https://gcc.gnu.org/bugzilla/show_bug.cgi?id=88698
Alexander Monakov <amonakov at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
CC| |amonakov at gcc dot gnu.org
--- Comment #6 from Alexander Monakov <amonakov at gcc dot gnu.org> ---
My recommendation is to use a union like below; this allows writing code using
both generic vectors and intrinsics without casts, and having each operation
show exactly what lane types it operates on:
typedef unsigned char u8v __attribute__((vector_size(16)));
typedef unsigned short u16v __attribute__((vector_size(16)));
typedef unsigned int u32v __attribute__((vector_size(16)));
typedef union {
u8v u8;
u16v u16;
u32v u32;
__m128i m;
} uv;
Example use:
uv x, t, lo_nib, hi_nib;
memcpy(&x, ptr, sizeof x);
t.u32 = x.u32 >> 4;
lo_nib.u8 = x.u8 & 15;
hi_nib.u8 = t.u8 & 15;
lo_nib.m = _mm_shuffle_epi8(lut.m, lo_nib.m);
hi_nib.m = _mm_shuffle_epi8(lut.m, hi_nib.m);
This also allows writing 256-bit and 128-bit versions together when appropriate
(with help of extra macros for using the right intrinsic function).
Would you like to see the documentation mention this pattern?
