src/amd/compiler/aco_validate.cpp

   1 /*
   2  * Copyright © 2018 Valve Corporation
   3  *
   4  * Permission is hereby granted, free of charge, to any person obtaining a
   5  * copy of this software and associated documentation files (the "Software"),
   6  * to deal in the Software without restriction, including without limitation
   7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   8  * and/or sell copies of the Software, and to permit persons to whom the
   9  * Software is furnished to do so, subject to the following conditions:
  10  *
  11  * The above copyright notice and this permission notice (including the next
  12  * paragraph) shall be included in all copies or substantial portions of the
  13  * Software.
  14  *
  15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  16  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  17  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  18  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  19  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
  20  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
  21  * IN THE SOFTWARE.
  22  *
  23  */
  24
  25 #include "aco_ir.h"
  26
  27 #include <array>
  28 #include <map>
  29
  30 namespace aco {
  31
  32 static void aco_log(Program *program, enum radv_compiler_debug_level level,
  33                     const char *prefix, const char *file, unsigned line,
  34                     const char *fmt, va_list args)
  35 {
  36    char *msg;
  37
  38    msg = ralloc_strdup(NULL, prefix);
  39
  40    ralloc_asprintf_append(&msg, "    In file %s:%u\n", file, line);
  41    ralloc_asprintf_append(&msg, "    ");
  42    ralloc_vasprintf_append(&msg, fmt, args);
  43
  44    if (program->debug.func)
  45       program->debug.func(program->debug.private_data, level, msg);
  46
  47    fprintf(stderr, "%s\n", msg);
  48
  49    ralloc_free(msg);
  50 }
  51
  52 void _aco_perfwarn(Program *program, const char *file, unsigned line,
  53                    const char *fmt, ...)
  54 {
  55    va_list args;
  56
  57    va_start(args, fmt);
  58    aco_log(program, RADV_COMPILER_DEBUG_LEVEL_PERFWARN,
  59            "ACO PERFWARN:\n", file, line, fmt, args);
  60    va_end(args);
  61 }
  62
  63 void _aco_err(Program *program, const char *file, unsigned line,
  64               const char *fmt, ...)
  65 {
  66    va_list args;
  67
  68    va_start(args, fmt);
  69    aco_log(program, RADV_COMPILER_DEBUG_LEVEL_ERROR,
  70            "ACO ERROR:\n", file, line, fmt, args);
  71    va_end(args);
  72 }
  73
  74 bool validate_ir(Program* program)
  75 {
  76    bool is_valid = true;
  77    auto check = [&program, &is_valid](bool check, const char * msg, aco::Instruction * instr) -> void {
  78       if (!check) {
  79          char *out;
  80          size_t outsize;
  81          FILE *memf = open_memstream(&out, &outsize);
  82
  83          fprintf(memf, "%s: ", msg);
  84          aco_print_instr(instr, memf);
  85          fclose(memf);
  86
  87          aco_err(program, out);
  88          free(out);
  89
  90          is_valid = false;
  91       }
  92    };
  93
  94    auto check_block = [&program, &is_valid](bool check, const char * msg, aco::Block * block) -> void {
  95       if (!check) {
  96          aco_err(program, "%s: BB%u", msg, block->index);
  97          is_valid = false;
  98       }
  99    };
 100
 101    for (Block& block : program->blocks) {
 102       for (aco_ptr<Instruction>& instr : block.instructions) {
 103
 104          /* check base format */
 105          Format base_format = instr->format;
 106          base_format = (Format)((uint32_t)base_format & ~(uint32_t)Format::SDWA);
 107          base_format = (Format)((uint32_t)base_format & ~(uint32_t)Format::DPP);
 108          if ((uint32_t)base_format & (uint32_t)Format::VOP1)
 109             base_format = Format::VOP1;
 110          else if ((uint32_t)base_format & (uint32_t)Format::VOP2)
 111             base_format = Format::VOP2;
 112          else if ((uint32_t)base_format & (uint32_t)Format::VOPC)
 113             base_format = Format::VOPC;
 114          else if ((uint32_t)base_format & (uint32_t)Format::VINTRP) {
 115             if (instr->opcode == aco_opcode::v_interp_p1ll_f16 ||
 116                 instr->opcode == aco_opcode::v_interp_p1lv_f16 ||
 117                 instr->opcode == aco_opcode::v_interp_p2_legacy_f16 ||
 118                 instr->opcode == aco_opcode::v_interp_p2_f16) {
 119                /* v_interp_*_fp16 are considered VINTRP by the compiler but
 120                 * they are emitted as VOP3.
