The correct way of writing this would be to catch the exception and throw it again, however this also fails. If it is not thrown again, then the behavior is correct.

<?php

class MyException extends Exception
{
    public function __construct($message = "", $code = 0, Throwable $previous = null)
    {
        parent::__construct($message, $code, $previous);
        print "2. Exception constructed\n";
    }
}

function throw_exception() {
    try {
        throw new MyException("5. Exception message\n");
    } catch(Exception $exception) {
        print "3. Exception caught and thrown again\n";
        throw $exception;
    } finally {
        return "4. Return statement\n";
    }
}

try {
    print "1. Start\n";
    print throw_exception();
} catch (Exception $e) {
    print $e->getMessage();
}
The flow should be: exception is thrown, caught by the inner block, thrown again, caught by the outer block and then the finally block should execute, returning the flow to 
print throw_exception();

> but it MUST NOT be silently dropped out.

That's a conclusion without rigorous justification.

btw The behaviour in PHP is the same as that which happens in Java: https://web.archive.org/web/20070922061412/http://weblogs.java.net/blog/staufferjames/archive/2007/06/_dont_return_in.html

It'd be a lot easier to say "Don't put returns in finally blocks, yo".

> The interpreter should trigger an E_NOTICE or something similar here.

This would almost certainly be better done with code analysis tools, rather than giving warnings on valid (though highly surprising) code.

This behaviour is also that used in Javascript.

function foo() {
  try {
	throw 42; 
  }
  finally {
    return "edge cases are hard";
  }
};

alert(foo());

> btw The behaviour in PHP is the same as that which happens in Java
> This behaviour is also that used in Javascript.

Sorry to hear that, but AFAIK nor Java nor JavaScript (ECMAScript) has no conventional (standartized) way to produce non critical warnings at the run time.

So PHP is better in this way and why can't we employ it?

> "Don't put returns in finally"
I believe that the good tool must help its users not to "shoot themselves" as much as possible. So the user becomes free from a lot of "Don't do something somewhere". If you don't agree you are free to use an assembler to do anything you want and remember a lot of such restrictions.

I can imagine another more expectable an clear way to handle this case. But it will break a BC so can be implemented only in a new major release of PHP.
Triggering E_NOTICE is the most convenient way for now.

BC is broken with any solution. The fact that javascript behaves in the same way worries me. It's probably very hard to implement. I agree with the E_NOTICE.

2 danack@php.net

Sorry, I don't want to be mean. My English skills are not very good and it is sometimes hard to express my thoughts in a clear way.

I've just realized that the finally statement is some-kind of antipattern for me. Because finally is a shortcut for catching and re-throwing an \Exception class eg.

<?php
try {
  throw new CustomException();
} catch (CustomException $ex) {
  handleException();
} catch (\Exception $ex) {
  doThingsInFinally();
  throw $ex;
}
doThingsInFinally();
?>

Or with finally:

<?php
try {
  throw new CustomException();
} catch (CustomException $ex) {
  handleException();
} finally {
  doThingsInFinally();
}
?>

Of cause if we have a return statement instead of a doTihingsInFinally() method, we will lose an exception. But we doing it in a clear way. On the other hand finally statement hides out from developer's eyes what actually is going.

So for my mind a little copy-paste is a lesser evil than an unclear behavior. For all other guys the E_NOTICE will be enough. (:

It seems to me this is simply a documentation issue.

The Php Inspections (EA Extended) guy has added this as code smell to that tool.

You could consider using that tool and giving them some money for being so nice: https://www.indiegogo.com/projects/php-inspections-ea-extended-a-code-analyzer-security#/

This faulty behaviour of the return statement in the finally{} of the the try{} block extends beyond the catchable error/exceptions.
1%0;  // No warning or error.
unknown_function(); // not reported. Program not aborted
It took me days to locate a misspelled function, because no error is reported at all and program would normally abort, now continue with undesirable behaviour.
Example code:
<?php
error_reporting(E_ALL|E_STRICT);
function do_test(): int {
  try {
    1/0;                    // a warning in php 7 - reported as documented
    nonexistent_function(); // a fatal error  - not reported - program continues
    1%0;                    // a fatal error
    return 0;
  } catch (Exception $e) {
    printf("CATCH\n");
    return 1;
  } finally {
    printf("FINALLY\n");
    return 2;
  }
  return 3;
}
printf("%d\n", do_test());

@anhlephuoc: The behavior you're seeing is because the "fatal error" from the undefined function call is actually an exception that extends Error, not an exception that extends Exception. You aren't catching it which means normally it would have been thrown - if not for the fact that your finally block overrides that behavior and returns a value instead.

As of PHP 7, "catch (Exception $e)" does not catch every single exception that could be thrown. Use the Throwable interface for that.
https://3v4l.org/6N4d9