On Wed, Apr 24, 2002 at 01:42:07PM +0100, Richard Earnshaw wrote:
> > On Mon, Apr 22, 2002 at 04:48:22PM +0100, Richard Earnshaw wrote:
> > > Dump of assembler code for function FontCouldBeTerminal:
> > > 0x10f7c>: stmdb   sp!, {r4, r5, r6, r7, lr}
> > > 0x10f80>: sub     sp, sp, #4      ; 0x4
> > ...
> > > 0x11028>: mov     r0, #0  ; 0x0
> >             add     sp, sp, #4
> > > 0x1102c>: ldmia   sp!, {r4, r5, r6, r7, pc}
> > 
> > Isn't is likely to be faster to align the stack by ensuring
> > that an even number of registers are saved/restored, rather
> > than having 2 extra instructions to modify the sp?
> > 
> > The single stack location is more likely to be in the
> > D-cache than the two instructions are in the I-cache.
> 
> Probably, but getting correct code at this point is more important than 
> the most efficient.  Doing this by locating a register that isn't pushed, 
> but is safe to push (ie doesn't interfere with the entry or exit sequence) 
> is far more complicated, and thus, more likely to be incorrect.

r3?

	David

-- 
David Laight: david@l8s.co.uk