 121                 */
 122                base_format = Format::VOP3;
 123             } else {
 124                base_format = Format::VINTRP;
 125             }
 126          }
 127          check(base_format == instr_info.format[(int)instr->opcode], "Wrong base format for instruction", instr.get());
 128
 129          /* check VOP3 modifiers */
 130          if (((uint32_t)instr->format & (uint32_t)Format::VOP3) && instr->format != Format::VOP3) {
 131             check(base_format == Format::VOP2 ||
 132                   base_format == Format::VOP1 ||
 133                   base_format == Format::VOPC ||
 134                   base_format == Format::VINTRP,
 135                   "Format cannot have VOP3A/VOP3B applied", instr.get());
 136          }
 137
 138          /* check SDWA */
 139          if (instr->isSDWA()) {
 140             check(base_format == Format::VOP2 ||
 141                   base_format == Format::VOP1 ||
 142                   base_format == Format::VOPC,
 143                   "Format cannot have SDWA applied", instr.get());
 144
 145             check(program->chip_class >= GFX8, "SDWA is GFX8+ only", instr.get());
 146
 147             SDWA_instruction *sdwa = static_cast<SDWA_instruction*>(instr.get());
 148             check(sdwa->omod == 0 || program->chip_class >= GFX9, "SDWA omod only supported on GFX9+", instr.get());
 149             if (base_format == Format::VOPC) {
 150                check(sdwa->clamp == false || program->chip_class == GFX8, "SDWA VOPC clamp only supported on GFX8", instr.get());
 151                check((instr->definitions[0].isFixed() && instr->definitions[0].physReg() == vcc) ||
 152                      program->chip_class >= GFX9,
 153                      "SDWA+VOPC definition must be fixed to vcc on GFX8", instr.get());
 154             }
 155
 156             if (instr->operands.size() >= 3) {
 157                check(instr->operands[2].isFixed() && instr->operands[2].physReg() == vcc,
 158                      "3rd operand must be fixed to vcc with SDWA", instr.get());
 159             }
 160             if (instr->definitions.size() >= 2) {
 161                check(instr->definitions[1].isFixed() && instr->definitions[1].physReg() == vcc,
 162                      "2nd definition must be fixed to vcc with SDWA", instr.get());
 163             }
 164
 165             check(instr->opcode != aco_opcode::v_madmk_f32 &&
 166                   instr->opcode != aco_opcode::v_madak_f32 &&
 167                   instr->opcode != aco_opcode::v_madmk_f16 &&
 168                   instr->opcode != aco_opcode::v_madak_f16 &&
 169                   instr->opcode != aco_opcode::v_readfirstlane_b32 &&
 170                   instr->opcode != aco_opcode::v_clrexcp &&
 171                   instr->opcode != aco_opcode::v_swap_b32,
 172                   "SDWA can't be used with this opcode", instr.get());
 173             if (program->chip_class != GFX8) {
 174                check(instr->opcode != aco_opcode::v_mac_f32 &&
 175                      instr->opcode != aco_opcode::v_mac_f16 &&
 176                      instr->opcode != aco_opcode::v_fmac_f32 &&
 177                      instr->opcode != aco_opcode::v_fmac_f16,
 178                      "SDWA can't be used with this opcode", instr.get());
 179             }
 180
 181             for (unsigned i = 0; i < MIN2(instr->operands.size(), 2); i++) {
 182                if (instr->operands[i].hasRegClass() && instr->operands[i].regClass().is_subdword())
 183                   check((sdwa->sel[i] & sdwa_asuint) == (sdwa_isra | instr->operands[i].bytes()), "Unexpected SDWA sel for sub-dword operand", instr.get());
 184             }
 185             if (instr->definitions[0].regClass().is_subdword())
 186                check((sdwa->dst_sel & sdwa_asuint) == (sdwa_isra | instr->definitions[0].bytes()), "Unexpected SDWA sel for sub-dword definition", instr.get());
 187          }
 188
 189          /* check opsel */
 190          if (instr->isVOP3()) {
 191             VOP3A_instruction *vop3 = static_cast<VOP3A_instruction*>(instr.get());
 192             check(vop3->opsel == 0 || program->chip_class >= GFX9, "Opsel is only supported on GFX9+", instr.get());
 193
 194             for (unsigned i = 0; i < 3; i++) {
 195                if (i >= instr->operands.size() ||
 196                    (instr->operands[i].hasRegClass() && instr->operands[i].regClass().is_subdword() && !instr->operands[i].isFixed()))
 197                   check((vop3->opsel & (1 << i)) == 0, "Unexpected opsel for operand", instr.get());
 198             }
 199             if (instr->definitions[0].regClass().is_subdword() && !instr->definitions[0].isFixed())
 200                check((vop3->opsel & (1 << 3)) == 0, "Unexpected opsel for sub-dword definition", instr.get());
 201          }
 202
 203          /* check for undefs */
 204          for (unsigned i = 0; i < instr->operands.size(); i++) {
 205             if (instr->operands[i].isUndefined()) {
 206                bool flat = instr->format == Format::FLAT || instr->format == Format::SCRATCH || instr->format == Format::GLOBAL;
 207                bool can_be_undef = is_phi(instr) || instr->format == Format::EXP ||
 208                                    instr->format == Format::PSEUDO_REDUCTION ||
 209                                    instr->opcode == aco_opcode::p_create_vector ||
 210                                    (flat && i == 1) || (instr->format == Format::MIMG && i == 1) ||
 211                                    ((instr->format == Format::MUBUF || instr->format == Format::MTBUF) && i == 1);
