David noticed this failing on AIX. It doesn't work on any 32-bit BE, not so easily noticed because BE Linux will not run these tests (many of the tests require a Power8 although the test does not need it, and there aren't many BE Linux Power8 installations).
It turns out the 32-bit implementation of this function is for LE only. This patch fixes it. Tested on Linux, and by David on AIX. Installing. Segher 2019-03-23 Segher Boessenkool <seg...@kernel.crashing.org> * config/rs6000/xmmintrin.h (_mm_movemask_pi8): Implement for 32-bit big endian. --- gcc/config/rs6000/xmmintrin.h | 6 ++++++ 1 file changed, 6 insertions(+) diff --git a/gcc/config/rs6000/xmmintrin.h b/gcc/config/rs6000/xmmintrin.h index 71e4bd4..f9474b6 100644 --- a/gcc/config/rs6000/xmmintrin.h +++ b/gcc/config/rs6000/xmmintrin.h @@ -1586,9 +1586,15 @@ _mm_movemask_pi8 (__m64 __A) #endif return __builtin_bpermd (p, __A); #else +#ifdef __LITTLE_ENDIAN__ unsigned int mask = 0x20283038UL; unsigned int r1 = __builtin_bpermd (mask, __A) & 0xf; unsigned int r2 = __builtin_bpermd (mask, __A >> 32) & 0xf; +#else + unsigned int mask = 0x38302820UL; + unsigned int r1 = __builtin_bpermd (mask, __A >> 32) & 0xf; + unsigned int r2 = __builtin_bpermd (mask, __A) & 0xf; +#endif return (r2 << 4) | r1; #endif } -- 1.8.3.1