By default the following code, for an array X,

X = X.astype('float64')

will trigger a memory copy, including when X is of dtype float64, which can have a significant impact on runtime performance (cf. e.g. #11966 (comment))

In #11966 @massich made a list of places where this may be relevant.

Overall when we potentially cast to an array to the same dtype as it currently is,

• for dense arrays we should probably always use X.atype(dtype, copy=False), unless a copy is explicitly needed.
• for sparse arrays, copy argument was added to astype only a year ago unfortunately, so we may have to do something along the lines of,

  if sp_version >= (1, 1) or not sparse.issparse(X):
     copy_args = {'copy': False}
  else:
     copy_args = {}
  X.astype(dtype, *copy_args)

  or possibly to factorize that in utils/fixes.py

#11966 addressed a part of this.