src/amd/compiler/aco_print_asm.cpp

   1 #include <array>
   2 #include <iomanip>
   3 #include "aco_ir.h"
   4 #include "llvm-c/Disassembler.h"
   5 #include "ac_llvm_util.h"
   6
   7 #include <llvm/ADT/StringRef.h>
   8 #if LLVM_VERSION_MAJOR >= 11
   9 #include <llvm/MC/MCDisassembler/MCDisassembler.h>
  10 #endif
  11
  12 namespace aco {
  13
  14 /* LLVM disassembler only supports GFX8+, try to disassemble with CLRXdisasm
  15  * for GFX6-GFX7 if found on the system, this is better than nothing.
  16 */
  17 void print_asm_gfx6_gfx7(Program *program, std::vector<uint32_t>& binary,
  18                          std::ostream& out)
  19 {
  20    char path[] = "/tmp/fileXXXXXX";
  21    char line[2048], command[128];
  22    const char *gpu_type;
  23    FILE *p;
  24    int fd;
  25
  26    /* Dump the binary into a temporary file. */
  27    fd = mkstemp(path);
  28    if (fd < 0)
  29       return;
  30
  31    for (uint32_t w : binary)
  32    {
  33       if (write(fd, &w, sizeof(w)) == -1)
  34          goto fail;
  35    }
  36
  37    /* Determine the GPU type for CLRXdisasm. Use the family for GFX6 chips
  38     * because it doesn't allow to use gfx600 directly.
  39     */
  40    switch (program->chip_class) {
  41    case GFX6:
  42       switch (program->family) {
  43       case CHIP_TAHITI:
  44          gpu_type = "tahiti";
  45          break;
  46       case CHIP_PITCAIRN:
  47          gpu_type = "pitcairn";
  48          break;
  49       case CHIP_VERDE:
  50          gpu_type = "capeverde";
  51          break;
  52       case CHIP_OLAND:
  53          gpu_type = "oland";
  54          break;
  55       case CHIP_HAINAN:
  56          gpu_type = "hainan";
  57          break;
  58       default:
  59          unreachable("Invalid GFX6 family!");
  60       }
  61       break;
  62    case GFX7:
  63       gpu_type = "gfx700";
  64       break;
  65    default:
  66       unreachable("Invalid chip class!");
  67    }
  68
  69    sprintf(command, "clrxdisasm --gpuType=%s -r %s", gpu_type, path);
  70
  71    p = popen(command, "r");
  72    if (p) {
  73       while (fgets(line, sizeof(line), p))
  74          out << line;
  75       pclose(p);
  76    }
  77
  78 fail:
  79    close(fd);
  80    unlink(path);
  81 }
  82
  83 void print_asm(Program *program, std::vector<uint32_t>& binary,
  84                unsigned exec_size, std::ostream& out)
  85 {
  86    if (program->chip_class <= GFX7) {
  87       print_asm_gfx6_gfx7(program, binary, out);
  88       return;
  89    }
  90
  91    std::vector<bool> referenced_blocks(program->blocks.size());
  92    referenced_blocks[0] = true;
  93    for (Block& block : program->blocks) {
  94       for (unsigned succ : block.linear_succs)
  95          referenced_blocks[succ] = true;
  96    }
  97
  98    #if LLVM_VERSION_MAJOR >= 11
  99    std::vector<llvm::SymbolInfoTy> symbols;
 100    #else
 101    std::vector<std::tuple<uint64_t, llvm::StringRef, uint8_t>> symbols;
 102    #endif
 103    std::vector<std::array<char,16>> block_names;
 104    block_names.reserve(program->blocks.size());
 105    for (Block& block : program->blocks) {
 106       if (!referenced_blocks[block.index])
 107          continue;
 108       std::array<char, 16> name;
 109       sprintf(name.data(), "BB%u", block.index);
 110       block_names.push_back(name);
 111       symbols.emplace_back(block.offset * 4, llvm::StringRef(block_names[block_names.size() - 1].data()), 0);
 112    }
 113
 114    const char *features = "";
 115    if (program->chip_class >= GFX10 && program->wave_size == 64) {
 116       features = "+wavefrontsize64";
 117    }
 118
 119    LLVMDisasmContextRef disasm = LLVMCreateDisasmCPUFeatures("amdgcn-mesa-mesa3d",
 120                                                              ac_get_llvm_processor_name(program->family),
 121                                                              features,
 122                                                              &symbols, 0, NULL, NULL);
