[x86-64] Broadcasting an element of a vector should not use `vpermb` or `vpshufb`

[Validark]

I have code like so:

export fn foo(x: @Vector(8, u8)) @Vector(64, u8) {
    return @splat(x[1]);
}

Here is the LLVM version:

define dso_local <64 x i8> @foo(<8 x i8> %0) local_unnamed_addr {
Entry:
  %1 = shufflevector <8 x i8> %0, <8 x i8> poison, <64 x i32> <i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1>
  ret <64 x i8> %1
}

Here is how it lowers on Zen 5:

.LCPI0_1:
        .byte   1
foo:
        vpbroadcastb    zmm1, byte ptr [rip + .LCPI0_1]
        vpermb  zmm0, zmm1, zmm0
        ret

Here is how I think it should lower:

foo:
        vpsrlq  xmm0, xmm0, 8
        vpbroadcastb    zmm0, xmm0
        ret

---

Same applies to broadcasting into an xmm0:

.LCPI0_0:
        .zero   16,1
foo:
        vpshufb xmm0, xmm0, xmmword ptr [rip + .LCPI0_0]
        ret

I would much rather avoid the trip to memory:

foo:
        vpsrlq  xmm0, xmm0, 8
        vpbroadcastb    xmm0, xmm0
        ret

[llvmbot]

@llvm/issue-subscribers-backend-x86

Author: Niles Salter (Validark)

I have code like so:

export fn foo(x: @<!-- -->Vector(8, u8)) @<!-- -->Vector(64, u8) {
    return @<!-- -->splat(x[1]);
}

Here is the LLVM version:

define dso_local &lt;64 x i8&gt; @<!-- -->foo(&lt;8 x i8&gt; %0) local_unnamed_addr {
Entry:
  %1 = shufflevector &lt;8 x i8&gt; %0, &lt;8 x i8&gt; poison, &lt;64 x i32&gt; &lt;i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1&gt;
  ret &lt;64 x i8&gt; %1
}

Here is how it lowers on Zen 5:

.LCPI0_1:
        .byte   1
foo:
        vpbroadcastb    zmm1, byte ptr [rip + .LCPI0_1]
        vpermb  zmm0, zmm1, zmm0
        ret

Here is how I think it should lower:

foo:
        vpsrlq  xmm0, xmm0, 8
        vpbroadcastb    zmm0, xmm0
        ret

---

Same applies to broadcasting into an xmm0:

.LCPI0_0:
        .zero   16,1
foo:
        vpshufb xmm0, xmm0, xmmword ptr [rip + .LCPI0_0]
        ret

I would much rather avoid the trip to memory:

foo:
        vpsrlq  xmm0, xmm0, 8
        vpbroadcastb    xmm0, xmm0
        ret