https://gcc.gnu.org/bugzilla/show_bug.cgi?id=64370
--- Comment #5 from dave.anglin at bell dot net --- On 2015-01-09 2:31 PM, ubizjak at gmail dot com wrote: > https://gcc.gnu.org/bugzilla/show_bug.cgi?id=64370 > > > > --- Comment #4 from Uroš Bizjak <ubizjak at gmail dot com> --- > > (In reply to Mikhail Maltsev from comment #3) > >> In libgfortran/intrinsics/c99_functions.c, scalbn is implemented like thi > s: > >> double >> scalbn (double x, int y) > > > You should be extra careful here, we are using scalbln to avoid overflows. > > > Please see PR64503, there are some ipa tests in the testsuite (e.g. > > gcc.dg/ipa/iinline-4.c) that report -inf for estimated badness when other > > functions are used. The change to using scalbln breaks hppa*-hpux11* as well: g++ -c -g -DIN_GCC -fno-exceptions -fno-rtti -fasynchronous-unwind-tables - W -Wall -Wno-narrowing -Wwrite-strings -Wcast-qual -Wno-format -Wmissing-format- attribute -Woverloaded-virtual -pedantic -Wno-long-long -Wno-variadic-macros -Wn o-overlength-strings -fno-common -DHAVE_CONFIG_H -I. -I. -I../../gcc/gcc -I../. ./gcc/gcc/. -I../../gcc/gcc/../include -I../../gcc/gcc/../libcpp/include -I/opt/ gnu/gcc/gmp/include -I../../gcc/gcc/../libdecnumber -I../../gcc/gcc/../libdecnu mber/dpd -I../libdecnumber -I../../gcc/gcc/../libbacktrace -o sreal.o -MT srea l.o -MMD -MP -MF ./.deps/sreal.TPo ../../gcc/gcc/sreal.c ../../gcc/gcc/sreal.c: In member function 'double sreal::to_double() const': ../../gcc/gcc/sreal.c:125:30: error: 'scalbln' was not declared in this scope val = scalbln (val, m_exp); I am puzzled why scalbln was used as m_exp has type "signed int". The code also seems to be specific to a radix of 2. So, ldexp (x, y) or x * pow (2, y) seem like the correct functions to use. Also note that an int type should hold the full range of exponents for double on every machine that I'm aware of, so it doesn't seem like there is any benefit in using scalbln over ldexp. I have read that there are problems with overflow detection on alpha. So, is that the issue? Dave
