gcc -O3 , data pointer seems to get lost

Question

Depending on optimization level the output differ as follows:

With unexpected output:

$ gcc -Wall -O3  otest.c -o otest                   
$ otest                                             
*x: 0 
y: 2048.899902 
y: 0.000000

With expected output:

$ gcc -Wall -O2  otest.c -o otest 
$ otest                           
*x: 45000e66 
y: 0.000000 
y: 2048.899902

source code :

#include <stdio.h>
int main(void)
{
   float y = 2048.9;
   void *p = &y;
   unsigned int *x = p;
   printf(" *x: %x \n",*x);
   *x = 0; 
   printf(" y: %f \n",y);
   *x = 0x45000e66;
   printf(" y: %f \n",y);  
   return 0;
 }

gcc version is 4.2.1.

Am I missing any important directive ?

@user3477950 I think that's a good keyword to look for further information on the net. If the OP still wants to use it, that's his problem... — Marc Glisse, Apr 30 '14 at 14:49
@MarcGlisse yes, but your comment suggests that it's a **solution.** It is not. — The Paramagnetic Croissant, Apr 30 '14 at 14:50
@ole in case it wasn't clear, a random word in a comment is not a "solution" or a "recommendation", it is advice on what to type in your search engine to get more information and understand in what ways your code is wrong. — Marc Glisse, Apr 30 '14 at 14:53
@LưuVĩnhPhúc: No, it is not. Reading union via different type than you set it is just as undefined behaviour as this. The only exception is if one of the types is `char`. — Jan Hudec, Apr 30 '14 at 15:48
@JanHudec It's not undefined behavior. **You are confusing pointer-based type punning** (which is UB unless the aliasing type is signed or unsigned `char`) **with union-based type punning** (which is **not** undefined behavior since C99). — The Paramagnetic Croissant, May 01 '14 at 05:50
@JanHudec Yes, it is. However, the question is about C, not C++. — The Paramagnetic Croissant, May 01 '14 at 17:48

score 4 · Answer 1 · answered Apr 30 '14 at 14:47

4

Yes. Your code is violating the strict aliasing rule (when you have a float, but you access it through a pointer to unsigned int, which is an incompatible type), invoking undefined behavior, so the compiler is allowed to do anything it pleases with your code, including entirely eliminating parts of it.

answered Apr 30 '14 at 14:47

The Paramagnetic Croissant

9,223
3
25
38

It might be worth pointing out that compilers these days do try to do the right thing -- they'll generate their code presuming circumstances that would lead to undefined behavior don't exist, which can open up quite a lot of optimization opportunities... – jthill Apr 30 '14 at 15:01

gcc -O3 , data pointer seems to get lost

1 Answers1