Re: [AArch64] PR18668, Generate long branch veneer if call to plt stub is out of range

[Jiong Wang <jiong dot wang at arm dot com>]

Jiong Wang writes:

> (add the patch)
>
> If text section size is really big, when the linker redirect function
> call to plt stub, the branch offset may be out of range.
>
> Currently, AArch64 BFD has long branch veneer infrastructure ready, but
> it's not enabled for call to plt stub. Only absolute call long branch
> implemented while for PIC/.so, call to PLT stub is the common case and
> we need to support this.
>
> This patch copied some code from ARM32, and cleaned up code logic in
> final_link code where BFD_RELOC_AARCH64_JUMP26/CALL26 handled.
>
> The code logic for handling CALL26/JUMP26 is simplied into:
>
>   * if it call via plt stub, then update relocation value to plt stub address.
>   * use the updated relocation value to check whether it's fit into range.
>     And if the offset don't fit into the range then check whether any
>     long branch veneer register for it.
>   * if long branch veneer register, then udpate final relocation value
>     to the address of veneer, otherwise don't touch relocation value.
>
> test done
> ===
>   native ld check OK.
>   linking the huge OpenLoop reported on PR 18668 successfully which
>   failed on the old bfd linker.
>   
>
> OK for trunk?
>
> 2015-07-15  Jiong Wang  <jiong.wang@arm.com>
> bfd/
>   PR ld/18668
>   * elfnn-aarch64.c (aarch64_type_of_stub): Update destination for
>   calls go through plt stub.
>   (elfNN_aarch64_final_link_relocate): Adjust code logic for CALL26,
>   JUMP26 relocation to support inserting veneer for call to plt stub.
>
> ld/testsuite/
>   * ld-aarch64/farcall-b-gsym.s: New test.
>   * ld-aarch64/farcall-b-plt.s: Ditto.
>   * ld-aarch64/farcall-bl-plt.s: Ditto.
>   * ld-aarch64/farcall-b-gsym.d: New expect file.
>   * ld-aarch64/farcall-b-plt.d: Ditto.
>   * ld-aarch64/farcall-bl-plt.d: Ditto.
>
> OK for trunk?

Ping~

I understand this patch modify some sensitive part of relocation
handling.

But the only functional changes of this patch are

  * in aarch64_type_of_stub, if the branch relocation is plt stub
    related, then we should calculate the distance by

      distance = branch location - PLT stub location

    instead of
    
      distance = branch location - symbol value.

  * "if (h && h->type == STT_FUNC)" in CALL/JMP26 relocation handling
    removed, because it's unncessary.

    if there is no branch stub, then it's not registered in the hash
    table, then the later get_sub_entry will return NULL. we should not
    differenciate local, global and symbol type here, as that should be
    done in stub type check and stub registration.

-- 
Regards,
Jiong