I believe that MaybeUninit::uninit currently zeroes memory rather than doing nothing. I was wondering if there is a good reason for this, or whether it is simply that the “do nothing” optimisation has not yet been implemented.
I cannot think of a good reason, but I could very well be missing something. Any ideas?
Well, your belief may just be wrong?
use std::mem::MaybeUninit;
#[unsafe(no_mangle)]
fn test1(x: &mut MaybeUninit<u8>) {
*x = MaybeUninit::new(0);
}
#[unsafe(no_mangle)]
fn test2(x: &mut MaybeUninit<u8>) {
*x = MaybeUninit::uninit();
}
test1:
movb $0, (%rdi)
retq
test2:
retq
Given this observation, I’ll just skip all your subsequent questions in the original post.
If you have some evidence that MaybeUninit::uninit does “zero memory” in certain situations (obviously the above example seems to be not one of them), you should share your code for such a concrete situation 
I will try to provide a simple example, but I encountered it in my own code.
Edit: well I have not been able to make a simple example yet, but at least the answer so far seems to be that it doesn’t need to zero the memory ( which is what I originally expected ).
Ok, here is a non-simple example, the program:
use pstd::collections::BTreeSet;
fn main() {
let set = BTreeSet::from([2, 4, 6]);
let mut x = 1234;
for _rep in 0..100_000_000 {
for e in set.iter() { x += e; }
}
println!("x={}", x);
}
```
The assembler for main:
.section .text.btest::main,"ax",@progbits
.hidden btest::main
.globl btest::main
.p2align 4
.type btest::main,@function
btest::main:
.cfi_startproc
.cfi_personality 155, DW.ref.rust_eh_personality
.cfi_lsda 27, .Lexception3
push rbp
.cfi_def_cfa_offset 16
push r15
.cfi_def_cfa_offset 24
push r14
.cfi_def_cfa_offset 32
push r13
.cfi_def_cfa_offset 40
push r12
.cfi_def_cfa_offset 48
push rbx
.cfi_def_cfa_offset 56
sub rsp, 1480
.cfi_def_cfa_offset 1536
.cfi_offset rbx, -56
.cfi_offset r12, -48
.cfi_offset r13, -40
.cfi_offset r14, -32
.cfi_offset r15, -24
.cfi_offset rbp, -16
mov qword ptr [rsp + 56], 0
lea r12, [rsp + 64]
mov qword ptr [rsp + 64], 1
mov dword ptr [rsp + 72], 0
lea r15, [rsp + 80]
mov word ptr [rsp + 84], 0
mov dword ptr [rsp + 80], 1048640
mov byte ptr [rsp + 808], 1
mov qword ptr [rsp + 768], 0
mov qword ptr [rsp + 800], r15
lea rdx, [rsp + 768]
mov rdi, r12
mov esi, 2
call <pstd::collections::btree_map::Tree<i32, ()>>::insert::<pstd::collections::btree_map::CustomTuning>
cmp dword ptr [rsp + 768], 1
jne .LBB10_30
mov ebx, dword ptr [rsp + 776]
mov rbp, qword ptr [rsp + 784]
mov r13, qword ptr [rsp + 792]
call qword ptr [rip + __rustc::__rust_no_alloc_shim_is_unstable_v2@GOTPCREL]
mov edi, 32
mov esi, 8
call qword ptr [rip + __rustc::__rust_alloc@GOTPCREL]
test rax, rax
je .LBB10_3
mov r14, rax
mov qword ptr [rax], 1
mov dword ptr [rax + 8], 0
mov qword ptr [rax + 16], 8
mov dword ptr [rax + 24], 0
mov esi, 1
mov rdi, rax
mov rdx, r15
call <pstd::collections::btree_map::vecs::PairVec<i32, ()>>::set_alloc::<pstd::collections::btree_map::CustomTuning>
movzx eax, word ptr [r14 + 8]
cmp ax, word ptr [r14 + 10]
jae .LBB10_13
mov rdi, r14
add rdi, 16
mov rcx, qword ptr [r14]
