[PATCH] Have clang list the imported modules in the debug info

[David Blaikie]

On Wed, Mar 18, 2015 at 4:53 PM, Adrian Prantl <aprantl at apple.com> wrote:

>
> On Mar 18, 2015, at 4:41 PM, David Blaikie <dblaikie at gmail.com> wrote:
>
>
>
> On Wed, Mar 18, 2015 at 4:31 PM, Adrian Prantl <aprantl at apple.com> wrote:
>
>>
>> On Mar 18, 2015, at 4:02 PM, David Blaikie <dblaikie at gmail.com> wrote:
>>
>>
>>
>> On Wed, Mar 18, 2015 at 3:50 PM, Adrian Prantl <aprantl at apple.com> wrote:
>>
>>>
>>> On Mar 17, 2015, at 6:44 PM, David Blaikie <dblaikie at gmail.com> wrote:
>>>
>>>
>>>
>>> On Tue, Mar 17, 2015 at 3:47 PM, Adrian Prantl <aprantl at apple.com>
>>> wrote:
>>>
>>>>
>>>> > On Mar 17, 2015, at 10:03 AM, Greg Clayton <gclayton at apple.com>
>>>> wrote:
>>>> >
>>>> >
>>>> >> On Mar 17, 2015, at 9:46 AM, David Blaikie <dblaikie at gmail.com>
>>>> wrote:
>>>> >>
>>>> >>
>>>> >>
>>>> >> On Tue, Mar 17, 2015 at 9:42 AM, Greg Clayton <gclayton at apple.com>
>>>> wrote:
>>>> >>
>>>> >>> On Mar 16, 2015, at 6:47 PM, David Blaikie <dblaikie at gmail.com>
>>>> wrote:
>>>> >>>
>>>> >>>
>>>> >>>
>>>> >>> On Mon, Mar 16, 2015 at 5:14 PM, Adrian Prantl <aprantl at apple.com>
>>>> wrote:
>>>> >>>
>>>> >>> Thanks for the explanation David, I missed that it is entirely the
>>>> linker's (or some dwarf post-processor's) responsibility to find the module
>>>> files and link in the debug info from the .pcm files, so debugger doesn’t
>>>> notice a difference.
>>>> >>>
>>>> >>> I think there's still some confusion here. Sorry if I'm rehashing
>>>> something, but I'll try to explain how this all works.
>>>> >>>
>>>> >>> Normal split DWARF:
>>>> >>>
>>>> >>> Compiler generates two files: .o and .dwo.
>>>> >>> .dwo has static, non-relocatable debug info.
>>>> >>> .o has a skeleton compile_unit that has the name of the .dwo file
>>>> and a hash to verify that the .dwo file isn't stale when the debugger reads
>>>> it.
>>>> >>> The .o files are all linked together, the .dwo files stay where
>>>> they are.
>>>> >>> The debugger reads the linked executable, finds the skeleton
>>>> compile_units contained therein, and find/loads the .dwo files
>>>> >>>
>>>> >>> The scenario I have in mind for module debug info is this:
>>>> >>> Module is compiled as an object file with debug info (this file is
>>>> actually a .dwo file, even if it has some other extension - it has the
>>>> non-relocatable debug info in it)
>>>> >>> .o file has a comdat'd skeleton compile_unit describing the
>>>> .dwo/module file
>>>> >>> <from here on no extra work is required, the linker and debugger
>>>> just act as normal>
>>>> >>> The .o files are linked together, the skeleton compile_units get
>>>> deduplicated by the linker (comdat sections)
>>>> >>
>>>> >> One issue I can think of is we will need to figure out a way to make
>>>> COMDAT work with mach-o. COMDAT requires large number of sections and
>>>> mach-o can only have 255.
>>>> >>
>>>> >> Ah, fair enough - how does MachO handle inline functions (the most
>>>> common use of comdat) currently, then?
>>>> >
>>>> > Currently mach-o relies on symbols in the symbol table being marked
>>>> as weak and I believe the data for these symbols are in special sections
>>>> that are marked as containing items that can be coalesced.
>>>> >
>>>> That’s not necessarily an issue that needs to be solved on Darwin, or
>>>> am I maybe missing something? The linker leaves all debug info in the .o
>>>> (as it currently does) and llvm-dsymutil is resolving all the external
>>>> module type references while creating the .dSYM bundle.
>>>>
>>>
>>> Yeah, with a debug aware linker (or in the case of dsymutil, a
>>> debug-only linker) you would just know that since you're looking at object
>>> files, module references will be redundant across objects and should be
>>> deduplicated (by the dwo hash, most likely).
>>>
>>> If you're not teaching your debugger to read modules, and want to link
>>> the debug info in from the .dwos - at that point you can probably drop the
>>> skeleton stuff entirely (you'd still need to teach your debugger about .dwo
>>> sections and some of the esoteric things there - like str_index and the
>>> extra/special line table just for file names (decl_file, etc, uses this))
>>> and just put the contents of the module debug info straight in the dsym.
>>> It'd be a bit weird, but do-able without too much work, I'd imagine. You
