http://gcc.gnu.org/bugzilla/show_bug.cgi?id=47274
--- Comment #28 from Dave Korn <davek at gcc dot gnu.org> 2011-02-01 06:59:58
UTC ---
It looks like the problem is much earlier than the linker; it looks like the IR
symtabs in the input object files are being generated incorrectly. Here are
the real symbol tables:
$ ./obj/binutils/.libs/nm-new.exe *.o
abs-1-lib.o:
00000001 C __gnu_lto_v1
U abort
00000000 T abs
U abs_called
U inside_main
00000024 T labs
abs-1.o:
00000001 C __gnu_lto_v1
U abort
U abs
00000000 B abs_called
00000000 T main_test
main.o:
00000001 C __gnu_lto_v1
00000004 C inside_main
00000000 T main
U main_test
And here are the IR symtabs (I built a cross GCC just far enough to get the
liblto_plugin):
$ ./obj/binutils/.libs/nm-new.exe --plugin ./obj-xgcc/lto-plugin/.libs/cyglto_p
lugin-0.dll *.o
abs-1-lib.o:
00000000 T abs
00000000 T abs_called
00000000 T inside_main
abs-1.o:
00000000 T abs
00000000 T abs_called
00000000 T main_test
main.o:
00000000 T inside_main
00000000 T main
00000000 T main_test
To confirm that the symtabs really are wrong in the object files, rather than
that the plugin is misreading them, here's a binary dump from one:
$ ./obj/binutils/.libs/objdump.exe -h abs-1-lib.o
abs-1-lib.o: file format elf32-hppa-linux
Sections:
Idx Name Size VMA LMA File off Algn
0 .text 00000058 00000000 00000000 00000034 2**2
CONTENTS, ALLOC, LOAD, RELOC, READONLY, CODE
1 .data 00000000 00000000 00000000 0000008c 2**0
CONTENTS, ALLOC, LOAD, DATA
2 .bss 00000000 00000000 00000000 0000008c 2**0
ALLOC
3 .gnu.lto_.jmpfuncs.29f392e1 00000019 00000000 00000000 0000008c 2**0
CONTENTS, READONLY, EXCLUDE
4 .gnu.lto_.pureconst.29f392e1 00000016 00000000 00000000 000000a5 2**0
CONTENTS, READONLY, EXCLUDE
5 .gnu.lto_abs.29f392e1 00000192 00000000 00000000 000000bb 2**0
CONTENTS, READONLY, EXCLUDE
6 .gnu.lto_labs.29f392e1 00000189 00000000 00000000 0000024d 2**0
CONTENTS, READONLY, EXCLUDE
7 .gnu.lto_.cgraph.29f392e1 0000003c 00000000 00000000 000003d6 2**0
CONTENTS, READONLY, EXCLUDE
8 .gnu.lto_.vars.29f392e1 00000018 00000000 00000000 00000412 2**0
CONTENTS, READONLY, EXCLUDE
9 .gnu.lto_.refs.29f392e1 0000001b 00000000 00000000 0000042a 2**0
CONTENTS, READONLY, EXCLUDE
10 .gnu.lto_.statics.29f392e1 00000014 00000000 00000000 00000445 2**0
CONTENTS, READONLY, EXCLUDE
11 .gnu.lto_.decls.29f392e1 000002ab 00000000 00000000 00000459 2**0
CONTENTS, READONLY, EXCLUDE
12 .gnu.lto_.symtab.29f392e1 00000048 00000000 00000000 00000704 2**0
CONTENTS, READONLY, EXCLUDE
13 .gnu.lto_.opts 00000012 00000000 00000000 0000074c 2**0
CONTENTS, READONLY, EXCLUDE
14 .PARISC.unwind 00000020 00000000 00000000 00000760 2**2
CONTENTS, ALLOC, LOAD, RELOC, READONLY, DATA
15 .comment 00000042 00000000 00000000 00000780 2**0
CONTENTS, READONLY
$ ./obj/binutils/.libs/objcopy.exe -j .gnu.lto_.symtab.29f392e1 abs-1-lib.o sym
s-abs-1-lib.o
$ ./obj/binutils/.libs/objdump.exe -s syms-abs-1-lib.o
syms-abs-1-lib.o: file format elf32-hppa-linux
Contents of section .gnu.lto_.symtab.29f392e1:
0000 61627300 00000000 00000000 00000000 abs.............
0010 00004669 6e736964 655f6d61 696e0000 ..Finside_main..
0020 00000000 00000000 00000000 00656162 .............eab
0030 735f6361 6c6c6564 00000000 00000000 s_called........
0040 00000000 00000067 .......g
The format of the IR symtab entries is:
(variable-length) NUL-terminated name
(variable-length) NUL-terminated comdat group
(1 byte) symbol kind from LDPK_ enumeration
(1 byte) symbol visibility from LDPV_ enumeration
(8 bytes) 64-bit symbol size
(4 bytes) 32-bits symbol aux info.
As you can see, this symtab claims that abs-1-lib.o defines inside_main
0000 61627300 00000000 00000000 00000000 abs.............
0010 00004669 6e736964 655f6d61 696e0000 ..Finside_main..
\/ \/\/\/\/ \/\/\/\/ \/\/\/\/
name--------++++++++++++++++++++++++++ +
comdat---------------------------------+
0020 00000000 00000000 00000000 00656162 .............eab
\/\/
kind--+ + 0 = LDPK_DEF
vis-----+ 0 = LDPV_DEFAULT
0030 735f6361 6c6c6564 00000000 00000000 s_called........
0040 00000000 00000067 .......g
So the problem is that the object files are being written incorrectly in the
first place, I think. Indeed, the contents of abs-1-lib.s look that way:
.section .gnu.lto_.symtab.29f392e1,"",@progbits
.stringz "abs"
.stringz ""
.stringz ""
.stringz ""
.stringz ""
.stringz ""
.stringz ""
.stringz ""
.stringz ""
.stringz ""
.stringz ""
.stringz ""
.stringz ""
.stringz ""
.stringz ""
.stringz "Finside_main"
.stringz ""
.stringz ""
.stringz ""
.stringz ""
.stringz ""
.stringz ""
.stringz ""
.stringz ""
.stringz ""
.stringz ""
.stringz ""
.stringz ""
.stringz ""
.stringz ""
.stringz "eabs_called"
.stringz ""
.stringz ""
.stringz ""
.stringz ""
.stringz ""
.stringz ""
.stringz ""
.stringz ""
.stringz ""
.stringz ""
.stringz ""
.stringz ""
.stringz ""
.stringz ""
.ascii "g"
.text
No attempt there to put anything but zeros for the symbol kind. So, LTO is
streaming out symtabs that say every object file defines every symbol... no
wonder the linker/plugin get confused!
This now needs somebody with a full compiler for the target to debug; I've only
got (cross) binutils and lto-plugin.
