468
 469         /* To apply an upper destination override, we "shift" the alphabet.
 470          * E.g. with an upper override on 32-bit, instead of xyzw, print efgh.
 471          * For upper 16-bit, instead of xyzwefgh, print ijklmnop */
 472
 473         const char *alphabet = components;
 474
 475         if (override == midgard_dest_override_upper) {
 476                 unsigned components = 128 / bits;
 477                 alphabet += components;
 478         }
 479
 480         for (unsigned i = 0; i < 8; i += skip) {
 481                 bool a = (mask & (1 << i)) != 0;
 482
 483                 for (unsigned j = 1; j < skip; ++j) {
 484                         bool dupe = (mask & (1 << (i + j))) != 0;
 485                         tripped |= (dupe != a);
 486                 }
 487
 488                 if (a) {
 489                         char c = alphabet[i / skip];
 490
 491                         if (uppercase)
 492                                 c = toupper(c);
 493
 494                         printf("%c", c);
 495                 }
 496         }
 497
 498         if (tripped)
 499                 printf(" /* %X */", mask);
 500 }
 501
 502 /* Prints the 4-bit masks found in texture and load/store ops, as opposed to
 503  * the 8-bit masks found in (vector) ALU ops */
 504
 505 static void
 506 print_mask_4(unsigned mask)
 507 {
 508         if (mask == 0xF) return;
 509
 510         printf(".");
 511
 512         for (unsigned i = 0; i < 4; ++i) {
 513                 bool a = (mask & (1 << i)) != 0;
 514                 if (a)
 515                         printf("%c", components[i]);
 516         }
 517 }
 518
 519 static void
 520 print_vector_field(const char *name, uint16_t *words, uint16_t reg_word,
 521                    unsigned tabs)
 522 {
 523         midgard_reg_info *reg_info = (midgard_reg_info *)&reg_word;
 524         midgard_vector_alu *alu_field = (midgard_vector_alu *) words;
 525         midgard_reg_mode mode = alu_field->reg_mode;
 526         unsigned override = alu_field->dest_override;
 527
 528         /* For now, prefix instruction names with their unit, until we
 529          * understand how this works on a deeper level */
 530         printf("%s.", name);
 531
 532         print_alu_opcode(alu_field->op);
 533
 534         /* Postfix with the size to disambiguate if necessary */
 535         char postfix = prefix_for_bits(bits_for_mode(mode));
 536         bool size_ambiguous = override != midgard_dest_override_none;
 537
 538         if (size_ambiguous)
 539                 printf("%c", postfix ? postfix : 'r');
 540
 541         /* Print the outmod, if there is one */
 542         print_outmod(alu_field->outmod,
 543                      midgard_is_integer_out_op(alu_field->op));
 544
 545         printf(" ");
 546
 547         /* Mask denoting status of 8-lanes */
 548         uint8_t mask = alu_field->mask;
 549
 550         /* First, print the destination */
 551         print_dest(reg_info->out_reg, mode, alu_field->dest_override);
 552
 553         if (override != midgard_dest_override_none) {
 554                 bool modeable = (mode != midgard_reg_mode_8);
 555                 bool known = override != 0x3; /* Unused value */
 556
 557                 if (!(modeable && known))
 558                         printf("/* do%u */ ", override);
 559         }
 560
 561         print_mask(mask, bits_for_mode(mode), override);
 562
 563         printf(", ");
 564
 565         bool is_int = midgard_is_integer_op(alu_field->op);
 566         print_vector_src(alu_field->src1, mode, reg_info->src1_reg, override, is_int);
 567
 568         printf(", ");
 569
 570         if (reg_info->src2_imm) {
 571                 uint16_t imm = decode_vector_imm(reg_info->src2_reg, alu_field->src2 >> 2);
 572                 print_immediate(imm);
 573         } else {
 574                 print_vector_src(alu_field->src2, mode,
 575                                  reg_info->src2_reg, override, is_int);
 576         }
 577
 578         midg_stats.instruction_count++;
 579         printf("\n");
 580 }
 581
 582 static void
 583 print_scalar_src(unsigned src_binary, unsigned reg)
 584 {
 585         midgard_scalar_alu_src *src = (midgard_scalar_alu_src *)&src_binary;
 586
 587         if (src->negate)
 588                 printf("-");
 589
 590         if (src->abs)
 591                 printf("abs(");
 592
 593         print_reg(reg, src->full ? 32 : 16);
 594
 595         unsigned c = src->component;
 596
 597         if (src->full) {
 598                 assert((c & 1) == 0);
 599                 c >>= 1;
 600         }
 601
 602         printf(".%c", components[c]);
