[AVX-512] Concatenating k-masks via shift+or should produce a `kunpckdq` instruction

[Validark]

export fn foo(a: @Vector(32, u8), b: @Vector(32, u8), c: @Vector(32, u8)) u64 {
    const b1: u32 = @bitCast(a == b);
    const b2: u32 = @bitCast(a == c);
    return (@as(u64, b1) << 32) | b2;
}

define dso_local i64 @foo(<32 x i8> %0, <32 x i8> %1, <32 x i8> %2) local_unnamed_addr {
Entry:
  %3 = icmp eq <32 x i8> %0, %1
  %4 = bitcast <32 x i1> %3 to i32
  %5 = icmp eq <32 x i8> %0, %2
  %6 = bitcast <32 x i1> %5 to i32
  %7 = zext i32 %4 to i64
  %8 = shl nuw i64 %7, 32
  %9 = zext i32 %6 to i64
  %10 = or disjoint i64 %8, %9
  ret i64 %10
}

This compiles like so: (Godbolt link)

foo:
        vpcmpeqb        k0, ymm0, ymm1
        kmovd   ecx, k0
        vpcmpeqb        k0, ymm0, ymm2
        kmovd   eax, k0
        shl     rcx, 32
        or      rax, rcx
        vzeroupper
        ret

I would think it should be:

foo:
        vpcmpeqb        k0, ymm0, ymm1
        vpcmpeqb        k1, ymm0, ymm2
        kunpckdq        k0, k1, k0
        kmovd   rax, k0
        vzeroupper
        ret

[RKSimon]

why aren't you able to use a shufflevector directly?

define i64 @foo(<32 x i8> %0, <32 x i8> %1, <32 x i8> %2) {
Entry:
  %3 = icmp eq <32 x i8> %0, %1
  %4 = icmp eq <32 x i8> %0, %2
  %5 = shufflevector <32 x i1> %3, <32 x i1> %4, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63>
  %6 = bitcast <64 x i1> %5 to i64
  ret i64 %6
}

[llvmbot]

@llvm/issue-subscribers-backend-x86

Author: Niles Salter (Validark)

```zig export fn foo(a: @Vector(32, u8), b: @Vector(32, u8), c: @Vector(32, u8)) u64 { const b1: u32 = @bitCast(a == b); const b2: u32 = @bitCast(a == c); return (@as(u64, b1) << 32) | b2; } ```

define dso_local i64 @<!-- -->foo(&lt;32 x i8&gt; %0, &lt;32 x i8&gt; %1, &lt;32 x i8&gt; %2) local_unnamed_addr {
Entry:
  %3 = icmp eq &lt;32 x i8&gt; %0, %1
  %4 = bitcast &lt;32 x i1&gt; %3 to i32
  %5 = icmp eq &lt;32 x i8&gt; %0, %2
  %6 = bitcast &lt;32 x i1&gt; %5 to i32
  %7 = zext i32 %4 to i64
  %8 = shl nuw i64 %7, 32
  %9 = zext i32 %6 to i64
  %10 = or disjoint i64 %8, %9
  ret i64 %10
}

This compiles like so: (Godbolt link)

foo:
        vpcmpeqb        k0, ymm0, ymm1
        kmovd   ecx, k0
        vpcmpeqb        k0, ymm0, ymm2
        kmovd   eax, k0
        shl     rcx, 32
        or      rax, rcx
        vzeroupper
        ret

I would think it should be:

foo:
        vpcmpeqb        k0, ymm0, ymm1
        vpcmpeqb        k1, ymm0, ymm2
        kunpckdq        k0, k0, k1
        kmovd   rax, k0
        vzeroupper
        ret

[Validark]

why aren't you able to use a shufflevector directly?

define i64 @foo(<32 x i8> %0, <32 x i8> %1, <32 x i8> %2) {
Entry:
  %3 = icmp eq <32 x i8> %0, %1
  %4 = icmp eq <32 x i8> %0, %2
  %5 = shufflevector <32 x i1> %3, <32 x i1> %4, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63>
  %6 = bitcast <64 x i1> %5 to i64
  ret i64 %6
}

I can! I just didn't think to write it that way. This gets me that LLVM code, and therefore the right assembly as well:

const std = @import("std");

export fn foo(a: @Vector(32, u8), b: @Vector(32, u8), c: @Vector(32, u8)) u64 {
    return @bitCast(std.simd.join(a == b, a == c));
}

Still, I think LLVM should identify opportunities like the one I mentioned, since I would normally write the code the way I originally stated.

[RKSimon]

This might be best handled in VectorCombine and not the backend