Unaligned floating point access generation for powerpc

[Olivier Hainque]

David Edelsohn wrote:
> The rs6000 MD already lists SFmode and DFmode as slow unaligned access.
 
 Right.

> Maybe there are more places in the common parts of the compiler that need to
> use that cost metric.

 Maybe. rs6000_emit_move is already partly dealing with such situations,
 for MEM to MEM DI moves.

 FWIW, GCC also emits the unaligned fp load for the alternate definition
 of the in-struct data type below, not violating the ABI alignment rules for
 long long anymore:

    typedef struct { char x [8]; } data_t;

 I'm actually first trying to determine if this is legitimate or if this
 should be considered a bug somehow.

 Thanks for your feedback,

 Olivier