[PATCH] Add new target-hook truncated_to_mode

[Joern RENNECKE]

Richard Sandiford wrote:

>  
>
>And you replied with what I thought to be examples of where that might
>not hold, and where defining the hook above would therefore be wrong.
>Did I misunderstand?
>
>In summary, could the hook be defined for SH too, in the way described above?
>  
>
I really want to fix the endian-issue for floating point registers so 
that vectors and
unions can be put into floating point registers where this makes the 
most sense
for code generation.  That means if a sign extension is a no-op will 
depend on the
register class of the hard register that you are doing the operation on.

You have to perserve the operation anyway, if you spill the register to 
memory.
So what the target could tell the middle-end is that the sign-extension 
operation
for a particular mode pair is expected to be a no-op, but that this is 
subject to register
allocation; and the register classs for which this is a no-op.

If we want to describe what operation a SUBREG is equivalent to, we 
should include
an register class argument to be more specific.  If the argument is 
ALL_REGS, a
definite operation should only be returned if it is true universally, 
otherwise UNKNOWN.
For a true noop conversion we could use the code SUBREG or CONST_INT (as 
in NOP,
which is represented with the pattern (const_int 0)).
The optimizers will probably be most interested in the behaviour for 
GENERAL_REGS
in particular.