Too much info is bad? :) Ok, I have no idea for this bug, but I guess, that we're just not using abbrevs everywhere.

It seems this is only true for c++ code.

Ok, looks like we are emitting debug info for symbols that are not used:
@.str560 = internal constant [10 x i8] c"operator=\00", section "llvm.metadata" ; <[10 x i8]> [#uses=1]
@.str561 = internal constant [82 x i8] c"std::basic_string<char, std::char_traits, std::allocator >::operator=\00", section "llvm.metadata" ; <[82 x i8]> [#uses
=1]
@.str562 = internal constant [13 x i8] c"_ZNSsaSERKSs\00", section "llvm.metadata" ; <[13 x i8]*> [#uses=1]

@​llvm.dbg.subprogram1045 = internal constant %llvm.dbg.subprogram.type {
i32 393262,
{ }* null,
{ }* bitcast (%llvm.dbg.compile_unit.type* @​llvm.dbg.compile_unit22 to { }),
i8 getelementptr ([10 x i8]* @.str560, i32 0, i32 0),
i8* getelementptr ([82 x i8]* @.str561, i32 0, i32 0),
i8* getelementptr ([13 x i8]* @.str562, i32 0, i32 0),
{ }* bitcast (%llvm.dbg.compile_unit.type* @​llvm.dbg.compile_unit22 to { }),
i32 2487,
{ } bitcast (%llvm.dbg.compositetype.type* @​llvm.dbg.compositetype563 to { }),
i1 false,
i1 false }, section "llvm.metadata" ; <%llvm.dbg.subprogram.type> [#uses=1]

=>
.byte 0x1c
.long 0x28ad
.asciz "operator="
.asciz "_ZNSsaSERKSs"
.byte 0xd
.short 0x9b7
.long 0x1f61
.byte 0x1
.byte 0x1

The real problem is there are a ton of duplicates.

@​llvm.dbg.global_variable = internal constant %llvm.dbg.global_variable.type {
i32 393268,
{ }* null,
{ }* bitcast (%llvm.dbg.compile_unit.type* @​llvm.dbg.compile_unit22 to { }),
i8 getelementptr ([5 x i8]* @.str44, i32 0, i32 0),
i8* getelementptr ([77 x i8]* @.str45, i32 0, i32 0),
i8* getelementptr ([12 x i8]* @.str46, i32 0, i32 0),
{ }* bitcast (%llvm.dbg.compile_unit.type* @​llvm.dbg.compile_unit47 to { }),
i32 261,
{ } bitcast (%llvm.dbg.derivedtype.type* @​llvm.dbg.derivedtype11 to { }),
i1 false,
i1 false,
{ } null }, section "llvm.metadata" ; <%llvm.dbg.global_variable.type*> [#uses=1]

@​llvm.dbg.global_variable189 = internal constant %llvm.dbg.global_variable.type {
i32 393268,
{ }* null,
{ }* bitcast (%llvm.dbg.compile_unit.type* @​llvm.dbg.compile_unit22 to { }),
i8 getelementptr ([5 x i8]* @.str44, i32 0, i32 0),
i8* getelementptr ([77 x i8]* @.str45, i32 0, i32 0),
i8* getelementptr ([12 x i8]* @.str46, i32 0, i32 0),
{ }* bitcast (%llvm.dbg.compile_unit.type* @​llvm.dbg.compile_unit47 to { }),
i32 261,
{ } bitcast (%llvm.dbg.derivedtype.type* @​llvm.dbg.derivedtype11 to { }),
i1 false,
i1 false,
{ } null }, section "llvm.metadata" ; <%llvm.dbg.global_variable.type*> [#uses=1]

g++ does like to output an initial form of a global variable
then later "fix it up" by replacing it with a new version with
more details. Perhaps debug info is being generated each time?

After this patch:
http://lists.cs.uiuc.edu/pipermail/llvm-commits/Week-of-Mon-20080623/064201.html

I got:

Segment : 1796897
Section (__TEXT, __text): 290857
Section (__DWARF, __debug_frame): 80204
Section (__DWARF, __debug_info): 823455
Section (__DWARF, __debug_abbrev): 1490
Section (__DWARF, __debug_aranges): 0
Section (__DWARF, __debug_macinfo): 0
Section (__DWARF, __debug_line): 149753
Section (__DWARF, __debug_loc): 0
Section (__DWARF, __debug_pubnames): 165012
Section (__DWARF, __debug_str): 0
Section (__DWARF, __debug_ranges): 0
Section (__DATA, __data): 76
Section (__TEXT, __StaticInit): 166
Section (__TEXT, __textcoal_nt): 168920
Section (__DATA, __bss): 65
Section (__TEXT, __cstring): 15574
Section (__DATA, __const): 136
Section (__DATA, __mod_init_func): 4
Section (__IMPORT, __jump_table): 12410
Section (__IMPORT, __pointers): 128
Section (__TEXT, __eh_frame): 88600
total 1796850
total 1796897

My gcc-4.2 numbers are different from Chris':

Segment : 1751265
Section (__TEXT, __text): 254569
Section (__DWARF, __debug_frame): 82612
Section (__DWARF, __debug_info): 841164
Section (__DWARF, __debug_abbrev): 4535
Section (__DWARF, __debug_aranges): 48
Section (__DWARF, __debug_macinfo): 0
Section (__DWARF, __debug_line): 125613
Section (__DWARF, __debug_loc): 107932
Section (__DWARF, __debug_pubnames): 159752
Section (__DWARF, __debug_pubtypes): 37242
Section (__DWARF, __debug_str): 13480
Section (__DWARF, __debug_ranges): 48
Section (__DATA, __mod_init_func): 4
Section (__TEXT, __textcoal_nt): 109696
Section (__TEXT, __cstring): 13968
Section (__TEXT, __StaticInit): 134
Section (__DATA, __bss): 33
Section (__DATA, __const): 136
Section (__TEXT, __const_coal): 71
Section (__DATA, __data): 60
Section (__TEXT, __literal8): 8
Section (__IMPORT, __pointers): 128
Section (__TEXT, __constructor): 0
Section (__TEXT, __destructor): 0
total 1751233
total 1751265

We are now generating less debug info than gcc 4.2.

I'll close this now. There are additional refinements to be made but they are no as critical as this one.