 603
 604         if (src->abs)
 605                 printf(")");
 606
 607 }
 608
 609 static uint16_t
 610 decode_scalar_imm(unsigned src2_reg, unsigned imm)
 611 {
 612         uint16_t ret;
 613         ret = src2_reg << 11;
 614         ret |= (imm & 3) << 9;
 615         ret |= (imm & 4) << 6;
 616         ret |= (imm & 0x38) << 2;
 617         ret |= imm >> 6;
 618         return ret;
 619 }
 620
 621 static void
 622 print_scalar_field(const char *name, uint16_t *words, uint16_t reg_word,
 623                    unsigned tabs)
 624 {
 625         midgard_reg_info *reg_info = (midgard_reg_info *)&reg_word;
 626         midgard_scalar_alu *alu_field = (midgard_scalar_alu *) words;
 627
 628         if (alu_field->unknown)
 629                 printf("scalar ALU unknown bit set\n");
 630
 631         printf("%s.", name);
 632         print_alu_opcode(alu_field->op);
 633         print_outmod(alu_field->outmod,
 634                      midgard_is_integer_out_op(alu_field->op));
 635         printf(" ");
 636
 637         bool full = alu_field->output_full;
 638         update_dest(reg_info->out_reg);
 639         print_reg(reg_info->out_reg, full ? 32 : 16);
 640         unsigned c = alu_field->output_component;
 641
 642         if (full) {
 643                 assert((c & 1) == 0);
 644                 c >>= 1;
 645         }
 646
 647         printf(".%c, ", components[c]);
 648
 649         print_scalar_src(alu_field->src1, reg_info->src1_reg);
 650
 651         printf(", ");
 652
 653         if (reg_info->src2_imm) {
 654                 uint16_t imm = decode_scalar_imm(reg_info->src2_reg,
 655                                                  alu_field->src2);
 656                 print_immediate(imm);
 657         } else
 658                 print_scalar_src(alu_field->src2, reg_info->src2_reg);
 659
 660         midg_stats.instruction_count++;
 661         printf("\n");
 662 }
 663
 664 static void
 665 print_branch_op(unsigned op)
 666 {
 667         switch (op) {
 668         case midgard_jmp_writeout_op_branch_uncond:
 669                 printf("uncond.");
 670                 break;
 671
 672         case midgard_jmp_writeout_op_branch_cond:
 673                 printf("cond.");
 674                 break;
 675
 676         case midgard_jmp_writeout_op_writeout:
 677                 printf("write.");
 678                 break;
 679
 680         case midgard_jmp_writeout_op_tilebuffer_pending:
 681                 printf("tilebuffer.");
 682                 break;
 683
 684         case midgard_jmp_writeout_op_discard:
 685                 printf("discard.");
 686                 break;
 687
 688         default:
 689                 printf("unk%u.", op);
 690                 break;
 691         }
 692 }
 693
 694 static void
 695 print_branch_cond(int cond)
 696 {
 697         switch (cond) {
 698         case midgard_condition_write0:
 699                 printf("write0");
 700                 break;
 701
 702         case midgard_condition_false:
 703                 printf("false");
 704                 break;
 705
 706         case midgard_condition_true:
 707                 printf("true");
 708                 break;
 709
 710         case midgard_condition_always:
 711                 printf("always");
 712                 break;
 713
 714         default:
 715                 printf("unk%X", cond);
 716                 break;
 717         }
 718 }
 719
 720 static void
 721 print_compact_branch_writeout_field(uint16_t word)
 722 {
 723         midgard_jmp_writeout_op op = word & 0x7;
 724
 725         switch (op) {
 726         case midgard_jmp_writeout_op_branch_uncond: {
 727                 midgard_branch_uncond br_uncond;
 728                 memcpy((char *) &br_uncond, (char *) &word, sizeof(br_uncond));
 729                 printf("br.uncond ");
 730
 731                 if (br_uncond.unknown != 1)
 732                         printf("unknown:%u, ", br_uncond.unknown);
 733
 734                 if (br_uncond.offset >= 0)
 735                         printf("+");
 736
 737                 printf("%d -> ", br_uncond.offset);
 738                 print_tag_short(br_uncond.dest_tag);
 739                 printf("\n");
 740
 741                 break;
 742         }
 743
 744         case midgard_jmp_writeout_op_branch_cond:
 745         case midgard_jmp_writeout_op_writeout:
 746         case midgard_jmp_writeout_op_discard:
 747         default: {
 748                 midgard_branch_cond br_cond;
 749                 memcpy((char *) &br_cond, (char *) &word, sizeof(br_cond));
 750
 751                 printf("br.");
 752
