combination of unary minus and float conversion

Question

Consider the following C statements:

unsigned long x = 1;
float a = -x;
double b = -x;

I would expect the unary minus term to yield an unsigned long value equal to ULONG_MAX and a and b to be set to single and double precision representations of ULONG_MAX, respectively.

This is the result I obtain with gcc 4.4.7 on 32-bit Linux and with the Intel and PGI compilers on 64-bit Linux. With gcc (tested versions 4.4.7, 4.7.2 and 4.8.0, both with -O0 and -O2) on 64-bit Linux, however, the double variable b has the expected value, but float a becomes equal to -1 instead.

By contrast, the following statements will set both a and b to floating point representations of ULONG_MAX on all compilers and systems I tested:

unsigned long x = 1;
unsigned long y = -x;
float a = y;
double b = y;

If I use unsigned int instead of unsigned long, I get the expected result on all systems, too.

Is this some kind of undefined behaviour or a compiler error?

Answer 1

This is due to a bug in GCC -- the type conversion happens before the negation.

The issue seems to be have been around for a while. Bug 55771 - Negation and type conversion incorrectly exchanged

In your second example, the negation happens before the type conversion. As such, you see expected results.

Answer 2

What you describe is a compiler bug.

(and there is no undefined behavior in the program below to excuse the compiler)

unsigned long x = 1;
float a = -x;
double b = -x;