[Bug gold/31415] New: Section coalescing should ignore (and merge) SHF_GNU_RETAIN

[address@hidden]

https://sourceware.org/bugzilla/show_bug.cgi?id=31415

            Bug ID: 31415
           Summary: Section coalescing should ignore (and merge)
                    SHF_GNU_RETAIN
           Product: binutils
           Version: 2.43 (HEAD)
            Status: UNCONFIRMED
          Severity: normal
          Priority: P2
         Component: gold
          Assignee: ccoutant at gmail dot com
          Reporter: alastair@alastairs-place.net
                CC: ian at airs dot com
  Target Milestone: ---

In layout.cc, the code avoids combining sections whose flags differ, while
ignoring `SHF_WRITE` and `SHF_EXECINSTR`.  (This seems an odd choice, frankly;
those are the flags I probably *wouldn't* want to ignore.)

Unfortunately, it doesn't ignore `SHF_GNU_RETAIN`; this is a particularly
awkward problem because programs tend to use `SHF_GNU_RETAIN` when they're
storing data in a section and using the linker-generated `__start` and `__stop`
symbols to access it.  Since the latter are not regarded as GC roots, the
retain flag is needed to stop the linker discarding the section, *but* if any
object file in the link doesn't have `SHF_GNU_RETAIN` set for that section,
then `gold` will generate another output section, with the same name, but of
course there's only one `__start` and `__stop` symbol pair and it only points
at a single section.

It seems to me that:

1. `gold` should ignore `SHF_GNU_RETAIN` when looking up sections to coalesce,
but

2. If it finds a section with `SHF_GNU_RETAIN` and coalesces it with a section
that did not have that flag, the resulting section *should* have the flag set. 
i.e. setting it on *any* input section that gets coalesced into an output
section should cause it to be set for the output section.

-- 
You are receiving this mail because:
You are on the CC list for the bug.