Flickr: PHP4 vs PHP5 drop in CPU comparison

We wrote a Compiler that does 95% of the work for us. Took a few weeks to write and it's made our 4->5 upgrade jobs a very profitable part of the business.

To the best of my knowledge, there was no changes to function parameters order. It's nearly all object-model stuff.

Some might find this interesting.. The 1.0 of the Compiler introduced bugs into the compiled (5.0 version) of the codebase in cases where the orginal dev relied on the Assign-By-Copy behavior of v4.

In most cases by-ref behavior is what was wanted, and many probably didn't notice the distinction. But there is a situation where the dev may have taken advantage of the fact that eh was getting a copy, eg:

function x($o) { $o->foo = 'bar'; $o->save(); unset($o); }

$y = new o(); $y->foo = 'baz'; x($y); $y->save();

This code in PHP4 works as the developer intended. In PHP5 it throws an error on the $y->save line because the object was passed by reference to x() and it was garbage collected inside x.

It's rare, but it exists, and it was a little tough to fix because we're running a single-pass Compiler.

The nice thing is that it's also able to introduce optimizations in the code, as well as use a Static Analysis approach to detecting probably bugs and security holes (for example, unfiltered use of $_REQUEST values, or assigning to $this) and it fixes them as well.

The end result is an app that always runs faster, sometimes significantly so.

IMO, the real reason 4->5 was a slow process is because it's biggest selling point was a "real-ish" object model and most PHP developers are Morts and just don't need to care about that.

The problem is finding all the places that the language has changed and all the places you relied on the old behavior. The documented stuff was easy to find, it was the undocumented (or unintended?) changes that were hard. For example, strtotime($null) returns now in php4 but returns null in php5.

QA and automated testing will only get you so far in this process. Nobody had thought to test how we display the "Date Taken" on photos from before December 14th 1901 until a member pointed out they were displaying wrong (caused by another change to strtotime() behavior).

At some point you have to upgrade a subset of your servers, watch the error logs, fix the bugs as they come up and be prepared to rollback quickly. It wasn't a fun experience...