bug018

gistfile1.txt

#include <amp.h>
#include <stdio.h>
#include <stdlib.h>
#include <vector>
using namespace Concurrency;

template <typename _type>
int test_debug() __GPU {
    const int rank = _type::rank;
    int data1[] = {1, 2, 3, 4, 5, 6, 7, 8, 9, 10};

    int correct_size = 1;
    extent<rank> e1(data1);
    const _type g1(e1);
//bad
    for (int i = 0; i < 2; i++)
        correct_size = data1[i];
//okay
//    correct_size = data1[1]; 
    return correct_size; 
}

int main()
{
    int gpu_result;
	concurrency::array_view<int, 1> gpu_resultv(1, &gpu_result); 
    gpu_resultv.discard_data(); 
	concurrency::parallel_for_each(gpu_resultv.get_extent() 
		, [=](concurrency::index<1> idx) restrict(amp) { 
		gpu_resultv[idx] =  []() restrict(amp,cpu)->int{
							return test_debug<extent<4>>();
							}();
	}); 
	gpu_resultv.synchronize(); 

    printf("%d", gpu_result);
    return 0;
}

I spot the same issue using both pathscale's branch and llvm svn trunk. I am targeting to southern island processors.

the problem instruction is buffer_load_dword in BB#2. you can reproduce it using 'llc -march=r600 -mcpu=tahiti --filetype=asm bug018.ll -o bug018.s'. the sbase s[8:11] is wrong because s[8:9] is assigned as 0. it should point to the scratch. the values of data1 are on it.