Use PIC_PROLOGUE_NOSAVE in long jumps, where we don't need to preserve r12.
While here, swap the order in which r4 and r5 saved on the stack so
that r4 is restored further away from its first use.

Oops^2. Init pointer to the saved registers area.

Oops, correct numeric label in branch.

Use PLT for PIC calls to avoids text relocs in the shared library.
Use constants from <machine/setjmp.h> for better readability.
Provide .size directives.

Switch arch/sh3/gen to the new compat code layout.