Remove UB from int-to-float conversion for INT_MIN.

[pnkfelix]

The fixed->float conversion intrinsics were carefully written to handle the case where the input is INT_MIN; e.g. avoiding sign extension via an intermediate cast to an unsigned int.

Unfortunately it wasn't quite careful enough; I am seeing evidence in rust-lang/rust#36518 that one cannot rely on -a to behave in the same manner as ~(unsigned)a+1, because it is simply undefined behavior.

• I found a useful discussion of that detail of INT_MIN here: http://stackoverflow.com/questions/37301078/is-negating-int-min-undefined-behaviour

Anyway, the fix I've chosen is simple enough: cast to unsigned int earlier in the computation and use ~a + 1 to do the negation.

[pnkfelix]

Oh wait, during more review, it looks like the code in https://github.com/llvm-mirror/compiler-rt/blob/master/lib/builtins/floatsitf.c is doing the right thing. So it would probably be better form for me to adapt that code rather than introducing my own variant..

[alexcrichton]

Oh if this is fixed upstream maybe we should just update to upstream?

[pnkfelix]

@alexcrichton Sorry, my previous commit was super unclear.

Its not fixed upstream.

floatsitf.c handles converting single-word fixed-point to quad-word floating-point

It is a more recent file and I guess its author took care to avoid the UB here.

In any case, I figured the best thing for me to do would be to adopt the coding style used by the author of floatsitf.c, because I figure that has a much better chance of being accepted upstream, since I can point to the precedent.

(the end effect in terms of code execution is the same as what I wrote; its just a matter of taste in terms of whether it is easier to miss bugs where one writes a when they should have written aAbs in this new version).

[pnkfelix]

(In case its unclear: commit b2676d8 is the one that has adapted the style used in floatsitf.c, so it can be reviewed and hopefully merged.)

[alexcrichton]

Ah ok, sounds good to me!