Bug 1372959

Description of problem:
In an attempt to optimize a recursive search algorithm using cilkplus and C++ templates to avoid cilk overhead for small sub-trees, the compiler gave an internal error.

Version-Release number of selected component (if applicable):
g++ (GCC) 6.1.1 20160621 (Red Hat 6.1.1-3)

How reproducible:
always

Steps to Reproduce:
1. Save the following code in a file test.C:

template<bool spawn>
int fib(int n) {
    if (n < 2)
        return n;
    bool deep = n > 30;
    int x, y;
    if (spawn && deep) {
        x = _Cilk_spawn fib<true>(n-1);
        y = fib<true>(n-2);
    } else {
        x = fib<false>(n-1);
        y = fib<false>(n-2);
    }
    if (spawn && deep)
        _Cilk_sync;
    return x + y;
}

int main() {
    return fib<true>(42) % 2;
}

2. Compile with the following command:
g++ -fcilkplus -O2 -Wall -o test test.C

Actual results:
test.C: In function ‘int fib(int) [with bool spawn = true]’:
test.C:10:38: internal compiler error: in cp_gimplify_expr, at cp/cp-gimplify.c:737
         x = _Cilk_spawn fib<true>(n-1);
                                      ^

Expected results:
A valid executable file.

Additional info:
I can reproduce this error also with g++ 6.2.0 from gcc.gnu.org.

A workaround is to introduce an extra variable, i.e. change this:

  x = _Cilk_spawn fib<true>(n-1);

to this:

  int x1 = _Cilk_spawn fib<true>(n-1);
  x = x1;

This however exposes a second problem. The resulting program is slower than the corresponding serial version, which suggests that the overhead introduced by the cilk keywords can not be optimized out when the <false> version of the template function is instantiated. If I manually create two separate functions I do get the expected speedup, but this is not practical for more complicated real world cases.