These are some general questions I am facing while designing the error handling for an algorithm that is supposed to run in parallel using MPI (in C++):
Do Exceptions work inside code that is executed in parallel? Is the behaviour defined?
How do they work? Does that differ for different implementations?
Is it good practice - or should I use return codes?
In an ideal world, you can use them to do what you ask. By "ideal world" I mean one where you have your choice of MPI implementation and are able to administer it yourself (instead of convincing the cluster owner to reconfigure it for you). The minimal configuration for exceptions will include the: --with-exceptions flag, and possibly a few more.
I've used LAM most often, and by default exceptions are disabled. I believe this is the default for other implementations as well.
They work in the same vein as 'vanilla' C++ exceptions. And they do work inside parallel executed code.
At some point in your startup code, you want to enable them:
MPI::COMM_WORLD.Set_errhandler ( MPI::ERRORS_THROW_EXCEPTIONS );
(if your library isn't configured to allow exceptions, this is probably a bad idea -- behaviour "undefined" according to LAM)
And then:
try { /* something that can fail */ }
catch ( MPI::Exception e ) {
cout << "Oops: " << e.Get_error_string() << e.Get_error_code();
MPI::COMM_WORLD.Abort (-1) ;
}
As for it being good or bad practice, I can't really say. I haven't seen extensive use of them in code written by hardened MPI hackers, but that may be because the code is generally more C than C++ in my experience.
A middle ground between error codes and exceptions may be error handlers, in a nutshell you can assign functions that will be called when a particular error (designated by code) occurs. This might be an option if you can't get your administrator on board with enabling exceptions.
Exceptions work the same in an MPI code as with a serial code, but you have to be extremely careful if it is possible for the exception is not raised on all processes in a communicator or you can easily end up with deadlock.
MPI_Barrier(comm); /* Or any synchronous call */
if (!rank) throw Exception("early exit on rank=0");
MPI_Barrier(comm); /* rank>0 deadlocks here because rank=0 exited early */
All error handling methods have this problem, it is difficult to recover from errors that do not occur consistently across a communicator. In the case above, you could perform an MPI_Allreduce so that all ranks choose the same branch.
My preference is for calling error handlers and propagating them up the stack since this tends to give me tho most useful/verbose error message and it's easy to catch with a breakpoint (or the error handler can attach a debugger to itself and send it to your workstation in an xterm).
Whether or not exceptions will work during parallel execution depends on your compiler and MPI library implementation. If you want portable behavior, I'd avoid throwing exceptions in that context.
If you want more detailed information about errors than just a numeric return code, you can of course return and/or pass around error strings or other objects (within the same process or through MPI, of course).
Related
I have a C++ program for which validated API users (not interactive users, so there is no major security/integrity exposure here) can pass a parameter that becomes a format string for printf() (actually, for vsnprintf()). Of course, bad format specifiers make the program blow up. I have documented that and it is acceptable -- but is there any way to trap printf() errors rather than having the C runtime assert?
Environment is pretty much standard Posix (z/OS XLC, FWIW).
Use sigsetjmp()/siglongjmp().
Call sigsetjmp() prior to using the potentially bad format, and install a custom handler for SIGSEGV and SIGBUS that calls siglongjmp() with the context from the sigsetjmp() call.
Just be sure to use restore the signal handlers after you don't need them lest a SIGSEGV somewhere else cause some unexpected results.
Example here:
http://www-01.ibm.com/support/knowledgecenter/ssw_ibm_i_72/apis/sigsetj.htm
Please see to BOOST format library - http://www.boost.org/doc/libs/1_58_0/libs/format/
The format library provides a class for formatting arguments according to a format-string, as does printf, but with two major differences :
format sends the arguments to an internal stream, and so is entirely type-safe and naturally supports all user-defined types.
You can use C++ exception with this library.
I was tempted to recommend a simple C++ style try/catch:
. . .
try
printf(...)
catch (...)
printf("An error happened handling your printf string!")
...
But I figure you tried that and have a reason why it doesn't meet your needs.