 212                check(can_be_undef, "Undefs can only be used in certain operands", instr.get());
 213             } else {
 214                check(instr->operands[i].isFixed() || instr->operands[i].isTemp() || instr->operands[i].isConstant(), "Uninitialized Operand", instr.get());
 215             }
 216          }
 217
 218          /* check subdword definitions */
 219          for (unsigned i = 0; i < instr->definitions.size(); i++) {
 220             if (instr->definitions[i].regClass().is_subdword())
 221                check(instr->format == Format::PSEUDO || instr->definitions[i].bytes() <= 4, "Only Pseudo instructions can write subdword registers larger than 4 bytes", instr.get());
 222          }
 223
 224          if (instr->isSALU() || instr->isVALU()) {
 225             /* check literals */
 226             Operand literal(s1);
 227             for (unsigned i = 0; i < instr->operands.size(); i++)
 228             {
 229                Operand op = instr->operands[i];
 230                if (!op.isLiteral())
 231                   continue;
 232
 233                check(instr->format == Format::SOP1 ||
 234                      instr->format == Format::SOP2 ||
 235                      instr->format == Format::SOPC ||
 236                      instr->format == Format::VOP1 ||
 237                      instr->format == Format::VOP2 ||
 238                      instr->format == Format::VOPC ||
 239                      (instr->isVOP3() && program->chip_class >= GFX10),
 240                      "Literal applied on wrong instruction format", instr.get());
 241
 242                check(literal.isUndefined() || (literal.size() == op.size() && literal.constantValue() == op.constantValue()), "Only 1 Literal allowed", instr.get());
 243                literal = op;
 244                check(!instr->isVALU() || instr->isVOP3() || i == 0 || i == 2, "Wrong source position for Literal argument", instr.get());
 245             }
 246
 247             /* check num sgprs for VALU */
 248             if (instr->isVALU()) {
 249                bool is_shift64 = instr->opcode == aco_opcode::v_lshlrev_b64 ||
 250                                  instr->opcode == aco_opcode::v_lshrrev_b64 ||
 251                                  instr->opcode == aco_opcode::v_ashrrev_i64;
 252                unsigned const_bus_limit = 1;
 253                if (program->chip_class >= GFX10 && !is_shift64)
 254                   const_bus_limit = 2;
 255
 256                uint32_t scalar_mask = instr->isVOP3() ? 0x7 : 0x5;
 257                if (instr->isSDWA())
 258                   scalar_mask = program->chip_class >= GFX9 ? 0x7 : 0x4;
 259
 260                if ((int) instr->format & (int) Format::VOPC ||
 261                    instr->opcode == aco_opcode::v_readfirstlane_b32 ||
 262                    instr->opcode == aco_opcode::v_readlane_b32 ||
 263                    instr->opcode == aco_opcode::v_readlane_b32_e64) {
 264                   check(instr->definitions[0].getTemp().type() == RegType::sgpr,
 265                         "Wrong Definition type for VALU instruction", instr.get());
 266                } else {
 267                   check(instr->definitions[0].getTemp().type() == RegType::vgpr,
 268                         "Wrong Definition type for VALU instruction", instr.get());
 269                }
 270
 271                unsigned num_sgprs = 0;
 272                unsigned sgpr[] = {0, 0};
 273                for (unsigned i = 0; i < instr->operands.size(); i++)
 274                {
 275                   Operand op = instr->operands[i];
 276                   if (instr->opcode == aco_opcode::v_readfirstlane_b32 ||
 277                       instr->opcode == aco_opcode::v_readlane_b32 ||
 278                       instr->opcode == aco_opcode::v_readlane_b32_e64) {
 279                      check(i != 1 ||
 280                            (op.isTemp() && op.regClass().type() == RegType::sgpr) ||
 281                            op.isConstant(),
 282                            "Must be a SGPR or a constant", instr.get());
 283                      check(i == 1 ||
 284                            (op.isTemp() && op.regClass().type() == RegType::vgpr && op.bytes() <= 4),
 285                            "Wrong Operand type for VALU instruction", instr.get());
 286                      continue;
 287                   }
 288
 289                   if (instr->opcode == aco_opcode::v_writelane_b32 ||
 290                       instr->opcode == aco_opcode::v_writelane_b32_e64) {
 291                      check(i != 2 ||
 292                            (op.isTemp() && op.regClass().type() == RegType::vgpr && op.bytes() <= 4),
 293                            "Wrong Operand type for VALU instruction", instr.get());
 294                      check(i == 2 ||
 295                            (op.isTemp() && op.regClass().type() == RegType::sgpr) ||
 296                            op.isConstant(),
 297                            "Must be a SGPR or a constant", instr.get());
 298                      continue;
 299                   }
 300                   if (op.isTemp() && instr->operands[i].regClass().type() == RegType::sgpr) {
 301                      check(scalar_mask & (1 << i), "Wrong source position for SGPR argument", instr.get());
 302
 303                      if (op.tempId() != sgpr[0] && op.tempId() != sgpr[1]) {
 304                         if (num_sgprs < 2)
 305                            sgpr[num_sgprs++] = op.tempId();
 306                      }
 307                   }
 308
 309                   if (op.isConstant() && !op.isLiteral())
