Preserve metadata w/ Solaris-like linkers.

[luqmana]

#84468 moved the -zignore linker flag from the gc_sections method to add_as_needed which is more accurate but Solaris-style linkers will also end up removing an unreferenced ELF sections [1]. This had the unfortunate side effect of causing the .rustc section (which has the metada) to be removed which could cause issues when trying to link against the resulting crates or use proc macros.

Since the -zignore is positional, we fix this by moving the metadata objects to before the flag.

[1] Specifically a section is considered unreferenced if:

• The section is allocatable
• No other sections bind to (relocate) to this section
• The section provides no global symbols

https://docs.oracle.com/cd/E19683-01/817-3677/6mj8mbtbs/index.html#chapter4-19

[rust-highfive]

r? @matthewjasper

(rust-highfive has picked a reviewer for you, use r? to override)

[luqmana]

r? @petrochenkov

[petrochenkov]

@luqmana
Do you have access to a Solaris machine for testing? (Which version?)
The link in the original post is very old, and more up-to-date documentation (Solaris 11.4, https://docs.oracle.com/cd/E37838_01/html/E36783/man-rum.html) says that -z ignore/-z record only affect dynamic libraries ( –z discard-unused=dependencies, --as-needed-style), but not sections (–z discard-unused=sections, --gc-sections-style).

UPD: Looks like the old doc also talks about dynamic libraries

An object built with the -z ignore option, will have no unused dependencies recorded in it.

, it's just clarified better in later versions.

[petrochenkov]

Some relevant PRs:

• rustc: Store metadata-in-rlibs in object files #84449 changes how metadata is kept
• linker: Reorder linker arguments #85086 changes order of linker arguments in a more systematic way

[petrochenkov]

If -z ignore acts as --gc-sections and have effect on unused sections (to which I find no confirmation in docs), then perhaps it's better to wrap metadata only into -z record instead?

[luqmana]

Not solaris exactly but I do have access to an illumos system which is where I ran into the issue. From man ld:

       -z ignore | record

           Ignores, or records, dynamic dependencies that are not referenced
           as part of the link-edit. Ignores, or records, unreferenced ELF
           sections from the relocatable objects that are read as part of the
           link-edit. By default, -z record is in effect.

           If an ELF section is ignored, the section is eliminated from the
           output file being generated. A section is ignored when three
           conditions are true. The eliminated section must contribute to an
           allocatable segment. The eliminated section must provide no global
           symbols. No other section from any object that contributes to the
           link-edit, must reference an eliminated section.

[luqmana]

I'm fine with alternatively wrapping the metadata object with -z record <metadata.o> -z ignore but I just initially moved it a bit above to minimize the linker invocation a bit.

[petrochenkov]

Looks like I've been looking at a wrong doc all this time, the illumos doc (https://illumos.org/man/1/ld) explicitly says that -z ignore means both --as-needed and --gc-sections, which is different from classic Solaris.

[luqmana]

Definitely sure this applies to illumos, but I'm pretty sure this applies to Solaris itself too:

https://github.com/kofemann/opensolaris/blob/80192cd83bf665e708269dae856f9145f7190f74/usr/src/cmd/sgs/libld/common/sections.c#L164-L182

If -zignore has been in effect, scan all input files to determine if the file, or sections from the file, have been referenced. If not, the file or some of the files sections can be discarded.

[petrochenkov]

cc @jclulow

[petrochenkov]

One more idea (from #84449) is to mark the sections somehow (e.g. as debuginfo sections) so they are not thrown away when necessary.
(Given that #84449 is probably going to land, its compatibility with Solaris/illumos needs to be considered anyway.)

[luqmana]

One more idea (from #84449) is to mark the sections somehow (e.g. as debuginfo sections) so they are not thrown away when necessary.
(Given that #84449 is probably going to land, its compatibility with Solaris/illumos needs to be considered anyway.)

I can definitely give #84449 a try to make sure it works fine on illumos. But I don't think it addresses the issue in this PR as it seems to be focused on metadata in rlibs whereas this issue specifically is for dylibs/proc-macros. (See write_compressed_metadata which hasn't been updated in that PR. Edit: I see @bjorn3 has brought that up already.)

[petrochenkov]

Sorry for the delay, busy week.
@bors r+

I made myself an OpenIndiana virtual machine today and it seems to work, so I'll be able to experiment with fine-grained -z ignore/-z record to properly support as-needed modifier and -Clink-dead-code on Solaris (while being compatible with linker argument ordering from #85086).
In the meantime let's land this to fix dylibs/proc-macros.

[bors]

📌 Commit cffef33 has been approved by petrochenkov