hppa-linux TLS work, extra DIR32 reloc for errno@@GLIBC_PRIVATE?

[Carlos O'Donell]

binutils,

hppa-linux may eventually have TLS support if the group of us working on
the toolchain can get binutils into shape.

We're in the process of debugging the output relocations in binutils,
and we are having some problems tracking down a couple of bugs. Perhaps
someone can give some insight to this problem.

The first bug involves relocations against errno@@GLIBC_PRIVATE while
building libpthread in glibc.

By the time we get to elf32_hppa_finish_dynamic_symbol, we've already
output the correct TPREL32 reloc against the TLS reference of
errno@@GLIBC_PRIVATE.

However, at this point there is an extra DIR32 errno reloc. This reloc
is one more reloc than accounted for, and when processed it trashes
atleast one entry in the output .rela.dyn.

Example libpthread.so:
00000000  00011241 R_PARISC_PLABEL32 00000000   _Jv_RegisterClasses + 0
^^^^^^^^ Trashed.
00000000  0001fcc0 unrecognized: c0      bad symbol index: 000001fc
^^^^^^^^^^^^^^^^^^ Trashed.
...
0001fc8c  00003b99 R_PARISC_TPREL32  00000000   errno + 0
0001fc8c  00003b01 R_PARISC_DIR32    00000000   errno + 0
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ What created this entry?
...
(elf32_hppa_relocate_section): 
	sym errno@@GLIBC_PRIVATE,
	bfd_elf32_swap_reloca_out (TPREL32),
	r_info = 0x3b99, 
	r_addend = 0x0, 
	r_offset = 0x1fd0c, 
	loc = 0x44c608

(elf32_hppa_finish_dynamic_symbol): 
	sym errno@@GLIBC_PRIVATE 
	bfd_elf32_swap_reloca_out (DIR32),
	got.offset 0xf9 
	tls_type -1

The errno duplicate is the DIR32 and appears only at the
finish_dynamic_relocs phase. Writing it out trashes the first two relocs
you see in the example.

Where might we have gone wrong? Any clues or insight?

c.