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 };
  59
  60 static void print_reg_class(const RegClass rc, FILE *output)
  61 {
  62    switch (rc) {
  63       case RegClass::s1: fprintf(output, " s1: "); return;
  64       case RegClass::s2: fprintf(output, " s2: "); return;
  65       case RegClass::s3: fprintf(output, " s3: "); return;
  66       case RegClass::s4: fprintf(output, " s4: "); return;
  67       case RegClass::s6: fprintf(output, " s6: "); return;
  68       case RegClass::s8: fprintf(output, " s8: "); return;
  69       case RegClass::s16: fprintf(output, "s16: "); return;
  70       case RegClass::v1: fprintf(output, " v1: "); return;
  71       case RegClass::v2: fprintf(output, " v2: "); return;
  72       case RegClass::v3: fprintf(output, " v3: "); return;
  73       case RegClass::v4: fprintf(output, " v4: "); return;
  74       case RegClass::v5: fprintf(output, " v5: "); return;
  75       case RegClass::v6: fprintf(output, " v6: "); return;
  76       case RegClass::v7: fprintf(output, " v7: "); return;
  77       case RegClass::v8: fprintf(output, " v8: "); return;
  78       case RegClass::v1b: fprintf(output, " v1b: "); return;
  79       case RegClass::v2b: fprintf(output, " v2b: "); return;
  80       case RegClass::v3b: fprintf(output, " v3b: "); return;
  81       case RegClass::v4b: fprintf(output, " v4b: "); return;
  82       case RegClass::v6b: fprintf(output, " v6b: "); return;
  83       case RegClass::v8b: fprintf(output, " v8b: "); return;
  84       case RegClass::v1_linear: fprintf(output, " v1: "); return;
  85       case RegClass::v2_linear: fprintf(output, " v2: "); return;
  86    }
  87 }
  88
  89 void print_physReg(PhysReg reg, unsigned bytes, FILE *output)
  90 {
  91    if (reg == 124) {
  92       fprintf(output, ":m0");
  93    } else if (reg == 106) {
  94       fprintf(output, ":vcc");
  95    } else if (reg == 253) {
  96       fprintf(output, ":scc");
  97    } else if (reg == 126) {
  98       fprintf(output, ":exec");
  99    } else {
 100       bool is_vgpr = reg / 256;
 101       unsigned r = reg % 256;
 102       unsigned size = DIV_ROUND_UP(bytes, 4);
 103       fprintf(output, ":%c[%d", is_vgpr ? 'v' : 's', r);
 104       if (size > 1)
 105          fprintf(output, "-%d]", r + size -1);
 106       else
 107          fprintf(output, "]");
 108       if (reg.byte() || bytes % 4)
 109          fprintf(output, "[%d:%d]", reg.byte()*8, (reg.byte()+bytes) * 8);
 110    }
 111 }
 112
 113 static void print_constant(uint8_t reg, FILE *output)
 114 {
 115    if (reg >= 128 && reg <= 192) {
 116       fprintf(output, "%d", reg - 128);
 117       return;
 118    } else if (reg >= 192 && reg <= 208) {
 119       fprintf(output, "%d", 192 - reg);
 120       return;
 121    }
 122
 123    switch (reg) {
 124    case 240:
 125       fprintf(output, "0.5");
 126       break;
 127    case 241:
 128       fprintf(output, "-0.5");
 129       break;
 130    case 242:
 131       fprintf(output, "1.0");
 132       break;
 133    case 243:
 134       fprintf(output, "-1.0");
 135       break;
 136    case 244:
 137       fprintf(output, "2.0");
 138       break;
 139    case 245:
 140       fprintf(output, "-2.0");
 141       break;
 142    case 246:
 143       fprintf(output, "4.0");
 144       break;
 145    case 247:
 146       fprintf(output, "-4.0");
 147       break;
 148    case 248:
 149       fprintf(output, "1/(2*PI)");
 150       break;
 151    }
 152 }
 153
 154 static void print_operand(const Operand *operand, FILE *output)
 155 {
 156    if (operand->isLiteral()) {
 157       fprintf(output, "0x%x", operand->constantValue());
 158    } else if (operand->isConstant()) {
 159       print_constant(operand->physReg().reg(), output);
 160    } else if (operand->isUndefined()) {
 161       print_reg_class(operand->regClass(), output);
 162       fprintf(output, "undef");
 163    } else {
 164       if (operand->isLateKill())
