Created attachment 50160 [details] minimal example __gcc_qadd, which adds two long doubles on POWER, can emit NaN for non-NaN input: $ cat ldbl.c extern int printf(const char* restrict, ...); int main() { long double a = -0x1.1e282e024debdb875f47f6c85p+1021L; printf("%Le\n", a + 0x1.fffffffffffff7ffffffffffff8p+1023L); } $ gcc -O0 -o ldbl ldbl.c $ ./ldbl nan $ gcc -O1 -o ldbl ldbl.c $ ./ldbl 1.546510e+308 Examining the assembly from the second invocation shows that GCC has optimized out the entire computation. In the first invocation, however, the compiled program calls __gcc_qadd, which yields NaN. By IEEE 754, adding two non-NaN inputs should never produce a NaN output. Output of 'gcc -v': Using built-in specs. COLLECT_GCC=gcc COLLECT_LTO_WRAPPER=/usr/lib/gcc/powerpc64le-linux-gnu/10/lto-wrapper OFFLOAD_TARGET_NAMES=nvptx-none OFFLOAD_TARGET_DEFAULT=1 Target: powerpc64le-linux-gnu Configured with: ../src/configure -v --with-pkgversion='Debian 10.2.1-6' --with-bugurl=file:///usr/share/doc/gcc-10/README.Bugs --enable-languages=c,ada,c++,go,d,fortran,objc,obj-c++,m2 --prefix=/usr --with-gcc-major-version-only --program-suffix=-10 --program-prefix=powerpc64le-linux-gnu- --enable-shared --enable-linker-build-id --libexecdir=/usr/lib --without-included-gettext --enable-threads=posix --libdir=/usr/lib --enable-nls --enable-bootstrap --enable-clocale=gnu --enable-libstdcxx-debug --enable-libstdcxx-time=yes --with-default-libstdcxx-abi=new --enable-gnu-unique-object --enable-plugin --enable-default-pie --with-system-zlib --enable-libphobos-checking=release --with-target-system-zlib=auto --enable-objc-gc=auto --enable-secureplt --with-cpu=power8 --enable-targets=powerpcle-linux --disable-multilib --enable-multiarch --disable-werror --with-long-double-128 --enable-offload-targets=nvptx-none=/build/gcc-10-jGDcMs/gcc-10-10.2.1/debian/tmp-nvptx/usr --without-cuda-driver --enable-checking=release --build=powerpc64le-linux-gnu --host=powerpc64le-linux-gnu --target=powerpc64le-linux-gnu --with-build-config=bootstrap-lto-lean --enable-link-mutex Thread model: posix Supported LTO compression algorithms: zlib zstd gcc version 10.2.1 20210110 (Debian 10.2.1-6)
IBM long double ("double-double") is not an IEEE floating point format, so all these rules do not apply, but you are right it is surprising.
a = -2.5118280672773279e+307 c = 1.7976931348623157e+308 c is in particular the maximum normalized number, __DBL_MAX__, binary 0x7fefffffffffffff. 126 z = a + c; z = 1.546510328134583e+308 and that is finite. 144 q = a - z; But this is -inf. 145 zz = q + c + (a - (q + z)) + aa + cc; and everything goes downhill from there, as a, c, z, aa and cc are all finite, but q is -inf and a - (-inf + finite) is +inf and -inf + +inf is nan. So, the code: z = a + c; if (nonfinite (z)) { ... } else { q = a - z; would need another if (nonfinite (q)) check after computation of q (and I don't believe it could just compute q = a - z; right after z = a + c; and replace if (nonfinite (z)) with if (nonfinite (q)) because if a is +-inf and c is finite, z will be +-inf and q will be nan.
Yup, something like that. It should not have any infinities here afaics.