Gainz for dcache_purge_all_with_buffer().

The instruction scheduling for the hot inner loop portion of
dcache_purge_all_with_buffer() was suboptimal. I was just able to shave
off over a second of total runtime from the SH4ZAM unit test scenario by
simply reordering instructions for better superscalar dispatch and by
spacing the movca.l and the ocbi further away from one-another.

Author: gyrovorbis

kernel/arch/dreamcast/kernel/cache.s

@@ -299,10 +299,10 @@ _dcache_purge_all_with_buffer:
 .dpurge_all_buffer_loop:
     ! Allocate and then invalidate the dcache line
     movca.l  r0, @r4
+    cmp/hi   r4, r5
     ocbi     @r4
+    bt.s    .dpurge_all_buffer_loop
     add      #32, r4        ! Move on to next cache block
-    cmp/hi   r4, r5
-    bt       .dpurge_all_buffer_loop
 
     rts
     nop