Non-zero RELA section contents

[ghost]

@Hello71 reported that jwasm-family assemblers (jwasm, asmc, uasm) generate wrong relocation.

See also:
https://sourceforge.net/p/sevenzip/bugs/2338/
https://sourceware.org/pipermail/binutils/2022-April/120428.html

[Baron-von-Riedesel]

They're using rather outdated versions of jwasm - current release is v2.15.
Relocation generation was slightly updated in jwasm in recent versions, so jwasm may differ from uasm/asmc now...

[nidud]

GCC seems to add offset to the relocation and subtract it from the code.

https://github.com/nidud/asmc/issues/22

Added a hack for this by adding the offset and simply clearing the output.

Test case for v2.15:

.data
 a db 3
 d db 1,2
.code
    mov al,a
    mov al,d
    mov al,d[1]
    end
...
mov al, byte ptr [a]    ; 0000 _ 8A. 05, 00000000(rel)
mov al, byte ptr [d]    ; 0006 _ 8A. 05, 00000000(rel)
mov al, byte ptr [d+1H] ; 000C _ 8A. 05, 00000001(rel)
...
0000000000000002 R_X86_64_PC32     a-0x0000000000000004
0000000000000008 R_X86_64_PC32     d-0x0000000000000004
000000000000000e R_X86_64_PC32     d-0x0000000000000004

Asmc:

mov al, byte ptr [a]    ; 0000 _ 8A. 05, 00000000(rel)
mov al, byte ptr [d]    ; 0006 _ 8A. 05, 00000000(rel)
mov al, byte ptr [d+1H] ; 000C _ 8A. 05, 00000000(rel)
...
0000000000000002 R_X86_64_PC32     a-0x0000000000000004
0000000000000008 R_X86_64_PC32     d-0x0000000000000004
000000000000000e R_X86_64_PC32     d-0x0000000000000003

[Hello71]

the difference between rel and rela is that in rela, the addend is stored in the relocation table. the instruction stream contents are "unspecified" but i think everyone agrees that they should be set to zero. it's not gcc, it's x86_64 elf platforms (actually binutils, not gcc).

[nidud]

I guess it's more of a "correction" as the numbers actually adds up but there's an additional variable that fails:

lea rax,$