This patch was the cause of the compile-time performance regression in
#3796.  My guess is that it is due to the use of unsafePerformIO which
traverses the stack up to the first update frame, and perhaps we have
a deep stack when reading the dictionary from a .hi file.  In any
case, since we're not relying on thread safety for FastStrings, I
think the safest thing to do is back this out until we can investigate
further.