 310                      check(scalar_mask & (1 << i), "Wrong source position for constant argument", instr.get());
 311                }
 312                check(num_sgprs + (literal.isUndefined() ? 0 : 1) <= const_bus_limit, "Too many SGPRs/literals", instr.get());
 313             }
 314
 315             if (instr->format == Format::SOP1 || instr->format == Format::SOP2) {
 316                check(instr->definitions[0].getTemp().type() == RegType::sgpr, "Wrong Definition type for SALU instruction", instr.get());
 317                for (const Operand& op : instr->operands) {
 318                  check(op.isConstant() || op.regClass().type() <= RegType::sgpr,
 319                        "Wrong Operand type for SALU instruction", instr.get());
 320             }
 321          }
 322          }
 323
 324          switch (instr->format) {
 325          case Format::PSEUDO: {
 326             bool is_subdword = false;
 327             bool has_const_sgpr = false;
 328             bool has_literal = false;
 329             for (Definition def : instr->definitions)
 330                is_subdword |= def.regClass().is_subdword();
 331             for (unsigned i = 0; i < instr->operands.size(); i++) {
 332                if (instr->opcode == aco_opcode::p_extract_vector && i == 1)
 333                   continue;
 334                Operand op = instr->operands[i];
 335                is_subdword |= op.hasRegClass() && op.regClass().is_subdword();
 336                has_const_sgpr |= op.isConstant() || (op.hasRegClass() && op.regClass().type() == RegType::sgpr);
 337                has_literal |= op.isLiteral();
 338             }
 339
 340             check(!is_subdword || !has_const_sgpr || program->chip_class >= GFX9,
 341                   "Sub-dword pseudo instructions can only take constants or SGPRs on GFX9+", instr.get());
 342             check(!is_subdword || !has_literal, "Sub-dword pseudo instructions cannot take literals", instr.get());
 343
 344             if (instr->opcode == aco_opcode::p_create_vector) {
 345                unsigned size = 0;
 346                for (const Operand& op : instr->operands) {
 347                   size += op.bytes();
 348                }
 349                check(size == instr->definitions[0].bytes(), "Definition size does not match operand sizes", instr.get());
 350                if (instr->definitions[0].getTemp().type() == RegType::sgpr) {
 351                   for (const Operand& op : instr->operands) {
 352                      check(op.isConstant() || op.regClass().type() == RegType::sgpr,
 353                            "Wrong Operand type for scalar vector", instr.get());
 354                   }
 355                }
 356             } else if (instr->opcode == aco_opcode::p_extract_vector) {
 357                check((instr->operands[0].isTemp()) && instr->operands[1].isConstant(), "Wrong Operand types", instr.get());
 358                check((instr->operands[1].constantValue() + 1) * instr->definitions[0].bytes() <= instr->operands[0].bytes(), "Index out of range", instr.get());
 359                check(instr->definitions[0].getTemp().type() == RegType::vgpr || instr->operands[0].regClass().type() == RegType::sgpr,
 360                      "Cannot extract SGPR value from VGPR vector", instr.get());
 361             } else if (instr->opcode == aco_opcode::p_parallelcopy) {
 362                check(instr->definitions.size() == instr->operands.size(), "Number of Operands does not match number of Definitions", instr.get());
 363                for (unsigned i = 0; i < instr->operands.size(); i++) {
 364                   if (instr->operands[i].isTemp())
 365                      check((instr->definitions[i].getTemp().type() == instr->operands[i].regClass().type()) ||
 366                            (instr->definitions[i].getTemp().type() == RegType::vgpr && instr->operands[i].regClass().type() == RegType::sgpr),
 367                            "Operand and Definition types do not match", instr.get());
 368                }
 369             } else if (instr->opcode == aco_opcode::p_phi) {
 370                check(instr->operands.size() == block.logical_preds.size(), "Number of Operands does not match number of predecessors", instr.get());
 371                check(instr->definitions[0].getTemp().type() == RegType::vgpr, "Logical Phi Definition must be vgpr", instr.get());
 372             } else if (instr->opcode == aco_opcode::p_linear_phi) {
 373                for (const Operand& op : instr->operands)
 374                   check(!op.isTemp() || op.getTemp().is_linear(), "Wrong Operand type", instr.get());
 375                check(instr->operands.size() == block.linear_preds.size(), "Number of Operands does not match number of predecessors", instr.get());
 376             }
 377             break;
 378          }
 379          case Format::SMEM: {
 380             if (instr->operands.size() >= 1)
 381                check(instr->operands[0].isTemp() && instr->operands[0].regClass().type() == RegType::sgpr, "SMEM operands must be sgpr", instr.get());
 382             if (instr->operands.size() >= 2)
 383                check(instr->operands[1].isConstant() || (instr->operands[1].isTemp() && instr->operands[1].regClass().type() == RegType::sgpr),
 384                      "SMEM offset must be constant or sgpr", instr.get());
 385             if (!instr->definitions.empty())
 386                check(instr->definitions[0].getTemp().type() == RegType::sgpr, "SMEM result must be sgpr", instr.get());
 387             break;
 388          }
 389          case Format::MTBUF:
 390          case Format::MUBUF: {
