Created attachment 15381 [details] Snippet demonstrating log10l imprecision. hello @all, quite sure you'll tell me it's not a bug but the 'feature of imprecision' in bin-FP-math :-( log10l( 1E+07l ) produces 7.000000000000000000434..., while logl( 1E+07l ) / logl( 10.0l ) would hold with 7.0, Same as with pow10s, having them exact at integral powers of ten would help to form a reliable grid of stable points, and find better ways for correct rounding, which would help for better math. Thus if possible to improve without messing other points I'd appreciate, if not I'd like some explanation why and which calculation path is taken / preferred, and a code pointer to have a look myself. 'Me bad' is always an option, tried to evaluate and present with best intention and skills available. If 'notabug' pls. leave open or unconfirmed as hook for others who at some time in the future might have a good idea.
7.000000000000000000434 is still within 1 ULP.
@Andreas Schwab: thanks for the hint, it's 1 ULP off! add. observations: for 4933 positive integral powers of ten log10l has 1 ULP devia in 1730 cases, logl( 10^n ) / logl( 10 ) has 1 ULP devia in 486 cases, for 4950 negative integral powers of ten log10l has 1 ULP devia in 1745 cases, logl( 10^n ) / logl( 10 ) has 1 ULP devia in 501 cases, all these devias can be neutralized with simple recheck calculations, if! integral powers of ten are exact. #28474.