[BFD][AARCH64]Properly truncate no overflow checking relocation value for load/store immediate.

[Renlin Li]

Ping~



On 04/27/2018 10:46 PM, Renlin Li wrote:
> Hi Peter,
> 
> Thanks for the comment!
> (I noticed that when I am working on the fix. But I forgot why I didn't do it there)
> 
> I update the patch by using PG_OFFSET when resolving the relocation.
> Regression test Okay.
> 
> Regards,
> Renlin
> 
> bfd/ChangeLog:
> 
> 2018-04-27  Renlin Li  <renlin.li@arm.com>
> 
>  Â Â Â Â * elfxx-aarch64.c (_bfd_aarch64_elf_resolve_relocation): Use PG_OFFSET
>  Â Â Â Â to resolve BFD_RELOC_AARCH64_TLSLD_LDST16_DTPREL_LO12_NC,
>  Â Â Â Â BFD_RELOC_AARCH64_TLSLD_LDST32_DTPREL_LO12_NC,
>  Â Â Â Â BFD_RELOC_AARCH64_TLSLD_LDST64_DTPREL_LO12_NC,
>  Â Â Â Â BFD_RELOC_AARCH64_TLSLD_LDST8_DTPREL_LO12_NC,
>  Â Â Â Â BFD_RELOC_AARCH64_TLSLE_LDST16_TPREL_LO12_NC,
>  Â Â Â Â BFD_RELOC_AARCH64_TLSLE_LDST32_TPREL_LO12_NC,
>  Â Â Â Â BFD_RELOC_AARCH64_TLSLE_LDST64_TPREL_LO12_NC,
>  Â Â Â Â BFD_RELOC_AARCH64_TLSLE_LDST8_TPREL_LO12_NC.
> 
> ld/ChangeLog:
> 
> 2018-04-27 Renlin Li  <renlin.li@arm.com>
> 
>  Â Â Â Â  * testsuite/ld-aarch64/emit-relocs-115.d: Update test with new value.
>  Â Â Â Â  * testsuite/ld-aarch64/emit-relocs-534.d: Likewise.
>  Â Â Â Â  * testsuite/ld-aarch64/emit-relocs-555.d: Likewise.
> 
> On 04/26/2018 03:53 PM, Peter Smith wrote:
>> Hello Renlin,
>>
>> I've checked that the patch works on an example I had that previously
>> gave different results to Gold and an in-progress implementation in
>> LLD. To the best of my knowledge the bitsizes in the HOWTO and the
>> method to truncate the result will work. One thing that I'm a bit
>> curious about is that relocations such as
>> BFD_RELOC_AARCH64_LDST64_LO12 use "value = PG_OFFSET (value +
>> addend);" to calculate the addend, whereas relocations like
>> BFD_RELOC_AARCH64_TLSLD_LDST32_DTPREL_LO12_NC use "value = value +
>> addend". In the former case the truncation is done in
>> _bfd_aarch64_elf_resolve_relocation whereas in the latter you are
>> having to truncate in _bfd_aarch64_elf_put_addend; it seems like you
>> might be able to use PG_OFFSET for all the LO12 relocations and not
>> need to put an extra truncation step in? I'm no expert in the BFD
>> codebase, so there may be good reasons not to do that so just treat
>> that as an observation.
>>
>> Thanks for the fix.
>>
>> Peter
>>
>> On 26 April 2018 at 14:05, Renlin Li <renlin.li@foss.arm.com> wrote:
>>> Hi all,
>>>
>>>
>>> In aarch64, there are relocations for the unsigned immediate offset of
>>> load/store instruction.
>>> The size of the immediate field is 12-bit. In the scaled case, not all
>>> 12-bit will be used.
>>>
>>> For example: R_AARCH64_LD64_GOT_LO12_NC
>>> Set the LD/ST immediate field to bits [11:3] of relocated value. No overflow
>>> check.
>>>
>>> In this case, the top 3 bits of the immediate field of the load instruction
>>> should
>>> be all 0. Only 9 bits from the value should be written into the instruction
>>> encoding.
>>>
>>> For relocations with over-flow check, if the value is larger than 12-bits,
>>> it
>>> will be caught by the checking mechanism.
>>>
>>> But for relocations without no overflow check, the value could be larger
>>> than
>>> 12 bits, after scaling, it might take more bits than the specification
>>> specified. The value should be properly truncated.
>>>
>>> The patch here correct the bitsize field of a few relocations, and the
>>> bitsize
>>> field is used to properly truncate the value written into the instruction
>>> encoding.
>>>
>>> Okay to commit?
>>>
>>> bfd/ChangeLog:
>>>
>>> 2018-04-26  Renlin Li  <renlin.li@arm.com>
>>>
>>> Â Â Â Â Â Â Â Â  * elfnn-aarch64.c (elfNN_aarch64_howto_table): Correct the bitsize
>>> Â Â Â Â Â Â Â Â  field of R_AARCH64_LD64_GOT_LO12_NC, R_AARCH64_P32_LD32_GOT_LO12_NC,
>>> Â Â Â Â Â Â Â Â  R_AARCH64_TLSIE_LD64_GOTTPREL_LO12_NC,
>>> Â Â Â Â Â Â Â Â  R_AARCH64_P32_LD32_GOTTPREL_LO12_NC,
>>> Â Â Â Â Â Â Â Â  R_AARCH64_LDST16_ABS_LO12_NC,
>>> Â Â Â Â Â Â Â Â  R_AARCH64_P32_TLSDESC_LD32_LO12_NC.
>>> Â Â Â Â Â Â Â Â  R_AARCH64_P32_TLSDESC_LD32_LO12_NC.
>>> Â Â Â Â Â Â Â Â  R_AARCH64_TLSDESC_LD64_LO12_NC.
>>> Â Â Â Â Â Â Â Â  R_AARCH64_LDST32_ABS_LO12_NC,
>>> Â Â Â Â Â Â Â Â  R_AARCH64_LDST64_ABS_LO12_NC,
>>> Â Â Â Â Â Â Â Â  R_AARCH64_LDST128_ABS_LO12_NC,
>>> Â Â Â Â Â Â Â Â  * elfxx-aarch64.c (_bfd_aarch64_elf_put_addend): Truncate the
>>> Â Â Â Â Â Â Â Â  immediate value for load/store instruction.
>>>
>>> ld/ChangeLog:
>>>
>>> 2018-04-26 Renlin Li  <renlin.li@arm.com>
>>>
>>> Â Â Â Â Â Â Â Â  * testsuite/ld-aarch64/emit-relocs-115.d: Update test with new
>>> value.
>>> Â Â Â Â Â Â Â Â  * testsuite/ld-aarch64/emit-relocs-534.d: Likewise.
>>> Â Â Â Â Â Â Â Â  * testsuite/ld-aarch64/emit-relocs-555.d: Likewise.
>