Bug 1513 – try/catch/finally misbehavior on windows

bug.d -------- import std.stdio; void main() { try { try { try { } finally { writefln("throw ex1"); throw new Exception("ex 1"); } } finally { writefln("throw ex2"); throw new Exception("ex 2"); } } finally { writefln("throw ex3"); throw new Exception("ex 3"); } } ------- dmd 1.021 on windows outputs: $ ./bug throw ex1 throw ex2 Error: ex2 dmd 1.021 on linux outputs: $ ./bug throw ex1 throw ex2 throw ex3 Error: ex3

err, actually I prefer the windows under hood behavior. Windows prevents you replace the exception silently. For a language to fix such a problem would need to have the stacktrace to locate where the exception thrown. It's a bit tough to fix the problem.

if i put another try catch block wrap this , everything goes ok seems it's not SEH related. I tried some similar code with C __try __except __try __finally blocks. It works as expected

http://www.digitalmars.com/d/statement.html#TryStatement # # If an exception is raised in the FinallyStatement and is not caught before # the FinallyStatement is executed, the new exception replaces any existing exception: # Added to DStress as http://dstress.kuehne.cn/run/f/finally_12_A.d

I figured it'd been ages and maybe something had changed to maybe fix this bug.. nope: dmd 1.061 on windows: Error: ex 2 throw ex1 throw ex2 dmd 1.061 on linux: throw ex1 throw ex2 throw ex3 Error: ex 3 dmd 2.046 on windows: throw ex1 throw ex2 object.Exception: ex 2 dmd 2.047 (close to svn tip) on linux: throw ex1 throw ex2 throw ex3 object.Exception: ex 3 --- <cut the stack trace> --- So.. still different.

The behaviour of this bug changed in DMD2.048, (possibly as a result of the fix for bug 4339?) Now, on Windows DMD2.050 produces: throw ex1 throw ex2 object.Exception: ex 2 object.Exception: ex 1 Which I don't understand at all. On D1.065, it is still: throw ex1 throw ex2 Error: ex 2 Does it still work on D2 Linux?

building against tip of trunk for dmd/druntime/phobos, the output: throw ex1 throw ex2 throw ex3 object.Exception: ex 3 ---------------- <snip stack trace> object.Exception: ex 2 ---------------- <snip stack trace> object.Exception: ex 1 ---------------- <snip stack trace>

I think the D2 change was actually caused by druntime svn 358. Sean's comment: Changed how exception chaining works. Now, the original exception will be retained and continue propagating even if other exceptions are thrown during stack unwinding. These exceptions will be chained via the 'next' pointer and regarded as collateral damage, as per TDPL. Is that supposed to apply in this case? Either druntime is wrong, or (more likely) the spec needs to be updated.

I'd prefer Sean's description, though the real issue is the 3rd finally block is not executed.

Reduced test case. Only two finally clauses are required, provided that each try{} block contains a throw statement. This shows that nesting of finally statements doesn't work at all on Windows. Raising priority to critical. ---- import std.stdio; void main() { try { try { writefln("throw ex1"); throw new Exception("ex 1"); } finally { writefln("throw ex2"); throw new Exception("ex 2"); } } finally { writefln("finally"); // never reached } }

I'm not sure that this is a compiler bug. It may be druntime. With this change: ---- deh.c, _d_framehandler(), line 210 + else if (prev_ndx == -1) + { // Exception didn't get caught. + // Call all the finally blocks skipped in this frame + _d_local_unwind(handler_table, frame, ndx); + } } } return ExceptionContinueSearch; ---- the finally clauses are called correctly. This isn't a correct patch, but I think it demonstrates that the exception tables are set up reasonably correctly. I couldn't find anything wrong with them. My limited understanding of exception handling is based on this article: http://www.microsoft.com/msj/0197/exception/exception.aspx

This is definitely a druntime issue (though it could be an issue with the C stdlib). The bug must lie in the druntime/src/rt/deh.c In the test case, here's my understanding of what happens: The first exception is thrown. Every try block is queried (in _d_framehandler() )to see if it's handled. Nothing in main() handles it, but something else does. (I think in this case, it's a catch-all block around main). Then, the unwinding starts. _d_framehandler() runs this: // Have system call all finally blocks in intervening frames int retval = _global_unwind((ESTABLISHER_FRAME *)frame, exception_record); _global_unwind is a wrapper around the Windows system function RtlUnwind. That calls _dframehandler() again, with EXCEPTION_UNWIND this time. This calls _d_local_unwind(), to run all of the finally blocks _d_local_unwind() is also in druntime/deh.c. In the test case, the finally block contains a throw, so the procedure is repeated. Again it doesn't get caught until the catch-all. As before, global_unwind and then local_unwind get called. The first thing _d_local_unwind() does is set up a double-fault exception handler. When it begins the unwind, the double-fault handler gets triggered. The double-fault handler basically just does this: if (!(ExceptionRecord->ExceptionFlags & EXCEPTION_UNWIND)) return ExceptionContinueSearch; *((ESTABLISHER_FRAME **)DispatcherContext) = EstablisherFrame; return ExceptionCollidedUnwind; Because it returns ExceptionCollidedUnwind, Windows is supposed to abandon the original unwinding, and the new unwinding should take over. But that's doesn't seem to be what happens. No further unwinding occurs. To my suprise, it returns from _global_unwind into _d_framehandler, back in the original catch handler, and all the variables seem to be unchanged (so you have no way of detecting what happened). Dunno where the problem is. Maybe it could be in _global_unwind in the C runtime? This stuff is amazingly undocumented by Microsoft. This is the most detailed info that I've found: http://www.nynaeve.net/?p=106 See point 7. And here is more detail about collided unwinds. http://www.nynaeve.net/?p=107

On D2, the behaviour is wrong on all platforms, not just Windows. Roughly speaking, Linux does exception chaining incorrectly, and Windows doesn't do it at all. These two commits are a first step towards fixing the bug: the problematic code in the C runtime is replaced with a D implementation (the buggy behaviour isn't yet changed though). http://www.dsource.org/projects/druntime/changeset/458 http://www.dsource.org/projects/druntime/changeset/459

Fixed for D2: http://www.dsource.org/projects/druntime/changeset/482

And it's now fixed for D1. http://www.dsource.org/projects/phobos/changeset/2282