 753                 print_branch_op(br_cond.op);
 754                 print_branch_cond(br_cond.cond);
 755
 756                 printf(" ");
 757
 758                 if (br_cond.offset >= 0)
 759                         printf("+");
 760
 761                 printf("%d -> ", br_cond.offset);
 762                 print_tag_short(br_cond.dest_tag);
 763                 printf("\n");
 764
 765                 break;
 766         }
 767         }
 768
 769         midg_stats.instruction_count++;
 770 }
 771
 772 static void
 773 print_extended_branch_writeout_field(uint8_t *words)
 774 {
 775         midgard_branch_extended br;
 776         memcpy((char *) &br, (char *) words, sizeof(br));
 777
 778         printf("brx.");
 779
 780         print_branch_op(br.op);
 781
 782         /* Condition codes are a LUT in the general case, but simply repeated 8 times for single-channel conditions.. Check this. */
 783
 784         bool single_channel = true;
 785
 786         for (unsigned i = 0; i < 16; i += 2) {
 787                 single_channel &= (((br.cond >> i) & 0x3) == (br.cond & 0x3));
 788         }
 789
 790         if (single_channel)
 791                 print_branch_cond(br.cond & 0x3);
 792         else
 793                 printf("lut%X", br.cond);
 794
 795         if (br.unknown)
 796                 printf(".unknown%u", br.unknown);
 797
 798         printf(" ");
 799
 800         if (br.offset >= 0)
 801                 printf("+");
 802
 803         printf("%d -> ", br.offset);
 804         print_tag_short(br.dest_tag);
 805         printf("\n");
 806
 807         midg_stats.instruction_count++;
 808 }
 809
 810 static unsigned
 811 num_alu_fields_enabled(uint32_t control_word)
 812 {
 813         unsigned ret = 0;
 814
 815         if ((control_word >> 17) & 1)
 816                 ret++;
 817
 818         if ((control_word >> 19) & 1)
 819                 ret++;
 820
 821         if ((control_word >> 21) & 1)
 822                 ret++;
 823
 824         if ((control_word >> 23) & 1)
 825                 ret++;
 826
 827         if ((control_word >> 25) & 1)
 828                 ret++;
 829
 830         return ret;
 831 }
 832
 833 static float
 834 float_bitcast(uint32_t integer)
 835 {
 836         union {
 837                 uint32_t i;
 838                 float f;
 839         } v;
 840
 841         v.i = integer;
 842         return v.f;
 843 }
 844
 845 static void
 846 print_alu_word(uint32_t *words, unsigned num_quad_words,
 847                unsigned tabs)
 848 {
 849         uint32_t control_word = words[0];
 850         uint16_t *beginning_ptr = (uint16_t *)(words + 1);
 851         unsigned num_fields = num_alu_fields_enabled(control_word);
 852         uint16_t *word_ptr = beginning_ptr + num_fields;
 853         unsigned num_words = 2 + num_fields;
 854
 855         if ((control_word >> 16) & 1)
 856                 printf("unknown bit 16 enabled\n");
 857
 858         if ((control_word >> 17) & 1) {
 859                 print_vector_field("vmul", word_ptr, *beginning_ptr, tabs);
 860                 beginning_ptr += 1;
 861                 word_ptr += 3;
 862                 num_words += 3;
 863         }
 864
 865         if ((control_word >> 18) & 1)
 866                 printf("unknown bit 18 enabled\n");
 867
 868         if ((control_word >> 19) & 1) {
 869                 print_scalar_field("sadd", word_ptr, *beginning_ptr, tabs);
 870                 beginning_ptr += 1;
 871                 word_ptr += 2;
 872                 num_words += 2;
 873         }
 874
 875         if ((control_word >> 20) & 1)
 876                 printf("unknown bit 20 enabled\n");
 877
 878         if ((control_word >> 21) & 1) {
 879                 print_vector_field("vadd", word_ptr, *beginning_ptr, tabs);
 880                 beginning_ptr += 1;
 881                 word_ptr += 3;
 882                 num_words += 3;
 883         }
 884
 885         if ((control_word >> 22) & 1)
 886                 printf("unknown bit 22 enabled\n");
 887
 888         if ((control_word >> 23) & 1) {
 889                 print_scalar_field("smul", word_ptr, *beginning_ptr, tabs);
 890                 beginning_ptr += 1;
 891                 word_ptr += 2;
 892                 num_words += 2;
 893         }
 894
 895         if ((control_word >> 24) & 1)
 896                 printf("unknown bit 24 enabled\n");
 897
 898         if ((control_word >> 25) & 1) {
 899                 print_vector_field("lut", word_ptr, *beginning_ptr, tabs);
 900                 word_ptr += 3;
 901                 num_words += 3;
 902         }
 903
 904         if ((control_word >> 26) & 1) {
 905                 print_compact_branch_writeout_field(*word_ptr);
