Array Repeat operator should not use a loop for small sizes

[programmerjake]

This likely caused poor assembly for Simd::splat in some circumstances:
https://rust-lang.zulipchat.com/#narrow/stream/257879-project-portable-simd/topic/Very.20bad.20.60u16x8.3A.3Asplat.60.20codegen.20on.20x86_64

I opened a PR for a workaround: rust-lang/portable-simd#284

Reduced reproducer:
https://godbolt.org/z/rhvEjz5MW

#![feature(portable_simd)]
use core::{mem, simd::u8x16};

pub fn f(v: u16, p: &mut [u8; 16]) {
    unsafe {
        let v: u8x16 = mem::transmute([v; 8]);
        *p = mem::transmute(v);
    }
}

.LCPI0_0:
        .short  65535
        .short  65535
        .short  65535
        .short  0
        .short  65535
        .short  65535
        .short  65535
        .short  65535
.LCPI0_1:
        .short  65535
        .short  65535
        .short  65535
        .short  65535
        .short  65535
        .short  0
        .short  65535
        .short  65535
.LCPI0_2:
        .short  65535
        .short  65535
        .short  65535
        .short  65535
        .short  65535
        .short  65535
        .short  65535
        .short  0
example::f:
        movd    xmm0, edi
        pshuflw xmm1, xmm0, 0
        pshufd  xmm2, xmm0, 0
        punpckldq       xmm1, xmm2
        movdqa  xmm2, xmm0
        psllq   xmm2, 48
        movdqa  xmm3, xmmword ptr [rip + .LCPI0_0]
        pand    xmm1, xmm3
        pandn   xmm3, xmm2
        por     xmm3, xmm1
        movdqa  xmm1, xmm3
        punpcklqdq      xmm1, xmm0
        movdqa  xmm2, xmm0
        pslldq  xmm2, 10
        movdqa  xmm4, xmmword ptr [rip + .LCPI0_1]
        pand    xmm1, xmm4
        pandn   xmm4, xmm2
        por     xmm4, xmm1
        movsd   xmm4, xmm0
        shufps  xmm3, xmm4, 36
        andps   xmm3, xmmword ptr [rip + .LCPI0_2]
        pslldq  xmm0, 14
        por     xmm0, xmm3
        movdqu  xmmword ptr [rsi], xmm0
        ret

[thomcc]

I wonder if this would be something that could/should be done as an MIR optimization?

[programmerjake]

i did some research: in mir, array repeat is a single opcode, translation to a loop occurs in the llvm and gcc backends:

https://github.com/rust-lang/rust/search?q=write_operand_repeatedly

[nikic]

Please provide a self-contained godbolt reproducer.

[programmerjake]

Please provide a self-contained godbolt reproducer.

I added it to the top comment.

[nikic]

Final IR: https://llvm.godbolt.org/z/5vq8scP6q Pretty wild shuffle sequence.

[nikic]

The pattern gets recognized in the backend: https://llvm.godbolt.org/z/zPojvdWzs But the variant with the bitcast at the start only works in LLVM 15, not LLVM 14.

I've opened llvm/llvm-project#55994 for the missing middle-end fold.

[okaneco]

It looks like the original reproducer is fixed. Current output, the LLVM-IR emitted is different now.

f:
  movd xmm0, edi
  pshuflw xmm0, xmm0, 0
  pshufd xmm0, xmm0, 0
  movdqu xmmword ptr [rsi], xmm0
  ret

define void @f(i16 noundef %v, ptr noalias nocapture noundef writeonly align 1 dereferenceable(16) %p) unnamed_addr #0 {
  %_4.sroa.0.0.vec.insert = insertelement <8 x i16> poison, i16 %v, i64 0
  %_4.sroa.0.14.vec.insert = shufflevector <8 x i16> %_4.sroa.0.0.vec.insert, <8 x i16> poison, <8 x i32> zeroinitializer
  store <8 x i16> %_4.sroa.0.14.vec.insert, ptr %p, align 1
  ret void
}

I think that means Simd::splat can be made const again unless there are other blockers.