mov dword ptr [rcx + 4*rax], ebx
inc eax
mov word ptr [r14 + 8], ax
mov esi, 2
mov rdx, r15
call <pstd::collections::btree_map::vecs::ShortVec<pstd::collections::btree_map::Tree<i32, ()>>>::set_alloc::<pstd::collections::btree_map::CustomTuning>
mov rdi, r12
mov r12, rbp
mov rbp, r13
mov rbx, qword ptr [rsp + 64]
mov r13, qword ptr [rsp + 72]
mov qword ptr [rsp + 64], 1
mov dword ptr [rsp + 72], 0
movzx eax, word ptr [r14 + 24]
movzx edx, word ptr [r14 + 26]
cmp ax, dx
jae .LBB10_16
mov rcx, qword ptr [r14 + 16]
mov esi, eax
shl esi, 4
mov qword ptr [rcx + rsi], rbx
mov qword ptr [rcx + rsi + 8], r13
lea rsi, [rax + 1]
mov word ptr [r14 + 24], si
cmp si, dx
jae .LBB10_21
shl esi, 4
mov qword ptr [rcx + rsi], r12
mov qword ptr [rcx + rsi + 8], rbp
add eax, 2
mov word ptr [r14 + 24], ax
mov qword ptr [rsp + 64], 0
mov qword ptr [rsp + 72], r14
mov r12, rdi
.LBB10_30:
cmp byte ptr [rsp + 808], 0
jne .LBB10_32
inc qword ptr [rsp + 56]
.LBB10_32:
mov byte ptr [rsp + 808], 1
mov qword ptr [rsp + 768], 0
mov qword ptr [rsp + 800], r15
lea rdx, [rsp + 768]
mov rdi, r12
mov esi, 4
call <pstd::collections::btree_map::Tree<i32, ()>>::insert::<pstd::collections::btree_map::CustomTuning>
cmp dword ptr [rsp + 768], 1
jne .LBB10_41
mov ebx, dword ptr [rsp + 776]
mov rbp, qword ptr [rsp + 784]
mov r13, qword ptr [rsp + 792]
call qword ptr [rip + __rustc::__rust_no_alloc_shim_is_unstable_v2@GOTPCREL]
mov edi, 32
mov esi, 8
call qword ptr [rip + __rustc::__rust_alloc@GOTPCREL]
test rax, rax
je .LBB10_3
mov r14, rax
mov qword ptr [rax], 1
mov dword ptr [rax + 8], 0
mov qword ptr [rax + 16], 8
mov dword ptr [rax + 24], 0
mov esi, 1
mov rdi, rax
mov rdx, r15
call <pstd::collections::btree_map::vecs::PairVec<i32, ()>>::set_alloc::<pstd::collections::btree_map::CustomTuning>
movzx eax, word ptr [r14 + 8]
cmp ax, word ptr [r14 + 10]
jae .LBB10_13
mov rdi, r14
add rdi, 16
mov rcx, qword ptr [r14]
mov dword ptr [rcx + 4*rax], ebx
inc eax
mov word ptr [r14 + 8], ax
mov esi, 2
mov rdx, r15
call <pstd::collections::btree_map::vecs::ShortVec<pstd::collections::btree_map::Tree<i32, ()>>>::set_alloc::<pstd::collections::btree_map::CustomTuning>
mov rdi, r12
mov r12, rbp
mov rbp, r13
mov rbx, qword ptr [rsp + 64]
mov r13, qword ptr [rsp + 72]
mov qword ptr [rsp + 64], 1
mov dword ptr [rsp + 72], 0
movzx eax, word ptr [r14 + 24]
movzx edx, word ptr [r14 + 26]
cmp ax, dx
jae .LBB10_16
mov rcx, qword ptr [r14 + 16]
mov esi, eax
shl esi, 4
mov qword ptr [rcx + rsi], rbx
mov qword ptr [rcx + rsi + 8], r13
lea rsi, [rax + 1]
mov word ptr [r14 + 24], si
cmp si, dx
jae .LBB10_21
shl esi, 4
mov qword ptr [rcx + rsi], r12
mov qword ptr [rcx + rsi + 8], rbp
add eax, 2
mov word ptr [r14 + 24], ax
mov qword ptr [rsp + 64], 0
mov qword ptr [rsp + 72], r14
mov r12, rdi
.LBB10_41:
cmp byte ptr [rsp + 808], 0
jne .LBB10_43
inc qword ptr [rsp + 56]
.LBB10_43:
mov byte ptr [rsp + 808], 1
mov qword ptr [rsp + 768], 0
mov qword ptr [rsp + 800], r15
lea rdx, [rsp + 768]
mov rdi, r12
mov esi, 6
call <pstd::collections::btree_map::Tree<i32, ()>>::insert::<pstd::collections::btree_map::CustomTuning>
cmp dword ptr [rsp + 768], 1
jne .LBB10_52