 165          fprintf(output, "(latekill)");
 166
 167       fprintf(output, "%%%d", operand->tempId());
 168
 169       if (operand->isFixed())
 170          print_physReg(operand->physReg(), operand->bytes(), output);
 171    }
 172 }
 173
 174 static void print_definition(const Definition *definition, FILE *output)
 175 {
 176    print_reg_class(definition->regClass(), output);
 177    if (definition->isPrecise())
 178       fprintf(output, "(precise)");
 179    fprintf(output, "%%%d", definition->tempId());
 180
 181    if (definition->isFixed())
 182       print_physReg(definition->physReg(), definition->bytes(), output);
 183 }
 184
 185 static void print_barrier_reorder(bool can_reorder, barrier_interaction barrier, FILE *output)
 186 {
 187    if (can_reorder)
 188       fprintf(output, " reorder");
 189
 190    if (barrier & barrier_buffer)
 191       fprintf(output, " buffer");
 192    if (barrier & barrier_image)
 193       fprintf(output, " image");
 194    if (barrier & barrier_atomic)
 195       fprintf(output, " atomic");
 196    if (barrier & barrier_shared)
 197       fprintf(output, " shared");
 198    if (barrier & barrier_gs_data)
 199       fprintf(output, " gs_data");
 200    if (barrier & barrier_gs_sendmsg)
 201       fprintf(output, " gs_sendmsg");
 202 }
 203
 204 static void print_instr_format_specific(const Instruction *instr, FILE *output)
 205 {
 206    switch (instr->format) {
 207    case Format::SOPK: {
 208       const SOPK_instruction* sopk = static_cast<const SOPK_instruction*>(instr);
 209       fprintf(output, " imm:%d", sopk->imm & 0x8000 ? (sopk->imm - 65536) : sopk->imm);
 210       break;
 211    }
 212    case Format::SOPP: {
 213       const SOPP_instruction* sopp = static_cast<const SOPP_instruction*>(instr);
 214       uint16_t imm = sopp->imm;
 215       switch (instr->opcode) {
 216       case aco_opcode::s_waitcnt: {
 217          /* we usually should check the chip class for vmcnt/lgkm, but
 218           * insert_waitcnt() should fill it in regardless. */
 219          unsigned vmcnt = (imm & 0xF) | ((imm & (0x3 << 14)) >> 10);
 220          if (vmcnt != 63) fprintf(output, " vmcnt(%d)", vmcnt);
 221          if (((imm >> 4) & 0x7) < 0x7) fprintf(output, " expcnt(%d)", (imm >> 4) & 0x7);
 222          if (((imm >> 8) & 0x3F) < 0x3F) fprintf(output, " lgkmcnt(%d)", (imm >> 8) & 0x3F);
 223          break;
 224       }
 225       case aco_opcode::s_endpgm:
 226       case aco_opcode::s_endpgm_saved:
 227       case aco_opcode::s_endpgm_ordered_ps_done:
 228       case aco_opcode::s_wakeup:
 229       case aco_opcode::s_barrier:
 230       case aco_opcode::s_icache_inv:
 231       case aco_opcode::s_ttracedata:
 232       case aco_opcode::s_set_gpr_idx_off: {
 233          break;
 234       }
 235       case aco_opcode::s_sendmsg: {
 236          unsigned id = imm & sendmsg_id_mask;
 237          switch (id) {
 238          case sendmsg_none:
 239             fprintf(output, " sendmsg(MSG_NONE)");
 240             break;
 241          case _sendmsg_gs:
 242             fprintf(output, " sendmsg(gs%s%s, %u)",
 243                     imm & 0x10 ? ", cut" : "", imm & 0x20 ? ", emit" : "", imm >> 8);
 244             break;
 245          case _sendmsg_gs_done:
 246             fprintf(output, " sendmsg(gs_done%s%s, %u)",
 247                     imm & 0x10 ? ", cut" : "", imm & 0x20 ? ", emit" : "", imm >> 8);
 248             break;
 249          case sendmsg_save_wave:
 250             fprintf(output, " sendmsg(save_wave)");
 251             break;
 252          case sendmsg_stall_wave_gen:
 253             fprintf(output, " sendmsg(stall_wave_gen)");
 254             break;
 255          case sendmsg_halt_waves:
 256             fprintf(output, " sendmsg(halt_waves)");
 257             break;
 258          case sendmsg_ordered_ps_done:
 259             fprintf(output, " sendmsg(ordered_ps_done)");
 260             break;
