Array `map` with unexpectedly bad codegen

[JakobDegen]

Note: I am aware of #91521 and related issues for this function, and believe that this has a separate cause

I tried this code:

pub fn foo<'a>(n: [*const u8; 2]) -> [&'a u8; 2] {
    n.map(|p| unsafe { &*p })
}

Assembly (rustc -Copt-level=3 -Cdebuginfo=0 --edition 2021, godbolt):

example::foo:
        mov     rax, rdi
        mov     rcx, qword ptr [rsi]
        mov     rdx, qword ptr [rsi + 8]
        test    rdx, rdx
        mov     rsi, rcx
        cmove   rsi, rdx
        test    rcx, rcx
        cmove   rsi, rcx
        mov     qword ptr [rdi], rsi
        mov     qword ptr [rdi + 8], rdx
        ret

The interesting thing to note here is the comparisons, which should definitely not be there. This problem significantly worsens when you increase the number of elements. I believe the source of these comparisons is that map is implemented via an iterator, which means that we build an Option<&u8> at some point. Still though, the LLVM IR:

LLVM IR

define void @_ZN7example3foo17h2b4bd706abe6f6b0E([2 x i8*]* noalias nocapture noundef writeonly sret([2 x i8*]) dereferenceable(16) %0, [2 x i8*]* noalias nocapture noundef readonly dereferenceable(16) %n) unnamed_addr #0 personality i32 (i32, i32, i64, %"unwind::libunwind::_Unwind_Exception"*, %"unwind::libunwind::_Unwind_Context"*)* @rust_eh_personality {
  %_2.sroa.0.0..sroa_cast = bitcast [2 x i8*]* %n to i64*
  %_2.sroa.0.0.copyload = load i64, i64* %_2.sroa.0.0..sroa_cast, align 8
  %_2.sroa.4.0..sroa_idx2 = getelementptr inbounds [2 x i8*], [2 x i8*]* %n, i64 0, i64 1
  %_2.sroa.4.0..sroa_cast = bitcast i8** %_2.sroa.4.0..sroa_idx2 to i64*
  %_2.sroa.4.0.copyload = load i64, i64* %_2.sroa.4.0..sroa_cast, align 8
  %.not.i.i.i.not.i = icmp eq i64 %_2.sroa.0.0.copyload, 0
  %.not.1.i.i.i.i = icmp eq i64 %_2.sroa.4.0.copyload, 0
  %or.cond.i = select i1 %.not.i.i.i.not.i, i1 true, i1 %.not.1.i.i.i.i
  %1 = inttoptr i64 %_2.sroa.4.0.copyload to i8*
  %2 = inttoptr i64 %_2.sroa.0.0.copyload to {}*
  %_2.sroa.6.0.i.i.i = select i1 %or.cond.i, i8* undef, i8* %1
  %_2.sroa.0.0.i.i.i = select i1 %or.cond.i, {}* null, {}* %2
  %3 = icmp ne {}* %_2.sroa.0.0.i.i.i, null
  tail call void @llvm.assume(i1 %3) #3
  %_4.sroa.0.0..sroa_cast.i.i = bitcast [2 x i8*]* %0 to {}**
  store {}* %_2.sroa.0.0.i.i.i, {}** %_4.sroa.0.0..sroa_cast.i.i, align 8, !alias.scope !2, !noalias !7
  %_4.sroa.4.0..sroa_idx4.i.i = getelementptr inbounds [2 x i8*], [2 x i8*]* %0, i64 0, i64 1
  store i8* %_2.sroa.6.0.i.i.i, i8** %_4.sroa.4.0..sroa_idx4.i.i, align 8, !alias.scope !2, !noalias !7
  ret void
}

Does look to have all the information necessary to continue optimizing.

Meta

rustc --version --verbose:

rustc 1.62.0-nightly (ec77f2524 2022-04-17)
binary: rustc
commit-hash: ec77f252434a532fdb5699ae4f21a3072d211edd
commit-date: 2022-04-17
host: x86_64-unknown-linux-gnu
release: 1.62.0-nightly
LLVM version: 14.0.0

@rustbot labels +A-llvm +I-slow

[JakobDegen]

@rustbot label +A-llvm +I-slow

[memoryruins]

nightly's codegen has improved

pub fn foo<'a>(n: [*const u8; 2]) -> [&'a u8; 2] {
    n.map(|p| unsafe { &*p })
}

example::foo:
        mov     rax, rdi
        movups  xmm0, xmmword ptr [rsi]
        movups  xmmword ptr [rdi], xmm0
        ret

rustc 1.64.0-nightly (06754d885 2022-07-08)

[JakobDegen]

Very nice