[LoopVectorize] LLVM fails to vectorise loops with multi-exit

[vfdff]

• Since the RFC, the process is slow,
  so use a simple case to track the result, https://gcc.godbolt.org/z/r6KGKMYMd

  the case from above RFC, which is expect a base case, both gcc and llvm don't support it now.

int foo(float *a, int n, int M){
  int i;
  for (i=0; i<n; i++){
    if (a[i] == 0) break;
  }
  return i;
}

• kernel body of loop: clang -march=armv8.2-a+sve -O3

.LBB0_2:                                // =>This Inner Loop Header: Depth=1
        ldr     s0, [x8, x0, lsl #2]
        fcmp    s0, #0.0
        b.eq    .LBB0_5
        add     x0, x0, #1
        cmp     x9, x0
        b.ne    .LBB0_2

[vfdff]

• surprised to find the demo case with multi-exit is supported, https://gcc.godbolt.org/z/6dnfadM5E

this case is the 1st scenarios mentioned in above RFC

int foo(float *a, int n, int M){
  int i;
  for (i=0; i<n; i++){
    if (i > M) break;
    a[i] = i;
  }
  return i;
}

[vfdff]

• the 2nd scenarios mentioned in above RFC, https://gcc.godbolt.org/z/q1fb9bxsM

  gcc will also fail to sve when we use if (s[i] == 0) to replace if (sa[i] == 0)

static float sa[100];
int foo(float *a, int n, int M){
  int i;
  for (i=0; i<n; i++){
    if (sa[i] == 0) return 0;
  }
  return 1;
}

llvm just do a scalar loop:

.LBB0_2:                                // =>This Inner Loop Header: Depth=1
        ldr     s0, [x9]
        fcmp    s0, #0.0
        b.eq    .LBB0_5
        subs    x8, x8, #1
        add     x9, x9, #4
        b.ne    .LBB0_2

gcc generates sve: gcc -march=armv8.2-a+sve -O3

.L4:
        add     x0, x0, x3
        whilelo p7.s, w0, w1
        b.none  .L7
.L5:
        ld1w    z29.s, p7/z, [x2, x0, lsl 2]
        mov     z31.d, z30.d
        fcmeq   p7.s, p7/z, z29.s, #0.0
        incw    z30.s
        ptest   p15, p7.b
        b.none  .L4

[pinskia]

For the first testcase, that will require what is called first fault support (which I know is aimed to get into GCC 15; I don't know about LLVM though) and as far as I know can only be vectorized using SVE (and not the normal advanced SIMD support in ARMv8-a).

[vfdff]

Thanks @pinskia for your reminder, I find llvm don't really support the multi-exit case.
For the 1st case, it use a csel to update the loop trip count, so it seems a loop version

csel w9, w2, w9, lo // w9 = min(M, n-1)

int foo(float *a, int n, int M){
  int i;
  if (M < n-1) {
    for (i=0; i<M; i++){
      a[i] = i;
    }
   return i;
  }
  for (i=0; i<n; i++){
    a[i] = i;
  }
  return i;
}