/*
 * demo.c : testcase to show that icc 12.1.0 may generate incorrect code
 *          with -O2 -openmp
 *
 * Platform : Linux x86-64 (Debian 6.0)
 * Compiler : icc version 12.1.0 (l_ccompxe_intel64_2011.7.256)
 * Flags    : -O2 -openmp
 *
 * The bug shows only if compiled with both flags -O2 and -openmp together
 *
 * The result of the application built with both flags is always wrong and
 * additionally varying with each run.
 *
 * Compiling with '-O2' or '-O0 -openmp' flags results in correct code.
 *
 */

#include <stdio.h>
#include <stdint.h>

#ifdef _OPENMP
#include <omp.h>
#endif


typedef union {
  uint64_t  number;
  struct {
    uint32_t one     :13;
    uint32_t         : 3;
    uint32_t two     :13;
    uint32_t         : 3;
    uint32_t three   :13;
    uint32_t         : 3;
    uint32_t four    :13;
    uint32_t         : 3;
  } fields;
} MyUnion;


static __inline uint32_t foo(MyUnion h)
{
	return h.fields.four | h.fields.three | h.fields.two | h.fields.one;
}

static __inline uint32_t bar(MyUnion h0, MyUnion h1, MyUnion h2)
{
	MyUnion h;

	h.number = h0.number | h1.number;
	h.number = h.number | h2.number;

	// no bug if bitwise-or is written as
	// h.number = h0.number | h1.number | h2.number;

	// no bug either, if foo() is 'inlined' here 'manually'
	return foo(h);
}


unsigned long calc(unsigned long n_iterations)
{
	MyUnion t1, t2, t3;
	unsigned long sum = 0;
	unsigned long k;

	t1.number = 0;
	t1.fields.one = 1;

	t2.number = 0;
	t2.fields.two = 2;

	t3.number = 0;
	t3.fields.three= 4;

#ifdef _OPENMP
	#pragma omp parallel for schedule(static) reduction(+:sum)
#endif
	for (k=n_iterations; k>0; --k) {
		sum += bar(t1, t2, t3);
	}

	return sum;
}


int main(int argc, char **argv)
{
	unsigned long sum, n_iterations = 100000;

	sum = calc(n_iterations);

	printf("sum = %lu (should be %lu)\n", sum, 7*n_iterations);

	return 0;
}