 906                 word_ptr += 1;
 907                 num_words += 1;
 908         }
 909
 910         if ((control_word >> 27) & 1) {
 911                 print_extended_branch_writeout_field((uint8_t *) word_ptr);
 912                 word_ptr += 3;
 913                 num_words += 3;
 914         }
 915
 916         if (num_quad_words > (num_words + 7) / 8) {
 917                 assert(num_quad_words == (num_words + 15) / 8);
 918                 //Assume that the extra quadword is constants
 919                 void *consts = words + (4 * num_quad_words - 4);
 920
 921                 if (is_embedded_constant_int) {
 922                         if (is_embedded_constant_half) {
 923                                 int16_t *sconsts = (int16_t *) consts;
 924                                 printf("sconstants %d, %d, %d, %d\n",
 925                                        sconsts[0],
 926                                        sconsts[1],
 927                                        sconsts[2],
 928                                        sconsts[3]);
 929                         } else {
 930                                 uint32_t *iconsts = (uint32_t *) consts;
 931                                 printf("iconstants 0x%X, 0x%X, 0x%X, 0x%X\n",
 932                                        iconsts[0],
 933                                        iconsts[1],
 934                                        iconsts[2],
 935                                        iconsts[3]);
 936                         }
 937                 } else {
 938                         if (is_embedded_constant_half) {
 939                                 uint16_t *hconsts = (uint16_t *) consts;
 940                                 printf("hconstants %g, %g, %g, %g\n",
 941                                        _mesa_half_to_float(hconsts[0]),
 942                                        _mesa_half_to_float(hconsts[1]),
 943                                        _mesa_half_to_float(hconsts[2]),
 944                                        _mesa_half_to_float(hconsts[3]));
 945                         } else {
 946                                 uint32_t *fconsts = (uint32_t *) consts;
 947                                 printf("fconstants %g, %g, %g, %g\n",
 948                                        float_bitcast(fconsts[0]),
 949                                        float_bitcast(fconsts[1]),
 950                                        float_bitcast(fconsts[2]),
 951                                        float_bitcast(fconsts[3]));
 952                         }
 953
 954                 }
 955         }
 956 }
 957
 958 static void
 959 print_varying_parameters(midgard_load_store_word *word)
 960 {
 961         midgard_varying_parameter param;
 962         unsigned v = word->varying_parameters;
 963         memcpy(&param, &v, sizeof(param));
 964
 965         if (param.is_varying) {
 966                 /* If a varying, there are qualifiers */
 967                 if (param.flat)
 968                         printf(".flat");
 969
 970                 if (param.interpolation != midgard_interp_default) {
 971                         if (param.interpolation == midgard_interp_centroid)
 972                                 printf(".centroid");
 973                         else
 974                                 printf(".interp%d", param.interpolation);
 975                 }
 976
 977                 if (param.modifier != midgard_varying_mod_none) {
 978                         if (param.modifier == midgard_varying_mod_perspective_w)
 979                                 printf(".perspectivew");
 980                         else if (param.modifier == midgard_varying_mod_perspective_z)
 981                                 printf(".perspectivez");
 982                         else
 983                                 printf(".mod%d", param.modifier);
 984                 }
 985         } else if (param.flat || param.interpolation || param.modifier) {
 986                 printf(" /* is_varying not set but varying metadata attached */");
 987         }
 988
 989         if (param.zero0 || param.zero1 || param.zero2)
 990                 printf(" /* zero tripped, %u %u %u */ ", param.zero0, param.zero1, param.zero2);
 991 }
 992
 993 static bool
 994 is_op_varying(unsigned op)
 995 {
 996         switch (op) {
 997         case midgard_op_st_vary_16:
 998         case midgard_op_st_vary_32:
 999         case midgard_op_st_vary_32i:
1000         case midgard_op_st_vary_32u:
1001         case midgard_op_ld_vary_16:
1002         case midgard_op_ld_vary_32:
1003         case midgard_op_ld_vary_32i:
