llvm is generating different move code for interior moves, and moving an argument

[erickt]

I'm not quite sure if this has a performance impact, but I noticed that rust seems to be generating different code for moves inside a function and a function argument. Here's an example:

struct Foo {
    x: Vec<int>,
}

#[inline(never)]
#[no_mangle]
fn interior(x: Vec<int>) -> Vec<int> {
    let Foo { x } = Foo { x: x };
    x
}

#[inline(never)]
#[no_mangle]
fn exterior(x: Vec<int>) -> Vec<int> {
    x
}

fn main() {
    let x = interior(Vec::new());
    println!("{}", x);

    let x = exterior(Vec::new());
    println!("{}", x);
}

At opt-level=3, llvm generates:

; Function Attrs: noinline uwtable
define internal fastcc void @interior(%"struct.collections::vec::Vec<[int]>[#6]"* noalias nocapture nonnull sret, %"struct.collections::vec::Vec<[int]>[#6]"* noalias nocapture nonnull) unnamed_addr #0 {
"_ZN32collections..vec..Vec$LT$int$GT$39glue_drop.$x22glue_drop$x22$LP$1381$RP$17h0c4580be04eb3673E.exit4":
  %2 = bitcast %"struct.collections::vec::Vec<[int]>[#6]"* %1 to i8*
  %3 = bitcast %"struct.collections::vec::Vec<[int]>[#6]"* %1 to <2 x i64>*
  %4 = load <2 x i64>* %3, align 8
  %.sroa.7.0..sroa_idx20 = getelementptr inbounds %"struct.collections::vec::Vec<[int]>[#6]"* %1, i64 0, i32 2
  %.sroa.7.0.copyload = load i64** %.sroa.7.0..sroa_idx20, align 8
  %5 = bitcast %"struct.collections::vec::Vec<[int]>[#6]"* %0 to <2 x i64>*
  store <2 x i64> %4, <2 x i64>* %5, align 8
  %x.sroa.6.0..sroa_idx13 = getelementptr inbounds %"struct.collections::vec::Vec<[int]>[#6]"* %0, i64 0, i32 2
  store i64* %.sroa.7.0.copyload, i64** %x.sroa.6.0..sroa_idx13, align 8
  tail call void @llvm.lifetime.end(i64 24, i8* %2)
  ret void
}

...

; Function Attrs: noinline uwtable
define internal fastcc void @exterior(%"struct.collections::vec::Vec<[int]>[#6]"* noalias nocapture nonnull sret, %"struct.collections::vec::Vec<[int]>[#6]"* noalias nocapture nonnull) unnamed_addr #0 {
"_ZN32collections..vec..Vec$LT$int$GT$39glue_drop.$x22glue_drop$x22$LP$1381$RP$17h0c4580be04eb3673E.exit":
  %2 = bitcast %"struct.collections::vec::Vec<[int]>[#6]"* %1 to i8*
  %3 = bitcast %"struct.collections::vec::Vec<[int]>[#6]"* %0 to i8*
  tail call void @llvm.memcpy.p0i8.p0i8.i64(i8* %3, i8* %2, i64 24, i32 8, i1 false)
  tail call void @llvm.lifetime.end(i64 24, i8* %2)
  ret void
}

cc: @pcwalton.

[dotdash]

This happens because in the interior case there's a call to the drop glue which triggers SROA to split the aggregate value. It's not reassembled to a memcpy, but the asm that's generated for both variants is pretty much the same. The differences in the IR should disappear once zeroing is gone.

[steveklabnik]

Triage: this still reproduces.

[arielb1]

triage: memcpy is generated for both cases.

[arielb1]

triage: memcpy is generated in both cases.

[frewsxcv]

Fixed by #40296 ?