This does not matter on newer cores (where we can use reciprocal estimates in fast-math mode anyway), but for older cores this allows us to generate better fast-math code where we have multiple FDIVs with a common divisor. llvm-svn: 222710
Hal Finkel