Bug fixes for ld/st elimination.

Bug fixes for ld/st elimination.

1) Only optimize each block once. Strictly speaking this is not a
correctness issue however it triggers the subsequent problem.

2) Ignore dead instructions.

   1: ldr     r2, [r5, #8]
   2: ldr     r3, [r5, #8](nop)
   3: str     r3, [r5, #4]
   4: movs    r3, r2

When using instruction 1 to initiate redundant ld/st eliminations, if
instruction 2 (which is already dead) is not ignored, it will be turned
into a "mov r3, r2" and that will clobber r3 used by the str.

Change-Id: I8eecd13eeb30e4a67ecf1f8fbad925b1e6e91fc8