AVR GCC 10 support for -mdouble=64

[ckormanyos]

AVR GCC 10 support for -mdouble=64 does not work on some floating-point benchmarks.
Understand why and support -mdouble=64 for AVR-GCC v10 and higher.

[ckormanyos]

Put this on hold for a while until either build my own AVR-GCC 10 or wait until Zak builds another one.

See also the comment reproduced in the image below:

[ckormanyos]

Recently tried Zak's build of avr-gcc 11.1.0. The 64-bit double type was doing better and a few tests, but not all, passed. One of the extended precision tests (with wide-decimal) faile, either the pi AGM or the wide-decimal sqrt.

Possibly see exactly which test faile and, if possible, figure out where/why?

[ckormanyos]

Zak's build of avr-gcc 11.1.0. The 64-bit double type was doing better and a few tests, but not all, passed

Errors were actually found in the self-written implementation of <cmath> regarding built-in assembly pneumonics for f, no-suffix, or l. By accident, pneumonics with no suffix (i.e., those for double) were erroneously used for float. When switching to 64-bit double, this old problem was revealed.

In summary, avr-gcc 11.1 does well with -mdouble=64 and passes all benchmark tests in real-time-cpp.

Closing this issue.

[ckormanyos]

See also #177

[ckormanyos]

Some problems remain such as finding (link stage) sinf, cosf, etc. in examples like trapezoid integral. Revert changes in <cmath>.

Trapezoid integral example still does not work when 64-bit double is active.

[ckormanyos]

Rework needed see also #180