"stéphane ducasse" ducasse@iam.unibe.ch wrote:

...

hi guys

Could one of you write a bunch of tests to record the correct behavior and add as comment the discussion of this email?

Stef

Attached you find a change set that contains additional tests for the class FloatTest. Comments are welcome.

Greetings, Boris

Hi, all:

I was temporarily bumped off the mailing list distribution, so I didn't see the flurry of Float NaN postings until now. I implemented the original "IEEE-ification" of the Float class, adding the NaNs, Infinities, stable printing, etc. The interaction of NaNs with Sets and Dictionary keys was a surprise, but it is interesting that it hadn't been discovered until recently (6+ years after the addition of NaNs with the behavior x == x, x ~= x).

Some comments on Boris' tests:

testNan2 is not portable; there is no reason for any particular floating-point library implementation to return NaN for sin / tan of values greater than 1e200. That is simply an artifact of Intel's x86 trancendental range reduction implementation.

You might also add a test for the difference of same-signed Infinities returning NaN:

Inf - Inf -> NaN Inf + -Inf -> NaN

In the testZero1 test:

self assert: (Float negativeZero at: 1) ~= (0 asFloat at: 1).

This can be written more understandably as:

Float negativeZero sign ~= 0 sign.

Additionally, you could add a testSign test which verified that sign works as above.

-- Tim Olson