MSVC brace initialization with doubles appears to

问题:

Check out this simple program:

int main() {
    float f2 = 7.2; // OK, with warning
    float f3 = 7.199999809265137; // OK, no warning
    float f4{ 7.2 }; // Fails
    float f5{ 7.199999809265137 }; // OK, no warning
    float f6 = { 7.2 }; // Fails
    float f7 = { 7.199999809265137 }; // OK, no warning
}

When compiled with MSVC 2015 using the default options (cl /W4, version 19.00.23918), I get the following messages:

FloatTest.cpp(2): warning C4305: 'initializing': truncation from 'double' to 'float'
FloatTest.cpp(4): error C2397: conversion from 'double' to 'float' requires a narrowing conversion
FloatTest.cpp(4): warning C4305: 'initializing': truncation from 'double' to 'float'
FloatTest.cpp(6): error C2397: conversion from 'double' to 'float' requires a narrowing conversion
FloatTest.cpp(6): warning C4305: 'initializing': truncation from 'double' to 'float'

This program compiles fine with Clang 3.0-3.8 and GCC 4.5.4-6.1.0 (tested with http://melpon.org/wandbox), with only warnings for unused variables. Further, removing/commenting out lines f4 and f6 result in successful compilation (with only the one warning for line f2).

Initially it looks like MSVC is just telling me that 7.2 can't be represented precisely as a float, so it's a narrowing conversion (which is illegal in brace initialization). However, the standard (draft N3337), section 8.5.4, note 7, says this:

A narrowing conversion is an implicit conversion...

• from long double to double or float, or from double to float, except where the source is a constant expression and the actual value after conversion is within the range of values that can be represented (even if it cannot be represented exactly)

Emphasis mine. Since 7.2 is within the range of values representable by float, its conversion to float should not be a narrowing conversion according to the standard. Is MSVC in the wrong here, and should I file a bug?

回答1:

It looks like a bug, indeed. For a workaround, the following appears to silence both errors and warnings in MSVC 2015.

#pragma float_control(precise, off, push)

float f2 = 7.2; // OK, with warning
//...

#pragma float_control(precise, pop)

The same works globally if using the /fp:fast compiler switch, though that one is incompatible with /Za which disables MS language extensions.

回答2:

Some floating point numbers can be exactly expressed in a float representation and some can't. If the number can be represented in the form x / 2^y where x is any integer and y is an integer 23 or less, it fits. Most decimal numbers can't be represented in this way, as a binary number they repeat forever. 7.2 is one example.

You can fix this easily by appending f to each number, to indicate to the compiler that this is a float constant rather than a double.

float f4{ 7.2f };