ld.lld -z start-stop-gc (GC of C identifier name sections)

[MaskRay]

https://lists.llvm.org/pipermail/llvm-dev/2021-February/148682.html I am thinking whether we can change the existing ld.lld --gc-sections rule about C identifier name sections.
https://maskray.me/blog/2021-01-31-metadata-sections-comdat-and-shf-link-order has more information. You may skip the paragraphs before "C identifier name sections".

A C identifier name section consists of isalnum characters and _. . is disallowed.

GNU ld's rule is

__start_/__stop_ references from a live input section retains all C identifier name sections.

(--gc-sections is based on input sections. A live input section indicates that the section is retained, due to a relocation chain from a GC root).

In LLD<=12, the rule is:

__start_/__stop_ references from a live input section retains all non-SHF_LINK_ORDER C identifier name sections.

(Example SHF_LINK_ORDER C identifier name sections: __patchable_function_entries (-fpatchable-function-entry), __sancov_guards (clang -fsanitize-coverage=trace-pc-guard, before clang 13))

In LLD>=13, the default rule is:

__start_/__stop_ references from a live input section retains all non-SHF_LINK_ORDER non-SHF_GROUP C identifier name sections.

(SHF_GROUP is to allow more __llvm_prf_cnts/__llvm_prf_data sections to be GCable.)

With ld.lld -z start-stop-gc (requires https://reviews.llvm.org/D96914), the (unhelpful in many cases) rule will be dropped entirely:

__start_/__stop_ references do not retain C identifier name sections.

i.e. All C identifier name sections will be GCable.

(Note, a KEEP(...) command in linker scripts makes all the matched input sections GC roots.)

HAVE_LD_DEAD_CODE_DATA_ELIMINATION (--gc-sections) is currently not very common, but the adoption is rising. The kernel has a large number of __start_/__stop_ sections, so understanding whether ld.lld -z start-stop-gc can cause breakage will be useful. We probably need a thorough test.

If a link has different sizes with ld.lld -z start-stop-gc, this means some C identifier name sections (and other sections reachable via relocation chain) can now be GCed.

• If nothing breaks => cool, we are now able to GC some unnecessarily retained sections.
• If something breaks => we should think whether we should use
  • KEEP(...) command in a linker script => no-brainer but it can unnecessarily retain sections which are otherwise unneeded.
  • Add a relocation from the user section to the metadata section (preferably R_*_NONE).

[nickdesaulniers]

@MaskRay is there anything more you'd like to do here? I'm going to close this out pre-emptively, but please do reopen with more info if you have more thoughts here.