Why does a floating-point multiplication compile more efficiently when I put it in an inline function?

c gcc floating-point arm abi