 261          case sendmsg_early_prim_dealloc:
 262             fprintf(output, " sendmsg(early_prim_dealloc)");
 263             break;
 264          case sendmsg_gs_alloc_req:
 265             fprintf(output, " sendmsg(gs_alloc_req)");
 266             break;
 267          }
 268          break;
 269       }
 270       default: {
 271          if (imm)
 272             fprintf(output, " imm:%u", imm);
 273          break;
 274       }
 275       }
 276       if (sopp->block != -1)
 277          fprintf(output, " block:BB%d", sopp->block);
 278       break;
 279    }
 280    case Format::SMEM: {
 281       const SMEM_instruction* smem = static_cast<const SMEM_instruction*>(instr);
 282       if (smem->glc)
 283          fprintf(output, " glc");
 284       if (smem->dlc)
 285          fprintf(output, " dlc");
 286       if (smem->nv)
 287          fprintf(output, " nv");
 288       print_barrier_reorder(smem->can_reorder, smem->barrier, output);
 289       break;
 290    }
 291    case Format::VINTRP: {
 292       const Interp_instruction* vintrp = static_cast<const Interp_instruction*>(instr);
 293       fprintf(output, " attr%d.%c", vintrp->attribute, "xyzw"[vintrp->component]);
 294       break;
 295    }
 296    case Format::DS: {
 297       const DS_instruction* ds = static_cast<const DS_instruction*>(instr);
 298       if (ds->offset0)
 299          fprintf(output, " offset0:%u", ds->offset0);
 300       if (ds->offset1)
 301          fprintf(output, " offset1:%u", ds->offset1);
 302       if (ds->gds)
 303          fprintf(output, " gds");
 304       break;
 305    }
 306    case Format::MUBUF: {
 307       const MUBUF_instruction* mubuf = static_cast<const MUBUF_instruction*>(instr);
 308       if (mubuf->offset)
 309          fprintf(output, " offset:%u", mubuf->offset);
 310       if (mubuf->offen)
 311          fprintf(output, " offen");
 312       if (mubuf->idxen)
 313          fprintf(output, " idxen");
 314       if (mubuf->addr64)
 315          fprintf(output, " addr64");
 316       if (mubuf->glc)
 317          fprintf(output, " glc");
 318       if (mubuf->dlc)
 319          fprintf(output, " dlc");
 320       if (mubuf->slc)
 321          fprintf(output, " slc");
 322       if (mubuf->tfe)
 323          fprintf(output, " tfe");
 324       if (mubuf->lds)
 325          fprintf(output, " lds");
 326       if (mubuf->disable_wqm)
 327          fprintf(output, " disable_wqm");
 328       print_barrier_reorder(mubuf->can_reorder, mubuf->barrier, output);
 329       break;
 330    }
 331    case Format::MIMG: {
 332       const MIMG_instruction* mimg = static_cast<const MIMG_instruction*>(instr);
 333       unsigned identity_dmask = !instr->definitions.empty() ?
 334                                 (1 << instr->definitions[0].size()) - 1 :
 335                                 0xf;
 336       if ((mimg->dmask & identity_dmask) != identity_dmask)
 337          fprintf(output, " dmask:%s%s%s%s",
 338                  mimg->dmask & 0x1 ? "x" : "",
 339                  mimg->dmask & 0x2 ? "y" : "",
 340                  mimg->dmask & 0x4 ? "z" : "",
 341                  mimg->dmask & 0x8 ? "w" : "");
 342       switch (mimg->dim) {
 343       case ac_image_1d:
 344          fprintf(output, " 1d");
 345          break;
 346       case ac_image_2d:
 347          fprintf(output, " 2d");
 348          break;
 349       case ac_image_3d:
 350          fprintf(output, " 3d");
 351          break;
 352       case ac_image_cube:
 353          fprintf(output, " cube");
 354          break;
 355       case ac_image_1darray:
 356          fprintf(output, " 1darray");
 357          break;
 358       case ac_image_2darray:
 359          fprintf(output, " 2darray");
 360          break;
 361       case ac_image_2dmsaa:
 362          fprintf(output, " 2dmsaa");
 363          break;
 364       case ac_image_2darraymsaa:
 365          fprintf(output, " 2darraymsaa");
 366          break;
