Hi Eric,
> Before the patch, gcc gave 9 of 16 tests failing on assertions, all
> dealing with NaN, when I ran:
>
> gnulib-tool --with-tests --test isfinite isnan ceill floorl frexpl isnanl
> ldexpl roundl truncl
>
> After the patch
Nice improvement!
> 2 tests still fail
For me:
- with CC="cc -O", all tests pass.
- with CC="gcc", a gcc 2.95.2 I get this failure
test-floorl.c:82: assertion failed
/bin/sh[9]: 255782 Abort(coredump)
FAIL: test-floorl
but this is due to a gcc bug: If floorl(-0.7L) is executed before
floorl(-1.0L), and both calls are compiled by gcc - regardless whether
in the same function, in different functions, or even in different
compilation
units -, then the floorl(-1.0L) call returns -2L instead of -1L. If at
least one
of the calls is in a compilation unit compiled by cc, no problem.
> Meanwhile, while running the tests, I noticed:
>
> checking whether HUGE_VAL works... yes
> checking whether roundl is declared... no
> checking whether ceill is declared... (cached) yes
> checking whether floorl is declared... (cached) yes
> (cached) (cached) checking for signbit macro... no
> checking for signbit compiler built-ins... no
This comes from the gl_FUNC_FLOORL_LIBS and gl_FUNC_CEILL_LIBS macros, which are
invoked by gl_FUNC_ROUNDL. Each of them contains an AC_CACHE_VAL invocation that
is not wrapped in AC_MSG_CHECKING.
> Ouch. That double (cached) at the beginning of a line means that we are
> doing one AC_CACHE_CHECK inside the body of another, and that it is likely
> that 'configure -C' will get things wrong on a second run.
'configure -C' does not mishandle AC_CACHE_VAL just because there is no AC_MSG_*
around it?
Bruno
