Dan Klishch

Well, I hadn't seen it when I was doing the initial research. Anyway, I believe that what is currently implemented would already be a significant improvement over a for loop...

Firstly, I've stress tested double conversion (checked 10^11 random bit patterns of double and all doubles with mantissa < 2^20 or mantissa > 2^52 - 2^20) and brute forced all...

@linusg Done @jk-jeon As for speed, it wasn't that bad, though I only tried running it for numbers up to 80 bits. With the "more likely to be correct algorithm"...

@jk-jeon Yes, I'm definitely not aligning table entries. > From which exponent did you get that? Oh, technically I have a bug in my script (I forgot about the counterintuitive...

@jk-jeon > Actually I'm quite doubtful about that. Are you sure you get more than 64-bits for IEEE-754 binary32? No, I'm not anymore. > You mean so your $B_0$ and...

@jk-jeon The value is somewhere between 39 and 40. As far as I can tell, they are using two different formulas for q: one with -1 (https://github.com/ulfjack/ryu/blob/75d5a85440ed356ad7b23e9e6002d71f62a6255c/ryu/d2s.c#L158) and one without...

@jk-jeon Hmm, earlier d2s also used formula without -1 if NICER_OUTPUT was not defined. https://github.com/ulfjack/ryu/blob/365dd77ddfca1e38ae88099b53d4c99c21b82a0e/ryu/d2s.c#L525

> My conspiracy theory: the formula without -1 is the correct one, but later they discovered that the one with -1 actually leads to the same output for `double` and...

I'm kinda losing hope for getting this merged. It was supposed to be a quick fix :)

I actually noticed this because Qt does a terrible job downscaling a framebuffer. Screenshots are taken on Sway 1.9 on a screen with 2.25 scale. Before:  After:  This...