Override Iterator::advance(_back)_by for array::IntoIter

[scottmcm]

Because I happened to notice that nth is currently getting codegen'd as a loop even for Copy types: https://rust.godbolt.org/z/fPqv7Gvs7

LLVM before and after

Rust:

#[no_mangle]
pub fn array_intoiter_nth(it: &mut std::array::IntoIter<i32, 100>, n: usize) -> Option<i32> {
    it.nth(n)
}

Current nightly:

define { i32, i32 } @array_intoiter_nth(%"core::array::iter::IntoIter<i32, 100_usize>"* noalias nocapture align 8 dereferenceable(416) %it, i64 %n) unnamed_addr #0 personality i32 (i32, i32, i64, %"unwind::libunwind::_Unwind_Exception"*, %"unwind::libunwind::_Unwind_Context"*)* @rust_eh_personality !dbg !6 {
start:
  %_3.i.i.i4.i.i = getelementptr inbounds %"core::array::iter::IntoIter<i32, 100_usize>", %"core::array::iter::IntoIter<i32, 100_usize>"* %it, i64 0, i32 0, i32 0
  %_4.i.i.i5.i.i = getelementptr inbounds %"core::array::iter::IntoIter<i32, 100_usize>", %"core::array::iter::IntoIter<i32, 100_usize>"* %it, i64 0, i32 0, i32 1
  %_4.i.i.i.i.i.i = load i64, i64* %_4.i.i.i5.i.i, align 8, !alias.scope !10
  %.not.i.i = icmp eq i64 %n, 0, !dbg !15
  %_3.i.i.i.i.pre.i = load i64, i64* %_3.i.i.i4.i.i, align 8, !dbg !40, !alias.scope !41
  br i1 %.not.i.i, label %bb4.i, label %bb4.preheader.i.i, !dbg !42

bb4.preheader.i.i:                                ; preds = %start
  %umax.i = tail call i64 @llvm.umax.i64(i64 %_3.i.i.i.i.pre.i, i64 %_4.i.i.i.i.i.i) #3, !dbg !43
  %0 = sub i64 %umax.i, %_3.i.i.i.i.pre.i, !dbg !43
  br label %bb4.i.i, !dbg !43

bb4.i.i:                                          ; preds = %bb3.i.i.i.i, %bb4.preheader.i.i
  %_3.i.i.i.i.i.i = phi i64 [ %2, %bb3.i.i.i.i ], [ %_3.i.i.i.i.pre.i, %bb4.preheader.i.i ], !dbg !52
  %iter.sroa.0.016.i.i = phi i64 [ %1, %bb3.i.i.i.i ], [ 0, %bb4.preheader.i.i ]
  %1 = add nuw i64 %iter.sroa.0.016.i.i, 1, !dbg !54
  %exitcond.not.i = icmp eq i64 %iter.sroa.0.016.i.i, %0, !dbg !52
  br i1 %exitcond.not.i, label %core::iter::traits::iterator::Iterator::nth.exit, label %bb3.i.i.i.i, !dbg !43

bb3.i.i.i.i:                                      ; preds = %bb4.i.i
  %2 = add nuw i64 %_3.i.i.i.i.i.i, 1, !dbg !63
  store i64 %2, i64* %_3.i.i.i4.i.i, align 8, !dbg !66, !alias.scope !75
  %exitcond.not.i.i = icmp eq i64 %1, %n, !dbg !15
  br i1 %exitcond.not.i.i, label %bb4.i, label %bb4.i.i, !dbg !42

bb4.i:                                            ; preds = %bb3.i.i.i.i, %start
  %_3.i.i.i.i.i = phi i64 [ %_3.i.i.i.i.pre.i, %start ], [ %2, %bb3.i.i.i.i ], !dbg !84
  %3 = icmp ult i64 %_3.i.i.i.i.i, %_4.i.i.i.i.i.i, !dbg !84
  br i1 %3, label %bb3.i.i.i, label %core::iter::traits::iterator::Iterator::nth.exit, !dbg !89

bb3.i.i.i:                                        ; preds = %bb4.i
  %4 = add nuw i64 %_3.i.i.i.i.i, 1, !dbg !90
  store i64 %4, i64* %_3.i.i.i4.i.i, align 8, !dbg !93, !alias.scope !96
  %5 = getelementptr inbounds %"core::array::iter::IntoIter<i32, 100_usize>", %"core::array::iter::IntoIter<i32, 100_usize>"* %it, i64 0, i32 1, i64 %_3.i.i.i.i.i, !dbg !105
  %6 = load i32, i32* %5, align 4, !dbg !131, !alias.scope !141, !noalias !144
  br label %core::iter::traits::iterator::Iterator::nth.exit, !dbg !149

core::iter::traits::iterator::Iterator::nth.exit: ; preds = %bb4.i.i, %bb4.i, %bb3.i.i.i
  %.sroa.3.0.i = phi i32 [ %6, %bb3.i.i.i ], [ undef, %bb4.i ], [ undef, %bb4.i.i ], !dbg !40
  %.sroa.0.0.i = phi i32 [ 1, %bb3.i.i.i ], [ 0, %bb4.i ], [ 0, %bb4.i.i ], !dbg !40
  %7 = insertvalue { i32, i32 } undef, i32 %.sroa.0.0.i, 0, !dbg !150
  %8 = insertvalue { i32, i32 } %7, i32 %.sroa.3.0.i, 1, !dbg !150
  ret { i32, i32 } %8, !dbg !151
}

