https://gcc.gnu.org/bugzilla/show_bug.cgi?id=106440
--- Comment #3 from Yuhao Yao <yaobig97 at gmail dot com> --- (In reply to Andrew Pinski from comment #1) > ubuntu@ubuntu:~/src/upstream-gcc-aarch64\# g++ t55.cc -ffp-contract=on -O2 > ubuntu@ubuntu:~/src/upstream-gcc-aarch64\# ./a.out > 3e+09 3e+09 > 0 > 0 > ubuntu@ubuntu:~/src/upstream-gcc-aarch64\# g++ t55.cc -ffp-contract=fast -O2 > ubuntu@ubuntu:~/src/upstream-gcc-aarch64\# ./a.out > 3e+09 3e+09 > 512 > 0 > > > GCC defaults to "-ffp-contract=fast" which enables the use of the Fused > multiple add instruction in some cases. > > What is happening in this case is the second print is being constant folded > (before the FMA instruction is formed) and the first print uses a fnmsub > instruction. > > That is there is no bug, you just don't understand floating point. Ah thanks! It seems that I know too few things on floating point and I should have done more research before making this post. Sorry for wasting your time.