 367       }
 368       if (mimg->unrm)
 369          fprintf(output, " unrm");
 370       if (mimg->glc)
 371          fprintf(output, " glc");
 372       if (mimg->dlc)
 373          fprintf(output, " dlc");
 374       if (mimg->slc)
 375          fprintf(output, " slc");
 376       if (mimg->tfe)
 377          fprintf(output, " tfe");
 378       if (mimg->da)
 379          fprintf(output, " da");
 380       if (mimg->lwe)
 381          fprintf(output, " lwe");
 382       if (mimg->r128 || mimg->a16)
 383          fprintf(output, " r128/a16");
 384       if (mimg->d16)
 385          fprintf(output, " d16");
 386       if (mimg->disable_wqm)
 387          fprintf(output, " disable_wqm");
 388       print_barrier_reorder(mimg->can_reorder, mimg->barrier, output);
 389       break;
 390    }
 391    case Format::EXP: {
 392       const Export_instruction* exp = static_cast<const Export_instruction*>(instr);
 393       unsigned identity_mask = exp->compressed ? 0x5 : 0xf;
 394       if ((exp->enabled_mask & identity_mask) != identity_mask)
 395          fprintf(output, " en:%c%c%c%c",
 396                  exp->enabled_mask & 0x1 ? 'r' : '*',
 397                  exp->enabled_mask & 0x2 ? 'g' : '*',
 398                  exp->enabled_mask & 0x4 ? 'b' : '*',
 399                  exp->enabled_mask & 0x8 ? 'a' : '*');
 400       if (exp->compressed)
 401          fprintf(output, " compr");
 402       if (exp->done)
 403          fprintf(output, " done");
 404       if (exp->valid_mask)
 405          fprintf(output, " vm");
 406
 407       if (exp->dest <= V_008DFC_SQ_EXP_MRT + 7)
 408          fprintf(output, " mrt%d", exp->dest - V_008DFC_SQ_EXP_MRT);
 409       else if (exp->dest == V_008DFC_SQ_EXP_MRTZ)
 410          fprintf(output, " mrtz");
 411       else if (exp->dest == V_008DFC_SQ_EXP_NULL)
 412          fprintf(output, " null");
 413       else if (exp->dest >= V_008DFC_SQ_EXP_POS && exp->dest <= V_008DFC_SQ_EXP_POS + 3)
 414          fprintf(output, " pos%d", exp->dest - V_008DFC_SQ_EXP_POS);
 415       else if (exp->dest >= V_008DFC_SQ_EXP_PARAM && exp->dest <= V_008DFC_SQ_EXP_PARAM + 31)
 416          fprintf(output, " param%d", exp->dest - V_008DFC_SQ_EXP_PARAM);
 417       break;
 418    }
 419    case Format::PSEUDO_BRANCH: {
 420       const Pseudo_branch_instruction* branch = static_cast<const Pseudo_branch_instruction*>(instr);
 421       /* Note: BB0 cannot be a branch target */
 422       if (branch->target[0] != 0)
 423          fprintf(output, " BB%d", branch->target[0]);
 424       if (branch->target[1] != 0)
 425          fprintf(output, ", BB%d", branch->target[1]);
 426       break;
 427    }
 428    case Format::PSEUDO_REDUCTION: {
 429       const Pseudo_reduction_instruction* reduce = static_cast<const Pseudo_reduction_instruction*>(instr);
 430       fprintf(output, " op:%s", reduce_ops[reduce->reduce_op]);
 431       if (reduce->cluster_size)
 432          fprintf(output, " cluster_size:%u", reduce->cluster_size);
 433       break;
 434    }
 435    case Format::FLAT:
 436    case Format::GLOBAL:
 437    case Format::SCRATCH: {
 438       const FLAT_instruction* flat = static_cast<const FLAT_instruction*>(instr);
 439       if (flat->offset)
 440          fprintf(output, " offset:%u", flat->offset);
 441       if (flat->glc)
 442          fprintf(output, " glc");
 443       if (flat->dlc)
 444          fprintf(output, " dlc");
 445       if (flat->slc)
 446          fprintf(output, " slc");
 447       if (flat->lds)
 448          fprintf(output, " lds");
 449       if (flat->nv)
 450          fprintf(output, " nv");
 451       if (flat->disable_wqm)
 452          fprintf(output, " disable_wqm");
 453       print_barrier_reorder(flat->can_reorder, flat->barrier, output);
 454       break;
 455    }
 456    case Format::MTBUF: {
 457       const MTBUF_instruction* mtbuf = static_cast<const MTBUF_instruction*>(instr);
 458       fprintf(output, " dfmt:");