 391             check(instr->operands.size() > 1, "VMEM instructions must have at least one operand", instr.get());
 392             check(instr->operands[1].hasRegClass() && instr->operands[1].regClass().type() == RegType::vgpr,
 393                   "VADDR must be in vgpr for VMEM instructions", instr.get());
 394             check(instr->operands[0].isTemp() && instr->operands[0].regClass().type() == RegType::sgpr, "VMEM resource constant must be sgpr", instr.get());
 395             check(instr->operands.size() < 4 || (instr->operands[3].isTemp() && instr->operands[3].regClass().type() == RegType::vgpr), "VMEM write data must be vgpr", instr.get());
 396             break;
 397          }
 398          case Format::MIMG: {
 399             check(instr->operands.size() == 3, "MIMG instructions must have exactly 3 operands", instr.get());
 400             check(instr->operands[0].hasRegClass() && (instr->operands[0].regClass() == s4 || instr->operands[0].regClass() == s8),
 401                   "MIMG operands[0] (resource constant) must be in 4 or 8 SGPRs", instr.get());
 402             if (instr->operands[1].hasRegClass() && instr->operands[1].regClass().type() == RegType::sgpr)
 403                check(instr->operands[1].regClass() == s4, "MIMG operands[1] (sampler constant) must be 4 SGPRs", instr.get());
 404             else if (instr->operands[1].hasRegClass() && instr->operands[1].regClass().type() == RegType::vgpr)
 405                check((instr->definitions.empty() || instr->definitions[0].regClass() == instr->operands[1].regClass() ||
 406                      instr->opcode == aco_opcode::image_atomic_cmpswap || instr->opcode == aco_opcode::image_atomic_fcmpswap),
 407                      "MIMG operands[1] (VDATA) must be the same as definitions[0] for atomics", instr.get());
 408             check(instr->operands[2].hasRegClass() && instr->operands[2].regClass().type() == RegType::vgpr,
 409                   "MIMG operands[2] (VADDR) must be VGPR", instr.get());
 410             check(instr->definitions.empty() || (instr->definitions[0].isTemp() && instr->definitions[0].regClass().type() == RegType::vgpr),
 411                   "MIMG definitions[0] (VDATA) must be VGPR", instr.get());
 412             break;
 413          }
 414          case Format::DS: {
 415             for (const Operand& op : instr->operands) {
 416                check((op.isTemp() && op.regClass().type() == RegType::vgpr) || op.physReg() == m0,
 417                      "Only VGPRs are valid DS instruction operands", instr.get());
 418             }
 419             if (!instr->definitions.empty())
 420                check(instr->definitions[0].getTemp().type() == RegType::vgpr, "DS instruction must return VGPR", instr.get());
 421             break;
 422          }
 423          case Format::EXP: {
 424             for (unsigned i = 0; i < 4; i++)
 425                check(instr->operands[i].hasRegClass() && instr->operands[i].regClass().type() == RegType::vgpr,
 426                      "Only VGPRs are valid Export arguments", instr.get());
 427             break;
 428          }
 429          case Format::FLAT:
 430             check(instr->operands[1].isUndefined(), "Flat instructions don't support SADDR", instr.get());
 431             /* fallthrough */
 432          case Format::GLOBAL:
 433          case Format::SCRATCH: {
 434             check(instr->operands[0].isTemp() && instr->operands[0].regClass().type() == RegType::vgpr, "FLAT/GLOBAL/SCRATCH address must be vgpr", instr.get());
 435             check(instr->operands[1].hasRegClass() && instr->operands[1].regClass().type() == RegType::sgpr,
 436                   "FLAT/GLOBAL/SCRATCH sgpr address must be undefined or sgpr", instr.get());
 437             if (!instr->definitions.empty())
 438                check(instr->definitions[0].getTemp().type() == RegType::vgpr, "FLAT/GLOBAL/SCRATCH result must be vgpr", instr.get());
 439             else
 440                check(instr->operands[2].regClass().type() == RegType::vgpr, "FLAT/GLOBAL/SCRATCH data must be vgpr", instr.get());
 441             break;
 442          }
 443          default:
 444             break;
 445          }
 446       }
 447    }
 448
 449    /* validate CFG */
 450    for (unsigned i = 0; i < program->blocks.size(); i++) {
 451       Block& block = program->blocks[i];
 452       check_block(block.index == i, "block.index must match actual index", &block);
 453
 454       /* predecessors/successors should be sorted */
 455       for (unsigned j = 0; j + 1 < block.linear_preds.size(); j++)
 456          check_block(block.linear_preds[j] < block.linear_preds[j + 1], "linear predecessors must be sorted", &block);
 457       for (unsigned j = 0; j + 1 < block.logical_preds.size(); j++)
 458          check_block(block.logical_preds[j] < block.logical_preds[j + 1], "logical predecessors must be sorted", &block);
 459       for (unsigned j = 0; j + 1 < block.linear_succs.size(); j++)
 460          check_block(block.linear_succs[j] < block.linear_succs[j + 1], "linear successors must be sorted", &block);
 461       for (unsigned j = 0; j + 1 < block.logical_succs.size(); j++)
 462          check_block(block.logical_succs[j] < block.logical_succs[j + 1], "logical successors must be sorted", &block);
 463
 464       /* critical edges are not allowed */
 465       if (block.linear_preds.size() > 1) {
 466          for (unsigned pred : block.linear_preds)
 467             check_block(program->blocks[pred].linear_succs.size() == 1, "linear critical edges are not allowed", &program->blocks[pred]);
