src/amd/compiler/aco_print_ir.cpp

   1 #include "aco_ir.h"
   2 #include "aco_builder.h"
   3
   4 #include "sid.h"
   5 #include "ac_shader_util.h"
   6
   7 namespace aco {
   8
   9 static const char *reduce_ops[] = {
  10    [iadd8] = "iadd8",
  11    [iadd16] = "iadd16",
  12    [iadd32] = "iadd32",
  13    [iadd64] = "iadd64",
  14    [imul8] = "imul8",
  15    [imul16] = "imul16",
  16    [imul32] = "imul32",
  17    [imul64] = "imul64",
  18    [fadd16] = "fadd16",
  19    [fadd32] = "fadd32",
  20    [fadd64] = "fadd64",
  21    [fmul16] = "fmul16",
  22    [fmul32] = "fmul32",
  23    [fmul64] = "fmul64",
  24    [imin8] = "imin8",
  25    [imin16] = "imin16",
  26    [imin32] = "imin32",
  27    [imin64] = "imin64",
  28    [imax8] = "imax8",
  29    [imax16] = "imax16",
  30    [imax32] = "imax32",
  31    [imax64] = "imax64",
  32    [umin8] = "umin8",
  33    [umin16] = "umin16",
  34    [umin32] = "umin32",
  35    [umin64] = "umin64",
  36    [umax8] = "umax8",
  37    [umax16] = "umax16",
  38    [umax32] = "umax32",
  39    [umax64] = "umax64",
  40    [fmin16] = "fmin16",
  41    [fmin32] = "fmin32",
  42    [fmin64] = "fmin64",
  43    [fmax16] = "fmax16",
  44    [fmax32] = "fmax32",
  45    [fmax64] = "fmax64",
  46    [iand8] = "iand8",
  47    [iand16] = "iand16",
  48    [iand32] = "iand32",
  49    [iand64] = "iand64",
  50    [ior8] = "ior8",
  51    [ior16] = "ior16",
  52    [ior32] = "ior32",
  53    [ior64] = "ior64",
  54    [ixor8] = "ixor8",
  55    [ixor16] = "ixor16",
  56    [ixor32] = "ixor32",
  57    [ixor64] = "ixor64",
  58    [gfx10_wave64_bpermute] = "gfx10_wave64_bpermute",
  59 };
  60
  61 static void print_reg_class(const RegClass rc, FILE *output)
  62 {
  63    switch (rc) {
  64       case RegClass::s1: fprintf(output, " s1: "); return;
  65       case RegClass::s2: fprintf(output, " s2: "); return;
  66       case RegClass::s3: fprintf(output, " s3: "); return;
  67       case RegClass::s4: fprintf(output, " s4: "); return;
  68       case RegClass::s6: fprintf(output, " s6: "); return;
  69       case RegClass::s8: fprintf(output, " s8: "); return;
  70       case RegClass::s16: fprintf(output, "s16: "); return;
  71       case RegClass::v1: fprintf(output, " v1: "); return;
  72       case RegClass::v2: fprintf(output, " v2: "); return;
  73       case RegClass::v3: fprintf(output, " v3: "); return;
  74       case RegClass::v4: fprintf(output, " v4: "); return;
  75       case RegClass::v5: fprintf(output, " v5: "); return;
  76       case RegClass::v6: fprintf(output, " v6: "); return;
  77       case RegClass::v7: fprintf(output, " v7: "); return;
  78       case RegClass::v8: fprintf(output, " v8: "); return;
  79       case RegClass::v1b: fprintf(output, " v1b: "); return;
  80       case RegClass::v2b: fprintf(output, " v2b: "); return;
  81       case RegClass::v3b: fprintf(output, " v3b: "); return;
  82       case RegClass::v4b: fprintf(output, " v4b: "); return;
  83       case RegClass::v6b: fprintf(output, " v6b: "); return;
  84       case RegClass::v8b: fprintf(output, " v8b: "); return;
  85       case RegClass::v1_linear: fprintf(output, " v1: "); return;
  86       case RegClass::v2_linear: fprintf(output, " v2: "); return;
  87    }
  88 }
  89
  90 void print_physReg(PhysReg reg, unsigned bytes, FILE *output)
  91 {
  92    if (reg == 124) {
  93       fprintf(output, ":m0");
  94    } else if (reg == 106) {
  95       fprintf(output, ":vcc");
  96    } else if (reg == 253) {
  97       fprintf(output, ":scc");
  98    } else if (reg == 126) {
  99       fprintf(output, ":exec");
 100    } else {
 101       bool is_vgpr = reg / 256;
 102       unsigned r = reg % 256;
 103       unsigned size = DIV_ROUND_UP(bytes, 4);
 104       fprintf(output, ":%c[%d", is_vgpr ? 'v' : 's', r);
 105       if (size > 1)
 106          fprintf(output, "-%d]", r + size -1);
 107       else
 108          fprintf(output, "]");
 109       if (reg.byte() || bytes % 4)
 110          fprintf(output, "[%d:%d]", reg.byte()*8, (reg.byte()+bytes) * 8);
