[PATCH] Change MIPS linker stubs to allow for more than 2^15 symbols.

[David Daney]

Thiemo Seufer wrote:
> David Daney wrote:
> 
>>As noted in this thread:
>>
>>http://sourceware.org/ml/binutils/2006-06/msg00086.html
>>
>>The mips linker would not reliably allow for more than 2^15 dynamic 
>>symbols.  This prevents GCC's libgcj java runtime library from being 
>>correctly linked.
>>
>>The consensus was to increase the size of the linker stubs by one 
>>instruction so that objects with many dynamic symbols could be supported.
>>
>>I increased the size of the stub from 16 to 20 bytes.
> 
> 
> It may be useful to use 24 bytes, this adheres to the ABI alignment
> rule of 2^3.

Which rule would that be?  Is there such a requirement for function 
entry points?

> 
> 
>>The new stubs 
>>support up to 2^31 dynamic symbols. That should be enough for at least 
>>the short term.  With the patch applied I can now successfully build and 
>>run libgcj from a recent GCC-4.2 snapshot on mipsel-linux.
> 
> 
> Hm, 2^32 symbols would nicely fit the maximum for NewABI.

I don't know mips64 well enough to know how to load an unsigned 32 bit 
constant.  So It had to be 2^31.


> 
> [snip]
> 
>>*** 631,642 ****
>>     ((ABI_64_P (abfd)						\
>>       ? 0x03e0782d		/* daddu t7,ra */		\
>>       : 0x03e07821))		/* addu t7,ra */
>>  #define STUB_JALR 0x0320f809	/* jalr t9,ra */
>>! #define STUB_LI16(abfd)                                         \
>>!   ((ABI_64_P (abfd)						\
>>!    ? 0x64180000			/* daddiu t8,zero,0 */		\
>>!    : 0x24180000))		/* addiu t8,zero,0 */
>>! #define MIPS_FUNCTION_STUB_SIZE (16)
>>  
>>  /* The name of the dynamic interpreter.  This is put in the .interp
>>     section.  */
>>--- 631,640 ----
>>     ((ABI_64_P (abfd)						\
>>       ? 0x03e0782d		/* daddu t7,ra */		\
>>       : 0x03e07821))		/* addu t7,ra */
>>+ #define STUB_LUI 0x3c180000     /* lui t8,0 */
>>  #define STUB_JALR 0x0320f809	/* jalr t9,ra */
>>! #define STUB_LI16 0x34180000	/* ori t8,zero,0 */
>>! #define MIPS_FUNCTION_STUB_SIZE (20)
> 
> 
> Why not the conventional expansion of lui/(d)addiu?

We are not using hi/lo relocations, so there is no requirement for this. 
  The calculation in the linker is easier this way.

> 
> [snip]
> 
>>*** 8013,8027 ****
>>  				   MIPS_ELF_STUB_SECTION_NAME (dynobj));
>>        BFD_ASSERT (s != NULL);
>>  
>>-       /* FIXME: Can h->dynindx be more than 64K?  */
>>-       if (h->dynindx & 0xffff0000)
>>- 	return FALSE;
>>- 
> 
> 
> When it is 2^31, shoudn't we continue to test for overflow?
> 
Probably, but I got lazy.

Please comment on my responses, and then I will generate a new patch.

David Daney.