Your fix adds worst-case rounding tests for math functions. It may also be useful to include worst-case rounding tests for number conversions (string-to-double).
I have explored string-to-double issues, including problems when the converted numbers are close to the breakpoints between doubles. This happens when the string is (m + 0.5 + eps)*2^exp I found three relevant test cases: 1) The shortest string with eps = 0 "1e23" = 0x1.52d02c7e14af6p76 + (0.5 + 0.0)*1p24 2) The minimum positive eps = 2.1546568143020914E-19 "3743626360493413e-165" = 0x1.8823a57adbef9p-497 + (-0.5 + 2.1546568143020914E-19)*1p-549 3) The maximum negative eps = -8.318563789335443E-19 "1864950924021923e213" = 0x1.3ae60753af6cap758 + (0.5 + -8.318563789335443E-19)*1p706 These examples may be useful for testing Double.parseDouble and Double.toString -Dmitry On Thu, Aug 11, 2011 at 1:24 AM, Joe Darcy <joe.da...@oracle.com> wrote: > Hello. > > Cleaning up my bug backlog, please review the addition of some new math > library tests: > > 4900206 "Include worst-case rounding tests for Math library functions" > > http://cr.openjdk.java.net/~**darcy/4900206.0/<http://cr.openjdk.java.net/%7Edarcy/4900206.0/> > > These tests probe at hard to round values generated by number-theoretic > techniques. I've verified the tests pass on Linux x86 and Solaris sparc. > > Thanks, > > -Joe >
