Shortest form I’ve found so far. Not a good odd/even balance on the pipeline usage though.

for (j = 0; j < num_indexes; j += 24) { qword* lower_qword = (qword*)&indexes[j]; qword i0 = lower_qword[0]; qword i1 = lower_qword[1]; qword i2 = lower_qword[2]; qword i0r = si_rotqmbyi(i0, -2); qword i1r = si_rotqmbyi(i1, -2); qword i2r = si_rotqmbyi(i2, -2); qword v0 = si_mpya(i0, vertex_sizes, verticess); qword v1 = si_mpya(i1, vertex_sizes, verticess); qword v2 = si_mpya(i2, vertex_sizes, verticess); qword v0r = si_mpya(i0r, vertex_sizes, verticess); qword v1r = si_mpya(i1r, vertex_sizes, verticess); qword v2r = si_mpya(i2r, vertex_sizes, verticess); // Little constant reuse here :\ qword vs7 = si_shufb(v2r, v2, SHUFB4(c,D,d,0)); qword vs6 = si_shufb(v2r, v2, SHUFB4(B,b,C,0)); qword vs5 = si_shufb(v1, v2r, SHUFB4(D,a,0,0)); vs5 = si_shufb(vs5, v2, SHUFB4(A,B,a,0)); qword vs4 = si_shufb(v1, v1r, SHUFB4(c,C,d,0)); qword vs3 = si_shufb(v1, v1r, SHUFB4(A,b,B,0)); qword vs2 = si_shufb(v0r, v0, SHUFB4(D,d,0,0)); vs2 = si_shufb(vs2, v1r,SHUFB4(A,B,a,0)); qword vs1 = si_shufb(v0r, v0, SHUFB4(b,C,c,0)); qword vs0 = si_shufb(v0r, v0, SHUFB4(A,a,B,0)); switch(num_indexes - j) { default: func(vs7); case 21: func(vs6); case 18: func(vs5); case 15: func(vs4); case 12: func(vs3); case 9: func(vs2); case 6: func(vs1); case 3: func(vs0); } }