The deeper alternative is an LE Condition Handler - it's essentially a callback routine that gives you control any time an exception occurs, and it includes a way to resume execution after an error. Read more by Googling CEEHDLR or see the "Handling error conditions, exceptions, and signals" section of the XL C/C++ Programming Guide.
In a pinch, you can also wrap your calls to printf() with ESTAE/ESPIE macros to trap the error before the runtime sees it...yes, it takes a few lines of assembler, but many times it's possible to get way more granular this way.
Best of luck!
Answering title:
if(printf("...") < 0)
{
perror("printf error message");
exit;
}
Addition.
Working example
wchar_t wide[2];
wide[0] = 129;
wide[1] = 0;
if(printf("%ls", wide) < 0)
perror("detect error");
I am writing a C++ code that controls a special device that is the fusion of several SDKs. My code looks like
#define sdk1SafeCall(err) __sdk1SafeCall(err,__FILE__,__LINE__)
int errorcode = 0;
sdk1SafeCall(sdk1_InitializeDevice());
errorcode=sdk2_InitializeDevice();
errorcode=sdk3_InitializeDevice();
if (some_parameter)
{
errorcode=sdk2_readDevice(true);
}
else
{
errorcode=sdk3_writeDevice();
}
label again: errorcode=sdk1_readDevice();
if (error) goto again;
errorcode=close_everything();
Use parameters will rearrange the control flow. My current method uses something like cudaSafeCall to wrap the error codes and exit. What I don't know how to do is where to store detailed explanations of these errors or how to recover from them (such as the goto example). At the end of the day the code looks really messy.
Edit
I do sometimes handle errors with a wrapper.
Writing error-safe code is hard. In your pseudocode, you don't actually deal with errors at all. errorcode will just have the result of "close_everything()" - which may have succeeded, even if everything else went wrong.
The typical way to solve this sort of thing in C++ is to have an object for each "resource" (e.g. a "device"), and throw an exception if something goes wrong that you don't expect to recover from. Wrap the whole function [or an outer set of functions] in a try/catch block.
Of course, if "failure is normal" (e.g. you try to read from a network port, and you get a timeout because there was no data available, that's not something you should throw an exception for). This should use a return-value.
Note that using objects to handle resources requires careful design of the code and especially that your destructor does a good job of cleaning it is run after an exception. Make sure you don't leave things behind if you throw an exception in a constructor - as that is how you get leaks.
There are naturally a whole host of other solutions - after all, we're talking programming, so there's always at least 11 different ways to solve something.
In C++, RAII is often advocated as a superior approach to exception handling: if an exception is thrown, the stack is unwound, all the destructors are called and resources are cleaned up.
However, this presents a problem with error reporting. Say a very generic function fails, the stack is unwound to the top level and all I see in the logs would be:
Couldn't read from socket: connection reset by peer.
...or any equally generic message. This doesn't say much about the context from which the exception is thrown. Especially if I'm running something like an event queue processing loop.
Of course I could wrap every call to socket reads with a try/catch block, catch the exception, construct a new one with more detailed context information and re-throw it, but it defeats the purpose of having RAII, and is slowly but surely becoming worse than handling return error codes.
What's a better way for detailed for error reporting in standard C++? I'm also open to suggestions involving Boost.
As James McNellis suggested here, there is a really neat trick involving a guard object and the std::uncaught_exception facility.
The idea is to write code like this:
void function(int a, int b)
{
STACK_TRACE("function") << "a: " << a << ", b: " << b;
// do anything
}
And have the message logged only in case an exception is actually thrown.
The class is very simple:
class StackTrace: boost::noncopyable // doesn't make sense to copy it
{
public:
StackTrace(): mStream() {}
~StackTrace()
{
if (std::uncaught_exception())
{
std::cout << mStream.str() << '\n';
}
}
std::ostream& set(char const* function, char const* file, unsigned int line)
{
return mStream << file << "#" << line << " - " << function << " - ";
}
private:
std::ostringstream mStream;
};
#define STACK_TRACE(func) \
StackTrace ReallyUnwieldyName; \
ReallyUnwieldyName.set(func, __FILE__, __LINE__)
One can use __PRETTY_FUNC__ or equivalent to avoid naming the function, but I have found in practice that it was too mangled / verbose for my own taste.
