Re: [SCM] master: Use libunwind cursor to get at CFA.

[Andrew Cagney <cagney at redhat dot com>]

Mark Wielaard wrote:
> Nice catch. Well that is somewhat of a bummer. Then we must fall back on
> the unwind one frame trick and get the SP again.
>   
That isn't as bad as it seems; computing the cfa requires running the 
unwind-program anyway.

I think having a method (perhaps called .getInnerCFA() though?) to 
differentiate between the sp and cfa is definitly a good idea.  Also, if 
possible, can the start code address of the CFI block be made 
available?  There's the theoretical possibility of further tuning the 
stepping code to use that to construct a quick-and-dirty 
frame-identifier when single stepping (letting us delay the request for 
the code's function).

(Your elimination of the uses of libunwind's lookup is wicked)
> What is somewhat worrying is that none of the tests caught this. So I
> added one that does fail with my change, and passes with that commit
> reverted.
>
>   
Yes.  Scary.  There's always one that gets away :-)  Nice simple 
portable test.

Andrew