LLVM Bugzilla is read-only and represents the historical archive of all LLVM issues filled before November 26, 2021. Use github to submit LLVM bugs

Bug 45370 - wrong/misleading "SHF_MERGE section size must be a multiple of sh_entsize"
Summary: wrong/misleading "SHF_MERGE section size must be a multiple of sh_entsize"
Status: RESOLVED FIXED
Alias: None
Product: lld
Classification: Unclassified
Component: ELF (show other bugs)
Version: unspecified
Hardware: All All
: P normal
Assignee: Unassigned LLVM Bugs
URL:
Keywords:
Depends on:
Blocks:
 
Reported: 2020-03-31 02:57 PDT by Andrew Gierth
Modified: 2020-04-03 15:03 PDT (History)
4 users (show)

See Also:
Fixed By Commit(s):

Attachments
trivial reproducer (5.00 KB, application/octet-stream)
2020-04-02 21:14 PDT, Andrew Gierth 		Details
Add an attachment (proposed patch, testcase, etc.)

Note You need to log in before you can comment on or make changes to this bug.
Description Andrew Gierth 2020-03-31 02:57:53 PDT 
Corresponding FreeBSD bug linked as "see also"; see that bug's own "see also" link for where this came from.

In ObjFile<ELFT>::shouldMerge in lld/ELF/InputFiles.cpp:

  if (sec.sh_size % entSize)
    fatal(toString(this) + ":(" + name + "): SHF_MERGE section size (" +
          Twine(sec.sh_size) + ") must be a multiple of sh_entsize (" +
          Twine(entSize) + ")");

  uint64_t flags = sec.sh_flags;
  if (!(flags & SHF_MERGE))
    return false;

Notice that the size is being checked _before_ the SHF_MERGE flag is checked. This means that either the logic is incorrect, or the error message is incorrect (since this error will then occur for sections without SHF_MERGE).

If this size check is only needed for mergeable sections, then surely it should be made _after_ SHF_MERGE is checked for.

Since this check is skipped when -O0 is in effect, I do not believe it is actually necessary to check that size is a multiple of sh_entsize when _not_ merging sections. If I'm wrong about that, though, then the check would need to be both moved to somewhere else and changed to not mention SHF_MERGE in the error message.
Comment 1 Andrew Gierth 2020-04-02 21:14:22 PDT 
Created attachment 23306 [details]
trivial reproducer

Added a reproducer. The .o file here is the output of `as --gstabs` (using the freebsd 12-stable base system version of `as`, which admittedly is very old, but the error is exactly the same when using a current binutils version) applied to the following trivial input:

        .text
        .align 16
        .global _start
_start:
        movl $0x01, %eax
        xorq %rdi,%rdi
        syscall
        int3
Comment 2 Fangrui Song 2020-04-02 22:23:53 PDT 
https://reviews.llvm.org/D77368(In reply to Andrew Gierth from comment #1)
> Created attachment 23306 [details]
> trivial reproducer
> 
> Added a reproducer. The .o file here is the output of `as --gstabs` (using
> the freebsd 12-stable base system version of `as`, which admittedly is very
> old, but the error is exactly the same when using a current binutils
> version) applied to the following trivial input:
> 
>         .text
>         .align 16
>         .global _start
> _start:
>         movl $0x01, %eax
>         xorq %rdi,%rdi
>         syscall
>         int3

https://sourceware.org/bugzilla/show_bug.cgi?id=25768
The issue has existed since "19990502 sourceware import".

Created https://reviews.llvm.org/D77368
Comment 3 Fangrui Song 2020-04-03 15:03:27 PDT 
commit 56decd982dc03a74d1796d9d4dbd7d9e0cea98dc (will be included in lld 11)

GNU as 2.35 will hopefully be fixed in 2.35 if my commit is merged.