1004         case midgard_op_ld_vary_32u:
1005                 return true;
1006         }
1007
1008         return false;
1009 }
1010
1011 static bool
1012 is_op_attribute(unsigned op)
1013 {
1014         switch (op) {
1015         case midgard_op_ld_attr_16:
1016         case midgard_op_ld_attr_32:
1017         case midgard_op_ld_attr_32i:
1018         case midgard_op_ld_attr_32u:
1019                 return true;
1020         }
1021
1022         return false;
1023 }
1024
1025 static void
1026 print_load_store_arg(uint8_t arg, unsigned index)
1027 {
1028         /* Try to interpret as a register */
1029         midgard_ldst_register_select sel;
1030         memcpy(&sel, &arg, sizeof(arg));
1031
1032         /* If unknown is set, we're not sure what this is or how to
1033          * interpret it. But if it's zero, we get it. */
1034
1035         if (sel.unknown) {
1036                 printf("0x%02X", arg);
1037                 return;
1038         }
1039
1040         unsigned reg = REGISTER_LDST_BASE + sel.select;
1041         char comp = components[sel.component];
1042
1043         printf("r%u.%c", reg, comp);
1044
1045         /* Only print a shift if it's non-zero. Shifts only make sense for the
1046          * second index. For the first, we're not sure what it means yet */
1047
1048         if (index == 1) {
1049                 if (sel.shift)
1050                         printf(" << %u", sel.shift);
1051         } else {
1052                 printf(" /* %X */", sel.shift);
1053         }
1054 }
1055
1056 static void
1057 update_stats(signed *stat, unsigned address)
1058 {
1059         if (*stat >= 0)
1060                 *stat = MAX2(*stat, address + 1);
1061 }
1062
1063 static void
1064 print_load_store_instr(uint64_t data,
1065                        unsigned tabs)
1066 {
1067         midgard_load_store_word *word = (midgard_load_store_word *) &data;
1068
1069         print_ld_st_opcode(word->op);
1070
1071         unsigned address = word->address;
1072
1073         if (is_op_varying(word->op)) {
1074                 print_varying_parameters(word);
1075
1076                 /* Do some analysis: check if direct cacess */
1077
1078                 if ((word->arg_2 == 0x1E) && midg_stats.varying_count >= 0)
1079                         update_stats(&midg_stats.varying_count, address);
1080                 else
1081                         midg_stats.varying_count = -16;
1082         } else if (is_op_attribute(word->op)) {
1083                 if ((word->arg_2 == 0x1E) && midg_stats.attribute_count >= 0)
1084                         update_stats(&midg_stats.attribute_count, address);
1085                 else
1086                         midg_stats.attribute_count = -16;
1087         }
1088
1089         printf(" r%u", word->reg);
1090         print_mask_4(word->mask);
1091
1092         if (!OP_IS_STORE(word->op))
1093                 update_dest(word->reg);
1094
1095         bool is_ubo = OP_IS_UBO_READ(word->op);
1096
1097         if (is_ubo) {
1098                 /* UBOs use their own addressing scheme */
1099
1100                 int lo = word->varying_parameters >> 7;
1101                 int hi = word->address;
1102
1103                 /* TODO: Combine fields logically */
1104                 address = (hi << 3) | lo;
1105         }
1106
1107         printf(", %u", address);
1108
1109         print_swizzle_vec4(word->swizzle, false, false);
1110
1111         printf(", ");
1112
1113         if (is_ubo) {
1114                 printf("ubo%u", word->arg_1);
1115                 update_stats(&midg_stats.uniform_buffer_count, word->arg_1);
1116         } else
1117                 print_load_store_arg(word->arg_1, 0);
1118
1119         printf(", ");
1120         print_load_store_arg(word->arg_2, 1);
1121         printf(" /* %X */\n", word->varying_parameters);
1122
1123         midg_stats.instruction_count++;
1124 }
1125
1126 static void
1127 print_load_store_word(uint32_t *word, unsigned tabs)
1128 {
1129         midgard_load_store *load_store = (midgard_load_store *) word;
1130
1131         if (load_store->word1 != 3) {
1132                 print_load_store_instr(load_store->word1, tabs);
1133         }
1134
1135         if (load_store->word2 != 3) {
1136                 print_load_store_instr(load_store->word2, tabs);
1137         }
1138 }
1139
1140 static void
1141 print_texture_reg(bool full, bool select, bool upper)
1142 {
1143         if (full)
1144                 printf("r%d", REG_TEX_BASE + select);
1145         else
1146                 printf("hr%d", (REG_TEX_BASE + select) * 2 + upper);
1147
1148         if (full && upper)
