Table of Contents

PHP RFC: Convert numeric keys in object/array casts



The PHP language has two core data types which are collections of key/value pairs.

The first of these, the array, is an ordered map of integer or string keys to arbitrary values. There is no overlap between integer and string keys in arrays; if a string fits the format /^(0|(-?[1-9][0-9]*))$/ and is small enough (PHP_INT_MIN ≤ n ≤ PHP_INT_MAX), it is converted to an integer key. Such strings are termed numeric strings.

The second of these, the object, is an ordered map of string property names to arbitrary values. Integer property names are not permitted, these are converted to string property names. Objects have some other attributes, but they do not concern us here.

In the Zend Engine, both PHP arrays and PHP objects are internally represented using the same data structure, the HashTable.1 The HashTable is, much like the array, an ordered map of integer or string keys to arbitrary values. However, unlike arrays, there is no guarantee that string and integer keys do not overlap; for instance, a HashTable can simultaneously contain the separate keys "123" and 123, which may correspond to different values.

Because arrays and objects have different restrictions versus the underlying HashTable type on what kinds of keys they can have, the Zend Engine must enforce their restrictions at a layer above HashTables, in the code implementing arrays and objects themselves. This means that if that code is bypassed and the underlying HashTables are modified directly, arrays and objects can exist with an invalid internal state.

1Nitpick: objects with only declared properties do not need their own HashTable, but even they use a HashTable for looking up properties, it's just part of the class rather than the object itself.

The Problem

Various edge cases in the Zend Engine exist where array HashTables can contain numeric string keys, and object HashTables can contain integer keys. In such cases, these keys are inaccessible from PHP code, because the code handling arrays will never look for numeric string keys in the HashTable (as arrays map those to integer keys), and the code handling objects will never look for integer keys in the HashTable (as objects map those to string keys).

This RFC focuses on a specific edge case, that of object-to-array casts and array-to-object casts. Currently, when using (object) or settype() to convert an object to an array, or when using (array) or settype() to convert an array to an object, the inner HashTable is naïvely copied or referenced without its keys being changed to reflect the restrictions on array keys and object property names, leading to inaccessible array keys or object properties in some cases. For example, $arr = [0 => 1, 1 => 2, 2 => 3]; $obj = (object)$arr; produces an object with inaccessible properties named 0, 1 and 2, while $obj = new stdClass; $obj->{'0'} = 1; $obj->{'1'} = 2; $obj->{'2'} = 3; $arr = (array)$obj; produces an array with the inaccessible keys "0", "1" and "2". The same issue also occurs when using get_object_vars().



This RFC proposes to fix this issue for object-to-array casts and array-to-object casts, both for the casting operators and for settype(), and also fix the same issue in get_object_vars(). This would be done by converting the keys of array or object HashTables as appropriate, so numeric string property names in objects would be converted to integer array keys, and vice-versa. Therefore, there would be no inaccessible properties. For example, $arr = [0 => 1, 1 => 2, 2 => 3]; $obj = (object)$arr; would now produce an object with accessible properties named "0", "1" and "2", and $obj = new stdClass; $obj->{'0'} = 1; $obj->{'1'} = 2; $obj->{'2'} = 3; $arr = (array)$obj; would now produce an array with the accessible keys 0, 1 and 2.


There have been attempts to fix this issue before, but there is a potential performance issue: naïvely copying the HashTable (or instead adding another reference to it if possible) without performing key conversion is much faster than creating a new HashTable and iterating over every key in the old HashTable to manually copy each key/value pair to the new HashTable, converting if necessary.

In order to minimise the potential performance impact, the proposed implementation would avoid expensively manually duplicating the whole HashTable wherever possible, by first checking if this is necessary, either by checking flags (for example, packed arrays are guaranteed to need conversion if being converted to an object), or by iterating over the HashTable checking for keys needing conversion. If conversion is not necessary, it will fall back to the faster zend_array_dup(), or merely copy the reference if possible. Because it only performs manual duplication where necessary, the most common cases (converting arrays with only string keys to objects, and converting objects with only non-numeric string property names to arrays) see minimal performance impact.

