Missed optimization when computing `max` on an iterator

[xTachyon]

From a conversation on the Discord Rust Community Server:

Impl 1:

pub fn f1(x: i32, y: i32, z: i32) -> i32 {
    [x, y, z].into_iter().max().unwrap()
}

Impl 2:

pub fn f2(x: i32, y: i32, z: i32) -> i32 {
    *[x, y, z].iter().max().unwrap()
}

Godbolt: https://godbolt.org/z/G37cqbfdE

These functions should be equivalent, but the first function generates way more code than it needs to. I expected the code to be equivalent.

There is another interesting combination: for u64, the code on 1.68 looks identical, but on nightly it's even bigger.

[zachs18]

This regressed in 1.65.0. It used to give the same almost the same asm for both functions in 1.64.0.
Possibly notable: Rust 1.65.0 was when Rust upgraded to LLVM 15.

(godbolt link):

pub fn f1(x: i32, y: i32, z: i32) -> i32 {
    [x, y, z].into_iter().max().unwrap()
}

pub fn f2(x: i32, y: i32, z: i32) -> i32 {
    *[x, y, z].iter().max().unwrap()
}

1.64.0 output:

example::f1:
  mov eax, esi
  cmp esi, edi
  cmovl eax, edi
  cmp eax, edx
  cmovle eax, edx
  ret

example::f2:
  mov eax, esi
  cmp edi, esi
  cmovg eax, edi
  cmp eax, edx
  cmovle eax, edx
  ret

1.65.0 through 1.69.0 output:

example::f1:
  sub rsp, 32
  mov dword ptr [rsp + 20], esi
  mov dword ptr [rsp + 24], edx
  movd xmm0, edi
  pshufd xmm0, xmm0, 80
  movq xmm1, qword ptr [rsp + 20]
  movdqa xmm2, xmm0
  pcmpgtd xmm2, xmm1
  pand xmm0, xmm2
  pandn xmm2, xmm1
  por xmm2, xmm0
  pshufd xmm0, xmm2, 85
  movdqa xmm1, xmm2
  pcmpgtd xmm1, xmm0
  pand xmm2, xmm1
  pandn xmm1, xmm0
  por xmm1, xmm2
  movd eax, xmm1
  add rsp, 32
  ret

example::f2:
  mov eax, esi
  cmp edi, esi
  cmovg eax, edi
  cmp eax, edx
  cmovle eax, edx
  ret

1.70.0 and later output:

example::f1:
        movd    xmm0, edx
        movd    xmm1, esi
        punpckldq       xmm1, xmm0
        movd    xmm0, edi
        pshufd  xmm0, xmm0, 80
        movdqa  xmm2, xmm0
        pcmpgtd xmm2, xmm1
        pand    xmm0, xmm2
        pandn   xmm2, xmm1
        por     xmm2, xmm0
        pshufd  xmm0, xmm2, 85
        movdqa  xmm1, xmm2
        pcmpgtd xmm1, xmm0
        pand    xmm2, xmm1
        pandn   xmm1, xmm0
        por     xmm1, xmm2
        movd    eax, xmm1
        ret

example::f2:
        mov     eax, esi
        cmp     edi, esi
        cmovg   eax, edi
        cmp     eax, edx
        cmovle  eax, edx
        ret

[SmartBoy84]

Could this be related?

[nikic]

The basic problem here is that by the time LoopFullUnroll runs we do not know the trip count yet, so we end up loop-vectorizing it, which of course is silly for a loop with three iterations.

Edit: There's a chance that https://reviews.llvm.org/D154157 might prevent vectorization here, depending on the exact costs involved.