Use zend_string for filenames and variable names

Most of Xdebug was written before PHP 7, and when updating Xdebug to work with both PHP 5 and PHP 7 certain PHP 5 APIs, wrapped in macros to make it work with PHP 7, remained in use by Xdebug. When dropping PHP 5 support, I did not update these APIs to use the PHP 7 variant. One of these APIs is that in PHP 5, strings were often passed around as a char string, size_t len pair. PHP 7 introduced a new data type zend_string str which encapsulates both elements of the pair, and also introduced reference counting for zend_string s, which reduces the need for duplication of memory. This gave about a ~2% performance boost to Wall Time (the time passage as indicated by a clock), and about a 5% reduction in CPU instructions.

Switch xdebug_sprintf to a new xdebug_str_add_fmt

Xdebug often needs to convert data into a single string, for either writing to file or network, storing function names with arguments in memory, or display purposes. It uses an API xdebug_sprintf which allocates new memory and formats the string according to format specifiers. Often it needs to add this to a xdebug_str buffer that makes up a longer piece of text. After it has added it to that buffer, it frees the allocated memory again.

By creating a new API that can add a formatted string to an xdebug_str buffer, I managed to reduce the amount of memory allocations and frees dramatically. This gave a 10-15% Wall Time performance boost, and a 10-25% reduction in CPU instructions.

Use a vector instead of a list for the stack

Xdebug's internals contain a stack of all the functions that are called. Every time a new function is called, a new stack level is allocated, and when the function ends this stack element is freed up. In applications which call a lot of small functions, such as Composer, this repeated allocation and freeing of memory, as well as the maintenance of the links in the list, is expensive.

I switched this list approach to use a Vector, which is a continuous amount of allocated memory, without the need for an allocation and free for every function that is being called. This gave a 2-6% Wall Time performance boost, most notably with Composer, and a 3-10% reduction in CPU instructions.

Buffering for the profiler

Most of the work on Xdebug's profiler were initiated by Michael Voříšek. The main performance boost was to use the new xdebug_str_add_fmt API, but also changing the initial size of xdebug_str buffers from 1 to 32, as most of the strings are of that length. This will use a bit more memory.

The other big part of the performance improvements for profiling is to no longer use fprintf and fflush a lot. Instead Xdebug now uses an xdebug_str, the new xdebug_str_add_fmt API, and also a new API function to format a number to an xdebug_str without having to use the xdebug_sprintf API call.

These, as well as the other mentioned changes reduced Wall Time, as well as CPU instructions, for profiling by about 40%.

I will continue to work on improving performance in September, but here are some graphs:

You might notice that the Xdebug 2.9.6 data is off the chart, here are they with Xdebug 2.9.6 included ;-)