I just ran the test suite with an x86/ARM cross-compiler and it looks like we have no cross-compiler specific problems!

There's one interesting failing test case: in runnable/mangle.d:

//UTF-8 chars
__gshared pragma(mangle, "test_эльфийские_письмена_9") ubyte test9_1;
__gshared extern pragma(mangle, "test_эльфийские_письмена_9") ubyte
test9_1_e;

void test9()
{
	import std.stdio;
    test9_1 = 42;
    writefln("test9_1 = %s", test9_1);
    writefln("test9_1_e = %s", test9_1_e);
    assert(test9_1_e == 42);
}

The assert fails with optimization enabled iff the writelns are removed. I haven't verified this yet but I guess that gcc reorders the read/write to the symbol(s) when optimization is enabled.

So: If we have symbols with the same mangled name, do we guarantee that reordering won't happen? What would be the best way to rewrite this test case so it doesn't rely on ordering or can't be reordered?

Placing the read/write into extra functions could work, but inlining could cause problems...