>>> could move them back into the original sections, if you wanted to avoid the
>>> weird .dwo +non-.dwo sections together... *shrug* not sure what exactly
>>> you'd want there.
>>>
>>>
>>> My plan was to have -gmodules to behave like the latter variant
>>> unless -gsplit-dwarf is also present; this way there wouldn't be any weird
>>> Darwin-specific code paths.
>>>
>>
>> Not sure I quite follow (mostly my fault given the rambling paragraph up
>> there) - given the lack of a dsymutil-like tool on other platforms as part
>> of the common tool path for debug info, I'm not sure module debug info
>> without split dwarf is viable in that world. There's no tool to read these
>> extra files at any point.
>>
>>
>> In theory someone could port llvm-dsymutil to a different platform, but
>> that scenario is a little far-fetched. I’m not sure what will happen if
>> LLDB is presented with linked, non-split debug info that contains module
>> references.
>>
>
> Linked non-split debug info should come out for free - all the debug info
> would be is a bunch of TUs in a single comdat - no skeleton CU, nothing
> else. It would look just like normal DWARF, except with one comdat instead
> of multiple, for each set of types from a module. (& there would be no real
> size gains - since you'd be redundantly including all the type information
> in every object file)
>
>
>>
>>
>> I suppose we could be creating one giant comdat for the module's debug
>> info (no skeleton unit, no distinct type unit comdats, just one big
>> comdat). But we'd probably want/need a tool to do the merging at compile
>> time (like the objcopy feature for split-dwarf, but in reverse - we'd
>> compile, then run a tool to smoosh all the comdats from the modules onto
>> the object we just generated). It wouldn't provide much in the way of space
>> savings, a little less stress on the linker (fewer comdats to handle), etc.
>> Not sure if there's a default mode of objcopy that would cope with this
>> straight out, or whether we'd need a new feature there (which wouldn't be a
>> priority for Google to implement, since we use fission, nor a priority for
>> you to implement since you have dsymutil, etc - so I'm not sure anyone
>> would bother)
>>
>> Long story short: maybe just error on -gmodules if -gsplit-dwarf isn't
>> specified or the platform isn't darwin? (& if it's darwin, dsymutil could
>> read the module skeletons to find which modules to link into the .dSYM?)
>>
>>
>> That’s reasonable, too :-)
>> The plan is for llvm-dsymutil to follow the references in the module
>> skeletons, copy the module CUs
>>
>
> TUs for now
>
>
>> into the .dSYM, and fixup the external type references to become
>> DW_FORM_ref_addrs.
>>
>
> Sounds good for you guys - the fixup work will be a bit non-trivial, since
> it'll need to remove the type skeletons in the CUs, move all the extra
> members from the skeletons into the type unit (& resolve any duplicates),
> etc... - does that make sense? (otherwise I can provide some DWARF snippets
> to explain better)
>
>
> Or we use a weird Darwin-specific code path to not emit the modules with
> -generate-type-units in the first place (bag of DWARF+index mapping hash to
> DIE),
>

bag-o-dwarf still doesn't address all the issues with type member merging I
described above. Certain things can't go in the type in the module because
they depend on context - most importantly/obviously, implicit special
members and member function template instatiations.

I suppose you could still have type references reference the type in the
bag-o-dwarf/type unit directly (DW_AT_type with DW_FORM_ref_sig8) while
having the partial type (the type declaration with its extra CU-specific
members) which would simplify the dwarf in the easy cases.


> which would make dsymutil's job really easy. As much as I’d like to get
> rid of platform-specific behavior, due to the automatic way that modules
> are generated on Darwin I don’t see an elegant way of making this
> switchable by the user.
>

Not sure I quite follow here how implicit modules impact this
functionality. We can still have a flag that you pass to the compiler that
dictates how debug info in modules is created/what schema we use.

- David


>
> -- adrian
>
>
>
>>
>> -- adrian
>>
>
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/cfe-commits/attachments/20150318/9b4793a7/attachment.html>