 459       switch (mtbuf->dfmt) {
 460       case V_008F0C_BUF_DATA_FORMAT_8: fprintf(output, "8"); break;
 461       case V_008F0C_BUF_DATA_FORMAT_16: fprintf(output, "16"); break;
 462       case V_008F0C_BUF_DATA_FORMAT_8_8: fprintf(output, "8_8"); break;
 463       case V_008F0C_BUF_DATA_FORMAT_32: fprintf(output, "32"); break;
 464       case V_008F0C_BUF_DATA_FORMAT_16_16: fprintf(output, "16_16"); break;
 465       case V_008F0C_BUF_DATA_FORMAT_10_11_11: fprintf(output, "10_11_11"); break;
 466       case V_008F0C_BUF_DATA_FORMAT_11_11_10: fprintf(output, "11_11_10"); break;
 467       case V_008F0C_BUF_DATA_FORMAT_10_10_10_2: fprintf(output, "10_10_10_2"); break;
 468       case V_008F0C_BUF_DATA_FORMAT_2_10_10_10: fprintf(output, "2_10_10_10"); break;
 469       case V_008F0C_BUF_DATA_FORMAT_8_8_8_8: fprintf(output, "8_8_8_8"); break;
 470       case V_008F0C_BUF_DATA_FORMAT_32_32: fprintf(output, "32_32"); break;
 471       case V_008F0C_BUF_DATA_FORMAT_16_16_16_16: fprintf(output, "16_16_16_16"); break;
 472       case V_008F0C_BUF_DATA_FORMAT_32_32_32: fprintf(output, "32_32_32"); break;
 473       case V_008F0C_BUF_DATA_FORMAT_32_32_32_32: fprintf(output, "32_32_32_32"); break;
 474       case V_008F0C_BUF_DATA_FORMAT_RESERVED_15: fprintf(output, "reserved15"); break;
 475       }
 476       fprintf(output, " nfmt:");
 477       switch (mtbuf->nfmt) {
 478       case V_008F0C_BUF_NUM_FORMAT_UNORM: fprintf(output, "unorm"); break;
 479       case V_008F0C_BUF_NUM_FORMAT_SNORM: fprintf(output, "snorm"); break;
 480       case V_008F0C_BUF_NUM_FORMAT_USCALED: fprintf(output, "uscaled"); break;
 481       case V_008F0C_BUF_NUM_FORMAT_SSCALED: fprintf(output, "sscaled"); break;
 482       case V_008F0C_BUF_NUM_FORMAT_UINT: fprintf(output, "uint"); break;
 483       case V_008F0C_BUF_NUM_FORMAT_SINT: fprintf(output, "sint"); break;
 484       case V_008F0C_BUF_NUM_FORMAT_SNORM_OGL: fprintf(output, "snorm"); break;
 485       case V_008F0C_BUF_NUM_FORMAT_FLOAT: fprintf(output, "float"); break;
 486       }
 487       if (mtbuf->offset)
 488          fprintf(output, " offset:%u", mtbuf->offset);
 489       if (mtbuf->offen)
 490          fprintf(output, " offen");
 491       if (mtbuf->idxen)
 492          fprintf(output, " idxen");
 493       if (mtbuf->glc)
 494          fprintf(output, " glc");
 495       if (mtbuf->dlc)
 496          fprintf(output, " dlc");
 497       if (mtbuf->slc)
 498          fprintf(output, " slc");
 499       if (mtbuf->tfe)
 500          fprintf(output, " tfe");
 501       if (mtbuf->disable_wqm)
 502          fprintf(output, " disable_wqm");
 503       print_barrier_reorder(mtbuf->can_reorder, mtbuf->barrier, output);
 504       break;
 505    }
 506    case Format::VOP3P: {
 507       if (static_cast<const VOP3P_instruction*>(instr)->clamp)
 508          fprintf(output, " clamp");
 509       break;
 510    }
 511    default: {
 512       break;
 513    }
 514    }
 515    if (instr->isVOP3()) {
 516       const VOP3A_instruction* vop3 = static_cast<const VOP3A_instruction*>(instr);
 517       switch (vop3->omod) {
 518       case 1:
 519          fprintf(output, " *2");
 520          break;
 521       case 2:
 522          fprintf(output, " *4");
 523          break;
 524       case 3:
 525          fprintf(output, " *0.5");
 526          break;
 527       }
 528       if (vop3->clamp)
 529          fprintf(output, " clamp");
 530       if (vop3->opsel & (1 << 3))
 531          fprintf(output, " opsel_hi");
 532    } else if (instr->isDPP()) {
 533       const DPP_instruction* dpp = static_cast<const DPP_instruction*>(instr);
 534       if (dpp->dpp_ctrl <= 0xff) {
 535          fprintf(output, " quad_perm:[%d,%d,%d,%d]",
 536                  dpp->dpp_ctrl & 0x3, (dpp->dpp_ctrl >> 2) & 0x3,
 537                  (dpp->dpp_ctrl >> 4) & 0x3, (dpp->dpp_ctrl >> 6) & 0x3);
 538       } else if (dpp->dpp_ctrl >= 0x101 && dpp->dpp_ctrl <= 0x10f) {
