Currently, the profiling harness generated by dmd -profile uses uint (32-bit) for counting the number of times a function gets called. For profiling the steady-state of long-running, CPU-intensive code, 32-bit counters are easily overflowed, causing the profile output to be useless because some counters may have wrapped around, thus distorting the output (counts, ordering of functions that consumed the most time, etc.).
Furthermore, due to the reliance on the rdtsc instruction which counts CPU cycles, given the high clock frequency of modern CPU's, even with a 64-bit counter a moderately long-running process will overflow this counter, causing time measurements to wrap around and again producing garbled output.
This makes dmd -profile useless in profiling long-running CPU-intensive processes, as there is no way to measure the steady state of the process before the output gets garbled by wrapped-around counters.