Valgrind must supervise the process from the start; it is not possible to attach it to already running process (or, equivalently, to ignore the process until some point in execution, then start emulating/checking).
The reverse is not true -- you can "detach" valgrind after some number of instructions; but I am guessing that's not what you want.
Please note that:
- the "output is very long" is a poor excuse -- Valgrind errors are usually
true positives (unless you are using optimized code, in which case: don't do that), and should really be addressed, and
- you can concentrate on the more serious problems (heap corruption) before addressing the use of uninitialized values, by using
--undef-value-errors=no