Test case fail only with optimized build

[BerengerBerthoul]

The test case update_ids_in_range_after_permutation__non_involutive fails with specific compilation options :

CXXFLAGS="-O3 -g -DNDEBUG -ftree-vectorize -ftree-loop-vectorize -fvect-cost-model=unlimited -ftree-loop-optimize -march=native -fopenmp-simd -mtune=intel -mprefer-vector-width=512"
CMAKE_BUILD_TYPE: Release

With CPU type : Sapphire rapids (Intel(R) Xeon(R) Platinum 8468)

Note that it does not fail on AMD Zen3 with similar compiler options.

Simplifying the test to narrow it down, we get :

TEST_CASE("narrowing the problem") {
  int* permutation = new int[3];
  permutation[0] = 2; 
  permutation[1] = 0; 
  permutation[2] = 1; 

  std::vector<int> ids = {-1, 0,2,1};
  //std::vector<int> ids(4);
  //ids[0] = -1;
  //ids[1] =  0;
  //ids[2] =  2;
  //ids[3] =  1;

  for (int i=0; i<4; ++i) {
    if ( ids[i] >= 0 ) {
      ids[i] = permutation[ids[i]];
    }
  }

  CHECK( ids[0] == -1 );
  CHECK( ids[1] ==  2 ); // Fail: returns 0
  CHECK( ids[2] ==  1 );
  CHECK( ids[3] ==  0 );
}

If the vector is initialized with an initializer list, the test fails. However, with an element-based initialization (such as with the commented lines), it works...