gcc 11.1.0: printf("%.43f\n", 0x1.52f8a8e32e982p-140): printed value is incorrectly rounded

[Keith Packard]

Corinna Vinschen <vinschen@redhat.com> writes:

> Is anybody here willing to take a look and change Newlib to gdtoa?

It might be useful for newlib to adopt the Ryū floating point conversion
code developed by Ulf Adams (currently at Google). This does correct
double conversion in 128 bits, avoiding need for arbitrary precision
math and any allocations.

        https://dl.acm.org/doi/10.1145/3360595

There is a BSD licensed implementation of the core algorithm available
on github; adopting that into complete printf/scanf implementations
isn't a huge amount of work. picolibc uses this implementation.

A friend of mine posited that this could be done, and got started on it
in grad school about 20 years ago. It was very nice to see someone
actually take the time to make it happen.

-- 
-keith
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 832 bytes
Desc: not available
URL: <https://sourceware.org/pipermail/newlib/attachments/20211104/b3cf7b85/attachment.sig>