Conflicting compilers and optimisers that broke this entry in 2023

As briefly noted in the index.html gcc and clang caused different problems with this entry depending on the optimiser being enabled or not, where if the optimiser was enabled it would work with one compiler and not the other. But if you then disabled the optimiser the opposite problem would occur: the working compiler would no longer work and the non-functional one would start to work! And then there’s the problem of linux versus macOS causing problems.

The alternate code (source code) demonstrates the problem so you can see if your compiler has the problem as described below.

In some versions of gcc and clang (this was first discovered in fedora linux 38) if the compiler was enabled one compiler would generate code that segfaulted but the other one would not. However, when then disabling the optimiser the compiler that had no problem suddenly did and the one that did suddenly did not. But segfaulting was not by any means the only problem.

Another problem was that depending on the optimiser an infinite loop would be entered. But then removing it would cause other problems as well, including a segfault.

It was not only one compiler that dumped core. Both dumped core with other behaviour, depending on what changes in code were made and the optimiser being on or off. Clang in macOS also dumped core!

The marshall.c has the original fix for clang which works with some compilers depending on the optimiser. When we refer to the code below we refer to the alternate code, marshall.alt.c.

The conflicting compiler problems:

Below we describe the conflicting options that resulted in the need to modify the original code.

Linux

With the optimiser disabled GCC compiled the code to enter an infinite loop. If the optimiser was enabled it worked fine. But if the optimiser is enabled clang generates code that first prints the text twice (it should print it once) and then immediately dumps core!

But then there’s that infinite loop. What happens if we remove it?

Once the loop is removed if we compile with clang and the optimiser is disabled all is okay. If however we compile with gcc and the optimiser is disabled the string is printed once and then the program dumps core! Why is this? It is because of the call to _exit() (see below).

Clang works fine with the loop removed and the optimiser disabled. If however we enable the optimiser clang will generate code that prints the string twice and dumps core just like before! What about the call to _exit()?

The call was:

    _exit(argv[(int)*argc-2/cc[1*(int)*argc]|-1<<4]);

Changing it to not refer to these variables and just exit 1 and the problems disappear!

macOS

But then there was macOS which had another problem!

If the infinite loop is not removed/commented out and the optimiser is enabled it works fine but if it’s disabled it prints out the string and then enters an infinite loop!

After we fix the infinite loop if we change the _exit() call in macOS and disable the optimiser we might get something like:

    $ ./marshall
          choo choo
    Segmentation fault: 11

Observe how it doesn’t print it twice. But of course it segfaults so it’s not right either.

If the optimiser is enabled, however, we will see that it works fine before the change to _exit().

Summary

And that is the tale of a little engine train that could…eventually! Or perhaps it’s the compilers, optimisers and OSes that eventually could?

Whichever you wish it to be, it now works with both clang and cc. Choo choo!

Jump to: top