 468          for (unsigned pred : block.logical_preds)
 469             check_block(program->blocks[pred].logical_succs.size() == 1, "logical critical edges are not allowed", &program->blocks[pred]);
 470       }
 471    }
 472
 473    return is_valid;
 474 }
 475
 476 /* RA validation */
 477 namespace {
 478
 479 struct Location {
 480    Location() : block(NULL), instr(NULL) {}
 481
 482    Block *block;
 483    Instruction *instr; //NULL if it's the block's live-in
 484 };
 485
 486 struct Assignment {
 487    Location defloc;
 488    Location firstloc;
 489    PhysReg reg;
 490 };
 491
 492 bool ra_fail(Program *program, Location loc, Location loc2, const char *fmt, ...) {
 493    va_list args;
 494    va_start(args, fmt);
 495    char msg[1024];
 496    vsprintf(msg, fmt, args);
 497    va_end(args);
 498
 499    char *out;
 500    size_t outsize;
 501    FILE *memf = open_memstream(&out, &outsize);
 502
 503    fprintf(memf, "RA error found at instruction in BB%d:\n", loc.block->index);
 504    if (loc.instr) {
 505       aco_print_instr(loc.instr, memf);
 506       fprintf(memf, "\n%s", msg);
 507    } else {
 508       fprintf(memf, "%s", msg);
 509    }
 510    if (loc2.block) {
 511       fprintf(memf, " in BB%d:\n", loc2.block->index);
 512       aco_print_instr(loc2.instr, memf);
 513    }
 514    fprintf(memf, "\n\n");
 515
 516    aco_err(program, out);
 517    free(out);
 518
 519    return true;
 520 }
 521
 522 bool validate_subdword_operand(chip_class chip, const aco_ptr<Instruction>& instr, unsigned index)
 523 {
 524    Operand op = instr->operands[index];
 525    unsigned byte = op.physReg().byte();
 526
 527    if (instr->opcode == aco_opcode::p_as_uniform)
 528       return byte == 0;
 529    if (instr->format == Format::PSEUDO && chip >= GFX8)
 530       return true;
 531    if (instr->isSDWA() && (static_cast<SDWA_instruction *>(instr.get())->sel[index] & sdwa_asuint) == (sdwa_isra | op.bytes()))
 532       return true;
 533    if (byte == 2 && can_use_opsel(chip, instr->opcode, index, 1))
 534       return true;
 535
 536    switch (instr->opcode) {
 537    case aco_opcode::v_cvt_f32_ubyte1:
 538       if (byte == 1)
 539          return true;
 540       break;
 541    case aco_opcode::v_cvt_f32_ubyte2:
 542       if (byte == 2)
 543          return true;
 544       break;
 545    case aco_opcode::v_cvt_f32_ubyte3:
 546       if (byte == 3)
 547          return true;
 548       break;
 549    case aco_opcode::ds_write_b8_d16_hi:
 550    case aco_opcode::ds_write_b16_d16_hi:
 551       if (byte == 2 && index == 1)
 552          return true;
 553       break;
 554    case aco_opcode::buffer_store_byte_d16_hi:
 555    case aco_opcode::buffer_store_short_d16_hi:
 556       if (byte == 2 && index == 3)
 557          return true;
 558       break;
 559    case aco_opcode::flat_store_byte_d16_hi:
 560    case aco_opcode::flat_store_short_d16_hi:
 561    case aco_opcode::scratch_store_byte_d16_hi:
 562    case aco_opcode::scratch_store_short_d16_hi:
 563    case aco_opcode::global_store_byte_d16_hi:
 564    case aco_opcode::global_store_short_d16_hi:
 565       if (byte == 2 && index == 2)
 566          return true;
 567    default:
 568       break;
 569    }
 570
 571    return byte == 0;
 572 }
 573
 574 bool validate_subdword_definition(chip_class chip, const aco_ptr<Instruction>& instr)
 575 {
 576    Definition def = instr->definitions[0];
 577    unsigned byte = def.physReg().byte();
 578
 579    if (instr->format == Format::PSEUDO && chip >= GFX8)
 580       return true;
 581    if (instr->isSDWA() && static_cast<SDWA_instruction *>(instr.get())->dst_sel == (sdwa_isra | def.bytes()))
 582       return true;
 583    if (byte == 2 && can_use_opsel(chip, instr->opcode, -1, 1))
 584       return true;
 585
 586    switch (instr->opcode) {
 587    case aco_opcode::buffer_load_ubyte_d16_hi:
 588    case aco_opcode::buffer_load_short_d16_hi:
 589    case aco_opcode::flat_load_ubyte_d16_hi:
 590    case aco_opcode::flat_load_short_d16_hi:
 591    case aco_opcode::scratch_load_ubyte_d16_hi:
 592    case aco_opcode::scratch_load_short_d16_hi:
 593    case aco_opcode::global_load_ubyte_d16_hi:
 594    case aco_opcode::global_load_short_d16_hi:
 595    case aco_opcode::ds_read_u8_d16_hi:
 596    case aco_opcode::ds_read_u16_d16_hi:
 597       return byte == 2;
 598    default:
 599       break;
 600    }
 601
 602    return byte == 0;
 603 }
 604
 605 unsigned get_subdword_bytes_written(Program *program, const aco_ptr<Instruction>& instr, unsigned index)
 606 {
 607    chip_class chip = program->chip_class;
 608    Definition def = instr->definitions[index];
 609
 610    if (instr->format == Format::PSEUDO)
 611       return chip >= GFX8 ? def.bytes() : def.size() * 4u;
 612    if (instr->isSDWA() && static_cast<SDWA_instruction *>(instr.get())->dst_sel == (sdwa_isra | def.bytes()))
 613       return def.bytes();
 614
 615    switch (instr->opcode) {
 616    case aco_opcode::buffer_load_ubyte_d16:
 617    case aco_opcode::buffer_load_short_d16:
 618    case aco_opcode::flat_load_ubyte_d16:
 619    case aco_opcode::flat_load_short_d16:
 620    case aco_opcode::scratch_load_ubyte_d16:
 621    case aco_opcode::scratch_load_short_d16:
 622    case aco_opcode::global_load_ubyte_d16:
 623    case aco_opcode::global_load_short_d16:
 624    case aco_opcode::ds_read_u8_d16:
 625    case aco_opcode::ds_read_u16_d16:
 626    case aco_opcode::buffer_load_ubyte_d16_hi:
 627    case aco_opcode::buffer_load_short_d16_hi:
 628    case aco_opcode::flat_load_ubyte_d16_hi:
 629    case aco_opcode::flat_load_short_d16_hi:
 630    case aco_opcode::scratch_load_ubyte_d16_hi:
