[PATCH 2/2] riscv: Fix R_RISCV_IRELATIVE overwrite and order issues

[Nelson Chu]

On Thu, May 16, 2024 at 2:16 PM Hau Hsu <hau.hsu@sifive.com> wrote:

> Hi Nelson,
>
> Sorry for the late reply.
>
> On May 8, 2024, at 9:00 AM, Nelson Chu <nelson@rivosinc.com> wrote:
>
> Can you provide the testcase to show the case which I mentioned in the
> pr13302?  Which is that an ifunc with a dynamic jump slot calls another
> ifunc, and are not in the rel.dyn.
>
> I am not quite sure what's the issue you mentioned in PR13302.
> You mean the order issue of cause by one ifunc (generates
> JUMP_SLOT[32|64]) calls another JUMP_SLOT[32|64] ifunc?
>
> Since according to the testcase below, it seems no requirement to apply
> this fix though.
>
> The new test case uses two methods to call ifuncs:
> 1. Through a normal function call: PLT + GOT
> 2. Through a global function pointer: GOT only
>
> Without the fix, the relocation of the first method overwrites the
> second's.
> The relocation section of my test case would be:
>
> Relocation section '.rela.plt' at offset 0x94 contains 2 entries:
>  Offset     Info    Type                Sym. Value  Symbol's Name + Addend
> 000110dc  0000003a R_RISCV_IRELATIVE                 100c0
> 00000000  00000000 R_RISCV_NONE                      0
>
>
> With the fix, it becomes
>
> Relocation section '.rela.plt' at offset 0x94 contains 2 entries:
>  Offset     Info    Type                Sym. Value  Symbol's Name + Addend
> 000110e0  0000003a R_RISCV_IRELATIVE                 100c0
> 000110dc  0000003a R_RISCV_IRELATIVE                 100c0
>
>
So it seems like the overwrite problem, not the order problem we were
discussing in the pr13302...


> --- /dev/null
>> +++ b/ld/testsuite/ld-riscv-elf/ifunc-macro.s
>> @@ -0,0 +1,19 @@
>> +/* Define macros to handle similar behaviors for rv32/rv64.
>> +   Assumes macro "__64_bit__" defined for rv64.
>> +   The macro is specifically defined for ifunc tests in
>> ld-riscv-elf.exp. */
>> +
>> +.macro PTR_DATA name
>> +.ifdef __64_bit__
>> +       .quad   \name
>> +.else
>> +       .long   \name
>> +.endif
>> +.endm
>> +
>> +.macro LOAD rd, rs, offset
>> +.ifdef __64_bit__
>> +       ld      \rd, \offset (\rs)
>> +.else
>> +       lw      \rd, \offset (\rs)
>> +.endif
>> +.endm
>
>
Btw, can we not use these macroes?

Nelson
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://sourceware.org/pipermail/binutils/attachments/20240516/54b848c7/attachment.htm>