[PATCH] D23924: [ELF] - Linkerscript: allow add MergeInputSection to regular OutputSection.

George Rimar via llvm-commits llvm-commits at lists.llvm.org
Tue Aug 30 01:46:28 PDT 2016

>Looking at the list George provided in
>I think we have to make one of two hard choices on how we handle this
>if we are to claim to properly support linker scripts. What is
>happening is that initially the sections do *not* have the same name:
>.rodata (No merge) in 1/scsi_all.o
>.rodata.str1.1 (SHF_MERGE) in 1/scsi_all.o
>.rodata.cst16 (SHF_MERGE) in 1/scsi_all.o
>They have different names and each one has a consistent set of flags.
>The problem is that it is very common for a linker script to have
>.rodata         : { *(.rodata .rodata.*) }
>And we are still expected to merge the bits that can be merged. For
>example, including the attached file twice in a link causes the string
>and the constant 0x42 to be merged. Only 0x43 is duplicated.
>One way to fix this is to go back to producing multiple output
>sections. It is somewhat confusing, but I don't expect any major
>problems with it. We would still be able to handle things like
> .rodata  : {
>   foo = .;
>  *(.rodata .rodata.*)
>  bar = .;
>  }
>With the VA of foo being in one section and the VA of bar in another.

FreeBSD kernel does not use that. 
Probably we can postpone the support for such cases.

>The other option is to introduce another data structure in the middle
>of input and output sections (Chunk?). Input sections are mapped to
>chunks of various types (Merge, Eh, Plain) and those are mapped to
>output sections which are then just a list of chunks.
>It is my impression that the complexity of adding another level is not
>worth it and it should be ok to produce multiple output sections when
>SHF_MERGE doesn't match.

I would probably try to produce multiple output sections then.
My consern is that for non script case we currently have Alignment as a part of key for these sections:

  // For SHF_MERGE we create different output sections for each alignment.
  // This makes each output section simple and keeps a single level mapping from
  // input to output.
  uintX_t Alignment = 0;
  if (isa<MergeInputSection<ELFT>>(C))
    Alignment = std::max(H->sh_addralign, H->sh_entsize);

Following above logic we can have many of them.
That returns us to the point where 7 .rodata sections was produced by lld fro FreeBSD kernel.
So is it fine ?


More information about the llvm-commits mailing list