A few months ago I wrote a blog post where I described some minimal changes to GHC's exception infrastructure which would allow us to use DWARF stack information to annotate exceptions.

Today @domenkozar brought up the matter with me on IRC and said that tracking down exceptions is his greatest challenge in production.

Proposal

We modify SomeException as follows:

data SomeException where
    SomeExceptionWithLocation
      :: forall e. Exception e
      => Maybe Backtrace   -- ^ backtrace, if available
      -> e                 -- ^ the exception
      -> SomeException

and add a bidirectional pattern synonym for compatibility:

pattern SomeException e <- SomeExceptionWithLocation _ e
  where
    SomeException e = SomeExceptionWithLocation Nothing e

Backtrace is a simple type capturing the various backtrace

data Backtrace
    = CostCenterBacktrace GHC.Stack.CCS.CostCentreStack
      -- ^ a cost center profiler backtrace
    | HasCallStackBacktrace GHC.Stack.CallStack
      -- ^ a stack from HasCallStack
    | ExecutionBacktrace [GHC.ExecutionStack.Location]
      -- ^ a stack unwinding (e.g. DWARF) backtrace

We can then implement a variety of throwIO actions which attach stack traces and a catch operation which provides access to the backtrace:

-- | Throws an exception with no stack trace.
throwIOWithoutBacktrace :: Exception e => e -> IO a

-- | Throws an exception with the current cost center stack.
throwIOWithCostCenterStack :: Exception e => e -> IO a

-- | Throws an exception with a stack trace captured via
-- 'GHC.Stack.getStackTrace' with the given stack limit
-- (or unlimited if 'Nothing').
throwIOWithExecutionStack :: Exception e => Maybe Int -> e -> IO a

-- | Throws an exception with a `HasCallStack` stack trace.
throwIOWithCallStack :: HasCallStack => e -> IO a

-- | Catch an exception, providing access to the exception's backtrace:
catchWithLocation :: IO a -> (Maybe Backtrace -> e -> IO a) -> IO a

throwIO :: Exception e => e -> IO a
throwIO = throwIOWithoutBacktrace

Appropriate analogues would be provided for the pure throw function.

Making backtraces ubiquitous

So far we have discussed a mechanism for attaching backtraces to exceptions in a backwards-compatible manner. However, this doesn't help existing code which call throwIO. This is where we need to be a bit careful as some of these backtrace-collection mechanisms are potentially expensive. For instance, throwIO = throwIOWithExecutionStack may open the user to DOS attacks due to the slow nature of DWARF unwinding.

Given that backtraces are primarily a debugging tool, I think a pragmatic, albeit somewhat ugly, solution using global state is appropriate. That is, we expose:

-- | How to collect a backtrace when an exception is thrown.
data BacktraceMechanism
    = NoBacktrace
      -- ^ don't collect a backtrace
    | CostCenterBacktrace
      -- ^ collect a cost center stacktrace (only available when built with profiling)
    | ExecutionStackBacktrace (Maybe Int)
      -- ^ use execution stack unwinding with given limit

setGlobalBacktraceMechanism :: BacktraceMechanism -> IO ()

The default global backtrace mechanism will be NoBacktrace.

We can then redefine throwIO as:

throwIO :: Exception e => e -> IO a
throwIO exc = do
    m <- getGlobalBacktraceMechanism
    case m of
      NoBacktraces                  -> throwIONoBacktrace exc
      CostCenterBacktrace           -> throwIOWithCostCenterStack exc
      ExecutionStackBacktrace limit -> throwIOWithExecutionStack limit exc

This leaves the policy question of backtrace mechanism selection to the user while preserving backwards compatibility.

The only wrinkle here is that of toException, since SomeException values created in this way won't get a Backtrace. One option here would be to rely on unsafePerformIO to collect a backtrace when constructing SomeException. That is rewrite the SomeException pattern synonym as:

pattern SomeException e <- SomeExceptionWithLocation _ e
  where
    SomeException e = mkSomeException e

mkSomeException :: Exception e => e -> SomeException
mkSomeException e = unsafePerformIO $ do
    bt <- collectBacktrace
    return (SomeExceptionWithLocation bt e)

Setting mechanism from environment

Some languages require that the user set an environment variable to enable backtrace reporting (e.g. Rust requires RUST_BACKTRACE=1). In principle we could expose such a mechanism. I'm a bit unsure of whether this would be a good thing or not.