Improve the coalescing of instruction cache flushes on R8000 in order

Improve the coalescing of instruction cache flushes on R8000 in order
not to perform redundant work. There is no direct way to flush the
cache; rather, the kernel evicts unwanted bits from the cache by
executing a block of instructions, which is expensive.

With this diff, make build time decreases about 0.4%.

Diff from miod@