1149                 printf("// error: out full / upper mutually exclusive\n");
1150
1151 }
1152
1153 static void
1154 print_texture_reg_triple(unsigned triple)
1155 {
1156         bool full = triple & 1;
1157         bool select = triple & 2;
1158         bool upper = triple & 4;
1159
1160         print_texture_reg(full, select, upper);
1161 }
1162
1163 static void
1164 print_texture_reg_select(uint8_t u)
1165 {
1166         midgard_tex_register_select sel;
1167         memcpy(&sel, &u, sizeof(u));
1168
1169         if (!sel.full)
1170                 printf("h");
1171
1172         printf("r%u", REG_TEX_BASE + sel.select);
1173
1174         unsigned component = sel.component;
1175
1176         /* Use the upper half in half-reg mode */
1177         if (sel.upper) {
1178                 assert(!sel.full);
1179                 component += 4;
1180         }
1181
1182         printf(".%c", components[component]);
1183
1184         assert(sel.zero == 0);
1185 }
1186
1187 static void
1188 print_texture_format(int format)
1189 {
1190         /* Act like a modifier */
1191         printf(".");
1192
1193         switch (format) {
1194                 DEFINE_CASE(MALI_TEX_1D, "1d");
1195                 DEFINE_CASE(MALI_TEX_2D, "2d");
1196                 DEFINE_CASE(MALI_TEX_3D, "3d");
1197                 DEFINE_CASE(MALI_TEX_CUBE, "cube");
1198
1199         default:
1200                 unreachable("Bad format");
1201         }
1202 }
1203
1204 static bool
1205 midgard_op_has_helpers(unsigned op, bool gather)
1206 {
1207         if (gather)
1208                 return true;
1209
1210         switch (op) {
1211         case TEXTURE_OP_NORMAL:
1212         case TEXTURE_OP_DFDX:
1213         case TEXTURE_OP_DFDY:
1214                 return true;
1215         default:
1216                 return false;
1217         }
1218 }
1219
1220 static void
1221 print_texture_op(unsigned op, bool gather)
1222 {
1223         /* Act like a bare name, like ESSL functions */
1224
1225         if (gather) {
1226                 printf("textureGather");
1227
1228                 unsigned component = op >> 4;
1229                 unsigned bottom = op & 0xF;
1230
1231                 if (bottom != 0x2)
1232                         printf("_unk%u", bottom);
1233
1234                 printf(".%c", components[component]);
1235                 return;
1236         }
1237
1238         switch (op) {
1239                 DEFINE_CASE(TEXTURE_OP_NORMAL, "texture");
1240                 DEFINE_CASE(TEXTURE_OP_LOD, "textureLod");
1241                 DEFINE_CASE(TEXTURE_OP_TEXEL_FETCH, "texelFetch");
1242                 DEFINE_CASE(TEXTURE_OP_DFDX, "dFdx");
1243                 DEFINE_CASE(TEXTURE_OP_DFDY, "dFdy");
1244
1245         default:
1246                 printf("tex_%X", op);
1247                 break;
1248         }
1249 }
1250
1251 static bool
1252 texture_op_takes_bias(unsigned op)
1253 {
1254         return op == TEXTURE_OP_NORMAL;
1255 }
1256
1257 static char
1258 sampler_type_name(enum mali_sampler_type t)
1259 {
1260         switch (t) {
1261         case MALI_SAMPLER_FLOAT:
1262                 return 'f';
1263         case MALI_SAMPLER_UNSIGNED:
1264                 return 'u';
1265         case MALI_SAMPLER_SIGNED:
1266                 return 'i';
1267         default:
1268                 return '?';
1269         }
1270
1271 }
1272
1273 #undef DEFINE_CASE
1274
1275 static void
1276 print_texture_word(uint32_t *word, unsigned tabs)
1277 {
1278         midgard_texture_word *texture = (midgard_texture_word *) word;
1279
1280         midg_stats.helper_invocations |=
1281                 midgard_op_has_helpers(texture->op, texture->is_gather);
1282
1283         /* Broad category of texture operation in question */
1284         print_texture_op(texture->op, texture->is_gather);
1285
1286         /* Specific format in question */
1287         print_texture_format(texture->format);
1288
1289         /* Instruction "modifiers" parallel the ALU instructions. */
1290
1291         if (texture->shadow)
1292                 printf(".shadow");
1293
1294         if (texture->cont)
1295                 printf(".cont");
1296
1297         if (texture->last)
1298                 printf(".last");
1299
1300         /* Output modifiers are always interpreted floatly */
1301         print_outmod(texture->outmod, false);
1302
1303         printf(" ");
1304
1305         print_texture_reg(texture->out_full, texture->out_reg_select, texture->out_upper);
1306         print_mask_4(texture->mask);
1307         printf(", ");
1308
1309         /* Depending on whether we read from textures directly or indirectly,
1310          * we may be able to update our analysis */