 111    }
 112 }
 113
 114 static void print_constant(uint8_t reg, FILE *output)
 115 {
 116    if (reg >= 128 && reg <= 192) {
 117       fprintf(output, "%d", reg - 128);
 118       return;
 119    } else if (reg >= 192 && reg <= 208) {
 120       fprintf(output, "%d", 192 - reg);
 121       return;
 122    }
 123
 124    switch (reg) {
 125    case 240:
 126       fprintf(output, "0.5");
 127       break;
 128    case 241:
 129       fprintf(output, "-0.5");
 130       break;
 131    case 242:
 132       fprintf(output, "1.0");
 133       break;
 134    case 243:
 135       fprintf(output, "-1.0");
 136       break;
 137    case 244:
 138       fprintf(output, "2.0");
 139       break;
 140    case 245:
 141       fprintf(output, "-2.0");
 142       break;
 143    case 246:
 144       fprintf(output, "4.0");
 145       break;
 146    case 247:
 147       fprintf(output, "-4.0");
 148       break;
 149    case 248:
 150       fprintf(output, "1/(2*PI)");
 151       break;
 152    }
 153 }
 154
 155 static void print_operand(const Operand *operand, FILE *output)
 156 {
 157    if (operand->isLiteral()) {
 158       fprintf(output, "0x%x", operand->constantValue());
 159    } else if (operand->isConstant()) {
 160       print_constant(operand->physReg().reg(), output);
 161    } else if (operand->isUndefined()) {
 162       print_reg_class(operand->regClass(), output);
 163       fprintf(output, "undef");
 164    } else {
 165       if (operand->isLateKill())
 166          fprintf(output, "(latekill)");
 167
 168       fprintf(output, "%%%d", operand->tempId());
 169
 170       if (operand->isFixed())
 171          print_physReg(operand->physReg(), operand->bytes(), output);
 172    }
 173 }
 174
 175 static void print_definition(const Definition *definition, FILE *output)
 176 {
 177    print_reg_class(definition->regClass(), output);
 178    fprintf(output, "%%%d", definition->tempId());
 179
 180    if (definition->isFixed())
 181       print_physReg(definition->physReg(), definition->bytes(), output);
 182 }
 183
 184 static void print_barrier_reorder(bool can_reorder, barrier_interaction barrier, FILE *output)
 185 {
 186    if (can_reorder)
 187       fprintf(output, " reorder");
 188
 189    if (barrier & barrier_buffer)
 190       fprintf(output, " buffer");
 191    if (barrier & barrier_image)
 192       fprintf(output, " image");
 193    if (barrier & barrier_atomic)
 194       fprintf(output, " atomic");
 195    if (barrier & barrier_shared)
 196       fprintf(output, " shared");
 197    if (barrier & barrier_gs_data)
 198       fprintf(output, " gs_data");
 199    if (barrier & barrier_gs_sendmsg)
 200       fprintf(output, " gs_sendmsg");
 201 }
 202
 203 static void print_instr_format_specific(const Instruction *instr, FILE *output)
 204 {
 205    switch (instr->format) {
 206    case Format::SOPK: {
 207       const SOPK_instruction* sopk = static_cast<const SOPK_instruction*>(instr);
 208       fprintf(output, " imm:%d", sopk->imm & 0x8000 ? (sopk->imm - 65536) : sopk->imm);
 209       break;
 210    }
 211    case Format::SOPP: {
 212       const SOPP_instruction* sopp = static_cast<const SOPP_instruction*>(instr);
 213       uint16_t imm = sopp->imm;
 214       switch (instr->opcode) {
 215       case aco_opcode::s_waitcnt: {
 216          /* we usually should check the chip class for vmcnt/lgkm, but
 217           * insert_waitcnt() should fill it in regardless. */
 218          unsigned vmcnt = (imm & 0xF) | ((imm & (0x3 << 14)) >> 10);
 219          if (vmcnt != 63) fprintf(output, " vmcnt(%d)", vmcnt);
 220          if (((imm >> 4) & 0x7) < 0x7) fprintf(output, " expcnt(%d)", (imm >> 4) & 0x7);
 221          if (((imm >> 8) & 0x3F) < 0x3F) fprintf(output, " lgkmcnt(%d)", (imm >> 8) & 0x3F);
 222          break;
 223       }
 224       case aco_opcode::s_endpgm:
 225       case aco_opcode::s_endpgm_saved:
 226       case aco_opcode::s_endpgm_ordered_ps_done:
 227       case aco_opcode::s_wakeup:
 228       case aco_opcode::s_barrier:
