This post is archived and probably outdated.


2010-08-22 23:52:00

In case you ever heard me talking about PHP internals I certainly mentioned something along the lines of "Everything in PHP is a HashTable" that's not true, but next to a zval the HashTable is one of the most important structures in PHP. While preparing my "PHP Under The Hood" talk for the Dutch PHP Conference there was a question on IRC about extension_loaded() being faster than function_exists(), which might be strange as both of them are simple hash lookups and a hash lookup is said to be O(1). I started to write some slides for it but figured out that I won't have the time to go through it during that presentation, so I'm doing this now, here:

You all should know PHP arrays. They allow you to create a list of elements where every element may be identified by a key. This key may either be an integer or a string. Now we need a way to store this association in an effective way in memory. An efficient way to store a collection of data in memory is a "real" array, an array of elements indexed by a sequence of numbers, starting at 0, up to n. As memory is essentially a sequence of numbered storage blocks (this obviously is simplified, there might be segments and offsets, there might be virtual pages, etc.) we can efficiently access an element by knowing the address of the first element, the size of the elements (we assume all have the same size, which is the case here), and the index: The address of the n-th element is start_address + (n * size_of_element). That's basically what C arrays are.

Now we're not dealing with C and C arrays but also want to use string offsets so we need a function to calculate a numeric value, a hash, for each key. An hash function you most likely know is MD5, MD5 is creating a 128 bit numeric value which is often represented using 32 hexadecimal characters. For our purpose 128 bit is a bit much and MD5 is too slow so the PHP developers have chosen the "DJBX33A (Daniel J. Bernstein, Times 33 with Addition)" algorithm. This hash function is relatively fast and gives us an integer of the value, the trouble with this algorithm is that it is more likely to produce conflicts, that means to string values which create the same numeric value.

Now back to our C array: For being able to safely read any element, to see whether it is used, we need to pre-allocate the entire array with space for all possible elements. Given our hash function returning a system dependent (32 bit or 64 bit) integer this is quite a lot (size of an element multiplied wit the max int value), so PHP does another trick: It throws some digits away. For this a table mask is calculated. The a table mask is a number which is a power of two and then one subtracted and ideally higher than the number elements in the hash table. If one looks at this in binary representation this is a number where all bits are set. Doing a binary AND operation of our hash value and the table this gives us a number which is smaller than our table mask. Let's look at an example: The hash value of "foobar" equals, in decimal digits, to 3127933054. We assume a table mask of 2047 (2นน-1).

  3127933054    10111010011100000111100001111110
&       2047    00000000000000000000011111111111
=        126    00000000000000000000000001111110

Wonderful - we have an array Index, 126, for this string and can set the value in memory!

If life were that easy. Remember: We used a hashing function which is by far not collision free and then dropped almost two thirds of the binary digits by using the table mask. This makes it rather likely that some collisions appear. These collisions are handled by storing values with the same hash in a linked list. So for accessing the an element one has to

  1. Calculate the hash
  2. Apply the table mask
  3. locate the memory offset
  4. check whether this is the element we need, traverse the linked list.

Now the question initially asked was why extension_loaded() might be faster than function_exists() and we can tell: For many random reasons and since you have chosen a value which probably conflicts in the first, not in the second. So now the question is how often such collisions happen for this I did a quick analysis: Given the function table of a random PHP build of mine with 1106 functions listed I have 634 unique hash values and 210 hash values calculated from different functions. Worst is the value of 471 which represents 6 functions.

Full results are online but please mind: These results are very specific to my environment. Also mind that the code actually works on a copy of my function table so the table mask might be different, which changes results. Also note that the given PHP code won't work for you as I added special functions exporting the table mask and hash function to mz version of PHP. And then yet another note: doing performance optimizations on this level is the by far worst place as to many unknown factors go in. And you don't have any measurable performance win. Mind readability. Clear code is worth way more than the 2 CPU cycles you probably gain! But you may have learned how hash tables work.