Link error when compiling gcc atomic operation in 32-bit mode

Question

I have the following program:

~/test> cat test.cc
int main()
{
  int i = 3;
  int j = __sync_add_and_fetch(&i, 1);
  return 0;
}

I\'

Answer 1

The answer from Dan Udey was close, close enough in fact to allow me to find the real solution.

According to the man page "-mcpu" is a deprecated synonym for "-mtune" and just means "optimize for a particular CPU (but still run on older CPUs, albeit less optimal)". I tried this, and it did not solve the issue.

However, "-march=" means "generate code for a particular CPU (and don't run on older CPUs)". When I tried this it solved the problem: specifying a CPU of i486 or better got rid of the link error.

~/test> /share/tools/gcc-4.2.2/bin/g++ -m32  test.cc
/tmp/ccYnYLj6.o(.text+0x27): In function `main':
: undefined reference to `__sync_add_and_fetch_4'
collect2: ld returned 1 exit status

~/test> /share/tools/gcc-4.2.2/bin/g++ -m32 -march=i386 test.cc
/tmp/ccOr3ww8.o(.text+0x22): In function `main':
: undefined reference to `__sync_add_and_fetch_4'
collect2: ld returned 1 exit status

~/test> /share/tools/gcc-4.2.2/bin/g++ -m32 -march=i486 test.cc

~/test> /share/tools/gcc-4.2.2/bin/g++ -m32 -march=pentium test.cc

Answer 2

From the GCC page on Atomic Builtins:

Not all operations are supported by all target processors. If a particular operation cannot be implemented on the target processor, a warning will be generated and a call an external function will be generated. The external function will carry the same name as the builtin, with an additional suffix `_n' where n is the size of the data type.

Judging from your compiler output, which refers to __sync_add_and_fetch_4, this is what's happening. For some reason, GCC is not generating the external function properly.

This is likely why you're only getting an error in 32-bit mode - when compiling for 64-bit mode, it compiles for your processor more closely. When compiling for 32-bit, it may well be using a generic arch (i386, for example) which does not natively support those features. Try specifying a specific architecture for your chip family (Xeon, Core 2, etc.) via -mcpu and see if that works.

If not, you'll have to figure out why GCC isn't including the appropriate function that it should be generating.