On 07/31/2010 01:45 PM, Richard Earnshaw wrote:
> This is part of the reload_{in,out}hi sequences.  These have always
> blown my mind away and I've never fully understood when and why these
> are really needed.

Essentially, whenever reload may have to move something from A to B,
but can't do it without a scratch reg.

> In thumb1 you potentially need some sort of recovery sequence if the
> compiler needs to spill a register held in r8-r14 (since these registers
> can't be stored directly to memory), but why it really needs a secondary
> reload escapes me -- the compiler should just copy the register to a
> lo-reg and then store that (or vice-versa for a load).

And I think that happens in all normal cases.  For example, arithmetic
operations use "l" constraints which means reload regs are LO_REGS and
no scratch reg is required to move anything into or out of it.

> I suggest we get rid of the pattern entirely, and the code chain that
> can end up calling it (ie thumb_reload_out_hi.

As below?  I've had this in my last test run (before checking in the
other patches) and that was successful.  It has a slight change in that
it uses TARGET_32BIT instead of TARGET_ARM now.  It may be worthwhile as
an experiment trying to delete the arm branch as well to see if that
causes problems.


Bernd