Please Cc me in replies. As part of the rebootstrap[1] project, I am trying to cross build gcc-4.9 for x32. This fails with the following message: | g++ -c -DIN_GCC -DGENERATOR_FILE -I. -Ibuild -I../../src/gcc -I../../src/gcc/build -I../../src/gcc/../include -I../../src/gcc/../libcpp/include \ | -o build/genpreds.o ../../src/gcc/genpreds.c | In file included from ../../src/gcc/rtl.h:26:0, | from ../../src/gcc/genpreds.c:27: | ../../src/gcc/real.h:76:76: error: size of array 'test_real_width' is negative | [sizeof (REAL_VALUE_TYPE) <= REAL_WIDTH * sizeof (HOST_WIDE_INT) ? 1 : -1]; | ^ | Makefile:2356: recipe for target 'build/genpreds.o' failed The sources used are Debian gcc-4.9 4.9.2-10 sources, which should mostly correspond to svn revision r218987 on branch gcc-4_9-branch. In this build, the build architecture is amd64. The host and target architectures are set to x32. An x32 cross compiler is built and installed from the same sources. Looking at the gcc/real.h header in question, I am wondering whether the various architectures are somehow mixed in unfortunate ways. We can see that the build arch (amd64) compiler is used. The struct real_value (which lives behind REAL_VALUE_TYPE) has an "unsigned long sig[SIGSZ];" member. The SIGSZ macro is #defined as (SIGNIFICAND_BITS / HOST_BITS_PER_LONG). HOST_BITS_PER_LONG becomes 32. Thus I am wondering whether the code handles the case where build arch sizeof(long) > host arch sizeof(long) correctly. Of course, this can be a red herring and some other aspects of my build may be wrong. I also note that an earlier attempt of cross building Debian gcc-4.9 4.9.0-1 (which corresponds to svn revision r209695) did not exhibit this issue. What other aspects should I check with my build environment? Thanks for considering Helmut [1] http://wiki.debian.org/HelmutGrohne/rebootstrap