 229       case aco_opcode::s_icache_inv:
 230       case aco_opcode::s_ttracedata:
 231       case aco_opcode::s_set_gpr_idx_off: {
 232          break;
 233       }
 234       case aco_opcode::s_sendmsg: {
 235          unsigned id = imm & sendmsg_id_mask;
 236          switch (id) {
 237          case sendmsg_none:
 238             fprintf(output, " sendmsg(MSG_NONE)");
 239             break;
 240          case _sendmsg_gs:
 241             fprintf(output, " sendmsg(gs%s%s, %u)",
 242                     imm & 0x10 ? ", cut" : "", imm & 0x20 ? ", emit" : "", imm >> 8);
 243             break;
 244          case _sendmsg_gs_done:
 245             fprintf(output, " sendmsg(gs_done%s%s, %u)",
 246                     imm & 0x10 ? ", cut" : "", imm & 0x20 ? ", emit" : "", imm >> 8);
 247             break;
 248          case sendmsg_save_wave:
 249             fprintf(output, " sendmsg(save_wave)");
 250             break;
 251          case sendmsg_stall_wave_gen:
 252             fprintf(output, " sendmsg(stall_wave_gen)");
 253             break;
 254          case sendmsg_halt_waves:
 255             fprintf(output, " sendmsg(halt_waves)");
 256             break;
 257          case sendmsg_ordered_ps_done:
 258             fprintf(output, " sendmsg(ordered_ps_done)");
 259             break;
 260          case sendmsg_early_prim_dealloc:
 261             fprintf(output, " sendmsg(early_prim_dealloc)");
 262             break;
 263          case sendmsg_gs_alloc_req:
 264             fprintf(output, " sendmsg(gs_alloc_req)");
 265             break;
 266          }
 267          break;
 268       }
 269       default: {
 270          if (imm)
 271             fprintf(output, " imm:%u", imm);
 272          break;
 273       }
 274       }
 275       if (sopp->block != -1)
 276          fprintf(output, " block:BB%d", sopp->block);
 277       break;
 278    }
 279    case Format::SMEM: {
 280       const SMEM_instruction* smem = static_cast<const SMEM_instruction*>(instr);
 281       if (smem->glc)
 282          fprintf(output, " glc");
 283       if (smem->dlc)
 284          fprintf(output, " dlc");
 285       if (smem->nv)
 286          fprintf(output, " nv");
 287       print_barrier_reorder(smem->can_reorder, smem->barrier, output);
 288       break;
 289    }
 290    case Format::VINTRP: {
 291       const Interp_instruction* vintrp = static_cast<const Interp_instruction*>(instr);
 292       fprintf(output, " attr%d.%c", vintrp->attribute, "xyzw"[vintrp->component]);
 293       break;
 294    }
 295    case Format::DS: {
 296       const DS_instruction* ds = static_cast<const DS_instruction*>(instr);
 297       if (ds->offset0)
 298          fprintf(output, " offset0:%u", ds->offset0);
 299       if (ds->offset1)
 300          fprintf(output, " offset1:%u", ds->offset1);
 301       if (ds->gds)
 302          fprintf(output, " gds");
 303       break;
 304    }
 305    case Format::MUBUF: {
 306       const MUBUF_instruction* mubuf = static_cast<const MUBUF_instruction*>(instr);
 307       if (mubuf->offset)
 308          fprintf(output, " offset:%u", mubuf->offset);
 309       if (mubuf->offen)
 310          fprintf(output, " offen");
 311       if (mubuf->idxen)
 312          fprintf(output, " idxen");
 313       if (mubuf->addr64)
 314          fprintf(output, " addr64");
 315       if (mubuf->glc)
 316          fprintf(output, " glc");
 317       if (mubuf->dlc)
 318          fprintf(output, " dlc");
 319       if (mubuf->slc)
 320          fprintf(output, " slc");
 321       if (mubuf->tfe)
 322          fprintf(output, " tfe");
 323       if (mubuf->lds)
 324          fprintf(output, " lds");
 325       if (mubuf->disable_wqm)
 326          fprintf(output, " disable_wqm");
 327       print_barrier_reorder(mubuf->can_reorder, mubuf->barrier, output);
 328       break;
 329    }
 330    case Format::MIMG: {
 331       const MIMG_instruction* mimg = static_cast<const MIMG_instruction*>(instr);
 332       unsigned identity_dmask = !instr->definitions.empty() ?
