dead code emits redundant instructions

[MkazemAkhgary]

I was testing a few things and noticed this. the _x variable is not used here so it shouldn't be compiled. (target AVX2)

uint8 test(uniform uint8<TARGET_WIDTH> x)
{
    uint8 _x;
    unmasked
    {
        _x = x[programIndex]+1;
    }
    return  x[programIndex]+1;
}

test___unT_3C_8_3E_:                    # @test___unT_3C_8_3E_
        vmovmskps       eax, ymm1
        cmp     eax, 255
        je      .LBB1_2
        not     al
        test    al, -127
.LBB1_2:
        vpcmpeqd        xmm1, xmm1, xmm1
        vpsubb  xmm0, xmm0, xmm1
        vzeroupper
        ret

---

by removing the redundant part:

uint8 test(uniform uint8<TARGET_WIDTH> x)
{
    return  x[programIndex]+1;
}

test___unT_3C_8_3E_:                    # @test___unT_3C_8_3E_
        vpcmpeqd        xmm1, xmm1, xmm1
        vpsubb  xmm0, xmm0, xmm1
        ret 

I'm not sure what those extra instructions do, looks redundant to me.

[nurmukhametov]

Yes, indeed. I think that unmasked region doesn't matter here. The following code also produces similar redundancy

uint8 test(uniform uint8<TARGET_WIDTH> x)
{
    uint8 y = x[0];
    return x[programIndex]+1;
}

test___unT_3C_8_3E_:                    # @test___unT_3C_8_3E_
        vmovmskps       eax, ymm1
        cmp     eax, 255
        je      .LBB1_2
        not     al
        test    al, -127
.LBB1_2:
        vpcmpeqd        xmm1, xmm1, xmm1
        vpsubb  xmm0, xmm0, xmm1
        vzeroupper
        ret

[dbabokin]

I would experiment with SimplifyCFG and DCE to see if they are able to solve the problem and if they trigger in the right place in the opt pipeline.