Re: backtrace() vs. _Unwind_Backtrace()

[Bryce McKinlay <bryce at mckinlay dot net dot nz>]

On Dec 5, 2003, at 5:26 PM, David Mosberger wrote:

> > > > > > On Fri, 5 Dec 2003 15:39:52 +1300, Bryce McKinlay 
> > > > > > <bryce@mckinlay.net.nz> said:
>
>   Bryce> But on Windows, breaking stack traces would be a regression
>   Bryce> because they do partially work there, although the current
>   Bryce> implementation is pretty bogus when it comes to the
>   Bryce> calling-classloader checks etc.
>
> Is this win32.cc:backtrace()?

Well, there's a bit more to it - somehow we need to get from the 
frame's IP to the Class and Method ID it belongs to.  Using the DWARF 
info that can be done with _Unwind_GetRegionStart / 
_Unwind_FindEnclosingFunction and a hashtable. The existing code 
(StackTrace and NameFinder stuff in gnu/gcj/runtime) does it by looking 
up symbols and demangling, but thats not robust and not a safe solution 
for security checks.

DWARF EH was previously enabled for Win32, but problems occurred where 
exceptions need to be thrown through OS callback frames (that don't 
have unwind info), which is a common requirement for Win32 code, so it 
uses SJLJ exceptions for now. So, it would need a way to unwind through 
info-less OS frames so that the DWARF unwinder can resume unwinding on 
the other side (ie a generalized MD_FALLBACK_FRAME_STATE_FOR).

Regards

Bryce.