[LLVMdev] Regarding exceptions (new to this list)

[Aristid Breitkreuz]

Am Sonntag, den 30.04.2006, 22:24 -0500 schrieb Chris Lattner:
> On Sun, 30 Apr 2006, Aristid Breitkreuz wrote:
> > I am, as said in the subject, new to this list. I have not actually used
> > LLVM, I tried with an older version but failed. Recently I noticed that
> > 1.7 came out and looked at the release notes. And there was something
> > that stroke me: llvm-g++ adheres to the IA64 ABI but _not for
> > exceptions_. I was not able to find information why this is so and why
> > it is necessary, so I hope you don't mind me asking here. Why? It would
> > make me a whole lot more interested in LLVM.
> 
> Primarily because it hasn't been implemented in the native code generators 
> yet.  I hope to get to this in the next month or two.

Cool. I'm relieved it's not primarily because of "considerations" :).

> 
> > I have another question: You said (I read in the archive) all your
> > libraries are static because of start-up time becoming too high. Have
> > you thought about using -fvisibility=hidden +
> > __attribute__((visibility(default))) (GCC 4.0 had this wonderful change
> > making C++ a better language to use)? Or is this currently impossible
> > with the LLVM infrastructure? Just curious.
> 
> Noone has looked into it.  If it were a relatively unobtrusive change to 
> the LLVM codebase, we would welcome it.  On the other hand, GCC has a 
> number of significant bugs that impact us: for example LLVM has many 
> classes declared in anonymous namespaces.  GCC does not give the methods 
> internal linkage, it just "mangles" them.  This is bad for static 
> and dynamic link times, but can only really be resolved in the GCC C++ 
> front-end.

The -fvisibility=hidden fixes the very problem for dynamic linktimes at
least. Still, it's certainly a partial solution only. I know that
supporting -fvisibility=hidden requires intrusive changes in library
code but for the implementation, I plain don't know. I'm just a curious
developer... I do know that using -fvisibility=hidden is worth it, let's
take a look at an example from my current project which uses this flag
(Linux x86-64, GCC 4.0.3):

libevents.so 216KB (unstripped)

readelf -s says:
Symbol table '.dynsym' contains 269 entries
Symbol table '.symtab' contains 702 entries

=> 156K events/libevents.so (stripped)
(the stripped version has no .symtab any more, the .dynsym is left in,
of course)

For another library, the .dynsym/.symtab ratio is 397/2032 and it's
592KB unstripped and 400KB stripped.

Hope to have got you interested.

> 
> -Chris
> 

Aristid