How to implement the trace cache?
After the machine identifies the start of a hot path is on the fly, it will
predict the target and direction of a branch according to the hot
path. Load the right instructions in cache. The replacing rule replaces
unexecuted instructions when there is a conflict. Required some hw suport.
That is my guess.

-- Profiling papers
Profiling people faces the issues as: exact run-time info collection and low
overhead.  There is a tradoff between them. The minimum coding is
prefered at the same time.

Shaojie