I recently attended a code optimization workshop for the IBM PPC compiler for Mac OSX. The compiler designer stressed that hand optimized code (i.e. unrolled loops, register variables, stupid pointer tricks etc.) only confused the optimizing compiler and would usually result in slower code overall when -O4 and higher optimizations are enabled. He provided a number of examples why this was the case and convinced us that modern compilers are much better at optimizing than humans.
He also stressed the need to profile code and look for things the compiler cannot optimize. Examples are using double floats where single precision will suffice and in the case of the PPC unecessary conversions between floats and integers.
Slashdot Mirror
I recently attended a code optimization workshop for the IBM PPC compiler for Mac OSX. The compiler designer stressed that hand optimized code (i.e. unrolled loops, register variables, stupid pointer tricks etc.) only confused the optimizing compiler and would usually result in slower code overall when -O4 and higher optimizations are enabled. He provided a number of examples why this was the case and convinced us that modern compilers are much better at optimizing than humans. He also stressed the need to profile code and look for things the compiler cannot optimize. Examples are using double floats where single precision will suffice and in the case of the PPC unecessary conversions between floats and integers.