[PATCH] D15814: Implicit conversion from float->bool

[Aaron Ballman via cfe-commits]

On Wed, Dec 30, 2015 at 12:30 PM, Richard Smith <richard at metafoo.co.uk> wrote:
> On Dec 30, 2015 6:34 AM, "Aaron Ballman" <aaron.ballman at gmail.com> wrote:
>>
>> aaron.ballman closed this revision.
>> aaron.ballman added a comment.
>>
>> Thanks! I've commit in r256643.
>>
>>
>> ================
>> Comment at: test/SemaCXX/warn-literal-conversion.cpp:49-50
>> @@ +48,4 @@
>> +  // values.
>> +  bool b3 = 0.0f;
>> +  bool b4 = 0.0;
>> +}
>> ----------------
>> rsmith wrote:
>> > What about
>> >
>> >   bool b5 = 1.0;
>> >   bool b6 = 2.0;
>> >
>> > ? Arguably any `float` -> `bool` conversion changes the value (because
>> > `true` and `false` are not values of type `float`), so it wouldn't be
>> > completely unreasonable to warn even if the literal is `0.0`.
>> Except those conversions won't cause confusion to the user, so I'm not
>> certain what we gain by diagnosing. Given that some mental models expect
>> 0.99 to convert to 0, which converts to false (because bool is an integral
>> type, so it "must" do the usual integral truncation dance), it makes sense
>> to tell the user "no no no, that converts to true." I'm less convinced about
>> the utility of warning on things like `bool b = 1.99f' where it changes the
>> value from 1.99 to true. Perhaps this should be changed to only diagnose
>> when converting through an integer would result in a different value that
>> converting through the float?
>
> Let me ask a slightly different question: when would it ever be reasonable
> and intentional to initialize a bool from a floating-point *literal*? For
> integers, people use things like
>
>   int n = 1e6;
>
> which is why we don't complain if the literal is exactly representable in
> the integer type. I can't imagine any similar cases for bool.

I see what you're getting at now, and I agree that it wouldn't be
particularly reasonable. However, I think that there are two things to
warn about: initializing from a literal in general, and initializing
from a literal where the resulting Boolean value may be surprising.
While we're at it: is there ever a time where it's reasonable to
initialize a bool from a literal of type other than an integer or
bool? For instance:

bool b1 = L'1';
bool b2 = '1';
bool b3 = 1.0f;
bool b4 = nullptr; // -Wnull-conversion catches this

I think it makes sense for all of these to diagnose under
-Wliteral-conversion, but drop the "changes value" clause.

bool b5 = 0.01;
bool b6 = '0';
bool b7 = L'0';

I think it makes sense for these to diagnose under
-Wliteral-conversion, but with the changes value clause.

~Aaron

>
>>
>> http://reviews.llvm.org/D15814
>>
>>
>>