Note that you need a named object if you wish it to live till the end of the scope, which is the purpose here. We could think of tricky ways to generate a unique identifier, but I have never needed it, even when protecting narrower scope within a function, the rules of name hiding play in our favor.
If you combine this with an ExceptionManager (something where exceptions thrown register themselves), then you can get a reference to the latest exception and in case of logging you can rather decide to setup your stack within the exception itself. So that it's printed by what and ignored if the exception is discarded.
This is a matter of taste.
Note that in the presence of ExceptionManager you must remain aware that not all exceptions can be retrieved with it --> only the ones you have crafted yourself. As such you still need a measure of protection against std::out_of_range and 3rd party exceptions.
Say a very generic function fails, the stack is unwound to the top level and all I see in the logs would be [...]
What's a better way for detailed for error reporting in standard C++?
Error handling isn't local to a class or library - it is a application level concern.
Best I can answer you question is that the error reporting should be always implemented by looking first and foremost at the error handling. (And the error handling also including the handling of the error by the user.) Error handling is the decision making about what has to be done about the error.
That is one of the reasons why error reporting is an application level concern and strongly depends on the application workflow. In one application the "connection reset by peer" is a fatal error - in another it is a norm of life, error should be silently handled, connection should be reestablished and pending operation retried.
Thus the approach you mention - catch the exception, construct a new one with more detailed context information and re-throw it - is a valid one too: it is up to the top level application logic (or even user configuration) to decide whether the error is really an error or some special (re)action has to taken upon the condition.
What you have encountered is one of the weaknesses of so called out-of-line error handling (aka exceptions). And I'm not aware of any way to do it better. Exceptions create a secondary code path in the application and if error reporting is vital the design of the secondary code path has to treated just like the main code path.
Obvious alternative to the out-of-line error handling is the in-line error handling - good ol' return codes and log messages on the error conditions. That allows for a trick where application saves all low-severity log messages into internal (circular) buffer (of fixed or configurable size) and dumps them into the log only if high-severity error happens. This way more context information is available and different layers of application do not have to be actively aware of each other. That is also standard (and sometimes literally "standard" - mandated by law) way of error reporting in application fields like safety and mission critical software, were one is not allowed to miss errors.
I've never actually done this, but you could roll your own "stacktrace":
struct ErrorMessage {
const char *s;
ErrorMessage(const char *s) : msg(s) {}
~ErrorMessage() { if (s) std::cout << s << "\n"; }
void done() { s = 0; }
};
void someOperation() {
ErrorMessage msg("Doing the first bit");
// do various stuff that could throw
msg = "Doing the second bit";
// do more stuff that could throw
msg.done();
}
You can have several levels of this (although not necessarily use it at every level):
void handleFoo() {
ErrorMessage msg("Handling foo event");
someOperation();
msg.done();
}
And add more constructors and members:
void handleBar(const BarEvent &b) {
ErrorMessage msg(std::stringstream("Handling bar event ") << b.id);
someOperation();
msg.done();
}
And you needn't write the messages to std::cout. It could be to some logging object, and the object could queue them, and not actually output them to the log unless the catch site tells it to. That way, if you catch an exception that doesn't warrant logging, nothing is written.
It's not pretty, but it's prettier than try/catch/throw or checking return values. And if you forget to call done on success (for example if your function has multiple returns and you miss one), then you will at least see your mistake in the logs, unlike a resource leak.
[Edit: oh, and with a suitable macro you can store __FILE__ and __LINE__ in the ErrorMessage.]
You could add a call stack to your exception. I'm not sure how good this works for release builds but works like a charm with debug. You could do this in the constructor of your exception (to encapsulate it). See here for a starting point. This is possible as well on Linux - eventhough I dont remember how exactly.
I use RAII and exceptions and just have various unit-test-like assert statements throughout the code - if they fail, the the stack unwinds to where I can catch and handle them.
#define APP_ASSERT_MSG(Class,Assertion,szDescription) \
if ( !(Assertion) ) \
{ \
throw Class(__LINE__,__FILE__,szDescription); \
}
For most of my particular use case, all I care about is logging debug information, so my exception has the file and line number in it along with an error message (message is optional as I have an assert without it as well). You could easily add FUNCTION or an error code of some type for better handling.
