-ffloat-store is probably the wrong way to fix this, instead numpy requires SSE[2] anyway (so strictly doesn't qualify for the baseline used for our i586 builds), which means the proper "fix" would likely be to use -mfpmath=sse instead.