SPARC: optimize window overflow trap entry

By using a optimized trap entry we can move instructions from
the window overflow function into the trap entry vector. By
using the saved locals instead of g1 we don't need to save
that register temporarily. Also spead out non store instructions
inbetween stores to use the write buffer better.

c/src/lib/libcpu/sparc/reg_win/window.S

@@ -60,6 +60,7 @@ SYM(window_overflow_trap_handler):
         srl  %l3, 1, %l3                 ! l3  = WIM >> 1
         wr   %l3, %l2, %wim              ! WIM = (WIM >> 1) ^
                                          !       (WIM << (Number Windows - 1))
+                                         ! 3 instruction delay not needed here
         std  %i0, [%sp + 0x20]           ! save input register set
         std  %i2, [%sp + 0x28]
         std  %i4, [%sp + 0x30]