I can then use it like this:
int nRet = download_file(...);
APP_ASSERT_MSG(DownloadException == ERR_OK, "Download failed");
This makes error handling and reporting much easier.
For really nasty debugging, I used GCC's function instrumentation to keep a trace list of what's going on. It works well, but slows down the app quite a bit.
What I do regularly, FWIW, is not use exceptions, but rather explicit error handling in a standard format (ie: using a macro). For example:
result = DoSomething();
CHECK_RESULT_AND_RETURN_ON_ERROR( result );
Now, obviously, this has many limitations design-wise:
Your return codes may need to be somewhat uniform
Code is cluttered with macros
You may need many macros for various check conditions
The upside, though, is as Dummy00001 said: you can effectively generate a stack trace on demand in the event of a serious error, by simply caching the results. I also use this paradigm to log all unexpected error conditions, so I can adjust the code later to handle unexpected conditions which occur "in the wild".
That's my 2c.
Here's how I handle error reporting in my libraries (this may or may not suit your situation).
First, as part of your design you want a "core" or "system" library that all this common logic will sit in. All you other libraries will then link to the core and use its error reporting APIs, so your whole system has a single compact chunk of logic for handling this mess.
Inside the core, provide a set of logging MACROS such as "LogWarning" and "LogFatal" with documented behavior and expected usage- for example, LogFatal should trigger a hard abort of the process but anything lower than "LogError" is simply advisory (does nothing). The macros can provide a "printf" interface that automatically appends the "LINE", "FILE", and "FUNC" macros as arguments to the underlying singleton object that handles your error reporting.
For the object itself I'll defer to you. However, you want public APIs that configure your "drains"- e.g. log to stderr, log to logfile, log to MS Services log, etc. You also want the underlying singleton to be thread safe, re-entrant where possible, and very robust.
With this system, you can make "exception reporting" ONE MORE DRAIN TYPE. Just add an internal API to that error manager object that packages your logged message as an exception, then throws it. Users can then enable AND DISABLE exceptions-on-error behavior in your code with ONE LINE in their apps. You can put try-catches around 3rd party or system code in you libraries that then calls a "Log..." macro in the catch clauses to enable clean re-throw behavior (with certain platforms and compiler options you can even catch things like segfaults using this).
I am just learning how to handle errors in my C++ code. I wrote this example that looks for a text file called some file, and if its not found will throw an exception.
#include <iostream>
#include <fstream>
using namespace std;
int main()
{
int array[90];
try
{
ifstream file;
file.open("somefile.txt");
if(!file.good())
throw 56;
}
catch(int e)
{
cout<<"Error number "<<e<<endl;
}
return 0;
}
Now I have two questions. First I would like to know if I am using Exceptions correctly. Second, (assuming the first is true) what is the benefit to using them vs an If else statement?
"Correctly" is a value judgment, but (unlike other classes) there's a major benefit from exceptions classes being a monolithic hierarchy, so I'd generally advise throwing something derived from std::exception, not simply an int.
Second, it's open to question whether an incorrect file name is sufficiently unexpected to qualify as a good reason to throw an exception at all.
As to benefits vs. an if/else statement: there are a couple. First, exceptions let you segregate the code that deals with errors, so the main idea and readability of the code don't get lost in a maze of error handling. Second, when you have several layers of code between throwing and catching the exception, the code that throws the exception may not know how it should really be handled. Your code, for example, uses std::cout to report the problem -- but most such code would report errors on std::cerr instead. You can change from one to the other without any change to the code that tried to open the file (which might be deep in a library, and have no clue of which should be used for this application -- and might be used in an application where both are wrong, and MessageBox was preferred).
First I would like to know if I am using Exceptions correctly.
Yes, though generally you want your exceptions to derive from std::exception.
Second, (assuming the first is true) what is the benefit to using them vs an If else statement?
For the given example, nothing. The benefit of exceptions comes when you have
many deep nested functions, like this.
