Frames are recorded by the Profiler in a time-linear fashion. While profiling, the profiler builds a list of frame stacks, with the frames having in format:
function_name <null> filename <null> function_line_number
When profiling is complete, this list is turned into a tree structure of Frame objects. This tree contains all the information as gathered by the profiler, suitable for a flame render.
The frames are assembled to a call tree by the profiler session. The time-linearity is retained at this stage.
Before rendering, the call tree is then fed through a sequence of 'processors' to transform the tree for output.
The most interesting is aggregate_repeated_calls, which combines different
instances of function calls into the same frame. This is intuitive as a
summary of where time was spent during execution.
The rest of the processors focus on removing or hiding irrelevant Frames from the output.
Self time nodes exist to record time spent in a node, but not in its children. But normal frame objects can have self_time too. Why? frame.self_time is used to store the self_time of any nodes that were removed during processing.