DWARF line table rows with line 0 are dropped, so code gets attributed to the wrong line

[deadprogram]

While working on some debugger related fixes in TinyGo, I discovered what I think is an issue, and also a possible fix. The following is an edited version of the automated analysis.

Thank you for looking it over.

Analysis

When wasm-opt rewrites the DWARF line table, it drops every row that has line 0, even when no passes run. In DWARF, line 0 means the code has no source line, for example code LLVM merged from two places. Once the line 0 row is gone, the previous row covers that code, so debuggers and profilers attribute it to the wrong line.

The check is in LineState::needToEmit (wasm-debug.cpp):

bool needToEmit() {
  // Zero values imply we can ignore this line.
  // https://github.com/WebAssembly/debugging/issues/9#issuecomment-567720872
  return line != 0 && addr != 0;
}

The linked comment explains why address 0 is reserved for dead code removed by the linker. It does not mention line 0. The code is the same on main.

Repro

t.c:

int g(int);

int f(int x) {
  int r = 0;
  if (x > 10)
    r = g(x);
  else
    r = g(x + 1);
  return r * 3;
}

clang --target=wasm32 -O2 -g -c t.c -o t.o
wasm-ld --no-entry --export=f --allow-undefined t.o -o t.wasm
wasm-opt -g t.wasm -o out.wasm
llvm-dwarfdump --debug-line t.wasm
llvm-dwarfdump --debug-line out.wasm

LLVM merges the two calls to g and marks the call with line 0. Input line table:

Address            Line   Column File   ISA Discriminator OpIndex Flags
0x0000000000000002      3      0      1   0             0       0  is_stmt
0x0000000000000007      5      9      1   0             0       0  is_stmt prologue_end
0x0000000000000008      5      7      1   0             0       0
0x000000000000000b      0      0      1   0             0       0
0x0000000000000013      9     12      1   0             0       0  is_stmt
0x0000000000000014      9      3      1   0             0       0
0x0000000000000015      9      3      1   0             0       0  end_sequence

Output line table after wasm-opt -g with no passes:

Address            Line   Column File   ISA Discriminator OpIndex Flags
0x0000000000000002      3      0      1   0             0       0  is_stmt
0x0000000000000007      5      9      1   0             0       0  is_stmt prologue_end
0x0000000000000008      5      7      1   0             0       0
0x000000000000000f      9     12      1   0             0       0  is_stmt
0x0000000000000010      9      3      1   0             0       0
0x0000000000000011      9      3      1   0             0       0  end_sequence

The call is at 0x0b in both files. llvm-dwarfdump --lookup=0x0b gives line 0 for t.wasm and line 5, column 7 (the if condition) for out.wasm.

Tested with wasm-opt version 133 and clang 18.1.2 from wasi-sdk.

Impact

We found this in TinyGo, where a no-op round trip through wasm-opt -g removes all 21,756 line 0 rows from a small program. It gets worse with --asyncify, since code it adds in the middle of a function also ends up under the previous source line. See tinygo-org/tinygo#5736.

Possible fix

Keep line 0 rows and only skip rows with address 0. Line 0 rows would still need their address mapped like any other row.