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