These intrinsics use the __builtin_shuffle() function to extract the
low and high half, respectively, of a 128-bit NEON vector. Currently,
they're defined to use bitcasts to simplify the emitter, so we get code
like:
uint16x4_t vget_low_u32(uint16x8_t __a) {
return (uint32x2_t) __builtin_shufflevector((int64x2_t) __a,
(int64x2_t) __a,
0);
}
While this works, it results in those bitcasts going all the way through
to the IR, resulting in code like:
%1 = bitcast <8 x i16> %in to <2 x i64>
%2 = shufflevector <2 x i64> %1, <2 x i64> undef, <1 x i32>
%zeroinitializer
%3 = bitcast <1 x i64> %2 to <4 x i16>
We can instead easily perform the operation directly on the input vector
like:
uint16x4_t vget_low_u16(uint16x8_t __a) {
return __builtin_shufflevector(__a, __a, 0, 1, 2, 3);
}
Not only is that much easier to read on its own, it also results in
cleaner IR like:
%1 = shufflevector <8 x i16> %in, <8 x i16> undef,
<4 x i32> <i32 0, i32 1, i32 2, i32 3>
This is both easier to read and easier for the back end to reason
about effectively since the operation is obfuscating the source with
bitcasts.
rdar://13894163
llvm-svn: 181865
Jim Grosbach