RFD: hookizing BITS_PER_UNIT in tree optimizers / frontends

[Joern Rennecke]

Quoting "Joseph S. Myers" <joseph@codesourcery.com>:

> If something relates to an interface to a lower-level part of the compiler
> then BITS_PER_UNIT is probably right - but if somethis relates to whether
> a type is a variant of char, or to alignment of a non-bit-field object
> (you can't have smaller than char alignment), or things like that, then
> TYPE_PRECISION (char_type_node) may be better.

Yes, I see examples for both in the C++ front end.
The tree optimizers seem mostly (or entirely?) concerned with the
addressable unit size.

> Note that BITS_PER_UNIT is used in code built for the target (libgcc2.c,
> dfp-bit.h, fixed-bit.h, fp-bit.h, libobjc/encoding.c, ...), and converting
> it to a hook requires eliminating those uses.

Full conversion does.  For the moment I would be content with a  
partial conversion so that not every tree optimizer that currently uses
BITS_PER_UNIT has to include tm.h itself once the bogus tm.h includes
from target.h / function.h / gimple.h are gone.