That is strange:
It is supposed to be fixed very recently.
I can only test on 64 bit Windows system, but compiled to 32 bit. (Not a true 32 bit Windows)
That works for the code given.
That means you can re- open the bug report if you want (And can prove it is still there, or again there with a simple example).
I can not reproduce it, though (Windows 10, 64 to 32 cross-compiler from today)
It may also depend on optimization settings, I have no room to test that. I suspect another screw up in recent i386 assembler optimizations, which are good but very risky.
I will try to reproduce the issue.