 333                                 (1 << instr->definitions[0].size()) - 1 :
 334                                 0xf;
 335       if ((mimg->dmask & identity_dmask) != identity_dmask)
 336          fprintf(output, " dmask:%s%s%s%s",
 337                  mimg->dmask & 0x1 ? "x" : "",
 338                  mimg->dmask & 0x2 ? "y" : "",
 339                  mimg->dmask & 0x4 ? "z" : "",
 340                  mimg->dmask & 0x8 ? "w" : "");
 341       switch (mimg->dim) {
 342       case ac_image_1d:
 343          fprintf(output, " 1d");
 344          break;
 345       case ac_image_2d:
 346          fprintf(output, " 2d");
 347          break;
 348       case ac_image_3d:
 349          fprintf(output, " 3d");
 350          break;
 351       case ac_image_cube:
 352          fprintf(output, " cube");
 353          break;
 354       case ac_image_1darray:
 355          fprintf(output, " 1darray");
 356          break;
 357       case ac_image_2darray:
 358          fprintf(output, " 2darray");
 359          break;
 360       case ac_image_2dmsaa:
 361          fprintf(output, " 2dmsaa");
 362          break;
 363       case ac_image_2darraymsaa:
 364          fprintf(output, " 2darraymsaa");
 365          break;
 366       }
 367       if (mimg->unrm)
 368          fprintf(output, " unrm");
 369       if (mimg->glc)
 370          fprintf(output, " glc");
 371       if (mimg->dlc)
 372          fprintf(output, " dlc");
 373       if (mimg->slc)
 374          fprintf(output, " slc");
 375       if (mimg->tfe)
 376          fprintf(output, " tfe");
 377       if (mimg->da)
 378          fprintf(output, " da");
 379       if (mimg->lwe)
 380          fprintf(output, " lwe");
 381       if (mimg->r128 || mimg->a16)
 382          fprintf(output, " r128/a16");
 383       if (mimg->d16)
 384          fprintf(output, " d16");
 385       if (mimg->disable_wqm)
 386          fprintf(output, " disable_wqm");
 387       print_barrier_reorder(mimg->can_reorder, mimg->barrier, output);
 388       break;
 389    }
 390    case Format::EXP: {
 391       const Export_instruction* exp = static_cast<const Export_instruction*>(instr);
 392       unsigned identity_mask = exp->compressed ? 0x5 : 0xf;
 393       if ((exp->enabled_mask & identity_mask) != identity_mask)
 394          fprintf(output, " en:%c%c%c%c",
 395                  exp->enabled_mask & 0x1 ? 'r' : '*',
 396                  exp->enabled_mask & 0x2 ? 'g' : '*',
 397                  exp->enabled_mask & 0x4 ? 'b' : '*',
 398                  exp->enabled_mask & 0x8 ? 'a' : '*');
 399       if (exp->compressed)
 400          fprintf(output, " compr");
 401       if (exp->done)
 402          fprintf(output, " done");
 403       if (exp->valid_mask)
 404          fprintf(output, " vm");
 405
 406       if (exp->dest <= V_008DFC_SQ_EXP_MRT + 7)
 407          fprintf(output, " mrt%d", exp->dest - V_008DFC_SQ_EXP_MRT);
 408       else if (exp->dest == V_008DFC_SQ_EXP_MRTZ)
 409          fprintf(output, " mrtz");
 410       else if (exp->dest == V_008DFC_SQ_EXP_NULL)
 411          fprintf(output, " null");
 412       else if (exp->dest >= V_008DFC_SQ_EXP_POS && exp->dest <= V_008DFC_SQ_EXP_POS + 3)
 413          fprintf(output, " pos%d", exp->dest - V_008DFC_SQ_EXP_POS);
 414       else if (exp->dest >= V_008DFC_SQ_EXP_PARAM && exp->dest <= V_008DFC_SQ_EXP_PARAM + 31)
 415          fprintf(output, " param%d", exp->dest - V_008DFC_SQ_EXP_PARAM);
 416       break;
 417    }
 418    case Format::PSEUDO_BRANCH: {
 419       const Pseudo_branch_instruction* branch = static_cast<const Pseudo_branch_instruction*>(instr);
 420       /* Note: BB0 cannot be a branch target */
 421       if (branch->target[0] != 0)
 422          fprintf(output, " BB%d", branch->target[0]);
 423       if (branch->target[1] != 0)
 424          fprintf(output, ", BB%d", branch->target[1]);
 425       break;
 426    }
 427    case Format::PSEUDO_REDUCTION: {
 428       const Pseudo_reduction_instruction* reduce = static_cast<const Pseudo_reduction_instruction*>(instr);
 429       fprintf(output, " op:%s", reduce_ops[reduce->reduce_op]);
 430       if (reduce->cluster_size)
 431          fprintf(output, " cluster_size:%u", reduce->cluster_size);