1311
1312         if (texture->texture_register) {
1313                 printf("texture[");
1314                 print_texture_reg_select(texture->texture_handle);
1315                 printf("], ");
1316
1317                 /* Indirect, tut tut */
1318                 midg_stats.texture_count = -16;
1319         } else {
1320                 printf("texture%u, ", texture->texture_handle);
1321                 update_stats(&midg_stats.texture_count, texture->texture_handle);
1322         }
1323
1324         /* Print the type, GL style */
1325         printf("%csampler", sampler_type_name(texture->sampler_type));
1326
1327         if (texture->sampler_register) {
1328                 printf("[");
1329                 print_texture_reg_select(texture->sampler_handle);
1330                 printf("]");
1331
1332                 midg_stats.sampler_count = -16;
1333         } else {
1334                 printf("%u", texture->sampler_handle);
1335                 update_stats(&midg_stats.sampler_count, texture->sampler_handle);
1336         }
1337
1338         print_swizzle_vec4(texture->swizzle, false, false);
1339         printf(", ");
1340
1341         print_texture_reg(texture->in_reg_full, texture->in_reg_select, texture->in_reg_upper);
1342         print_swizzle_vec4(texture->in_reg_swizzle, false, false);
1343
1344         /* There is *always* an offset attached. Of
1345          * course, that offset is just immediate #0 for a
1346          * GLES call that doesn't take an offset. If there
1347          * is a non-negative non-zero offset, this is
1348          * specified in immediate offset mode, with the
1349          * values in the offset_* fields as immediates. If
1350          * this is a negative offset, we instead switch to
1351          * a register offset mode, where the offset_*
1352          * fields become register triplets */
1353
1354         if (texture->offset_register) {
1355                 printf(" + ");
1356                 print_texture_reg_triple(texture->offset_x);
1357
1358                 /* The less questions you ask, the better. */
1359
1360                 unsigned swizzle_lo, swizzle_hi;
1361                 unsigned orig_y = texture->offset_y;
1362                 unsigned orig_z = texture->offset_z;
1363
1364                 memcpy(&swizzle_lo, &orig_y, sizeof(unsigned));
1365                 memcpy(&swizzle_hi, &orig_z, sizeof(unsigned));
1366
1367                 /* Duplicate hi swizzle over */
1368                 assert(swizzle_hi < 4);
1369                 swizzle_hi = (swizzle_hi << 2) | swizzle_hi;
1370
1371                 unsigned swiz = (swizzle_lo << 4) | swizzle_hi;
1372                 unsigned reversed = util_bitreverse(swiz) >> 24;
1373                 print_swizzle_vec4(reversed, false, false);
1374
1375                 printf(", ");
1376         } else if (texture->offset_x || texture->offset_y || texture->offset_z) {
1377                 /* Only select ops allow negative immediate offsets, verify */
1378
1379                 bool neg_x = texture->offset_x < 0;
1380                 bool neg_y = texture->offset_y < 0;
1381                 bool neg_z = texture->offset_z < 0;
1382                 bool any_neg = neg_x || neg_y || neg_z;
1383
1384                 if (any_neg && texture->op != TEXTURE_OP_TEXEL_FETCH)
1385                         printf("/* invalid negative */ ");
1386
1387                 /* Regardless, just print the immediate offset */
1388
1389                 printf(" + <%d, %d, %d>, ",
1390                        texture->offset_x,
1391                        texture->offset_y,
1392                        texture->offset_z);
1393         } else {
1394                 printf(", ");
1395         }
1396
1397         char lod_operand = texture_op_takes_bias(texture->op) ? '+' : '=';
1398
1399         if (texture->lod_register) {
1400                 printf("lod %c ", lod_operand);
1401                 print_texture_reg_select(texture->bias);
1402                 printf(", ");
1403
1404                 if (texture->bias_int)
1405                         printf(" /* bias_int = 0x%X */", texture->bias_int);
1406         } else if (texture->op == TEXTURE_OP_TEXEL_FETCH) {
1407                 /* For texel fetch, the int LOD is in the fractional place and
1408                  * there is no fraction / possibility of bias. We *always* have
1409                  * an explicit LOD, even if it's zero. */
1410
1411                 if (texture->bias_int)
1412                         printf(" /* bias_int = 0x%X */ ", texture->bias_int);
1413
1414                 printf("lod = %u, ", texture->bias);
1415         } else if (texture->bias || texture->bias_int) {
