本人在A72上做一个数组累加：
s += a[i]，测试的时候数组a是一个10KB的数组，循环100000次，保证整个过程的数据都是l1cache上，排除cachemiss的可能。
发现效率怎么都达不到最优，按理说这个数组每4个值需要一个LD1指令和一个ADD指令，按照F0/F1和L/S四个单元并行原则，理论上是1个cycle就可以搞定了，但实际上需要1.7cycle。
具体实现如下：
"2:\n"
"subs x1 , x1, #1\n"
"ld1 {v8.4s},[x0],#16\n"
"ld1 {v9.4s},[x0],#16\n"
"ld1 {v10.4s},[x0],#16\n"
"ld1 {v11.4s},[x0],#16\n"
"ld1 {v12.4s},[x0],#16\n"
"ld1 {v13.4s},[x0],#16\n"
"add v0.4s,v0.4s,v8.4s\n"
"ld1 {v14.4s},[x0],#16\n"
"add v1.4s,v1.4s,v9.4s\n"
"ld1 {v15.4s},[x0],#16\n"
"add v2.4s,v2.4s,v10.4s\n"
"ld1 {v16.4s},[x0],#16\n"
"add v3.4s,v3.4s,v11.4s\n"
"ld1 {v17.4s},[x0],#16\n"
"add v4.4s,v4.4s,v12.4s\n"
"ld1 {v18.4s},[x0],#16\n"
"add v5.4s,v5.4s,v13.4s\n"
"ld1 {v19.4s},[x0],#16\n"
"add v6.4s,v6.4s,v14.4s\n"
"ld1 {v20.4s},[x0],#16\n"
"add v7.4s,v7.4s,v15.4s\n"
"ld1 {v21.4s},[x0],#16\n"
"add v24.4s,v24.4s,v16.4s\n"
"ld1 {v22.4s},[x0],#16\n"
"add v25.4s,v25.4s,v17.4s\n"
"ld1 {v23.4s},[x0],#16\n"
"add v26.4s,v26.4s,v18.4s\n"
"add v27.4s,v27.4s,v19.4s\n"
"add v28.4s,v28.4s,v20.4s\n"
"add v29.4s,v29.4s,v21.4s\n"
"add v30.4s,v30.4s,v22.4s\n"
"add v31.4s,v31.4s,v23.4s\n"
"bne 2b\n"
第二种：
"2:\n"
"subs x1 , x1, #1\n"
"ld1 { v8.4s},[x0],#16\n"
"ld1 { v9.4s},[x0],#16\n"
"ld1 {v10.4s},[x0],#16\n"
"ld1 {v11.4s},[x0],#16\n"
"ld1 {v12.4s},[x0],#16\n"
"ld1 {v13.4s},[x0],#16\n"
"ld1 {v14.4s},[x0],#16\n"
"ld1 {v15.4s},[x0],#16\n"
"add v0.4s,v0.4s, v8.4s\n"
"add v1.4s,v1.4s, v9.4s\n"
"add v2.4s,v2.4s,v10.4s\n"
"add v3.4s,v3.4s,v11.4s\n"
"add v4.4s,v4.4s,v12.4s\n"
"add v5.4s,v5.4s,v13.4s\n"
"add v6.4s,v6.4s,v14.4s\n"
"add v7.4s,v7.4s,v15.4s\n"
"bne 2b\n"这两种的效率都是1.7cycle。
求解无法达到理论值的原因
