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