1416                 signed bias_int = texture->bias_int;
1417                 float bias_frac = texture->bias / 256.0f;
1418                 float bias = bias_int + bias_frac;
1419
1420                 bool is_bias = texture_op_takes_bias(texture->op);
1421                 char sign = (bias >= 0.0) ? '+' : '-';
1422                 char operand = is_bias ? sign : '=';
1423
1424                 printf("lod %c %f, ", operand, fabsf(bias));
1425         }
1426
1427         printf("\n");
1428
1429         /* While not zero in general, for these simple instructions the
1430          * following unknowns are zero, so we don't include them */
1431
1432         if (texture->unknown4 ||
1433             texture->unknownA ||
1434             texture->unknown8) {
1435                 printf("// unknown4 = 0x%x\n", texture->unknown4);
1436                 printf("// unknownA = 0x%x\n", texture->unknownA);
1437                 printf("// unknown8 = 0x%x\n", texture->unknown8);
1438         }
1439
1440         midg_stats.instruction_count++;
1441 }
1442
1443 struct midgard_disasm_stats
1444 disassemble_midgard(uint8_t *code, size_t size)
1445 {
1446         uint32_t *words = (uint32_t *) code;
1447         unsigned num_words = size / 4;
1448         int tabs = 0;
1449
1450         bool prefetch_flag = false;
1451
1452         int last_next_tag = -1;
1453
1454         unsigned i = 0;
1455
1456         /* Stats for shader-db */
1457         memset(&midg_stats, 0, sizeof(midg_stats));
1458         midg_ever_written = 0;
1459
1460         while (i < num_words) {
1461                 unsigned tag = words[i] & 0xF;
1462                 unsigned next_tag = (words[i] >> 4) & 0xF;
1463                 unsigned num_quad_words = midgard_word_size[tag];
1464
1465                 /* Check the tag */
1466                 if (last_next_tag > 1) {
1467                         if (last_next_tag != tag) {
1468                                 printf("/* TAG ERROR got ");
1469                                 print_tag_short(tag);
1470                                 printf(" expected ");
1471                                 print_tag_short(last_next_tag);
1472                                 printf(" */ ");
1473                         }
1474                 } else {
1475                         /* TODO: Check ALU case */
1476                 }
1477
1478                 last_next_tag = next_tag;
1479
1480                 switch (midgard_word_types[tag]) {
1481                 case midgard_word_type_texture:
1482                         print_texture_word(&words[i], tabs);
1483                         break;
1484
1485                 case midgard_word_type_load_store:
1486                         print_load_store_word(&words[i], tabs);
1487                         break;
1488
1489                 case midgard_word_type_alu:
1490                         print_alu_word(&words[i], num_quad_words, tabs);
1491
1492                         /* Reset word static analysis state */
1493                         is_embedded_constant_half = false;
1494                         is_embedded_constant_int = false;
1495
1496                         break;
1497
1498                 default:
1499                         printf("Unknown word type %u:\n", words[i] & 0xF);
1500                         num_quad_words = 1;
1501                         print_quad_word(&words[i], tabs);
1502                         printf("\n");
1503                         break;
1504                 }
1505
1506                 if (prefetch_flag && midgard_word_types[tag] == midgard_word_type_alu)
1507                         break;
1508
1509                 printf("\n");
1510
1511                 unsigned next = (words[i] & 0xF0) >> 4;
1512
1513                 /* We are parsing per bundle anyway */
1514                 midg_stats.bundle_count++;
1515                 midg_stats.quadword_count += num_quad_words;
1516
1517                 /* Break based on instruction prefetch flag */
1518
1519                 if (i < num_words && next == 1) {
1520                         prefetch_flag = true;
1521
1522                         if (midgard_word_types[words[i] & 0xF] != midgard_word_type_alu)
1523                                 break;
1524                 }
1525
1526                 i += 4 * num_quad_words;
1527         }
1528
1529         /* We computed work_count as max_work_registers, so add one to get the
1530          * count. If no work registers are written, you still have one work
1531          * reported, which is exactly what the hardware expects */
1532
1533         midg_stats.work_count++;
1534
1535         return midg_stats;
1536 }