Phalcon\Di Optimizations

[ghost]

• use return_value wherever possible instead of temporary variables;
• use phalcon_fetch_property() instead of phalcon_read_property() (return value then does not have to be observed)
• use phalcon_array_isset_fetch() to combine isse()/fetch() operations and avoid extra search pass
• validate arguments and throw exceptions before the memory frame is created
• do not convert NULLs to zvals unless absolutely necessary

[ronaldcastillo]

Sorry to bring this here, but, do you happen to have a benchmark with the optimizations you have made lately? I'd love to see the difference in performance.

[ghost]

I run the official test suite under callgrind before and after optimization and compare the results.

Something like this:

ZEND_MM_ALLOC=0 valgrind --tool=callgrind $(phpenv which php) unit-tests/manual-unit.php

or, if I need to test only specific class:

ZEND_MM_ALLOC=0 valgrind --tool=callgrind $(phpenv which php) unit-tests/manual-unit.php TestFile.php ClassName

Usually I take the most expensive function and try to reduce its cost (the lower the cost is, the less time is spent in the function — you can see the figures in my pull requests).

Some optimizations are somewhat evident — for example, the code will run faster without allocating/freeing a memory frame (because we save class to emalloc()/efree() and memory allocation is always costly); same when the code uses pure (constant) references instead of copies (see the recent commits that introduced phalco_read_(n)property(_this) and phalcon_array_isset_fetch() functions) as this saves at least one memory allocation and copy; every such a reference saves at least 32 bytes (sizeof(zval*) on a 64-bit platform) etc.

If I look at the code, I can usually tell how much the optimization allowed one to save memory but I do not record such figures.

[ghost]

For example, for this commit:

• chunk 1: memory frame (48 bytes), 1 zval (32 bytes), 1 emalloc(), 1 efree() => 80 bytes
• chunk 2: 1 zval* (8 bytes), 1 zval (32 bytes) => 40 bytes
• setShared(): same as above (40 bytes)
• attempt(): 1 zval_, 2 zval (72 bytes) + 1 memory frame (48 bytes) and 40 bytes (zval_ + zval) if the service exists + 1 array search pass if the service exists
• getRaw(): 2 zval (64 bytes) plus 1 array search pass if the service exists
• getService(): 2 zval (64 bytes); if the service exists: one memory frame (48 bytes), one array search pass
• get(): 2 zval (64 bytes), 1 array search in most cases
• has(): 1 zval*, 2 zval, 1 memory frame (120 bytes)

etc

[phalcon]

Thanks!