 432       break;
 433    }
 434    case Format::FLAT:
 435    case Format::GLOBAL:
 436    case Format::SCRATCH: {
 437       const FLAT_instruction* flat = static_cast<const FLAT_instruction*>(instr);
 438       if (flat->offset)
 439          fprintf(output, " offset:%u", flat->offset);
 440       if (flat->glc)
 441          fprintf(output, " glc");
 442       if (flat->dlc)
 443          fprintf(output, " dlc");
 444       if (flat->slc)
 445          fprintf(output, " slc");
 446       if (flat->lds)
 447          fprintf(output, " lds");
 448       if (flat->nv)
 449          fprintf(output, " nv");
 450       if (flat->disable_wqm)
 451          fprintf(output, " disable_wqm");
 452       print_barrier_reorder(flat->can_reorder, flat->barrier, output);
 453       break;
 454    }
 455    case Format::MTBUF: {
 456       const MTBUF_instruction* mtbuf = static_cast<const MTBUF_instruction*>(instr);
 457       fprintf(output, " dfmt:");
 458       switch (mtbuf->dfmt) {
 459       case V_008F0C_BUF_DATA_FORMAT_8: fprintf(output, "8"); break;
 460       case V_008F0C_BUF_DATA_FORMAT_16: fprintf(output, "16"); break;
 461       case V_008F0C_BUF_DATA_FORMAT_8_8: fprintf(output, "8_8"); break;
 462       case V_008F0C_BUF_DATA_FORMAT_32: fprintf(output, "32"); break;
 463       case V_008F0C_BUF_DATA_FORMAT_16_16: fprintf(output, "16_16"); break;
 464       case V_008F0C_BUF_DATA_FORMAT_10_11_11: fprintf(output, "10_11_11"); break;
 465       case V_008F0C_BUF_DATA_FORMAT_11_11_10: fprintf(output, "11_11_10"); break;
 466       case V_008F0C_BUF_DATA_FORMAT_10_10_10_2: fprintf(output, "10_10_10_2"); break;
 467       case V_008F0C_BUF_DATA_FORMAT_2_10_10_10: fprintf(output, "2_10_10_10"); break;
 468       case V_008F0C_BUF_DATA_FORMAT_8_8_8_8: fprintf(output, "8_8_8_8"); break;
 469       case V_008F0C_BUF_DATA_FORMAT_32_32: fprintf(output, "32_32"); break;
 470       case V_008F0C_BUF_DATA_FORMAT_16_16_16_16: fprintf(output, "16_16_16_16"); break;
 471       case V_008F0C_BUF_DATA_FORMAT_32_32_32: fprintf(output, "32_32_32"); break;
 472       case V_008F0C_BUF_DATA_FORMAT_32_32_32_32: fprintf(output, "32_32_32_32"); break;
 473       case V_008F0C_BUF_DATA_FORMAT_RESERVED_15: fprintf(output, "reserved15"); break;
 474       }
 475       fprintf(output, " nfmt:");
 476       switch (mtbuf->nfmt) {
 477       case V_008F0C_BUF_NUM_FORMAT_UNORM: fprintf(output, "unorm"); break;
 478       case V_008F0C_BUF_NUM_FORMAT_SNORM: fprintf(output, "snorm"); break;
 479       case V_008F0C_BUF_NUM_FORMAT_USCALED: fprintf(output, "uscaled"); break;
 480       case V_008F0C_BUF_NUM_FORMAT_SSCALED: fprintf(output, "sscaled"); break;
 481       case V_008F0C_BUF_NUM_FORMAT_UINT: fprintf(output, "uint"); break;
 482       case V_008F0C_BUF_NUM_FORMAT_SINT: fprintf(output, "sint"); break;
 483       case V_008F0C_BUF_NUM_FORMAT_SNORM_OGL: fprintf(output, "snorm"); break;
 484       case V_008F0C_BUF_NUM_FORMAT_FLOAT: fprintf(output, "float"); break;
 485       }
 486       if (mtbuf->offset)
 487          fprintf(output, " offset:%u", mtbuf->offset);
 488       if (mtbuf->offen)
 489          fprintf(output, " offen");
 490       if (mtbuf->idxen)
 491          fprintf(output, " idxen");
 492       if (mtbuf->glc)
 493          fprintf(output, " glc");
 494       if (mtbuf->dlc)
 495          fprintf(output, " dlc");
 496       if (mtbuf->slc)
 497          fprintf(output, " slc");
 498       if (mtbuf->tfe)
 499          fprintf(output, " tfe");
 500       if (mtbuf->disable_wqm)
 501          fprintf(output, " disable_wqm");
 502       print_barrier_reorder(mtbuf->can_reorder, mtbuf->barrier, output);
 503       break;
 504    }
 505    case Format::VOP3P: {
 506       if (static_cast<const VOP3P_instruction*>(instr)->clamp)
 507          fprintf(output, " clamp");
 508       break;
 509    }
 510    default: {
 511       break;
 512    }
 513    }
 514    if (instr->isVOP3()) {
 515       const VOP3A_instruction* vop3 = static_cast<const VOP3A_instruction*>(instr);
 516       switch (vop3->omod) {
 517       case 1:
 518          fprintf(output, " *2");
 519          break;
 520       case 2:
