[vectorization] gcc generate better code for a loop body with scalar and array expansion

[zhongsir1]

test: https://godbolt.org/z/eMedWs4Ef

void s257(void) {
  for (int i = 1; i < LEN_2D; i++) {
    for (int j = 0; j < LEN_2D; j++) {
      aa[j][i] = a[i] + bb[j][i];
    }
  }
}

• gcc: generate a sve loop

  .L3:
  ld1d z31.d, p7/z, [x4, x0, lsl 3]
  ld1d z30.d, p7/z, [x3, x0, lsl 3]
  fadd z31.d, p7/m, z31.d, z30.d
  st1d z31.d, p7, [x1, x0, lsl 3]
  add x0, x0, x5
  whilelo p7.d, w0, w2
  b.any .L3

• llvm:

  .LBB0_2: // Parent Loop BB0_1 Depth=1
  // => This Inner Loop Header: Depth=2
  add x13, x8, x12
  ldr d1, [x13, #8]
  ldr d2, [x13, #1032]
  add x13, x9, x12
  add x12, x12, #2048
  fadd d1, d0, d1
  cmp x12, #32, lsl #12 // =131072
  fadd d2, d0, d2
  str d1, [x13, #8]
  str d2, [x13, #1032]
  b.ne .LBB0_2

[vfdff]

This issue seems related to the cost model of gather/scatter, https://godbolt.org/z/c3EaP7hYK