Re: [PATCH] Find tailcall frames before inline frames

[Tom Tromey <tom@tromey.com>]

>>>>> "Andrew" == Andrew Burgess <andrew.burgess@embecosm.com> writes:

>> However, in this scenario, what happened is that the DWARF unwinder
>> did find tailcall frames -- but then the PC of the first such frame
>> was recognized and claimed by the inline frame sniffer.

Andrew> I'm trying to understand the setup you have here in the hope I might
Andrew> be able to craft a test case for this - given that I'm not convinced
Andrew> the new placement of the tail call sniffer is safe.

Andrew> Was the setup something like:

Andrew>                     ,-- f3 tail calls to f4.
Andrew>                     |
Andrew>                     |
Andrew>   f1 --> f2 --> f3 --> f4 --> f5 --> f6

Andrew>   |_______________|
Andrew>   All inlined in f1

Andrew> Was there anything else special about this case?  I feel like there
Andrew> must have been, but I don't really understand the problem description.

Sorry about that.  While I do still have the test executable and core
file (and so I can easily try any patches), I can't send them out.  I'm
happy to test patches.

I only vaguely remember the details, and I can't really re-debug the
problem in the near term (I'm on PTO next week...).  However, I do also
have some notes, though not exhaustive ones.

Looking at the backtrace now, it does seem be like the above.
Here's some heavily trimmed down output:

    (gdb) frame apply 7 info frame
    [...]
    Stack level 0, frame at 0x7ffffffb07c0:
     called by frame at 0x7ffffffb07c0
    Stack level 1, frame at 0x7ffffffb07c0:
     tail call frame, caller of frame at 0x7ffffffb07c0
    Stack level 2, frame at 0x7ffffffb07c0:
     tail call frame, caller of frame at 0x7ffffffb07c0
    Stack level 3, frame at 0x7ffffffb0ab0:
     called by frame at 0x7ffffffb8c80, caller of frame at 0x7ffffffb07c0
    Stack level 4, frame at 0x7ffffffb8c80:
     inlined into frame 5, caller of frame at 0x7ffffffb0ab0
    Stack level 5, frame at 0x7ffffffb8c80:
     called by frame at 0x7ffffffb8cc0, caller of frame at 0x7ffffffb8c80
    Stack level 6, frame at 0x7ffffffb8cc0:
     called by frame at 0x7ffffffba490, caller of frame at 0x7ffffffb8c80

All I recall is what I wrote: dwarf2_tailcall_sniffer_first would note
that a frame came from tailcalls -- so tailcall_frame_sniffer would be
expected to notice this.  However, the inline sniffer would grab it
first, causing confusion.

One thing that would have helped here was the ability to turn off inline
and/or tail-call unwinding.  Perhaps we should add settings for this.

I wonder if I erred in my analysis and there's a combination of
tail-call and inlining happening.  That certainly seems like a
possibility and would explain the confusion ... whereas the "info frame"
stuff above seems innocuous to me now.

Tom