this optimization doesn't look good on SIMD

[MkazemAkhgary]

this function with AVX2

unmasked varying int32 RoundIndex(varying int32 Index)
{
    return (Index >> 2) << 2;
}

compiles to

.LCPI1_0:
        .long   4294967292                      # 0xfffffffc
vbroadcastss    ymm1, dword ptr [rip + .LCPI1_0] # ymm1 = [4294967292,4294967292,4294967292,4294967292,4294967292,4294967292,4294967292,4294967292]
vandps  ymm0, ymm0, ymm1
ret

wouldn't it be better to just do this?

vpsrad  ymm0, ymm0, 2
vpslld  ymm0, ymm0, 2
ret

Using shifts avoids loading from memory and results in less register pressure. The second approach should complete in about 2 cycles, whereas the broadcast alone takes about 5 to 8 cycles (if I'm not mistaken). am I missing something?

---

P.S. there is also a vpand instruction with 3rd operand as memory location. I'm not sure why ispc doesn't use this. (it does use vandps zmm, zmm, m512 for avx512)

If there are additional shift operations similar to this, or if the value 0xfffffffc is already in a register, it would be more efficient to use AND instead.

[MkazemAkhgary]

This issue could be generalized into use of broadcasting constants. I think some times it's better to just do simple computations rather than broadcasting a constant from memory.