Bug #899: The undefined reference to `__sync_bool_compare_and_swap_4' problem - Ryzom - Ryzom Core Development Site

Bug #899

« Previous | 212 of 267 | Next »

The undefined reference to `__sync_bool_compare_and_swap_4' problem

Added by rti over 8 years ago. Updated over 7 years ago.

Status:

New

Start date:

05/14/2010

Priority:

Normal

Due date:

Assignee:

% Done:

Category:

Build

Target version:

Description

Hi.

I just looked a bit further into the "undefined reference to `__sync_bool_compare_and_swap_4'" problem people are getting lately due to #874.

int main(int p_argc, char const** p_argv) {
    char val = 0;
    __sync_bool_compare_and_swap(&val, 0, 1);
    return 0;
}

I succeeded building this code with the following commands:

gcc/g++ -m32 -march=i486 ...

gcc/g++ -m64 ...

The building fails with the famous undefined reference to `__sync_bool_compare_and_swap_4' with the following commands:

gcc/g++ -m32 -march=i386 ...

I think the reason that brocifer was not able to compile the code (http://dev.ryzom.com/boards/17/topics/1648) is, that he only set the CFLAGS and not the CXXFLAGS variable. Since the call to __sync_bool_compare_and_swap resides in a C++ header, setting CFLAGS had simply no effect.

So I think the solution to this problem should be:

add the -march value on BuildForLinux
ensure that cmake generated build scripts contain the right -march setting on gcc

History

#1 Updated by vl over 8 years ago

Perhaps a more portable solution should be to detect (how?) if the function is available or not, and use the asm if the function is not available.

#2 Updated by rti over 8 years ago

Only versions of gcc >= 4.1 provide this function
http://stackoverflow.com/questions/2118992/how-do-i-use-gcc-builtin-sync-bool-compare-and-swap-in-g-on-macosx
Could be checked using the following defines provided by GCC:

  #define __GNUC__ 4
  #define __GNUC_MINOR__ 2

If the instruction is not available on the architecture
...but the compiler version is >= 4.1, the call is redirected:
"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." (taken from http://gcc.gnu.org/onlinedocs/gcc-4.1.1/gcc/Atomic-Builtins.html)

Thats where the __sync_bool_compare_and_swap_4 comes from. So implementing this function using the ASM Macro might be a way... But then again it does not make much sense, because GCC "thinks" the current architecture has no XCHG instruction, as a result the code would force the ASM to contain XCHG.

#3 Updated by vl over 8 years ago

Agree that i386 is quite old and every processor should have xchg so i think we can definitively change the to use at least i486

#4 Updated by vl over 8 years ago

Target version set to Version 0.8.0

#5 Updated by rti about 8 years ago

Just saw that the new PCH script does stuff like that:
IF(gcc_compiler_version MATCHES "4\\.[0-9]\\.[0-9]")

Just wanted to mention... On that basis we could decide whether to use the ASM or the gcc built-in, or, just complain that minimum version of GCC is 4.2 :)
But still the target processor arch needs to be checked / forced.