[thumb2, size] allocate local variables with fewer instructions

[Yao Qi]

GCC PR40657 is about this kind of problem, and was fixed. However, we can find similar problem when build toolchain with hard float.
Compiled test case like this,
$ arm-none-linux-gnueabi-gcc -Os -
fno-common -c -mthumb -mcpu=cortex-a9 -mfpu=neon -Os -fno-common -mfloat-abi=hard -S extral_sp_sub.c,

GCC generates code like this,
00000000 <diffmeasure>:
   0: b570 push {r4, r5, r6, lr}
   2: 2400 movs r4, #0
   4: ed2d 8b04 vpush {d8-d9}
   8: b082 sub sp, #8
...
  b0: b002 add sp, #8
  b2: ecbd 8b04 vpop {d8-d9}
  b6: bd70 pop {r4, r5, r6, pc}

This code sequence can be improved like this,
push {r2, r3, r4, r5, lr}
movs r4, #0
vpush {d8-d9}
...
vpop {d8-d9}
pop {r2, r3, r4, r5, pc}
 so, sub and add instruction to sp can be removed.