 521          fprintf(output, " *4");
 522          break;
 523       case 3:
 524          fprintf(output, " *0.5");
 525          break;
 526       }
 527       if (vop3->clamp)
 528          fprintf(output, " clamp");
 529       if (vop3->opsel & (1 << 3))
 530          fprintf(output, " opsel_hi");
 531    } else if (instr->isDPP()) {
 532       const DPP_instruction* dpp = static_cast<const DPP_instruction*>(instr);
 533       if (dpp->dpp_ctrl <= 0xff) {
 534          fprintf(output, " quad_perm:[%d,%d,%d,%d]",
 535                  dpp->dpp_ctrl & 0x3, (dpp->dpp_ctrl >> 2) & 0x3,
 536                  (dpp->dpp_ctrl >> 4) & 0x3, (dpp->dpp_ctrl >> 6) & 0x3);
 537       } else if (dpp->dpp_ctrl >= 0x101 && dpp->dpp_ctrl <= 0x10f) {
 538          fprintf(output, " row_shl:%d", dpp->dpp_ctrl & 0xf);
 539       } else if (dpp->dpp_ctrl >= 0x111 && dpp->dpp_ctrl <= 0x11f) {
 540          fprintf(output, " row_shr:%d", dpp->dpp_ctrl & 0xf);
 541       } else if (dpp->dpp_ctrl >= 0x121 && dpp->dpp_ctrl <= 0x12f) {
 542          fprintf(output, " row_ror:%d", dpp->dpp_ctrl & 0xf);
 543       } else if (dpp->dpp_ctrl == dpp_wf_sl1) {
 544          fprintf(output, " wave_shl:1");
 545       } else if (dpp->dpp_ctrl == dpp_wf_rl1) {
 546          fprintf(output, " wave_rol:1");
 547       } else if (dpp->dpp_ctrl == dpp_wf_sr1) {
 548          fprintf(output, " wave_shr:1");
 549       } else if (dpp->dpp_ctrl == dpp_wf_rr1) {
 550          fprintf(output, " wave_ror:1");
 551       } else if (dpp->dpp_ctrl == dpp_row_mirror) {
 552          fprintf(output, " row_mirror");
 553       } else if (dpp->dpp_ctrl == dpp_row_half_mirror) {
 554          fprintf(output, " row_half_mirror");
 555       } else if (dpp->dpp_ctrl == dpp_row_bcast15) {
 556          fprintf(output, " row_bcast:15");
 557       } else if (dpp->dpp_ctrl == dpp_row_bcast31) {
 558          fprintf(output, " row_bcast:31");
 559       } else {
 560          fprintf(output, " dpp_ctrl:0x%.3x", dpp->dpp_ctrl);
 561       }
 562       if (dpp->row_mask != 0xf)
 563          fprintf(output, " row_mask:0x%.1x", dpp->row_mask);
 564       if (dpp->bank_mask != 0xf)
 565          fprintf(output, " bank_mask:0x%.1x", dpp->bank_mask);
 566       if (dpp->bound_ctrl)
 567          fprintf(output, " bound_ctrl:1");
 568    } else if ((int)instr->format & (int)Format::SDWA) {
 569       const SDWA_instruction* sdwa = static_cast<const SDWA_instruction*>(instr);
 570       switch (sdwa->omod) {
 571       case 1:
 572          fprintf(output, " *2");
 573          break;
 574       case 2:
 575          fprintf(output, " *4");
 576          break;
 577       case 3:
 578          fprintf(output, " *0.5");
 579          break;
 580       }
 581       if (sdwa->clamp)
 582          fprintf(output, " clamp");
 583       switch (sdwa->dst_sel & sdwa_asuint) {
 584       case sdwa_udword:
 585          break;
 586       case sdwa_ubyte0:
 587       case sdwa_ubyte1:
 588       case sdwa_ubyte2:
 589       case sdwa_ubyte3:
 590          fprintf(output, " dst_sel:%sbyte%u", sdwa->dst_sel & sdwa_sext ? "s" : "u",
 591                  sdwa->dst_sel & sdwa_bytenum);
 592          break;
 593       case sdwa_uword0:
 594       case sdwa_uword1:
 595          fprintf(output, " dst_sel:%sword%u", sdwa->dst_sel & sdwa_sext ? "s" : "u",
 596                  sdwa->dst_sel & sdwa_wordnum);
 597          break;
 598       }
 599       if (sdwa->dst_preserve)
 600          fprintf(output, " dst_preserve");
 601    }
 602 }
 603
 604 void aco_print_instr(const Instruction *instr, FILE *output)
 605 {
 606    if (!instr->definitions.empty()) {
 607       for (unsigned i = 0; i < instr->definitions.size(); ++i) {
 608          print_definition(&instr->definitions[i], output);
 609          if (i + 1 != instr->definitions.size())
 610             fprintf(output, ", ");
 611       }
 612       fprintf(output, " = ");
 613    }
 614    fprintf(output, "%s", instr_info.name[(int)instr->opcode]);
 615    if (instr->operands.size()) {
 616       bool abs[instr->operands.size()];
 617       bool neg[instr->operands.size()];
 618       bool opsel[instr->operands.size()];
