[sve] gcc generate better code for a search loop

[zhongsir1]

test: https://godbolt.org/z/qjzhsY8c1

int s331(void)
{
// search loops
  int j = -1;
  for (int i = 0; i < LEN_1D; i++) {
    if (a[i] < (real_t)0.) {
      j = i;
    }
  }
  return j+1;
}

• gcc: generate a sve loop

  .L2:
  punpkhi p7.h, p15.b
  punpklo p5.h, p15.b
  ld1d z31.d, p7/z, [x4, x0, lsl 3]
  ld1d z28.d, p5/z, [x2, x0, lsl 3]
  fcmlt p14.d, p6/z, z31.d, #0.0
  add x0, x0, x3
  fcmlt p7.d, p6/z, z28.d, #0.0
  uzp1 p7.s, p7.s, p14.s
  and p7.b, p7/z, p15.b, p15.b
  whilelo p15.s, w0, w1
  clastb s29, p7, s29, z30.s
  incw z30.s
  b.any .L2

• llvm:

  .LBB0_1: // =>This Inner Loop Header: Depth=1
  ldr d0, [x10, x8, lsl #3]
  fcmp d0, #0.0
  csel w9, w8, w9, mi
  add x8, x8, #1
  cmp x8, x11
  b.ne .LBB0_1

[llvmbot]

@llvm/issue-subscribers-backend-aarch64

[davemgreen]

This may be addressed with https://reviews.llvm.org/D143465 or https://reviews.llvm.org/D150851 or one of the related reviews.