OK for me to close it as a compiler bug since there is a workaround.

It would be great if we could detect this compiler and deactivate this optimization automatically, but I am too lazy to search the xlc compiler documentation for a way to do that.

I suppose that the people who will be confronted to this problem (the 4 of us in the world who compile Python on AIX with xlc) will find this issue and the correct flag to make it work.

Also I am using xlc instead of gcc because it provides some better optimization for this architecture. If the computed gotos optimization only works with gcc, then it may be more interesting to use the gcc compiler instead (I will do some benchs).