Re: Using section flags to indicate stripable or persistent sections

[Fangrui Song <i@maskray.me>]

On 2022-11-07, Michael Matz via Gnu-gabi wrote:
>Hey,
>
>On Mon, 7 Nov 2022, Nick Clifton via Gnu-gabi wrote:
>
>>     We would like to suggest an extension the ELF section flags which can be
>>     used to indicate sections that should, or should not, be stripped when
>>     removing debug information.
>
>Okay, so stripping of sections intended for debug purposes ...
>
>>     The problem we are trying to address is that different stripping tools
>>     (strip, eu-strip, llvm-strip) have different heuristics for deciding
>>     which sections should be removed when stripping debug information.  In
>>     order to fix this we are proposing two new section flags:
>>
>>       GNU_SHF_CAN_BE_STRIPPED
>>       GNU_SHF_DO_NOT_STRIP
>
>... but here you have "can be stripped?".  The immediate question is: for
>which purposes?  A section might be strippable in other context than debug
>info.  And other sections might be non-strippable for still other
>purposes.  Just saying "please don't strip this one" conveys no useful
>information, except of course to inhibit each and all stripping.  But was
>it perhaps set for other reasons than debug info?  Even with those flags
>tools will continue to wonder if maybe this-and-that DO_NOT_STRIP option
>can actually be stripped in some cases.
>
>IOW: if you really want what you said, make the flags more specific (to
>debug info purposes).  But then of course it's unclear why you wouldn't
>want still other flags to mean "strip-for-something-else" purposes.  At
>which point you quickly come to the conclusion that one wants to specify
>some sort of type per section, and not lump everything under SHT_PROGBITS.
>
>So, I think two flag bits are not the right solution.  Maybe we could use
>some bits of sh_type to specify a more detailed type: the low bits would
>just be set to the ELF SHT_xxx as appropriate, and higher bits would
>contain a detailed type:
>
>  0x60XXYYZZ
>
>(ZZ would be the old SHT_xxx type, and XX/YY be a detail type).
>
>
>Ciao,
>Michael.

Agree that the different purposes can be difficult to capture with just
one or two bits. As a concrete example, llvm-strip/llvm-objcopy has this
piece of logic for various --strip-* modes: --strip-dwo, --strip-debug,
--strip-sections, --strip-all-gnu, --strip-all
https://github.com/llvm/llvm-project/blob/c997fe658655b42237f556c11c7f5bc8ec60658b/llvm/lib/ObjCopy/ELF/ELFObjcopy.cpp#L380-L446

--strip-all-gnu exists because it is determined GNU strip's default
--strip-all doesn't strip all :( https://github.com/llvm/llvm-project/commit/6ad72d05f5ebaada700487d3d0fe1d107b32468a

I feel that whether .ctf is strippable in various strip modes may be a
related question. (I tend to suggest a more intuitive rule instead of
doing some magic to ease the adoption of a new section. Let users do
--keep-section=xxx for new sections.)

>>
>>     These would be set by the assembler and/or linker to indicate sections
>>     that should be removed when stripping and sections which must not be
>>     removed when stripping.  It would be an error if both flags were present
>>     on a given section, and if neither flag is present then the stripping
>>     tool would fall back on its built in heuristics.
>>
>>     In addition we need new flags for the assembler's .section directive
>>     (suggestion: 'D': can be stripped, 'K' do not strip).
>>     This email is to ask if you think that this idea has merit, and if so,
>>     are there any guidelines for writing and submitting a formal specification
>> ?
>>
>>  Cheers
>>     Nick
>>