Extended Description

Two versions of the same 'select based on comparison' pattern, one using generic vector types and the other using SSE intrinsics.

The presence of the bitcasts between SSE types prevents the recognition of the select.

#include <x86intrin.h>

__v16qi cmp_i8_sel_v16i8(__v16qi a, __v16qi b, __v16qi c, __v16qi d) {
__v16qi cc = _mm_cmpeq_epi8(a, b);
return (c & ~cc) | (d & cc);
}

__m128i cmp_i8_sel_m128i(__m128i a, __m128i b, __m128i c, __m128i d) {
__m128i cc = _mm_cmpeq_epi8(a, b);
return _mm_or_si128(_mm_andnot_si128(cc, c), _mm_and_si128(cc, d));
}

define <16 x i8> @​cmp_i8_sel_v16i8(<16 x i8>, <16 x i8>, <16 x i8>, <16 x i8>) {
%5 = icmp eq <16 x i8> %0, %1
%6 = select <16 x i1> %5, <16 x i8> %3, <16 x i8> %2
ret <16 x i8> %6
}

define <2 x i64> @​cmp_i8_sel_m128i(<2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>) {
%5 = bitcast <2 x i64> %0 to <16 x i8>
%6 = bitcast <2 x i64> %1 to <16 x i8>
%7 = icmp eq <16 x i8> %5, %6
%8 = sext <16 x i1> %7 to <16 x i8>
%9 = bitcast <16 x i8> %8 to <2 x i64>
%10 = xor <2 x i64> %9, <i64 -1, i64 -1>
%11 = and <2 x i64> %10, %2
%12 = and <2 x i64> %9, %3
%13 = or <2 x i64> %11, %12
ret <2 x i64> %13
}

Mind you, on X86 at least it doesn't affect final codegen:

cmp_i8_sel_v16i8:
vpcmpeqb %xmm1, %xmm0, %xmm0
vpblendvb %xmm0, %xmm3, %xmm2, %xmm0
retq

cmp_i8_sel_m128i:
vpcmpeqb %xmm1, %xmm0, %xmm0
vpblendvb %xmm0, %xmm3, %xmm2, %xmm0
retq