feat: add cpptrace for better stack traces

[kdrienCG]

_{(Requires TPL PR #347)}

This PR aims to improve the stack trace output when an error is thrown.

When GEOS crashes from a C++ exception, the stack trace we log is not really useful: it points at main.cpp instead of the code that actually threw, missing on valuable informations for devs. This is not a bug in the error handler, it is how the C++ runtime works: by the time we catch, the stack between the throw and the handler has already been unwound, so we cannot recover it ourselves. And we can't instrument every throw site either, since many exceptions come from the standard library itself (like std::map::at for example), which we can't modify.

cpptrace solves this by hooking into the exception "machinery" to capture the stack at throw time, so we get a proper trace even when throw happens inside the std library, which is why we pull it as a dependency rather than rolling our own.

Here is an example defining a myThrowingFunction() incorrectly using a map directly (written in the main.cpp for convenience)

Without cpptrace:

***** Exception
***** Rank 0
***** Message :
map::at

***** StackTrace of 7 frames
  - Frame  0:  geos::ErrorLogger::initCurrentExceptionMessage(geos::MsgType, std::basic_string_view<char, std::char_traits<char> >, int) 
  - Frame  1:  main /somewhere/in/my/disk/geos/GEOS/src/main/main.cpp:272 (discriminator 3)
  - Frame  2:  __libc_start_main 
  - Frame  3:  _start 
  - Frame  4:  main /somewhere/in/my/disk/geos/GEOS/src/main/main.cpp:275 (discriminator 2)
  - Frame  5:  __libc_start_main 
  - Frame  6:  _start 
*****

With cpptrace:
(truncated to fit the PR description)

***** Exception
***** Rank 0
***** Message :
map::at

***** StackTrace of 15 frames
  - Frame  0: at libstdc++.so.6
  - Frame  1: in __gxx_personality_v0 at libstdc++.so.6
  - Frame  2: at libstdc++.so.6
  - Frame  3: in std::map<std::string, int, std::less<std::string>>::at(std::string const&) at stl_map.h:551:24
  - Frame  4: in myThrowingFunction() at main.cpp:40:24
  - Frame  5: in operator() at main.cpp:49:16
  - Frame  6: in do_try_catch<const geos::NotAnError&, main(int, char**)::<lambda()>&, main(int, char**)::<lambda...
  - Frame  7: in operator() at from_current.hpp:143:32
  - Frame  8: in operator() at from_current.hpp:143:32
  - Frame  9: in operator() at from_current.hpp:143:32
  - Frame 10: in try_catch_impl<main(int, char**)::<lambda()>&, main(int, char**)::<lambda(const geos::NotAnError...
  - Frame 11: in try_catch<main(int, char**)::<lambda()>&, main(int, char**)::<lambda(const geos::NotAnError&)>&...
  - Frame 12: in main at main.cpp:129:22
  - Frame 13: in __libc_start_main at libc.so.6
  - Frame 14: in _start at geosx
*****

We have a detailled stack trace with the name of the function that throws and the correct line and character position.

[rrsettgast]

@kdrienCG Is there a discussion somewhere about whether we want to add this dependency?

[herve-gross]

@rrsettgast yes, we discussed this internally. There is value in having this TPL for debugging deeper issues. We can talk if you wish.

[rrsettgast]

@rrsettgast yes, we discussed this internally. There is value in having this TPL for debugging deeper issues. We can talk if you wish.

@herve-gross That decision being made without consulting others on the core team is inappropriate. The inclusion of this package covers the fundamental problem in the approach to c++ exceptions in an HPC simulation code. There are simpler solutions...like output stack on the throw.

[jafranc]

Looks like the MACRO dispach will occur at each new lambda-exception. Maybe worth passing a callback through the lambda's interface and dispatch all below ? In case we add more of this and the duplication spread

[MelReyCG]

in order to not modify the stacktrace, do you agree that we should:

#ifdef GEOS_USE_CPPTRACE
  #define GEOS_FORMATTED_STACKTRACE StackTrace::formatStackTrace( cpptrace::from_current_exception() )
#else
  #define GEOS_FORMATTED_STACKTRACE LvArray::system::stackTrace( true )
#endif

and then,

    std::string const stacktrace = GEOS_FORMATTED_STACKTRACE;

[dkachuma]

Since all the calls involving cpptrace are static I would move anything related to choosing whether cpptrace is used to the .cpp file to leave the header unencumbered by these macros.

[MelReyCG]

Hi @rrsettgast,
As an example, how do you simply throw from a std::map exception?

[rrsettgast]

Hi @rrsettgast, As an example, how do you simply throw from a std::map exception?

I assume you mean in the case of the std wrappers, so I will restrict my response to this case. You don't need to throw a std::map exception. The wrapper gives flexibility so you can mimic what was done for bounds checks in LvArray which would result in a well located stack trace and an abort().