With this PR:

define { i32, i32 } @array_intoiter_nth(%"core::array::iter::IntoIter<i32, 100_usize>"* noalias nocapture align 8 dereferenceable(416) %it, i64 %n) unnamed_addr #0 personality i32 (...)* @__CxxFrameHandler3 {
start:
  %0 = getelementptr inbounds %"core::array::iter::IntoIter<i32, 100_usize>", %"core::array::iter::IntoIter<i32, 100_usize>"* %it, i64 0, i32 0, i32 1
  %_2.i.i.i.i = load i64, i64* %0, align 8, !alias.scope !6, !noalias !13
  %1 = getelementptr inbounds %"core::array::iter::IntoIter<i32, 100_usize>", %"core::array::iter::IntoIter<i32, 100_usize>"* %it, i64 0, i32 0, i32 0
  %_3.i.i.i.i = load i64, i64* %1, align 8, !alias.scope !16
  %2 = sub i64 %_2.i.i.i.i, %_3.i.i.i.i
  %3 = icmp ult i64 %2, %n
  %.0.sroa.speculated.i.i.i.i.i = select i1 %3, i64 %2, i64 %n
  %_10.i.i = add i64 %.0.sroa.speculated.i.i.i.i.i, %_3.i.i.i.i
  store i64 %_10.i.i, i64* %1, align 8, !alias.scope !16
  %.not.i = xor i1 %3, true
  %4 = icmp ult i64 %_10.i.i, %_2.i.i.i.i
  %or.cond.i = select i1 %.not.i, i1 %4, i1 false
  br i1 %or.cond.i, label %bb3.i.i.i, label %_ZN4core4iter6traits8iterator8Iterator3nth17hcbc727011e9e2a3bE.exit

bb3.i.i.i:                                        ; preds = %start
  %5 = add nuw i64 %_10.i.i, 1
  store i64 %5, i64* %1, align 8, !alias.scope !17
  %6 = getelementptr inbounds %"core::array::iter::IntoIter<i32, 100_usize>", %"core::array::iter::IntoIter<i32, 100_usize>"* %it, i64 0, i32 1, i64 %_10.i.i
  %7 = load i32, i32* %6, align 4, !alias.scope !26, !noalias !29
  br label %_ZN4core4iter6traits8iterator8Iterator3nth17hcbc727011e9e2a3bE.exit

_ZN4core4iter6traits8iterator8Iterator3nth17hcbc727011e9e2a3bE.exit: ; preds = %start, %bb3.i.i.i
  %.sroa.3.0.i = phi i32 [ undef, %start ], [ %7, %bb3.i.i.i ]
  %.sroa.0.0.i = phi i32 [ 0, %start ], [ 1, %bb3.i.i.i ]
  %8 = insertvalue { i32, i32 } undef, i32 %.sroa.0.0.i, 0
  %9 = insertvalue { i32, i32 } %8, i32 %.sroa.3.0.i, 1
  ret { i32, i32 } %9
}

[rust-highfive]

r? @Mark-Simulacrum

(rust-highfive has picked a reviewer for you, use r? to override)

[the8472]

Tangent: These dances seem a bit unergonomic and repeated in several places.

ptr::drop_in_place can take a *mut [T] (fat pointer), so for unsafe code it would be great if we had something like [T].as_slice_ptr().get_unchecked(range).cast() then we wouldn't have to refer to use those unwieldy associated methods on MaybeUninit since we're operating on pointers anyway. Generally there's a lack of helpers to work with *mut/const [T].

[scottmcm]

Hmm, there's assume_init_drop but no slice version of that, so I guess there could be assume_init_drop_range like there's slice_assume_init_mut and such -- one could always pass .. to drop everything easily.

Don't think I'll do that in this PR, though.

[Mark-Simulacrum]

@bors r+ rollup=never

[bors]

📌 Commit eb846db has been approved by Mark-Simulacrum

[bors]

⌛ Testing commit eb846db with merge ce0f7ba...

[bors]

☀️ Test successful - checks-actions
Approved by: Mark-Simulacrum
Pushing ce0f7ba to master...

[rust-timer]

Finished benchmarking commit (ce0f7ba): comparison url.

Summary: This benchmark run did not return any relevant changes.

If you disagree with this performance assessment, please file an issue in rust-lang/rustc-perf.

@rustbot label: -perf-regression