Add basic backtrace functionality

[alexcrichton]

There are a number of locations where it would be really nice if we had some very simple support for backtraces. By simple I just mean the ability to know the name of the function and possibly the instruction pointer inside the function.

Some possible use case:

• Traces for each gc allocation (the old C++ managed box allocator did this)
• Stack trace printing on failure
• Less codegen bloat from failure functions that pass around file/line number. (function name isn't equivalent, but maybe sufficient?)

This is obviously a tricky thing to do, and I don't think that we should bring in libunwind for just this (it's pretty simple), but it appears that each platform has their own solution for acquiring a simple backtrace, and we can take the least common denominator of all of them:

• windows
• osx
• linux

This would probably be a module in std::unstable like the way dynamic_lib is a module. This isn't really a high priority, but it would certainly be something nice to have!

[catamorphism]

+1

[alexcrichton]

I played around with this for a bit using the backtrace function on OSX and linux. OSX refused to give any symbols at all (the backtrace function returned 0), and linux gave awful symbols names because all of our symbols in executables are flagged as internal. When getting a backtrace from a library, the symbol names were all acquired, but they were also all mangled.

Not entirely sure how useful this would be in light of this, and it also turns out that -Z extra-debug-info doesn't improve symbols at all. Certainly worth the interesting project tag!

[jdm]

Firefox has stackwalking code that might be worth investigating: http://mxr.mozilla.org/mozilla-central/source/xpcom/base/nsStackWalk.cpp

[jdm]

However, it appears that the production stackwalker code for our in-app profiler that runs on release builds is built on top of Breakpad.

[pnkfelix]

cc me

[pnkfelix]

In a toy example in C on Mac OS X, it seems like the backtrace functionality works, sort of, for a compiled executable. (I say "sort of" because once I start marking functions in my toy example as static, gcc starts emitting tail calls, so of course that means those frames get left off the stack, so my toy example isn't as full an illustration as I'd like it to be.)

[flaper87]

cc me

[alexcrichton]

The ruby implementation looks like this:

https://github.com/ruby/ruby/blob/trunk/vm_dump.c#L679-L705

[jdm]

Most recent Gecko work for use by the built-in profiler: https://bugzilla.mozilla.org/show_bug.cgi?id=938157