[LLVMdev] RFC: debug_line Emission

[Dan Gohman]

On Jul 17, 2008, at 3:33 PM, Bill Wendling wrote:

> In CodeGen/DwarfWriter.cpp's EmitDebugLine file, these lines are
> causing havoc on Mac OS X systems:
>
>    // If there are no lines to emit (such as when we're using .loc
> directives
>    // to emit .debug_line information) don't emit a .debug_line
> header.
>    if (SectionSourceLines.empty())
>      return;
>
> Basically, if there's a file with only data in it, we still need the
> debug_line prologue generated.
>
> $ cat a.c
> const char ver[] __attribute__((used)) = "Hello world\n";
> $ gcc -g -O0 -c a.c -o a.gcc.o
> $ dwarfdump --debug-line=0 a.gcc.o
> ----------------------------------------------------------------------
> File: a.gcc.o (architecture i386)
> ----------------------------------------------------------------------
>
> .debug_line contents:
> ----------------------------------------------------------------------
> debug_line[0x00000000]
> ----------------------------------------------------------------------
> Line table prologue:
>   total_length: 0x00000027
>        version: 0x0002
> prologue_length: 0x00000017
> min_inst_length: 0x01
> default_is_stmt: 0x01
>      line_base: -10
>     line_range: 245
>    opcode_base: 0x0a
> standard_opcode_lengths[ DW_LNS_copy             ] = 0
> standard_opcode_lengths[ DW_LNS_advance_pc       ] = 1
> standard_opcode_lengths[ DW_LNS_advance_line     ] = 1
> standard_opcode_lengths[ DW_LNS_set_file         ] = 1
> standard_opcode_lengths[ DW_LNS_set_column       ] = 1
> standard_opcode_lengths[ DW_LNS_negate_stmt      ] = 0
> standard_opcode_lengths[ DW_LNS_set_basic_block  ] = 0
> standard_opcode_lengths[ DW_LNS_const_add_pc     ] = 0
> standard_opcode_lengths[ DW_LNS_fixed_advance_pc ] = 1
>                Dir  Mod Time   File Len   File Name
>                ---- ---------- ---------- ---------------------------
> file_names[  1]    0 0x00000000 0x00000000 a.c
> 0x00000021: DW_LNE_set_address( 0x00000000 )
> 0x00000028: DW_LNE_end_sequence
>            0x0000000000000000      1      1      0
>
> By removing the "early exit" from EmitDebugLine, I got LLVM to
> generate this data. However, Dan pointed out that assemblers that use
> the ".loc" directive can't have their debug_line generated by the
> compiler. My suggestion was to have a flag that indicated that ".loc"
> was used. If it was, then we would exit out of the EmitDebugLine
> method early. Something like so:
>
>    // If there are no lines to emit (such as when we're using .loc
> directives
>    // to emit .debug_line information) don't emit a .debug_line
> header.
>    if (LocDirectiveUsed())
>      return;
>
> Dan wasn't sure if this would work in all cases. We don't have a Linux
> box to test this on. Does this sound like a good idea? Those who work
> with Linux or other OSes, would this idea break things for you?

I've now tested how GCC does this on Ubuntu, at least. It uses .loc and
does not emit a separate .debug_line prologue. It apparently assumes
that the .file numbers will directly correspond with file indicies in
the debug line section, suitable for use with DW_AT_decl_file.

So I think the way to fix this is to make it dependent on whether
the target uses .loc or not. Something like this:

   // If the target is using .loc/.file, the assembler will be emitting
   // the .debug_line table automatically.
   if (TAI->hasDotLocAndDotFile())
     return;

Dan