[vectorize] gcc choose wider vector length than clang for float type

[vfdff]

• Test: https://gcc.godbolt.org/z/ebG5fdndY

          const float delx = xtmp - x[j].x;
          const float dely = ytmp - x[j].y;
          const float delz = ztmp - x[j].z;
          const float rsq = delx * delx + dely * dely + delz * delz;

• The kernel body of GCC: VL = 4, use loop vectorize

        fsub    v30.4s, v22.4s, v30.4s
        zip1    v28.4s, v28.4s, v29.4s
        zip1    v31.4s, v31.4s, v27.4s
        fmul    v29.4s, v30.4s, v30.4s
        fsub    v28.4s, v21.4s, v28.4s
        fsub    v31.4s, v23.4s, v31.4s
        fmla    v29.4s, v28.4s, v28.4s
        fmla    v29.4s, v31.4s, v31.4s
        fdiv    v27.4s, v16.4s, v29.4s   -- float r2inv = 1.0 / rsq;

• The kernel body of Clang: VL = 2, use SLP

.LBB0_8:                                //   Parent Loop BB0_5 Depth=1
        ldrsw   x16, [x15], #4
        add     x16, x6, x16, lsl #4
        ldr     d16, [x16]
        ldr     s17, [x16, #8]
        fsub    v16.2s, v6.2s, v16.2s
        fsub    s17, s7, s17
        fmul    v18.2s, v16.2s, v16.2s
        faddp   s18, v18.2s
        fmadd   s18, s17, s17, s18
        fcmp    s18, s0
        b.ge    .LBB0_7
        fdiv    s18, s3, s18   -- float r2inv = 1.0 / rsq;

[vfdff]

clang fail to loop vectorize as the cost mode of some high load

[vfdff]

Vectorization is possible but not beneficial because the cost of gather load and phi are high

• the cost of load can be adjust with option -neon-nonconst-stride-overhead
• the cost of phi should also be less with D152523

[llvmbot]

@llvm/issue-subscribers-backend-aarch64

[vfdff]

Now clang also choose VF=4 with option -mllvm -neon-nonconst-stride-overhead=3, https://gcc.godbolt.org/z/39a5zdWa5