 631    case aco_opcode::scratch_load_short_d16_hi:
 632    case aco_opcode::global_load_ubyte_d16_hi:
 633    case aco_opcode::global_load_short_d16_hi:
 634    case aco_opcode::ds_read_u8_d16_hi:
 635    case aco_opcode::ds_read_u16_d16_hi:
 636       return program->sram_ecc_enabled ? 4 : 2;
 637    case aco_opcode::v_mad_f16:
 638    case aco_opcode::v_mad_u16:
 639    case aco_opcode::v_mad_i16:
 640    case aco_opcode::v_fma_f16:
 641    case aco_opcode::v_div_fixup_f16:
 642    case aco_opcode::v_interp_p2_f16:
 643       if (chip >= GFX9)
 644          return 2;
 645    default:
 646       break;
 647    }
 648
 649    return MAX2(chip >= GFX10 ? def.bytes() : 4, instr_info.definition_size[(int)instr->opcode] / 8u);
 650 }
 651
 652 } /* end namespace */
 653
 654 bool validate_ra(Program *program, const struct radv_nir_compiler_options *options) {
 655    if (!(debug_flags & DEBUG_VALIDATE_RA))
 656       return false;
 657
 658    bool err = false;
 659    aco::live live_vars = aco::live_var_analysis(program, options);
 660    std::vector<std::vector<Temp>> phi_sgpr_ops(program->blocks.size());
 661
 662    std::map<unsigned, Assignment> assignments;
 663    for (Block& block : program->blocks) {
 664       Location loc;
 665       loc.block = &block;
 666       for (aco_ptr<Instruction>& instr : block.instructions) {
 667          if (instr->opcode == aco_opcode::p_phi) {
 668             for (unsigned i = 0; i < instr->operands.size(); i++) {
 669                if (instr->operands[i].isTemp() &&
 670                    instr->operands[i].getTemp().type() == RegType::sgpr &&
 671                    instr->operands[i].isFirstKill())
 672                   phi_sgpr_ops[block.logical_preds[i]].emplace_back(instr->operands[i].getTemp());
 673             }
 674          }
 675
 676          loc.instr = instr.get();
 677          for (unsigned i = 0; i < instr->operands.size(); i++) {
 678             Operand& op = instr->operands[i];
 679             if (!op.isTemp())
 680                continue;
 681             if (!op.isFixed())
 682                err |= ra_fail(program, loc, Location(), "Operand %d is not assigned a register", i);
 683             if (assignments.count(op.tempId()) && assignments[op.tempId()].reg != op.physReg())
 684                err |= ra_fail(program, loc, assignments.at(op.tempId()).firstloc, "Operand %d has an inconsistent register assignment with instruction", i);
 685             if ((op.getTemp().type() == RegType::vgpr && op.physReg().reg_b + op.bytes() > (256 + program->config->num_vgprs) * 4) ||
 686                 (op.getTemp().type() == RegType::sgpr && op.physReg() + op.size() > program->config->num_sgprs && op.physReg() < program->sgpr_limit))
 687                err |= ra_fail(program, loc, assignments.at(op.tempId()).firstloc, "Operand %d has an out-of-bounds register assignment", i);
 688             if (op.physReg() == vcc && !program->needs_vcc)
 689                err |= ra_fail(program, loc, Location(), "Operand %d fixed to vcc but needs_vcc=false", i);
 690             if (op.regClass().is_subdword() && !validate_subdword_operand(program->chip_class, instr, i))
 691                err |= ra_fail(program, loc, Location(), "Operand %d not aligned correctly", i);
 692             if (!assignments[op.tempId()].firstloc.block)
 693                assignments[op.tempId()].firstloc = loc;
 694             if (!assignments[op.tempId()].defloc.block)
 695                assignments[op.tempId()].reg = op.physReg();
 696          }
 697
 698          for (unsigned i = 0; i < instr->definitions.size(); i++) {
 699             Definition& def = instr->definitions[i];
 700             if (!def.isTemp())
 701                continue;
 702             if (!def.isFixed())
 703                err |= ra_fail(program, loc, Location(), "Definition %d is not assigned a register", i);
 704             if (assignments[def.tempId()].defloc.block)
 705                err |= ra_fail(program, loc, assignments.at(def.tempId()).defloc, "Temporary %%%d also defined by instruction", def.tempId());
 706             if ((def.getTemp().type() == RegType::vgpr && def.physReg().reg_b + def.bytes() > (256 + program->config->num_vgprs) * 4) ||
 707                 (def.getTemp().type() == RegType::sgpr && def.physReg() + def.size() > program->config->num_sgprs && def.physReg() < program->sgpr_limit))
 708                err |= ra_fail(program, loc, assignments.at(def.tempId()).firstloc, "Definition %d has an out-of-bounds register assignment", i);
 709             if (def.physReg() == vcc && !program->needs_vcc)
 710                err |= ra_fail(program, loc, Location(), "Definition %d fixed to vcc but needs_vcc=false", i);
 711             if (def.regClass().is_subdword() && !validate_subdword_definition(program->chip_class, instr))
 712                err |= ra_fail(program, loc, Location(), "Definition %d not aligned correctly", i);
 713             if (!assignments[def.tempId()].firstloc.block)
 714                assignments[def.tempId()].firstloc = loc;
 715             assignments[def.tempId()].defloc = loc;
 716             assignments[def.tempId()].reg = def.physReg();
 717          }
 718       }
 719    }
 720
 721    for (Block& block : program->blocks) {
 722       Location loc;
 723       loc.block = &block;
 724
 725       std::array<unsigned, 2048> regs; /* register file in bytes */
 726       regs.fill(0);
 727
 728       std::set<Temp> live;
 729       live.insert(live_vars.live_out[block.index].begin(), live_vars.live_out[block.index].end());
 730       /* remove killed p_phi sgpr operands */
