[LoopVectorize] Fails on OpenMP SIMD Prefix Sum

[orpheezt]

On clang 19.1.7

void prefix_scan_omp_simd(uint64_t n, const float *const __restrict src,
                          float *const __restrict dst, const double init) {
  auto acc = init;
#pragma omp simd reduction(inscan, + : acc)
  for (uint64_t i = 0; i < n; ++i) {
    acc += src[i];
#pragma omp scan inclusive(acc)
    dst[i] = acc;
  }
}

The following code is emitting the warning

warning: loop not vectorized: the optimizer was unable to perform the requested transformation; the transformation might be disabled or specified as part of an unsupported transformation ordering [-Wpass-failed=transform-warning]
   15 | #pragma omp simd reduction(inscan, + : acc)
      | ^
1 warning generated.

While on gcc it suceeds.

[llvmbot]

@llvm/issue-subscribers-openmp

Author: Isaac David (orpheezt)

On `clang 19.1.7`

void prefix_scan_omp_simd(uint64_t n, const float *const __restrict src,
                          float *const __restrict dst, const double init) {
  auto acc = init;
#pragma omp simd reduction(inscan, + : acc)
  for (uint64_t i = 0; i < n; ++i) {
    acc += src[i];
#pragma omp scan inclusive(acc)
    dst[i] = acc;
  }
}

The following code is emitting the warning

warning: loop not vectorized: the optimizer was unable to perform the requested transformation; the transformation might be disabled or specified as part of an unsupported transformation ordering [-Wpass-failed=transform-warning]
   15 | #pragma omp simd reduction(inscan, + : acc)
      | ^
1 warning generated.

While on gcc it suceeds.

[fhahn]

What's the behavior of #pragma omp scan inclusive(acc), how is it encoded in LLVM? cc @Meinersbur @alexey-bataev

[Meinersbur]

bd1c03d

I don't know what happens with a simd inscan though. One of the lowered implementation loops might just not be vectorizable.