Repository navigation
Moving all elements out of a Vec generates worse assembly since 1.45 #78918
Description
I was looking for the best way to move all elements out of a Vec and returning them, basically what I'd write in C++ as:
struct A {
std::vector<int> moveOutOfVec() {
return std::move(myV); // this empties the vector
}
std::vector<int> myV;
};
std::vector<int> test(A& a) {
return a.moveOutOfVec();
}This code generates very good assembly with clang (link to Godbolt)
mov rax, rdi
movups xmm0, xmmword ptr [rsi]
movups xmmword ptr [rdi], xmm0
mov rcx, qword ptr [rsi + 16]
mov qword ptr [rdi + 16], rcx
xorps xmm0, xmm0
movups xmmword ptr [rsi], xmm0
mov qword ptr [rsi + 16], 0
retso I tried to do the equivalent with Rust.
Code
NOTE: I always used rustc -O for all the following tests.
I tried this code:
pub struct A {
v: Vec<i32>
}
impl A {
#[inline]
pub fn move_out_of_vec(&mut self) -> Vec<i32> {
let mut v = Vec::default();
std::mem::swap(&mut v, &mut self.v);
v
}
}
pub fn test(a: &mut A) -> Vec<i32> {
a.move_out_of_vec()
}I expected to see similar assembly as C++.
Instead, this was the output: (link to Godbolt)
example::foo:
sub rsp, 24
mov rax, rdi
mov qword ptr [rdi], 4
xorps xmm0, xmm0
movups xmmword ptr [rdi + 8], xmm0
mov rcx, qword ptr [rdi + 16]
mov qword ptr [rsp + 16], rcx
mov rcx, qword ptr [rdi]
mov qword ptr [rsp], rcx
mov rcx, qword ptr [rdi + 8]
mov qword ptr [rsp + 8], rcx
mov rcx, qword ptr [rsi + 16]
mov qword ptr [rdi + 16], rcx
movups xmm0, xmmword ptr [rsi]
movups xmmword ptr [rdi], xmm0
mov rcx, qword ptr [rsp + 16]
mov qword ptr [rsi + 16], rcx
mov rcx, qword ptr [rsp]
mov qword ptr [rsi], rcx
mov rcx, qword ptr [rsp + 8]
mov qword ptr [rsi + 8], rcx
add rsp, 24
retVersion it worked on
The same code in 1.44.0 generates much better assembly (only if you mark the method as inline): (link to Godbolt)
example::foo:
sub rsp, 24
mov rax, rdi
mov rcx, qword ptr [rsi]
movups xmm0, xmmword ptr [rsi + 8]
movaps xmmword ptr [rsp], xmm0
mov qword ptr [rsi], 4
xorps xmm0, xmm0
movups xmmword ptr [rsi + 8], xmm0
mov qword ptr [rdi], rcx
movaps xmm0, xmmword ptr [rsp]
movups xmmword ptr [rdi + 8], xmm0
add rsp, 24
retVersion with regression
From 1.45.0 until current nightly, the generated code is worse (inline doesn't affect the output).
Notes
I also tried using drain and split_off to achieve the same thing, but they both generate far worse assembly, with jumps and all.
Activity
silverweed commentedon Nov 10, 2020
@rustbot modify labels: regression-untriaged
silverweed commentedon Nov 10, 2020
Other than Godbolt, I also tested on my machine, running Windows 10. I ran
rustc --emit=asm --crate-type=lib -O -C llvm-args=-x86-asm-syntax=intel test.rsand these are the results (trimmed down to the relevant parts):rustc 1.46.0:
rustc 1.44.0:
SkiFire13 commentedon Nov 10, 2020
Curiously
std::mem::take(&mut self.v)compiles to the better assembly even on newer rustc versions, even though it is the same as yourmove_out_of_vecif manually inlined.camelid commentedon Nov 10, 2020
Assigning
P-mediumand removingI-prioritizeas discussed in the prioritization working group.eggyal commentedon Apr 6, 2021
Bisected this regression to nightly-2020-05-22; alas CI artefacts from back then are no longer available (to me?), but looking at the commits that made it into that nightly release (c7813ff to 9310e3b, inclusive) the one that most immediately jumps out as likely responsible is 82911b3, which was the merge of #67759 (Update to LLVM 10).