 731       for (Temp tmp : phi_sgpr_ops[block.index])
 732          live.erase(tmp);
 733
 734       /* check live out */
 735       for (Temp tmp : live) {
 736          PhysReg reg = assignments.at(tmp.id()).reg;
 737          for (unsigned i = 0; i < tmp.bytes(); i++) {
 738             if (regs[reg.reg_b + i]) {
 739                err |= ra_fail(program, loc, Location(), "Assignment of element %d of %%%d already taken by %%%d in live-out", i, tmp.id(), regs[reg.reg_b + i]);
 740             }
 741             regs[reg.reg_b + i] = tmp.id();
 742          }
 743       }
 744       regs.fill(0);
 745
 746       for (auto it = block.instructions.rbegin(); it != block.instructions.rend(); ++it) {
 747          aco_ptr<Instruction>& instr = *it;
 748
 749          /* check killed p_phi sgpr operands */
 750          if (instr->opcode == aco_opcode::p_logical_end) {
 751             for (Temp tmp : phi_sgpr_ops[block.index]) {
 752                PhysReg reg = assignments.at(tmp.id()).reg;
 753                for (unsigned i = 0; i < tmp.bytes(); i++) {
 754                   if (regs[reg.reg_b + i])
 755                      err |= ra_fail(program, loc, Location(), "Assignment of element %d of %%%d already taken by %%%d in live-out", i, tmp.id(), regs[reg.reg_b + i]);
 756                }
 757                live.emplace(tmp);
 758             }
 759          }
 760
 761          for (const Definition& def : instr->definitions) {
 762             if (!def.isTemp())
 763                continue;
 764             live.erase(def.getTemp());
 765          }
 766
 767          /* don't count phi operands as live-in, since they are actually
 768           * killed when they are copied at the predecessor */
 769          if (instr->opcode != aco_opcode::p_phi && instr->opcode != aco_opcode::p_linear_phi) {
 770             for (const Operand& op : instr->operands) {
 771                if (!op.isTemp())
 772                   continue;
 773                live.insert(op.getTemp());
 774             }
 775          }
 776       }
 777
 778       for (Temp tmp : live) {
 779          PhysReg reg = assignments.at(tmp.id()).reg;
 780          for (unsigned i = 0; i < tmp.bytes(); i++)
 781             regs[reg.reg_b + i] = tmp.id();
 782       }
 783
 784       for (aco_ptr<Instruction>& instr : block.instructions) {
 785          loc.instr = instr.get();
 786
 787          /* remove killed p_phi operands from regs */
 788          if (instr->opcode == aco_opcode::p_logical_end) {
 789             for (Temp tmp : phi_sgpr_ops[block.index]) {
 790                PhysReg reg = assignments.at(tmp.id()).reg;
 791                for (unsigned i = 0; i < tmp.bytes(); i++)
 792                   regs[reg.reg_b + i] = 0;
 793             }
 794          }
 795
 796          if (instr->opcode != aco_opcode::p_phi && instr->opcode != aco_opcode::p_linear_phi) {
 797             for (const Operand& op : instr->operands) {
 798                if (!op.isTemp())
 799                   continue;
 800                if (op.isFirstKillBeforeDef()) {
 801                   for (unsigned j = 0; j < op.getTemp().bytes(); j++)
 802                      regs[op.physReg().reg_b + j] = 0;
 803                }
 804             }
 805          }
 806
 807          for (unsigned i = 0; i < instr->definitions.size(); i++) {
 808             Definition& def = instr->definitions[i];
 809             if (!def.isTemp())
 810                continue;
 811             Temp tmp = def.getTemp();
 812             PhysReg reg = assignments.at(tmp.id()).reg;
 813             for (unsigned j = 0; j < tmp.bytes(); j++) {
 814                if (regs[reg.reg_b + j])
 815                   err |= ra_fail(program, loc, assignments.at(regs[reg.reg_b + j]).defloc, "Assignment of element %d of %%%d already taken by %%%d from instruction", i, tmp.id(), regs[reg.reg_b + j]);
 816                regs[reg.reg_b + j] = tmp.id();
 817             }
 818             if (def.regClass().is_subdword() && def.bytes() < 4) {
 819                unsigned written = get_subdword_bytes_written(program, instr, i);
 820                /* If written=4, the instruction still might write the upper half. In that case, it's the lower half that isn't preserved */
 821                for (unsigned j = reg.byte() & ~(written - 1); j < written; j++) {
 822                   unsigned written_reg = reg.reg() * 4u + j;
 823                   if (regs[written_reg] && regs[written_reg] != def.tempId())
 824                      err |= ra_fail(program, loc, assignments.at(regs[written_reg]).defloc, "Assignment of element %d of %%%d overwrites the full register taken by %%%d from instruction", i, tmp.id(), regs[written_reg]);
 825                }
 826             }
 827          }
 828
 829          for (const Definition& def : instr->definitions) {
 830             if (!def.isTemp())
 831                continue;
 832             if (def.isKill()) {
 833                for (unsigned j = 0; j < def.getTemp().bytes(); j++)
 834                   regs[def.physReg().reg_b + j] = 0;
 835             }
 836          }
 837
 838          if (instr->opcode != aco_opcode::p_phi && instr->opcode != aco_opcode::p_linear_phi) {
 839             for (const Operand& op : instr->operands) {
 840                if (!op.isTemp())
 841                   continue;
 842                if (op.isLateKill() && op.isFirstKill()) {
 843                   for (unsigned j = 0; j < op.getTemp().bytes(); j++)
 844                      regs[op.physReg().reg_b + j] = 0;
 845                }
 846             }
 847          }
 848       }
 849    }
 850
 851    return err;
 852 }
 853 }