 539          fprintf(output, " row_shl:%d", dpp->dpp_ctrl & 0xf);
 540       } else if (dpp->dpp_ctrl >= 0x111 && dpp->dpp_ctrl <= 0x11f) {
 541          fprintf(output, " row_shr:%d", dpp->dpp_ctrl & 0xf);
 542       } else if (dpp->dpp_ctrl >= 0x121 && dpp->dpp_ctrl <= 0x12f) {
 543          fprintf(output, " row_ror:%d", dpp->dpp_ctrl & 0xf);
 544       } else if (dpp->dpp_ctrl == dpp_wf_sl1) {
 545          fprintf(output, " wave_shl:1");
 546       } else if (dpp->dpp_ctrl == dpp_wf_rl1) {
 547          fprintf(output, " wave_rol:1");
 548       } else if (dpp->dpp_ctrl == dpp_wf_sr1) {
 549          fprintf(output, " wave_shr:1");
 550       } else if (dpp->dpp_ctrl == dpp_wf_rr1) {
 551          fprintf(output, " wave_ror:1");
 552       } else if (dpp->dpp_ctrl == dpp_row_mirror) {
 553          fprintf(output, " row_mirror");
 554       } else if (dpp->dpp_ctrl == dpp_row_half_mirror) {
 555          fprintf(output, " row_half_mirror");
 556       } else if (dpp->dpp_ctrl == dpp_row_bcast15) {
 557          fprintf(output, " row_bcast:15");
 558       } else if (dpp->dpp_ctrl == dpp_row_bcast31) {
 559          fprintf(output, " row_bcast:31");
 560       } else {
 561          fprintf(output, " dpp_ctrl:0x%.3x", dpp->dpp_ctrl);
 562       }
 563       if (dpp->row_mask != 0xf)
 564          fprintf(output, " row_mask:0x%.1x", dpp->row_mask);
 565       if (dpp->bank_mask != 0xf)
 566          fprintf(output, " bank_mask:0x%.1x", dpp->bank_mask);
 567       if (dpp->bound_ctrl)
 568          fprintf(output, " bound_ctrl:1");
 569    } else if ((int)instr->format & (int)Format::SDWA) {
 570       const SDWA_instruction* sdwa = static_cast<const SDWA_instruction*>(instr);
 571       switch (sdwa->omod) {
 572       case 1:
 573          fprintf(output, " *2");
 574          break;
 575       case 2:
 576          fprintf(output, " *4");
 577          break;
 578       case 3:
 579          fprintf(output, " *0.5");
 580          break;
 581       }
 582       if (sdwa->clamp)
 583          fprintf(output, " clamp");
 584       switch (sdwa->dst_sel & sdwa_asuint) {
 585       case sdwa_udword:
 586          break;
 587       case sdwa_ubyte0:
 588       case sdwa_ubyte1:
 589       case sdwa_ubyte2:
 590       case sdwa_ubyte3:
 591          fprintf(output, " dst_sel:%sbyte%u", sdwa->dst_sel & sdwa_sext ? "s" : "u",
 592                  sdwa->dst_sel & sdwa_bytenum);
 593          break;
 594       case sdwa_uword0:
 595       case sdwa_uword1:
 596          fprintf(output, " dst_sel:%sword%u", sdwa->dst_sel & sdwa_sext ? "s" : "u",
 597                  sdwa->dst_sel & sdwa_wordnum);
 598          break;
 599       }
 600       if (sdwa->dst_preserve)
 601          fprintf(output, " dst_preserve");
 602    }
 603 }
 604
 605 void aco_print_instr(const Instruction *instr, FILE *output)
 606 {
 607    if (!instr->definitions.empty()) {
 608       for (unsigned i = 0; i < instr->definitions.size(); ++i) {
 609          print_definition(&instr->definitions[i], output);
 610          if (i + 1 != instr->definitions.size())
 611             fprintf(output, ", ");
 612       }
 613       fprintf(output, " = ");
 614    }
 615    fprintf(output, "%s", instr_info.name[(int)instr->opcode]);
 616    if (instr->operands.size()) {
 617       bool abs[instr->operands.size()];
 618       bool neg[instr->operands.size()];
 619       bool opsel[instr->operands.size()];
 620       uint8_t sel[instr->operands.size()];
 621       if ((int)instr->format & (int)Format::VOP3A) {
 622          const VOP3A_instruction* vop3 = static_cast<const VOP3A_instruction*>(instr);
 623          for (unsigned i = 0; i < instr->operands.size(); ++i) {
 624             abs[i] = vop3->abs[i];
 625             neg[i] = vop3->neg[i];
 626             opsel[i] = vop3->opsel & (1 << i);
 627             sel[i] = sdwa_udword;
 628          }
 629       } else if (instr->isDPP()) {