In the case of get_object_vars(), the object HashTable was always duplicated anyway, so the only change is that it now checks for numeric string property names in its main loop.

For the purpose of these conversions, new zend_symtable_to_proptable() (array-style HashTable to object-style HashTable) and zend_proptable_to_symtable() (object-style HashTable to array-style HashTable) functions are added to the Zend API.

Backward Incompatible Changes

The current behaviour, though arguably an unhelpful oversight, is documented. Therefore, fixing this issue means changing documented behaviour, and so breaks backwards-compatibility.

The justification for breaking backwards-compatibility here is that the existing behaviour is unintuitive and unhelpful. This is an uncommon edge case that is unlikely to be relied upon, because it prevents the user doing anything useful with the result.

Proposed PHP Version(s)

This is proposed to be changed in the next minor or major version of PHP, whichever comes first. At the present time, that would be 7.2.

RFC Impact


I do not believe there to be any impact on the behaviour of the SAPIs.

To Existing Extensions

Any extension using the convert_to_object() and convert_to_array() Zend API functions will now see the different behaviour described above.

To Opcache

The implementation does not fail any tests when compiled with OPcache.


I am yet to benchmark the current implementation, and I have done no benchmarks on real-world applications so far.

I performed a benchmark on an earlier version of the patch. As hoped, it found only small (at worst ~19%) slowdowns versus the master branch for the common cases (casting arrays with no integer keys to objects and casting objects with no numeric string property names to arrays). The uncommon cases (casting arrays with some or all keys integers to objects, and casting objects with some or all keys numeric to arrays), on the other hand, saw much larger slowdowns (as much as ~364%). I would argue that this is a bearable performance decrease, given that otherwise the uncommon cases produce an unusable result.

Open Issues

There are no open issues at this time.

Unaffected PHP Functionality

This has no impact on the behaviour of arrays or objects, nor the operators acting upon them (beyond the cast operators). It also does not deal with other edge cases that cause arrays and objects to contain invalid keys.

Future Scope

Object/array casts are not the only edge case which creates arrays and objects with invalid keys. It may be worth considering a comprehensive solution (for example, performing numeric string to integer normalisation universally, rather than solely for arrays) in future. That would be a much larger undertaking than this RFC, however, and has greater possible downsides (such as reduced performance for property and variable accesses).


This could be construed as a language change, so this RFC requires a 2/3 majority in voting to be accepted.

It is a single Yes/No vote on whether to accept the RFC and implement it in PHP 7.2. Voting started on 2016-11-05 and ended on 2016-11-14. The result was to accept the RFC for 7.2.

Accept the Convert numeric keys in object/array casts RFC for PHP 7.2?
Real name Yes No
ajf (ajf)  
bwoebi (bwoebi)  
cmb (cmb)  
colinodell (colinodell)  
cpriest (cpriest)  
daverandom (daverandom)  
guilhermeblanco (guilhermeblanco)  
hywan (hywan)  
laruence (laruence)  
lcobucci (lcobucci)  
malukenho (malukenho)  
marcio (marcio)  
mariano (mariano)  
mgocobachi (mgocobachi)  
mike (mike)  
mrook (mrook)  
nikic (nikic)  
ocramius (ocramius)  
sammyk (sammyk)  
santiagolizardo (santiagolizardo)  
svpernova09 (svpernova09)  
trowski (trowski)  
Final result: 21 1
This poll has been closed.

Patches and Tests

The pull request for the PHP interpreter is here:

There is no language specification patch, because none is required. The language specification did not specify or comment on this bug.


This is implemented in master, which will become PHP 7.2. The commit is:

After the project is implemented, this section should contain

  1. a link to the PHP manual entry for the feature


None at present.

Rejected Features