[Bug ld/31179] RISC-V: The SET/ADD/SUB fix breaks ABI compatibility with 2.41 objects

[palmer at gcc dot gnu.org]

https://sourceware.org/bugzilla/show_bug.cgi?id=31179

--- Comment #6 from Palmer Dabbelt <palmer at gcc dot gnu.org> ---
(In reply to Andreas Schwab from comment #5)
> There are a lot of SUB_ULEB128 relocations with a non-zero addend in
> installed static archives that will change behaviour with this commit.
> 
> $ readelf -Wr /usr/lib64/libc.a | grep -c 'SUB_ULEB128.*[+-] [^0]'
> 17043
> $ readelf -Wr /usr/lib64/gcc/riscv64-suse-linux/13/libgphobos.a | grep -c
> 'SUB_ULEB128.*[+-] [^0]'
> 55318

Awesome, thanks.  Nelson and I were getting some confusing error reports, but
this matches what I was trying to fix at the beginning.  We might have another
bug, but I think this one we can try to fix.

So I think we can do something like the following:
* Check for binaries with the bug -- I think we want a tag here, but I guess we
could also try and pattern match the +N/+N vs +M/+0.
* If the binary has the bug, then revert to the old linker behavior for
calculating the addresses from the input.
* Always emit correct addends into the output.

Then we can add some some warning when we detect the old inputs, along with
some commandline/autoconf arguments to turn the warnings into errors.

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