#include <stdexcept>
#include <iostream>
#include <string>
void anErrorFunc(const std::string& x)
{
ifstream file;
file.open(x);
if (!file)
throw std::runtime_error("Could not open file");
}
void someOtherFunction(const std::string& y)
{
//Do stuff
anErrorFunc(y);
//Do other stuff
}
int main()
{
try {
someOtherFunction("somefile.txt");
} catch (std::exception &ex) {
std::cout << "Ouch! That hurts, because: "
<< ex.what() << "!\n";
}
}
Note that the exception will be caught in main(), and someOtherFunction
does not have to worry about dealing with passing through failure return codes.
Well, you are using exception correctly in that there is nothing wrong with your code. That said, typically we do not throw primitive types around (even though you can). It is generally a better idea to throw an object that derives from std::exception, and even better to throw an std::exception that is also a boost::exception.
When things are very simple and the handling code and the throwing code are in the same function, then there really is no reason to use exceptions instead of if statements (and, indeed, it would be faster and more efficient to use if...else in that particular case). However, in most situations, the point where the error is discovered and you need to report it is far removed from the logic where the error is to be handled. In many cases the error-recovery logic is specific to the application in question and the logic where the error is discovered cannot make a sensible choice about how to recover from the error, hence the need to throw.
Another benefit of exception handling is that the type of the exception can be used to convey the type of error that occurred. Usually the types in exception hierarchies are much more meaningful than those error codes that end up being used in C code. Also, you cannot ignore an exception as easily as you can ignore an error code; while you can ignore an exception, it will cause the program to die a horrible death. By contrast, if a C function returns an error status code, and you ignore it, it's possible to continue executing and get silently wrong results... in that sense, the use of exceptions is much safer than using error codes.
You may also be interested in reading about exceptions and error handling from the C++ FAQ Lite.
You are not using exceptions correctly. Your code has a much simpler equivalent, without exceptions, which provides the same behaviour.
Exceptions are for when you can't test the results of a function call with other methods. In this case you can, so you shouldn't use exceptions.
As a corollary, you shouldn't throw and catch the exception in the same function body - just do whatever you want to do instead of throwing it.
Syntactically speaking, your code is correct. Idiomatically speaking, maybe not so much -- or at least this depends on the context. When a file can't open then we might do our handling right there inside that if( !file.good() ) if it's perfectly common and possible to happen. For example, if the user asks to open a file in a text editor then it's perfectly plausible and common that the file doesn't exist. On the other hand, if the editor can't find the spelling corpus file then that means something is (arguably) terribly wrong. The program wasn't probably installed, or the user messed around with that file -- anything is possible.
In C++ we use exceptions for exceptional cases. That is, cases that are really not meant to happen and that we don't "accept" to happen. This is as opposed to a user file not opening, or an invalid user input, or no internet connection: these are all examples of perfectly valid things happening, common situations, things we expect to happen sooner or later in a program's run. They aren't exceptional.
Now, what are the benefits of using exceptions compared to conditionals? Allow me to extend this question to any other jump (goto) mechanism: returns as well as conditionals. Exceptions are more expressive if that's what you want to say: if you are dealing with an exceptional case then use exceptions. Exceptions also get more done than plain conditionals, in an analogous way to that of virtual functions getting more done than conditionals. The right code-block will be executed depending on the exception, but also the right scope will handle the exception depending on the handlers.
There are other advantages to exceptions compared with conditionals: exceptions separate error-handling from other code. They allow associating arbitrary data and actions with an error state. They allow communicating success states (via a return) as well as error state (via a throw). And the list goes on...
Technically speaking and at the lowest level exceptions are a sophisticated jump mechanism. Back in the butterfly-days people invented the if conditional as a somewhat-sophisticated goto in order to enhance expressiveness (because a goto can be used for anything really) and to reduce programmer errors. The looping constructs such as the C for loop are also in essence a sophisticated jump with sparkles and rainbow colours, again for reducing errors and enhancing expressiveness. C++ introduced its new casting operators for the same reasons.
So there: exceptions aren't magic, just something somewhat new in the scene compared to conditionals and loops. Don't use them when you don't mean to, just like you don't use a loop when you really mean to use a conditional.
Syntactically what you're doing is right. Style-wise, as others have noted, you should be throwing something descended from std::exception.
