[X86] Invalid operand order for fp16 vector comparison

[azwolski]

Godbolt: https://godbolt.org/z/WhjWaWqM5

While compiling the following functions:

declare <16 x half> @test_call_16()

define <16 x i1> @test_cmp_v16half_olt(<16 x half> %rhs, <16 x i1> %mask) nounwind {
    %lhs = call <16 x half> @test_call_16()
    %comp = fcmp olt <16 x half> %lhs, %rhs
    %res = and <16 x i1> %comp, %mask
    ret <16 x i1> %res
}

define <16 x i1> @test_cmp_v16half_olt_rev(<16 x half> %rhs, <16 x i1> %mask) nounwind {
    %lhs = call <16 x half> @test_call_16()
    %comp = fcmp olt <16 x half> %rhs, %lhs
    %res = and <16 x i1> %comp, %mask
    ret <16 x i1> %res
}

Despite reversing the operands in the comparison (%lhs, %rhs vs %rhs, %lhs), both functions generate identical assembly

        sub     rsp, 56
        vmovups ymmword ptr [rsp + 16], ymm0
        vpsllw  xmm0, xmm1, 7
        vpmovb2m        k1, xmm0
        kmovw   word ptr [rsp + 14], k1
        call    test_call_16@PLT
        kmovw   k1, word ptr [rsp + 14]
        vcmpltph        k0 {k1}, ymm0, ymmword ptr [rsp + 16]
        vpmovm2b        xmm0, k0
        add     rsp, 56
        vzeroupper
        ret

The opcode remains vcmpltph. Changing the order of operands should reverse the comparison condition, but in this case, the generated assembly does not reflect that change.

I will submit a patch to address this issue.

[llvmbot]

@llvm/issue-subscribers-backend-x86

Author: None (azwolski)

Godbolt: https://godbolt.org/z/WhjWaWqM5

While compiling the following functions:

declare &lt;16 x half&gt; @<!-- -->test_call_16()

define &lt;16 x i1&gt; @<!-- -->test_cmp_v16half_olt(&lt;16 x half&gt; %rhs, &lt;16 x i1&gt; %mask) nounwind {
    %lhs = call &lt;16 x half&gt; @<!-- -->test_call_16()
    %comp = fcmp olt &lt;16 x half&gt; %lhs, %rhs
    %res = and &lt;16 x i1&gt; %comp, %mask
    ret &lt;16 x i1&gt; %res
}

define &lt;16 x i1&gt; @<!-- -->test_cmp_v16half_olt_rev(&lt;16 x half&gt; %rhs, &lt;16 x i1&gt; %mask) nounwind {
    %lhs = call &lt;16 x half&gt; @<!-- -->test_call_16()
    %comp = fcmp olt &lt;16 x half&gt; %rhs, %lhs
    %res = and &lt;16 x i1&gt; %comp, %mask
    ret &lt;16 x i1&gt; %res
}

Despite reversing the operands in the comparison (%lhs, %rhs vs %rhs, %lhs), both functions generate identical assembly

        sub     rsp, 56
        vmovups ymmword ptr [rsp + 16], ymm0
        vpsllw  xmm0, xmm1, 7
        vpmovb2m        k1, xmm0
        kmovw   word ptr [rsp + 14], k1
        call    test_call_16@<!-- -->PLT
        kmovw   k1, word ptr [rsp + 14]
        vcmpltph        k0 {k1}, ymm0, ymmword ptr [rsp + 16]
        vpmovm2b        xmm0, k0
        add     rsp, 56
        vzeroupper
        ret

The opcode remains vcmpltph. Changing the order of operands should reverse the comparison condition, but in this case, the generated assembly does not reflect that change.

I will submit a patch to address this issue.