Floating Point Exception

[zifnabbe]

Hi,

I've stumbled upon the following when importing a text file with floating points generated by a java application. The following one gives an invalid floating point operation when loading into a TFloatField

531.1861582473471

while the others don't give an exception:

529.9518191598611
534.4721380018583

Further testing showed that the fraction part .1861582 is still okay, but .18615824 isn't

now, I could solve this exception, by using Set8087CW to mask the exceptions while loading the file.

However I'm wondering why this exception is triggered in the first place. Is the fraction for that specific floating point invalid for Delphi and not for Java?

What is the issue behind this?

Thanks in advance!

 

[Griffyn]

I don't know, but I hazard a guess that it's to do with that particular floating point number not being able to be represented as read from the text file exactly as a floating point number. See

http://en.wikipedia.org/wiki/Floating_point#Accuracy_problems

for some information on this.

 

[Glenn9999]

Or my guess would be that the representation doesn't exactly fit the field that TFloatField is masking. As my documentation says, this is a field with 15 significant digits. If my count is correct, all your numbers have 16 significant digits, which would no doubt cause the problem. Perhaps this is related to some in-system text to float conversion routine, especially to try and make sure there is no data loss for truncation as it goes into the database.

Perhaps the reason that you're not seeing a consistent problem would be how the number is represented, that somehow it is able to store some numbers accurately and not others, for the field size. This is not unlike a similar problem with storing 5 digit numbers in a signed word. You can store +20000, but +70000 won't work.

If I remember floating-point types right, this sounds like TFloatField is a IEEE single type that is redefined via type clause. If you need 16 digits precision, you might try to see what you can do to use IEEE double types.