Agreed, if your program works properly with -O0 and bombs with any other optimization, it sounds like your program has some variables that should be marked volatile.
The optimizer is getting rid of code that does nothing useful (in its opinion), like repeatedly checking a flag. Why would you check a flag that never gets set? That's useless, get rid of that code.
Or setting a variable that never gets read, that's useless, get rid of that code.
Now if those variables happen to be device registers, they should be marked as volatile to tell the optimizer not to make assumptions about there being no behind the scenes activity.
Values that are being shared between interrupts and mainline code, or between tasks should be declared _Atomic (C11). And if you're doing complex things, the values should be guarded with a mutex.
