[llvm-commits] [llvm] r46088 - in /llvm/trunk: include/llvm/CodeGen/MachineModuleInfo.h lib/CodeGen/DwarfWriter.cpp lib/CodeGen/MachineModuleInfo.cpp

[Dale Johannesen]

On Jan 17, 2008, at 1:31 AM, Duncan Sands wrote:

> Hi Dale,
>
>> Do not mark EH tables no-dead-strip unless the
>> associated function is so marked.
>
> can you please explain what this means?

Dead stripping is a linktime optimization, where code not reachable  
from main is removed.  Functions can be marked no-dead-strip  
(__attribute__((used)) in gcc) to prevent the linker from removing  
them, even though they aren't obviously reachable.

On Darwin the EH tables do not have a magic association with the  
corresponding function; they're treated as separate symbols by the  
linker (and there are subtle differences between the Tiger and Leopard  
linkers, just to make things more interesting).   Generally, storage  
class and visibility have to match between function and table so the  
linker does not pick a function/table combination from different files  
(which doesn't work).   So, if a function is marked no-dead-strip, its  
table must be also.  The bug here was that tables were being marked no- 
dead-strip when the function was not; since the tables have references  
to the functions, this resulted in the function being kept as well;  
and if the functions have references to undefined externals that don't  
exist anywhere, you get a linktime (or, in some circumstances, dynamic  
linktime) crash.  (If I ran the universe, people with dead code that  
calls undefined externals would have to clean up their code.   I don't.)

> And is
> there a testcase?

Not now.  This only affects Darwin (and now SPU, but I'm pretty sure  
this is the right thing there too), fwiw.

> Thanks!
>
> D.