llvm dev - Feb 2009 - [LLVMdev] Question LowerSetJmp

Hi
    In llvm src the file llvm/lib/Transforms/IPO/LowerSetJmp.cpp,  it
seems for sjlj-eh ?  but this pass  has no effect about  c++ sjlj-eh.
this pass is for future extend?

zhangzw

Hi zhangzw,
>     In llvm src the file llvm/lib/Transforms/IPO/LowerSetJmp.cpp,  it
> seems for sjlj-eh ?
it's not for sjlj-eh.  The problem with setjmp/longjmp is that they
can jump from one part of a function to any other part, which means
that the control-flow graph is pointless; this makes everything more
complicated.  So how to deal with this?  LLVM does have constructs
that allow "jumping around": invoke and unwind, and all the optimizers
know how to deal with these correctly.  So LLVM turns setjmp/longjmp
into a bunch of invokes and unwinds, in order to generate correct code
while not having to teach all the optimizers about setjmp/longjmp.
Setjmp/longjmp eh is something different.
> but this pass  has no effect about  c++ sjlj-eh.
> this pass is for future extend?
Ciao,

Duncan.

On Feb 26, 2009, at 12:57 AM, Duncan Sands wrote:
>  So LLVM turns setjmp/longjmp
> into a bunch of invokes and unwinds, in order to generate correct code
> while not having to teach all the optimizers about setjmp/longjmp.
> Setjmp/longjmp eh is something different.
Or virtually identical.  Bout the only difference is the number of  
registers saved and the order of those registers.