Re: [avr-libc-dev] Test results

[Dmitry K.]

Hi.

1. signbit()
This is my omission.  With GCC 4.0 and early the signbit()
function work fine.  Since 4.1 branch the GCC replaces it
with themselves inline code with internal compile error in
result (GCC bug #30243).  So a long time this function was
untested.  Now this bug is corrected and we see a mismatch:
inline GCC's code does return any nonzero value, where
the old realization (both Avr-libc and GCC inline) was
return exactly 1.
   I must to exclude the note about "1" from documentation
(include/math.h).  IMHO, this API change is possible in
current branch, as it is known the signbit() function
was unusable (due to internal compiler error) a long time.

2. isinf()
See the GCC bug #35509.  Avr-libc's isinf() work fine.
Hovever, since 4.3 branch the GCC replaces it with
themselves inline code, which return the +1 value for
negative infinity.  This is mismatch to traditionaly
isinf() behavior.
   I think, it is needed to add note about this bug.

3. modf()
The modern GCC replaces it with inline code, which
does not permit to use NULL address for the avoiding an
integral part storing.
   I think, it is needed to add note about this.

Other is not needed a quick corrections:

4. bug-22828 and other EEPROM tests:  seems, this is a
Simulavr problem.  The 0.1.2.1 version is correct.  I will
look this in details.

5. Compilation of a few of printf() test is failed due
to degradation of modern GCC code-size quality.  Early
branches are fine.  I will split its in future.

--------------
In future, it is desirable to exclude inline replacing
of math functions for AVR back-end.  This will reduce the
code size (and, probably, enlarge the speed).  I remember
that it is not difficult.

--------------
I plan to correct docs today, 5 Mar, at the nearest 6 hours.
If it is not desirable, mail to 'avr-libc-dev', I will not
have a connection with the mail-box.

Dmitry.