[lldb-dev] LTO debug info

[Philippe Lavoie via lldb-dev]

What made me suspect a data corruption issue were asserts triggered in the VC++ vector implementation when we use an LTO binary with a DEBUG lldb build on Windows.

   For example, 
   at source/Plugins/SymbolFile/DWARF/DWARFCompileUnit.cpp:624,

   File: C:\Program Files (x86)\Microsoft Visual Studio 12.0\VC\INCLUDE\vector
   Line: 240
   Expression: vector iterators incompatible
 
Digging deeper, the assertion is triggered by accesses to DWARFCompileUnit::m_die_array 

In the 'parser_fn' lambda in SymbolFileDWARF::Index() that indexes each compile unit in parallel, DWARFCompileUnit::ExtractDIEsIfNeeded(..) , DWARFCompileUnit::Index(..) and DWARFCompileUnit::ClearDIEs() are called in sequence.

'ExtractDIEsIfNeeded' and 'ClearDIEs' respectively populates and clears DWARFCompileUnit::m_die_array.

But when DWARFCompileUnit::Index(..) looks up a DIE in another CU and that CU is running 'ExtractDIEsIfNeeded' or 'ClearDIEs', it will access a stale or invalid DWARFCompileUnit::m_die_array ...

-Philippe

________________________________________
From: Greg Clayton [gclayton at apple.com]
Sent: Monday, April 25, 2016 7:59 PM
To: Philippe Lavoie
Cc: lldb-dev at lists.llvm.org
Subject: Re: [lldb-dev] LTO debug info

> On Apr 25, 2016, at 1:37 PM, Philippe Lavoie via lldb-dev <lldb-dev at lists.llvm.org> wrote:
>
> Hello,
>
> We noticed an issue with dwarf parsing when debugging a large application compiled with LTO.
>
> Since an LTO application's debug info can have inter-compile-unit references, (for example: a type defined in one compile-unit and used in another) we noticed that the "m_type_index" vector in SymbolFileDWARF can get corrupted. This is because many index structures in SymbolFileDWARF are generated in parallel.
>
> Is is a bug or is it expected to be the compiler's/linker's job to generate self-contained DWARF data without inter-CU references?
>
> -Philippe

TEach SymbolFileDWARF should generate its own accelerator tables in a self contained way. No SymbolFileDWARF should be adding references to their accelerator maps that actually refer to DIEs from another SymbolFileDWARF. We might need to look at the indexing function that make sure if they see and DW_FORM_ref_addr attributes, that they don't add any of them to their index. That seems like the bug?

Greg