 630          const DPP_instruction* dpp = static_cast<const DPP_instruction*>(instr);
 631          for (unsigned i = 0; i < instr->operands.size(); ++i) {
 632             abs[i] = i < 2 ? dpp->abs[i] : false;
 633             neg[i] = i < 2 ? dpp->neg[i] : false;
 634             opsel[i] = false;
 635             sel[i] = sdwa_udword;
 636          }
 637       } else if (instr->isSDWA()) {
 638          const SDWA_instruction* sdwa = static_cast<const SDWA_instruction*>(instr);
 639          for (unsigned i = 0; i < instr->operands.size(); ++i) {
 640             abs[i] = i < 2 ? sdwa->abs[i] : false;
 641             neg[i] = i < 2 ? sdwa->neg[i] : false;
 642             opsel[i] = false;
 643             sel[i] = i < 2 ? sdwa->sel[i] : sdwa_udword;
 644          }
 645       } else {
 646          for (unsigned i = 0; i < instr->operands.size(); ++i) {
 647             abs[i] = false;
 648             neg[i] = false;
 649             opsel[i] = false;
 650             sel[i] = sdwa_udword;
 651          }
 652       }
 653       for (unsigned i = 0; i < instr->operands.size(); ++i) {
 654          if (i)
 655             fprintf(output, ", ");
 656          else
 657             fprintf(output, " ");
 658
 659          if (neg[i])
 660             fprintf(output, "-");
 661          if (abs[i])
 662             fprintf(output, "|");
 663          if (opsel[i])
 664             fprintf(output, "hi(");
 665          else if (sel[i] & sdwa_sext)
 666             fprintf(output, "sext(");
 667          print_operand(&instr->operands[i], output);
 668          if (opsel[i] || (sel[i] & sdwa_sext))
 669             fprintf(output, ")");
 670          if (!(sel[i] & sdwa_isra)) {
 671             if (sel[i] & sdwa_udword) {
 672                /* print nothing */
 673             } else if (sel[i] & sdwa_isword) {
 674                unsigned index = sel[i] & sdwa_wordnum;
 675                fprintf(output, "[%u:%u]", index * 16, index * 16 + 15);
 676             } else {
 677                unsigned index = sel[i] & sdwa_bytenum;
 678                fprintf(output, "[%u:%u]", index * 8, index * 8 + 7);
 679             }
 680          }
 681          if (abs[i])
 682             fprintf(output, "|");
 683
 684          if (instr->format == Format::VOP3P) {
 685             const VOP3P_instruction* vop3 = static_cast<const VOP3P_instruction*>(instr);
 686             if ((vop3->opsel_lo & (1 << i)) || !(vop3->opsel_hi & (1 << i))) {
 687                fprintf(output, ".%c%c",
 688                        vop3->opsel_lo & (1 << i) ? 'y' : 'x',
 689                        vop3->opsel_hi & (1 << i) ? 'y' : 'x');
 690             }
 691             if (vop3->neg_lo[i] && vop3->neg_hi[i])
 692                fprintf(output, "*[-1,-1]");
 693             else if (vop3->neg_lo[i])
 694                fprintf(output, "*[-1,1]");
 695             else if (vop3->neg_hi[i])
 696                fprintf(output, "*[1,-1]");
 697          }
 698       }
 699    }
 700    print_instr_format_specific(instr, output);
 701 }
 702
 703 static void print_block_kind(uint16_t kind, FILE *output)
 704 {
 705    if (kind & block_kind_uniform)
 706       fprintf(output, "uniform, ");
 707    if (kind & block_kind_top_level)
 708       fprintf(output, "top-level, ");
 709    if (kind & block_kind_loop_preheader)
 710       fprintf(output, "loop-preheader, ");
 711    if (kind & block_kind_loop_header)
 712       fprintf(output, "loop-header, ");
 713    if (kind & block_kind_loop_exit)
 714       fprintf(output, "loop-exit, ");
 715    if (kind & block_kind_continue)
 716       fprintf(output, "continue, ");
 717    if (kind & block_kind_break)
 718       fprintf(output, "break, ");
 719    if (kind & block_kind_continue_or_break)
 720       fprintf(output, "continue_or_break, ");
 721    if (kind & block_kind_discard)
 722       fprintf(output, "discard, ");
 723    if (kind & block_kind_branch)