As to part two of your question, I'd like to go into more detail on that.
The whole point of exceptions is to separate policy from implementation. As Billy ONeal said, you get no benefit at all from using exceptions within the same function that an if statement wouldn't make better. You need to be deeply nested in function calls for it to make sense.
In most code, your high level code has enough information and context to know what to do about errors, but no mechanism to detect them. Your low level code can detect the errors but has none of the information needed to deal with them.
The traditional means of coping with this -- returning error codes -- has a few problems:
It clutters up the code with error handling code to the point that the actual logic is obfuscated.
It relies on programmers not being lazy and checking EVERY error code return, an often foolhardy assumption. (C programmers, be honest here: when was the last time you checked the return value of printf?)
It adds the overhead of error checking and handling to EVERY function call whether there's an error or not.
Exceptions solve these issues (with varying degrees of success).
Exceptions solve #1 by only having exception-related code at the point of detection and at the point of handling. Intervening functions don't get cluttered with handling for obscure errors that they themselves have no interest in nor capability of dealing with.
They solve #2 by forcing handling. You can't ignore an exception. You have to take action on them. (Lazy programmers can still catch all exceptions and then ignore them, but here their crippling inability to program is now highlighted for all to see.)
They solve #3 (when not naively implemented) by having near-zero costs when not used, albeit often at a very, very high cost when actually used.
This is not to say that exceptions are the end-all/be-all of error handling. The disadvantages:
Exceptions are usually very costly when used. They have to be eschewed, despite their advantages, if performance is paramount.
Exceptions lead to very opaque code at times. They are non-local transfers of control -- effectively slightly safer versions of goto statements, but across functions. An exception can transfer control from hundreds of layers deep in your code in source files not even slightly related to the ones you're working on (and, in fact, quite possibly not even accessible to you). This kind of "spooky action at a distance" can make code very difficult to figure out.
"Checked exceptions" can actually be worse for noise generation than the old-style if handling. They tend to be more verbose, you see, than if or switch statements and the fact that you must handle checked exceptions for the code to even compile makes them a liability to many situations.
Because of their often high cost of use, carelessly using them for all error handling can make your code slow and bloated.
I've programmed C and C++ for a long time and so far I've never used exceptions and try / catch. What are the benefits of using that instead of just having functions return error codes?
Possibly an obvious point - a developer can ignore (or not be aware of) your return status and go on blissfully unaware that something failed.
An exception needs to be acknowledged in some way - it can't be silently ignored without actively putting something in place to do so.
The advantage of exceptions are two fold:
They can't be ignored. You must deal with them at some level, or they will terminate your program. With error code, you must explicitly check for them, or they are lost.
They can be ignored. If an error can't be dealt with at one level, it will automatically bubble up to the next level, where it can be. Error codes must be explicitly passed up until they reach the level where it can be dealt with.
The advantage is that you don't have to check the error code after each potentially failing call. In order for this to work though, you need to combine it with RAII classes so that everything gets automatically cleaned up as the stack unwinds.
With error messages:
int DoSomeThings()
{
int error = 0;
HandleA hA;
error = CreateAObject(&ha);
if (error)
goto cleanUpFailedA;
HandleB hB;
error = CreateBObjectWithA(hA, &hB);
if (error)
goto cleanUpFailedB;
HandleC hC;
error = CreateCObjectWithA(hB, &hC);
if (error)
goto cleanUpFailedC;
...
cleanUpFailedC:
DeleteCObject(hC);
cleanUpFailedB:
DeleteBObject(hB);
cleanUpFailedA:
DeleteAObject(hA);
return error;
}
With Exceptions and RAII
void DoSomeThings()
{
RAIIHandleA hA = CreateAObject();
RAIIHandleB hB = CreateBObjectWithA(hA);
RAIIHandleC hC = CreateCObjectWithB(hB);
...
}
struct RAIIHandleA
{
HandleA Handle;
RAIIHandleA(HandleA handle) : Handle(handle) {}
~RAIIHandleA() { DeleteAObject(Handle); }
}
...
On first glance, the RAII/Exceptions version seems longer, until you realize that the cleanup code needs to be written only once (and there are ways to simplify that). But the second version of DoSomeThings is much clearer and maintainable.
