
The widest block in a cold flame graph is the autoloader
The same endpoint profiled twice - 65.8 ms and 8.4 ms. What changed in the picture, and how to tell one slow call from ten thousand fast ones.
Tags
10 posts

The same endpoint profiled twice - 65.8 ms and 8.4 ms. What changed in the picture, and how to tell one slow call from ten thousand fast ones.

One workload, two profiling mechanisms. Instrumentation cost 13% and missed the function that was 55% of the request; sampling cost nothing measurable and found it.

The same function ran at 21.4 ns with no OPcache, 21.6 with OPcache and the optimizer off, and 14.8 with it on. What the passes removed, dumped one at a time.

The same loop ran 23% faster with int parameters, and only with OPcache on. Operand classes pick a handler when the file compiles; proven types pick a different one.

The same .= statement took 0.269 ms and 218.836 ms in two loops that differ by one assignment. What a zend_string costs, where literals live, and why implode is not the fix.

Plain, backed and virtual property reads compile to one identical FETCH_OBJ_R. What separates them is a function call the opcode does not show, measured at nine times.

One hundred thousand four-field records cost 143.54 bytes each as objects and 397.14 as associative arrays. Where the crossover is, and what one undeclared property costs.

Walking an array with foreach ($rows as &$row) wraps every element in a zend_reference. On a million integers that is 32 MB the loop never gives back.

OPcache removes compilation, not execution. The JIT reaches arithmetic-heavy code and little else. Where the remaining time lives.

Microbenchmarks measure exactly what you tell them to, which is rarely what you meant. Method, noise, and the numbers worth reporting.