Hi Helmut, On 28.08.21 13:53, Helmut Grohne wrote: > The issue is with the detection of the long double epsilon on powerpc > architectures. m4/floatparam.m4' macro CL_FLOATPARAM_CROSS attempts to > deduce the minimal value that can be added to 1 using compile time > bisection. That works, because gcc does constant folding of floating > point numbers, except when it does not not powerpc architectures. There > is a long standing gcc bug > https://gcc.gnu.org/bugzilla/show_bug.cgi?id=19779. > > Would there be any chance to use gcc constants such as LDBL_EPSILON for > this case (even opportunistically)? > > In any case, the current implementation is very bad, because the while > loop never finishes as every compilation attempt fails on the supposedly > non-constant expression. At minimum, there should be a bound on the > number of iterations.
How can I reproduce this? On my bullseye setup, this works: $ ./configure --host=powerpc64le-linux-gnu --disable-shared $ make $ qemu-ppc64le -L /usr/powerpc64le-linux-gnu ./tests/exam Tests passed. It surely produces the expected floatparam.h file, with long_double_mant_bits set to 106. -richy. -- Richard B. Kreckel <https://in.terlu.de/~kreckel/>
