This bug is cascading to another: https://bugs.php.net/bug.php?id=65915 -- consider this test script:

<?php
$tmp = '/tmp/testNg3hUy';
foreach (['a','b'] as $f) {
  file_put_contents($tmp, '<?php return function(){ return "'.$f.'"; };');
  echo file_get_contents($tmp), "\n";
  $$f = require $tmp;
}
printf( "%s, %s\n ", $a(), $b());
unlink($tmp);

The require assignment should generate the equivalent of 

$a = function(){ return "a"; };
$a = function(){ return "b"; };

so that the printf should generate 

a, b

However it actually generates

a, a

demonstrating that the "\0{closure}%s%p" naming convention does not generate unique closure names.

Sorry the second line above should read:

$b = function(){ return "b"; };

The issue here is that the runtime defined function keys should be unique, and the current algo does not generate unique keys.  There are many approaches that could be taken to remove such incorrect name clashes.  The patch that I've submitted add a sequence count to the key -- simple but sufficient to prevent this bug.

Just to note that my patch is not strong enough when used with OPcache, since a simple static running count can fail when the interpreter is forked.  Better alternatives include 

* some form of unique id, e.g. a UUID or uniquid(true)
* a content based hash, such as md5file(__FILE__) -- though if this were to be adopted then it would be better always to generate this as part of the compile creating say __FILE_MD5__, though this would add a few % to the compile time.

This all needs wider discussion of these issues on the Internals ML.

Your second script prints "a, a" only with OPCache, because it caches the included temporary file. Without OPCache it prints the expected "a, b".

I also don't think that the first script indicates a bug. The more functions you create the more memory it requires.

What threw me was the botch with the temporary entries "\0{closure}$filenane$offset" are used in the EG(function_table).  This is as clear as mud.  When a file is compiled, a function table entry is created for each closue in the source.  This entry is never executed directly, but is used by the ZEND_DECLARE_LAMBDA_FUNCTION to construct the closure object which contains a deep copy of this zend_function record.  It is this copy that used when the closure is called.  So long as the ZEND_DECLARE_LAMBDA_FUNCTION
are executed within the same scope as the compile, this should normally be unique, but it is quite easy to construct a test case which the unique assumption fails: 

--TEST--
ISSUE #65915 Temporary function entries for closures are not unique
--INI--
opcache.enable=0
--SKIPIF--
--FILE--
<?php
$tmp = tempnam(__DIR__, 'test');
foreach (['a','b'] as $f) {
  file_put_contents($tmp, "<?php function $f() {return function(){ return '$f'; };}");
  echo file_get_contents($tmp), "\n";
  require $tmp;
}
$a = a();
$b = b();
printf( "%s, %s\n ", $a(), $b());
unlink($tmp);
?>
--CLEAN--	
--EXPECT--
<?php function a() {return function(){ return 'a'; };}
<?php function b() {return function(){ return 'b'; };}
a, b

And here's the eval version:

--TEST--
ISSUE #64291 Temporary function entries for closures are not unique
--INI--
opcache.enable=1
--FILE--
<?php
foreach (['a','b'] as $f) {
  $tmp = "function $f() {return function(){ return '$f'; };}\n";
  eval($tmp);
  echo $tmp;
}
$a = a();
$b = b();
printf( "%s, %s\n ", $a(), $b());
?>
--EXPECT--
function a() {return function(){ return 'a'; };}
function b() {return function(){ return 'b'; };}
a, b

Sorry, the above should read opcache.enable=0 though the failure is the same for opcache enabled and not enabled in the eval case.

Dmitry, 

I have just realised that this "mangled names should be unique" issue applied to any runtime bound function or class as the following -- albeit perverse example shows:

--TEST--
ISSUE #65915A Temporary class entries are not unique
--INI--
opcache.enable=0
--SKIPIF--
--FILE--
<?php
$tmp = tempnam(__DIR__, 'test');
foreach (['a','b'] as $f) {
  file_put_contents($tmp, <<<END
<?php 
  function $f() {
    class Hello { const WORLD = "Hello world from $f\\n"; }
  }
END
);
  require $tmp;
}
a();
echo Hello::WORLD;
unlink($tmp);
?>
--CLEAN--	
--EXPECT--
Hello world from a

Here the two functions both compile a class with a mangled name "\0$class$filename$string_addr" which is the same for the a() and b() copies
so b() version overwrites the a() one, and the DECLARE_CLASS opcode in a() incorrecly binds to the wrong class, hence Hello::WORLD incorrectly prints out the "from b" version.

However, I suspect in practice that this is unlikely to manifest itself in real word apps.

I'm getting bitten by this bug, too. My use case is basically that I'm generating and eval()'ing code. Since this is not in a one-shot HTTP server process environment but in a longer running process that caches the generated code for performance reasons, this causes various failures that depend on the order in which requests are made.

Using Terry's patch on PHP 5.4.24 on Debian/AMD64 caused the problem to go away, my services now run stable. What's keeping you from applying this patch (apart from the fact that it mixes tabs/spaces)?

Closing this is a duplicate of bug #76982. PHP 8.1 mostly (but not completely) fixes the leak, and the RTD key collision issue has already been addressed previously (in some 7.x version I believe).