Any optimizations cause miscompiles and crashes

[Midar]

Building with everything above -O0 seems to cause miscompiles leading to crashes. I don't have any specific example, since it seems so many things are miscompiled. From what I've seen when compiling ObjFW with -O1, -O2 and -Os:

-O1: Mostly works, until some loop gets miscompiled. -O2: Immediately crashes -Os: Immediately has an assertion fail

I'm not sure where to start at all, given how much is being miscompiled. But being able to use something better than -O0 would benefit such old hardware a lot.

[bebbo]

this reproduced the bug: http://franke.ms/cex/z/85Ed9M

the lines 43-45 read

        move.l (a2),(a0)
        move.l (a2),a0
        move.l a2,(4,a0)

with the fix it's

        move.l (a2),a0
        move.l a0,(a0)
        move.l (a2),a0
        move.l a2,(4,a0)

[bebbo]

http://franke.ms/cex/z/85Ed9M now shows the correct code

[Midar]

With the recent commits, it seems to get a lot farther now.

GCC also gets confused about initialization when using -Os.

Here's (simplified) code:

+ (id)handler
{
        id handler;

        [globalMutex lock];
        @try {
                handler = somethingReturningTheHandler();                  
        } @finally {   
                [globalMutex unlock];
        }                                                                  

        if (handler == nil)
                @throw [SomeException exception];

        return handler;
}

GCC complains with -Os that on return handler, handler could be uninitialized. However, if handler isn't initialized, this code is never reached as it means an exception was thrown and the control flow stopped.

[bebbo]

GCC complains with -Os that on return handler, handler could be uninitialized. However, if handler isn't initialized, this code is never reached as it means an exception was thrown and the control flow stopped.

I can't help there

[bebbo]

and using double for seconds is also not a good choice on the Amiga. Use long long instead.

[Midar]

long long doesn't have more than second precision. double allows the precision to be scaled with how far away the date is, which is very desirable: Nobody cares about nanosecond granularity for a date in the year 10524.