Re: [PATCH v2 2/7] BaseTools-GenFw:Add new x86_64 Elf relocation types for PIC/PIE code

[Ard Biesheuvel <ard.biesheuvel@linaro.org>]

On 2 August 2016 at 13:40, Shi, Steven <steven.shi@intel.com> wrote:
>>
>> CoffAddFixup() must be used for absolute symbol references only. These
>> instructions contain relative symbol references, which are
>> recalculated in WriteSections64().
>>
>> The only absolute symbol reference is the GOT entry for 'n', and your
>> code (in WriteRelocations64()) calculates the address of the GOT entry
>> (which is always in .text BTW) and adds a fixup for it, i.e.,
>>
>> +              CoffAddFixup(
>> +                (UINT32)(UINTN)((UINT64)
>> mCoffSectionsOffset[RelShdr->sh_info] + GoTPcRelPtrOffset),
>> +                EFI_IMAGE_REL_BASED_DIR64);
>>
>> This code adds a fixup to the PE/COFF .reloc section for the GOT entry
>> containing the address of 'n', and the instructions perform a IP
>> relative load of the contents of the GOT entry to retrieve the address
>> of 'n'.
>>
>> By adding two fixups, the PE/COFF loader will apply the load offset
>> twice, resulting in an incorrect value.
>>
> OK, I get your point now. Yes, the current patch could generate multiple fixups for the same GOT relocation entry. How about we introduce a simple IsDuplicatedCoffFixup() to check whether a converting fixup offset is duplicated before we use CoffAddFixup() to really add it? If it is new, we add it, otherwise just skip it.

That could work, but you have to be aware that fixups are best emitted
in the order they need to be applied in the binary, or it will become
very inefficient. (Please refer to the PE/COFF spec section that
explains the layout of the .reloc section)

What it comes down to is that relocations are grouped by target page,
and for every place in the page that requires a relocation to be
applied, a 4 bit type is emitted followed by a 12-bit offset, which is
the offset into the current page. If you emit fixups for the current
instruction, followed by one for the GOT, it will basically take two
'page switches' every time.

So it would be better to simply emit the relocations, but introduce a
sorting pass that merges all duplicates as well.

Thanks,
Ard.