broken case

t5.cpp

#include <amp.h>
#include <vector>
#include <stdlib.h>
using namespace concurrency;
using namespace std;
const static int DOMAIN_SIZE = 1024;
const static int BLOCK_SIZE = 16;
typedef int TY;

template<typename type, int rank>
struct s1
{
    s1(Concurrency::array<type, rank> &a) __GPU : av_a(a) {}
    s1(Concurrency::extent<rank> e, type *pa) __GPU : av_a(e, pa){}
    ~s1() __GPU {}

    int placeholder1;
    array_view<type, rank> av_a;
    float placeholder2;
};

void bar0(index<1> idx, s1<TY, 1> *p) restrict(amp) {
  int * pa = &p->av_a[idx];
  *pa = 1; 
}

void foo2() {
    Concurrency::extent<1> g(DOMAIN_SIZE);
    vector<int> a(DOMAIN_SIZE);
    Concurrency::array<int, 1> a_a(g);
    Concurrency::array<int, 1> a_b(g);
    int b1 = 0;
 
    parallel_for_each(a_a.get_extent(), [&, b1] (index<1> idx) restrict(amp) {
      s1<int,1> o(a_a);
      s1<int,1> o_f(a_b);

      s1<int, 1> *p = !b1 ? &o : &(o_f); // broken
      //s1<int, 1> *p = &o;  //okay
      bar0(idx, p); 
    });

    a = a_a;

    for (int i=0; i<DOMAIN_SIZE; ++i) {
      if (a[i] != 1) {
        fprintf(stderr, "wrong answer idx = %d bad value= %d\n", i, a[i]);
        exit(-1);
      }
    }
}

int main() {
    foo2();
}

pathamp -device=tahiti -CG2:dump-opt-ir -Wb,-tt61:0x1 ./t5.cpp