 123
 124    char outline[1024];
 125    size_t pos = 0;
 126    bool invalid = false;
 127    unsigned next_block = 0;
 128    while (pos < exec_size) {
 129       while (next_block < program->blocks.size() && pos == program->blocks[next_block].offset) {
 130          if (referenced_blocks[next_block])
 131             out << "BB" << std::dec << next_block << ":" << std::endl;
 132          next_block++;
 133       }
 134
 135       /* mask out src2 on v_writelane_b32 */
 136       if (((program->chip_class == GFX8 || program->chip_class == GFX9) && (binary[pos] & 0xffff8000) == 0xd28a0000) ||
 137           (program->chip_class >= GFX10 && (binary[pos] & 0xffff8000) == 0xd7610000)) {
 138          binary[pos+1] = binary[pos+1] & 0xF803FFFF;
 139       }
 140
 141       size_t l = LLVMDisasmInstruction(disasm, (uint8_t *) &binary[pos],
 142                                        (exec_size - pos) * sizeof(uint32_t), pos * 4,
 143                                        outline, sizeof(outline));
 144
 145       size_t new_pos;
 146       const int align_width = 60;
 147       if (!l &&
 148           ((program->chip_class >= GFX9 && (binary[pos] & 0xffff8000) == 0xd1348000) || /* v_add_u32_e64 + clamp */
 149            (program->chip_class >= GFX10 && (binary[pos] & 0xffff8000) == 0xd7038000) || /* v_add_u16_e64 + clamp */
 150            (program->chip_class <= GFX9 && (binary[pos] & 0xffff8000) == 0xd1268000)) /* v_add_u16_e64 + clamp */) {
 151          out << std::left << std::setw(align_width) << std::setfill(' ') << "\tinteger addition + clamp";
 152          bool has_literal = program->chip_class >= GFX10 &&
 153                             (((binary[pos+1] & 0x1ff) == 0xff) || (((binary[pos+1] >> 9) & 0x1ff) == 0xff));
 154          new_pos = pos + 2 + has_literal;
 155       } else if (program->chip_class >= GFX10 && l == 4 && ((binary[pos] & 0xfe0001ff) == 0x020000f9)) {
 156          out << std::left << std::setw(align_width) << std::setfill(' ') << "\tv_cndmask_b32 + sdwa";
 157          new_pos = pos + 2;
 158       } else if (!l) {
 159          out << std::left << std::setw(align_width) << std::setfill(' ') << "(invalid instruction)";
 160          new_pos = pos + 1;
 161          invalid = true;
 162       } else {
 163          out << std::left << std::setw(align_width) << std::setfill(' ') << outline;
 164          assert(l % 4 == 0);
 165          new_pos = pos + l / 4;
 166       }
 167       out << std::right;
 168
 169       out << " ;";
 170       for (; pos < new_pos; pos++)
 171          out << " " << std::setfill('0') << std::setw(8) << std::hex << binary[pos];
 172       out << std::endl;
 173    }
 174    out << std::setfill(' ') << std::setw(0) << std::dec;
 175    assert(next_block == program->blocks.size());
 176
 177    LLVMDisasmDispose(disasm);
 178
 179    if (program->constant_data.size()) {
 180       out << std::endl << "/* constant data */" << std::endl;
 181       for (unsigned i = 0; i < program->constant_data.size(); i += 32) {
 182          out << '[' << std::setw(6) << std::setfill('0') << std::dec << i << ']';
 183          unsigned line_size = std::min<size_t>(program->constant_data.size() - i, 32);
 184          for (unsigned j = 0; j < line_size; j += 4) {
 185             unsigned size = std::min<size_t>(program->constant_data.size() - (i + j), 4);
 186             uint32_t v = 0;
 187             memcpy(&v, &program->constant_data[i + j], size);
 188             out << " " << std::setw(8) << std::setfill('0') << std::hex << v;
 189          }
 190          out << std::endl;
 191       }
 192    }
 193
 194    out << std::setfill(' ') << std::setw(0) << std::dec;
 195
 196    if (invalid) {
 197       /* Invalid instructions usually lead to GPU hangs, which can make
 198        * getting the actual invalid instruction hard. Abort here so that we
 199        * can find the problem.
 200        */
 201       abort();
 202    }
 203 }
 204
 205 }