Re: Anyone running -current?

[Martin Husemann <martin%duskware.de@localhost>]

The problem with -current looks like a toolchain bug, and I could use
help with some asm code...

The difference is in libc, and boils down to wether the global locale symbols
_lc_global_locale and _lc_C_locale are marked as protected or not.

This is from a failing libc.so:

Symbol table '.dynsym' contains 3636 entries:
   Num:    Value  Size Type    Bind   Vis      Ndx Name
  1399: 000efd00   260 OBJECT  GLOBAL PROTECTED   20 _lc_global_locale
  1825: 000ebb7c   260 OBJECT  GLOBAL PROTECTED   17 _lc_C_locale

Symbol table '.symtab' contains 3855 entries:
   Num:    Value  Size Type    Bind   Vis      Ndx Name
  1618: 000efd00   260 OBJECT  GLOBAL PROTECTED   20 _lc_global_locale
  2044: 000ebb7c   260 OBJECT  GLOBAL PROTECTED   17 _lc_C_locale

DYNAMIC RELOCATION RECORDS
OFFSET   TYPE              VALUE 
000eddc8 R_VAX_GLOB_DAT    _lc_global_locale
000ede34 R_VAX_GLOB_DAT    _lc_C_locale

Disassambly of isalpha:

000b2b14 <isalpha>:
   b2b14:       00 00           .word 0x0000 # Entry mask: < >
   b2b16:       c2 04 5e        subl2 $0x4,sp
   b2b19:       d0 ac 04 50     movl 0x4(ap),r0
   b2b1d:       d0 ff dd d1     movl *efd00 <_lc_global_locale>,r1
   b2b21:       03 00 51 
   b2b24:       3c 40 a1 02     movzwl 0x2(r1)[r0],r0
   b2b28:       50 
   b2b29:       ca 8f fe ff     bicl2 $0xfffffffe,r0
   b2b2d:       ff ff 50 
   b2b30:       04              ret
   b2b31:       01              nop

Relocation info is:
000eddc8  00057714 R_VAX_GLOB_DAT    000efd00   _lc_global_locale + 0

I find the indirect movl *efd00 a bit strange - I would have expected a
pc relative or immediate (with relocation).

Now the working version, as expected, goes indirectly via the GOT:

Symbol table '.dynsym' contains 3636 entries:
   Num:    Value  Size Type    Bind   Vis      Ndx Name
  1399: 000efca0   260 OBJECT  GLOBAL DEFAULT   20 _lc_global_locale
  1825: 000ebb1c   260 OBJECT  GLOBAL DEFAULT   17 _lc_C_locale

Symbol table '.symtab' contains 3855 entries:
   Num:    Value  Size Type    Bind   Vis      Ndx Name
  1618: 000efca0   260 OBJECT  GLOBAL DEFAULT   20 _lc_global_locale
  2044: 000ebb1c   260 OBJECT  GLOBAL DEFAULT   17 _lc_C_locale

DYNAMIC RELOCATION RECORDS
OFFSET   TYPE              VALUE 
000edd60 R_VAX_GLOB_DAT    _lc_global_locale
000eddcc R_VAX_GLOB_DAT    _lc_C_locale

Disassambly of isalpha:

000b2a9c <isalpha>:
   b2a9c:       00 00           .word 0x0000 # Entry mask: < >
   b2a9e:       c2 04 5e        subl2 $0x4,sp
   b2aa1:       d0 ff b9 b2     movl *edd60 <_GLOBAL_OFFSET_TABLE_+0x13bc>,r0
   b2aa5:       03 00 50 
   b2aa8:       d0 60 51        movl (r0),r1
   b2aab:       d0 ac 04 50     movl 0x4(ap),r0
   b2aaf:       3c 40 a1 02     movzwl 0x2(r1)[r0],r0
   b2ab3:       50 
   b2ab4:       ca 8f fe ff     bicl2 $0xfffffffe,r0
   b2ab8:       ff ff 50 
   b2abb:       04              ret


Can anyone explain the movl in the broken case?

Sounds to me like it dereferences the first 4 bytes of _lc_global_locale
to access _lc_global_locale, which causes a segfault.


Thanks,

Martin