 724       fprintf(output, "branch, ");
 725    if (kind & block_kind_merge)
 726       fprintf(output, "merge, ");
 727    if (kind & block_kind_invert)
 728       fprintf(output, "invert, ");
 729    if (kind & block_kind_uses_discard_if)
 730       fprintf(output, "discard_if, ");
 731    if (kind & block_kind_needs_lowering)
 732       fprintf(output, "needs_lowering, ");
 733    if (kind & block_kind_uses_demote)
 734       fprintf(output, "uses_demote, ");
 735    if (kind & block_kind_export_end)
 736       fprintf(output, "export_end, ");
 737 }
 738
 739 static void print_stage(Stage stage, FILE *output)
 740 {
 741    fprintf(output, "ACO shader stage: ");
 742
 743    if (stage == compute_cs)
 744       fprintf(output, "compute_cs");
 745    else if (stage == fragment_fs)
 746       fprintf(output, "fragment_fs");
 747    else if (stage == gs_copy_vs)
 748       fprintf(output, "gs_copy_vs");
 749    else if (stage == vertex_ls)
 750       fprintf(output, "vertex_ls");
 751    else if (stage == vertex_es)
 752       fprintf(output, "vertex_es");
 753    else if (stage == vertex_vs)
 754       fprintf(output, "vertex_vs");
 755    else if (stage == tess_control_hs)
 756       fprintf(output, "tess_control_hs");
 757    else if (stage == vertex_tess_control_hs)
 758       fprintf(output, "vertex_tess_control_hs");
 759    else if (stage == tess_eval_es)
 760       fprintf(output, "tess_eval_es");
 761    else if (stage == tess_eval_vs)
 762       fprintf(output, "tess_eval_vs");
 763    else if (stage == geometry_gs)
 764       fprintf(output, "geometry_gs");
 765    else if (stage == vertex_geometry_gs)
 766       fprintf(output, "vertex_geometry_gs");
 767    else if (stage == tess_eval_geometry_gs)
 768       fprintf(output, "tess_eval_geometry_gs");
 769    else if (stage == ngg_vertex_gs)
 770       fprintf(output, "ngg_vertex_gs");
 771    else if (stage == ngg_tess_eval_gs)
 772       fprintf(output, "ngg_tess_eval_gs");
 773    else if (stage == ngg_vertex_geometry_gs)
 774       fprintf(output, "ngg_vertex_geometry_gs");
 775    else if (stage == ngg_tess_eval_geometry_gs)
 776       fprintf(output, "ngg_tess_eval_geometry_gs");
 777    else
 778       fprintf(output, "unknown");
 779
 780    fprintf(output, "\n");
 781 }
 782
 783 void aco_print_block(const Block* block, FILE *output)
 784 {
 785    fprintf(output, "BB%d\n", block->index);
 786    fprintf(output, "/* logical preds: ");
 787    for (unsigned pred : block->logical_preds)
 788       fprintf(output, "BB%d, ", pred);
 789    fprintf(output, "/ linear preds: ");
 790    for (unsigned pred : block->linear_preds)
 791       fprintf(output, "BB%d, ", pred);
 792    fprintf(output, "/ kind: ");
 793    print_block_kind(block->kind, output);
 794    fprintf(output, "*/\n");
 795    for (auto const& instr : block->instructions) {
 796       fprintf(output, "\t");
 797       aco_print_instr(instr.get(), output);
 798       fprintf(output, "\n");
 799    }
 800 }
 801
 802 void aco_print_program(const Program *program, FILE *output)
 803 {
 804    print_stage(program->stage, output);
 805
 806    for (Block const& block : program->blocks)
 807       aco_print_block(&block, output);
 808
 809    if (program->constant_data.size()) {
 810       fprintf(output, "\n/* constant data */\n");
 811       for (unsigned i = 0; i < program->constant_data.size(); i += 32) {
 812          fprintf(output, "[%06d] ", i);
 813          unsigned line_size = std::min<size_t>(program->constant_data.size() - i, 32);
 814          for (unsigned j = 0; j < line_size; j += 4) {
 815             unsigned size = std::min<size_t>(program->constant_data.size() - (i + j), 4);
 816             uint32_t v = 0;
 817             memcpy(&v, &program->constant_data[i + j], size);
 818             fprintf(output, " %08x", v);
 819          }
 820          fprintf(output, "\n");
 821       }
 822    }
 823
 824    fprintf(output, "\n");
 825 }
 826
 827 }