DO NOT try and use exceptions in C++ without the RAII idiom, as you will leak resources and memory. All your cleanup needs to be done in destructors of stack-allocated objects.
I realize there are other ways to do the error code handling, but they all end up looking somewhat the same. If you drop the gotos, you end up repeating clean up code.
One point for error codes, is that they make it obvious where things can fail, and how they can fail. In the above code, you write it with the assumption that things are not going to fail (but if they do, you'll be protected by the RAII wrappers). But you end up paying less heed to where things can go wrong.
Exception handling is useful because it makes it easy to separate the error handling code from the code written to handle the function of the program. This makes reading and writing the code easier.
return an error code when an error condition is expected in some cases
throw an exception when an error condition is not expected in any cases
in the former case the caller of the function must check the error code for the expected failure; in the latter case the exception can be handled by any caller up the stack (or the default handler) as is appropriate
Aside from the other things that were mentioned, you can't return an error code from a constructor. Destructors either, but you should avoid throwing an exception from a destructor too.
I wrote a blog entry about this (Exceptions make for Elegant Code), which was subsequently published in Overload. I actually wrote this in response to something Joel said on the StackOverflow podcast!
Anyway, I strongly believe that exceptions are preferable to error codes in most circumstances. I find it really painful to use functions that return error codes: you have to check the error code after each call, which can disrupt the flow of the calling code. It also means you can't use overloaded operators as there is no way to signal the error.
The pain of checking error codes means that people often neglect to do so, thus rendering them completely pointless: at least you have to explicitly ignore exceptions with a catch statement.
The use of destructors in C++ and disposers in .NET to ensure that resources are correctly freed in the presence of exceptions can also greatly simplify code. In order to get the same level of protection with error codes you either need lots of if statements, lots of duplicated cleanup code, or goto calls to a common block of cleanup at the end of a function. None of these options are pleasant.
Here's a good explanation of EAFP ("Easier to Ask for Forgiveness than Permission."), which I think applies here even if it's a Python page in Wikipedia. Using exceptions leads to a more natural style of coding, IMO -- and in the opinion of many others, too.
When I used to teach C++, our standard explanation was that they allowed you to avoid tangling sunny-day and rainy-day scenarios. In other words, you could write a function as if everything would work ok, and catch the exception in the end.
Without exceptions, you would have to get a return value from each call and ensure that it is still legitimate.
A related benefit, of course, is that you don't "waste" your return value on exceptions (and thus allow methods that should be void to be void), and can also return errors from constructors and destructors.
Google's C++ Style Guide has a great, thorough analysis of the pros and cons of exception use in C++ code. It also indicates some of the larger questions you should be asking; i.e. do I intend to distribute my code to others (who may have difficulty integrating with an exception-enabled code base)?
Sometimes you really have to use an exception in order to flag an exceptional case. For example, if something goes wrong in a constructor and you find it makes sense to notify the caller about this then you have no choice but to throw an exception.
Another example: Sometimes there is no value your function can return to denote an error; any value the function may return denotes success.
int divide(int a, int b)
{
if( b == 0 )
// then what? no integer can be used for an error flag!
else
return a / b;
}
The fact that you have to acknowledge exceptions is correct but this can also be implemented using error structs.
You could create a base error class that checks in its dtor whether a certain method ( e.g. IsOk ) has been called. If not, you could log something and then exit, or throw an exception, or raise an assert, etc...
Just calling the IsOk on the error object without reacting to it, would then be the equivalent of writing catch( ... ) {}
Both statement would display the same lack of programmer good will.
The transport of the error code up to the correct level is a greater concern. You would basically have to make almost all methods return an error code for the sole reason of propagation.
But then again, a function or method should always be annotated with the exceptions it can generate. So basically you have to same problem, without an interface to support it.
As #Martin pointed out throwing exceptions forces the programmer to handle the error. For example, not checking return codes is one of the biggest sources of security holes in C programs. Exceptions make sure that you handle the error (hopefully) and provide some kind of recover path for your program. And if you choose to ignore an exception rather than introduce a security hole your program crashes.