gcc warning problem, strict aliasing

[MadCow108]

I'm currently trying to compile some foreign code on a newer gcc version (4.4.1)
I get following warning in code which reads a binary file:

tools/hist_draw.cc:128: error: dereferencing type-punned pointer will break strict-aliasing rules
in following code segment:

void
mystream::myget(float& f) { // mystream inherits from ifstream
char c[4];
get(c[0]);
get(c[1]);
get(c[2]);
get(c[3]);
f = *(float*)c; // line 128
}

I do not fully understand the error message yet, but changed the segment to this:

void
mystream::myget(float& f) {
read(reinterpret_cast<char*>(&f),sizeof(float));
}
and the warning disappears.

The question is:
Have I fixed the problem or just outsmarted gcc's warning system?

(I could also remove the warning by introducing a temporary float pointer but that probably does not solve the reason but just takes gcc ability to detect the problem)

a yes no answer is enough, but an explanation or link to the reason of the problem would be nice too.

[ibuclaw]

Are you able to reproduce this on a small scale program example?

I can't seem to get the same message you are getting.

Regards
Iain

[MadCow108]

here:
compile with -Wall -O2 in gcc 4.4.1 (I think it doesn't occur in gcc <= 4.3.3)

#include <iostream>
#include <fstream>

using namespace std;

void readfilewarning(ifstream & results, float &m)
{
//warning
char c[4];
results.get(c[0]);
results.get(c[1]);
results.get(c[2]);
results.get(c[3]);
m = *(float*)c;
//nowarning
//results.read(reinterpret_cast < char * > (&m), sizeof(m));
}

int main (int argc, char *argv[])
{
ofstream file("numbers.dat",ios::binary);
float x = 5.2342;
file.write(reinterpret_cast<char *>(&x),sizeof(float));
file.close();
float m;
ifstream results("numbers.dat",ios::binary);
readfilewarning(results,m);
std::cout << m << std::endl;
return 0;
}

both versions seem to work, I just would like to make sure I really solved the warning issue

[ibuclaw]

Had a small look around, and it looks like that the reason you get that output is because the behaviour of how the line is interpreted at the optimisation stage of the compiler is generally undefined.

Adding -fno-strict-aliasing will hide the warning message.

Regards
Iain

[MadCow108]

yes but it will also disable aliasing optimizations. Just disabling the warning with -Wno-strict-aliasing is also not wanted.

After some testing and reading I guess my solution is ok as the operation on type-punned pointers happen (if at all) in the (hopefully) safe standard library and not in my code.
If I'm wrong please say.

[dribeas]

For an interesting article on strict aliasing, follow this link (http://cellperformance.beyond3d.com/articles/2006/06/understanding-strict-aliasing.html), it explains it and some of the consequences better than I could reproduce here.

At the end, the simplest solution to your particular problem is creating an union:



union binary_float
{
char c[4];
float f;
};

void readfilewarning(ifstream & results, float &m)
{
union binary_float u;
results.get(u.c[0]);
results.get(u.c[1]);
results.get(u.c[2]);
results.get(u.c[3]);
m = u.f;
}

[MadCow108]

thanks for this nice article. It is very informative.

I don't actually need that char buffer, it was just some strange way it was implemented by someone else and in the article it says a cast to a char* is save so it should be fine.
So I'll stick with my solution.

Thank you for your help.

[tmcg]

If it is still of interest: If you'r strict even the union approach is not standard conforming since you write a char[] and read as float afterward.

On the other hand accessing any type as char (using a cast) is explicitly allowed. So your reinterpret_cast is the correct solution.