[PATCH] D74169: [WIP][LLD][ELF][DebugInfo] Skeleton implementation of removing obsolete debug info.

Mon Feb 24 14:29:26 PST 2020

dblaikie added a comment.

>>>> I mean, it'd be probably quite practical to extend it to be able to be done as a standalone/post-processing step.
>>> 
>>> That would be great. As accelerator tables evolve and get updated, it would be nice to have a path to update older accelerator tables to newer more complete versions.
>>> 
>>>> For the sake of object and linked executable size, I suspect we might end up trying to move the indexing out of the object files/linked executable - perhaps putting the names either in a third file, or in the .dwo files and generating an index from them instead of from the objects. Not sure. Some further experimentation expected there.
>>> 
>>> That is fine too. We make a dwarf linking tool that can link all of the .dwo files back into a normal DWARF file like with dsymutil and .o files on mac.
>> 
>> That's already done - DWP ( https://gcc.gnu.org/wiki/DebugFissionDWP ). But for an interactive scenario over a distributed filesystem shipping the whole linked DWP file back to the user for access by the debugger is a bit unfortunate/slow. Good for archival purposes, as you say - but it's still a bit of an open question as to what to do for interactive users if object file/linked executable size is a major concern - either generating a third file with just the index parts (so building the full index can be done in parallel/distributed separately from the link) or doing an index-only link of the .dwo files (not super great - still means a bottleneck of shipping all the .dwo files to one build machine, etc)...
> 
> For the linking, I was more thinking of converting the DWP format back into vanilla DWARF with no DWO/DWP artifacts (skeleton DIEs etc) or any other redirections (string table offsets, address tables, etc). Not sure if this would reduce the DWARF size of all of the DWO/DWP instrstructure was all removed?

Ah, sorry, I see what you mean. Well, the offsets systems in some cases save total size even once linked (eg: having a single 64 bit address with two short uleb indexes compared to two 64 bit addresses - even once the binary is linked and there's no extra bytes of relocation records for that address/addresses) & these features are no longer Split DWARF specific - Clang/LLVM generates almost (entirely, maybe?) all these constructs the same way with split or unsplit DWARF when emitting DWARFv5.

Might save a bit - certainly something someone could experiment with as improvements to DWARFLinker - for dsymutil, lld, and maybe dwp in the future.

(one of the cheap ones that'd be neat to see would be debug_rnglists optimization - wouldn't need to parse DWARF DIEs, etc, would be an entirely local optimization - may be useful for PROPELLER (~= -fbasic-block-sections) - because of the indirection in debug_rnglists (though the rnglist header's offset table) you could totally rewrite range lists to coalesce any contiguous ranges that end up together & rewriting the offsets in the offset table - though this optimization would benefit from being in both lld and dsymutil I think - not applicable to dwp unless address information is passed from the linker like dsymutil)

Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D74169/new/

https://reviews.llvm.org/D74169