 619       uint8_t sel[instr->operands.size()];
 620       if ((int)instr->format & (int)Format::VOP3A) {
 621          const VOP3A_instruction* vop3 = static_cast<const VOP3A_instruction*>(instr);
 622          for (unsigned i = 0; i < instr->operands.size(); ++i) {
 623             abs[i] = vop3->abs[i];
 624             neg[i] = vop3->neg[i];
 625             opsel[i] = vop3->opsel & (1 << i);
 626             sel[i] = sdwa_udword;
 627          }
 628       } else if (instr->isDPP()) {
 629          const DPP_instruction* dpp = static_cast<const DPP_instruction*>(instr);
 630          for (unsigned i = 0; i < instr->operands.size(); ++i) {
 631             abs[i] = i < 2 ? dpp->abs[i] : false;
 632             neg[i] = i < 2 ? dpp->neg[i] : false;
 633             opsel[i] = false;
 634             sel[i] = sdwa_udword;
 635          }
 636       } else if (instr->isSDWA()) {
 637          const SDWA_instruction* sdwa = static_cast<const SDWA_instruction*>(instr);
 638          for (unsigned i = 0; i < instr->operands.size(); ++i) {
 639             abs[i] = i < 2 ? sdwa->abs[i] : false;
 640             neg[i] = i < 2 ? sdwa->neg[i] : false;
 641             opsel[i] = false;
 642             sel[i] = i < 2 ? sdwa->sel[i] : sdwa_udword;
 643          }
 644       } else {
 645          for (unsigned i = 0; i < instr->operands.size(); ++i) {
 646             abs[i] = false;
 647             neg[i] = false;
 648             opsel[i] = false;
 649             sel[i] = sdwa_udword;
 650          }
 651       }
 652       for (unsigned i = 0; i < instr->operands.size(); ++i) {
 653          if (i)
 654             fprintf(output, ", ");
 655          else
 656             fprintf(output, " ");
 657
 658          if (neg[i])
 659             fprintf(output, "-");
 660          if (abs[i])
 661             fprintf(output, "|");
 662          if (opsel[i])
 663             fprintf(output, "hi(");
 664          else if (sel[i] & sdwa_sext)
 665             fprintf(output, "sext(");
 666          print_operand(&instr->operands[i], output);
 667          if (opsel[i] || (sel[i] & sdwa_sext))
 668             fprintf(output, ")");
 669          if (!(sel[i] & sdwa_isra)) {
 670             if (sel[i] & sdwa_udword) {
 671                /* print nothing */
 672             } else if (sel[i] & sdwa_isword) {
 673                unsigned index = sel[i] & sdwa_wordnum;
 674                fprintf(output, "[%u:%u]", index * 16, index * 16 + 15);
 675             } else {
 676                unsigned index = sel[i] & sdwa_bytenum;
 677                fprintf(output, "[%u:%u]", index * 8, index * 8 + 7);
 678             }
 679          }
 680          if (abs[i])
 681             fprintf(output, "|");
 682
 683          if (instr->format == Format::VOP3P) {
 684             const VOP3P_instruction* vop3 = static_cast<const VOP3P_instruction*>(instr);
 685             if ((vop3->opsel_lo & (1 << i)) || !(vop3->opsel_hi & (1 << i))) {
 686                fprintf(output, ".%c%c",
 687                        vop3->opsel_lo & (1 << i) ? 'y' : 'x',
 688                        vop3->opsel_hi & (1 << i) ? 'y' : 'x');
 689             }
 690             if (vop3->neg_lo[i] && vop3->neg_hi[i])
 691                fprintf(output, "*[-1,-1]");
 692             else if (vop3->neg_lo[i])
 693                fprintf(output, "*[-1,1]");
 694             else if (vop3->neg_hi[i])
 695                fprintf(output, "*[1,-1]");
 696          }
 697       }
 698    }
 699    print_instr_format_specific(instr, output);
 700 }
 701
 702 static void print_block_kind(uint16_t kind, FILE *output)
 703 {
 704    if (kind & block_kind_uniform)
 705       fprintf(output, "uniform, ");
 706    if (kind & block_kind_top_level)
 707       fprintf(output, "top-level, ");
 708    if (kind & block_kind_loop_preheader)
 709       fprintf(output, "loop-preheader, ");
 710    if (kind & block_kind_loop_header)
 711       fprintf(output, "loop-header, ");
 712    if (kind & block_kind_loop_exit)
 713       fprintf(output, "loop-exit, ");
 714    if (kind & block_kind_continue)
 715       fprintf(output, "continue, ");
 716    if (kind & block_kind_break)
 717       fprintf(output, "break, ");
