Re: MIPS textrel fix

[David Daney <ddaney@avtrex.com>]

Thiemo Seufer wrote:
> David Daney wrote:
> 
>>Daniel Jacobowitz wrote:
>>
>>>While stuck offline earlier today, I revisited the textrel-1 MIPS
>>>failure.  Eric originally tried setting DF_TEXTREL during section
>>>relocation, but it's too late: we add the entry in size_dynamic_sections.
>>>But I couldn't see any other way to get it right, since there's no
>>>hook to predict whether elf-eh-frame.c will eliminate a relocation.
>>>
>>>The easiest approach I found was to annul the DT_TEXTREL and DT_FLAGS
>>>changes in finish_dynamic_sections if no text relocations were
>>>actually generated.  This is not immensely pretty, but does work.
>>>OK?
>>>
>>>For background, the problem arises from the use of absolute addresses
>>>in .eh_frame.  In this case, they're being produced by gas CFI
>>>directives.  Is there a reason we can't mix and match encodings?
>>>i.e. why not have gas use a PC-relative format?  I'm sure there's
>>>a reason, but I can't think of it...
>>>
>>
>>Could it be related to the reason that GCC no longer generates a 
>>PC-relative .eh_frame?
>>
>>I never fully understood the reason, but I think it has to do with 
>>MIPS-ELF specifications not allowing the needed relocation types (even 
>>though they work well in binutils/glibc/linux).
> 
> 
> The needed relocation types were originally defined for embedded PIC
> MIPS and are not available in the ABI spec, their use for eh_frame
> was an accident. When the removal of embedded PIC support broke the
> toolchain, the fix chosen was to become ABI compatible again.
> 

Not to beat a dead horse, but it would be nice to have a gnu extension 
that allows us to have read-only eh_frame sections.  Large C++ shared 
libraries have way to many relocations.  It is a lot of dirty pages in 
the name of ABI compliance.

David Daney.