 718    if (kind & block_kind_continue_or_break)
 719       fprintf(output, "continue_or_break, ");
 720    if (kind & block_kind_discard)
 721       fprintf(output, "discard, ");
 722    if (kind & block_kind_branch)
 723       fprintf(output, "branch, ");
 724    if (kind & block_kind_merge)
 725       fprintf(output, "merge, ");
 726    if (kind & block_kind_invert)
 727       fprintf(output, "invert, ");
 728    if (kind & block_kind_uses_discard_if)
 729       fprintf(output, "discard_if, ");
 730    if (kind & block_kind_needs_lowering)
 731       fprintf(output, "needs_lowering, ");
 732    if (kind & block_kind_uses_demote)
 733       fprintf(output, "uses_demote, ");
 734    if (kind & block_kind_export_end)
 735       fprintf(output, "export_end, ");
 736 }
 737
 738 static void print_stage(Stage stage, FILE *output)
 739 {
 740    fprintf(output, "ACO shader stage: ");
 741
 742    if (stage == compute_cs)
 743       fprintf(output, "compute_cs");
 744    else if (stage == fragment_fs)
 745       fprintf(output, "fragment_fs");
 746    else if (stage == gs_copy_vs)
 747       fprintf(output, "gs_copy_vs");
 748    else if (stage == vertex_ls)
 749       fprintf(output, "vertex_ls");
 750    else if (stage == vertex_es)
 751       fprintf(output, "vertex_es");
 752    else if (stage == vertex_vs)
 753       fprintf(output, "vertex_vs");
 754    else if (stage == tess_control_hs)
 755       fprintf(output, "tess_control_hs");
 756    else if (stage == vertex_tess_control_hs)
 757       fprintf(output, "vertex_tess_control_hs");
 758    else if (stage == tess_eval_es)
 759       fprintf(output, "tess_eval_es");
 760    else if (stage == tess_eval_vs)
 761       fprintf(output, "tess_eval_vs");
 762    else if (stage == geometry_gs)
 763       fprintf(output, "geometry_gs");
 764    else if (stage == vertex_geometry_gs)
 765       fprintf(output, "vertex_geometry_gs");
 766    else if (stage == tess_eval_geometry_gs)
 767       fprintf(output, "tess_eval_geometry_gs");
 768    else if (stage == ngg_vertex_gs)
 769       fprintf(output, "ngg_vertex_gs");
 770    else if (stage == ngg_tess_eval_gs)
 771       fprintf(output, "ngg_tess_eval_gs");
 772    else if (stage == ngg_vertex_geometry_gs)
 773       fprintf(output, "ngg_vertex_geometry_gs");
 774    else if (stage == ngg_tess_eval_geometry_gs)
 775       fprintf(output, "ngg_tess_eval_geometry_gs");
 776    else
 777       fprintf(output, "unknown");
 778
 779    fprintf(output, "\n");
 780 }
 781
 782 void aco_print_block(const Block* block, FILE *output)
 783 {
 784    fprintf(output, "BB%d\n", block->index);
 785    fprintf(output, "/* logical preds: ");
 786    for (unsigned pred : block->logical_preds)
 787       fprintf(output, "BB%d, ", pred);
 788    fprintf(output, "/ linear preds: ");
 789    for (unsigned pred : block->linear_preds)
 790       fprintf(output, "BB%d, ", pred);
 791    fprintf(output, "/ kind: ");
 792    print_block_kind(block->kind, output);
 793    fprintf(output, "*/\n");
 794    for (auto const& instr : block->instructions) {
 795       fprintf(output, "\t");
 796       aco_print_instr(instr.get(), output);
 797       fprintf(output, "\n");
 798    }
 799 }
 800
 801 void aco_print_program(const Program *program, FILE *output)
 802 {
 803    print_stage(program->stage, output);
 804
 805    for (Block const& block : program->blocks)
 806       aco_print_block(&block, output);
 807
 808    if (program->constant_data.size()) {
 809       fprintf(output, "\n/* constant data */\n");
 810       for (unsigned i = 0; i < program->constant_data.size(); i += 32) {
 811          fprintf(output, "[%06d] ", i);
 812          unsigned line_size = std::min<size_t>(program->constant_data.size() - i, 32);
 813          for (unsigned j = 0; j < line_size; j += 4) {
 814             unsigned size = std::min<size_t>(program->constant_data.size() - (i + j), 4);
 815             uint32_t v = 0;
 816             memcpy(&v, &program->constant_data[i + j], size);
 817             fprintf(output, " %08x", v);
 818          }
 819          fprintf(output, "\n");
 820       }
 821    }
 822
 823    fprintf(output, "\n");
 824 }
 825
 826 }