mov ebx, dword ptr [rsp + 776]
mov rbp, qword ptr [rsp + 784]
mov r13, qword ptr [rsp + 792]
call qword ptr [rip + __rustc::__rust_no_alloc_shim_is_unstable_v2@GOTPCREL]
mov edi, 32
mov esi, 8
call qword ptr [rip + __rustc::__rust_alloc@GOTPCREL]
test rax, rax
je .LBB10_3
mov r14, rax
mov qword ptr [rax], 1
mov dword ptr [rax + 8], 0
mov qword ptr [rax + 16], 8
mov dword ptr [rax + 24], 0
mov esi, 1
mov rdi, rax
mov rdx, r15
call <pstd::collections::btree_map::vecs::PairVec<i32, ()>>::set_alloc::<pstd::collections::btree_map::CustomTuning>
movzx eax, word ptr [r14 + 8]
cmp ax, word ptr [r14 + 10]
jae .LBB10_13
mov rdi, r14
add rdi, 16
mov rcx, qword ptr [r14]
mov dword ptr [rcx + 4*rax], ebx
inc eax
mov word ptr [r14 + 8], ax
mov esi, 2
mov rdx, r15
call <pstd::collections::btree_map::vecs::ShortVec<pstd::collections::btree_map::Tree<i32, ()>>>::set_alloc::<pstd::collections::btree_map::CustomTuning>
mov r12, rbp
mov rbp, r13
mov rbx, qword ptr [rsp + 64]
mov r13, qword ptr [rsp + 72]
mov qword ptr [rsp + 64], 1
mov dword ptr [rsp + 72], 0
movzx eax, word ptr [r14 + 24]
movzx edx, word ptr [r14 + 26]
cmp ax, dx
jae .LBB10_16
mov rcx, qword ptr [r14 + 16]
mov esi, eax
shl esi, 4
mov qword ptr [rcx + rsi], rbx
mov qword ptr [rcx + rsi + 8], r13
lea rsi, [rax + 1]
mov word ptr [r14 + 24], si
cmp si, dx
jae .LBB10_21
shl esi, 4
mov qword ptr [rcx + rsi], r12
mov qword ptr [rcx + rsi + 8], rbp
add eax, 2
mov word ptr [r14 + 24], ax
mov qword ptr [rsp + 64], 0
mov qword ptr [rsp + 72], r14
.LBB10_52:
cmp byte ptr [rsp + 808], 0
jne .LBB10_54
inc qword ptr [rsp + 56]
.LBB10_54:
movups xmm0, xmmword ptr [rsp + 56]
movups xmm1, xmmword ptr [rsp + 72]
movaps xmmword ptr [rsp + 16], xmm0
movaps xmmword ptr [rsp + 32], xmm1
mov dword ptr [rsp + 12], 1234
mov r12, qword ptr [rsp + 16]
xor ebp, ebp
mov r13d, 1234
lea r14, [rsp + 768]
lea r15, [rsp + 56]
mov rbx, qword ptr [rip + memcpy@GOTPCREL]
jmp .LBB10_55
.p2align 4
.LBB10_70:
inc ebp
cmp ebp, 100000000
je .LBB10_71
.LBB10_55:
mov edx, 704
mov rdi, r14
xor esi, esi
call qword ptr [rip + memset@GOTPCREL]
mov qword ptr [rsp + 1472], r12
mov rdi, r14
lea rsi, [rsp + 24]
mov edx, 1
call <pstd::collections::btree_map::Range<i32, ()>>::push_tree
mov edx, 712
mov rdi, r15
mov rsi, r14
call rbx
mov rax, qword ptr [rsp + 760]
test rax, rax
je .LBB10_70
.p2align 4
.LBB10_58:
dec rax
mov qword ptr [rsp + 760], rax
mov rcx, qword ptr [rsp + 64]
cmp rcx, qword ptr [rsp + 72]
je .LBB10_69
mov rax, qword ptr [rsp + 56]
movzx edx, word ptr [rax + 8]
cmp rcx, rdx
jae .LBB10_60
mov rax, qword ptr [rax]
lea rax, [rax + 4*rcx]
inc rcx
mov qword ptr [rsp + 64], rcx
.LBB10_77:
test rax, rax
je .LBB10_70
add r13d, dword ptr [rax]
mov dword ptr [rsp + 12], r13d
mov rax, qword ptr [rsp + 760]
test rax, rax
jne .LBB10_58
jmp .LBB10_70
.LBB10_69:
mov rdi, r15
call <pstd::collections::btree_map::Range<i32, ()>>::next_cold
jmp .LBB10_77
.LBB10_71:
lea rax, [rsp + 12]
mov qword ptr [rsp + 768], rax
mov rax, qword ptr [rip + <i32 as core::fmt::Display>::fmt@GOTPCREL]
mov qword ptr [rsp + 776], rax
lea rdi, [rip + .Lanon.4e23a766b1161c58f315fd4481a40bbc.31]
lea rsi, [rsp + 768]
call qword ptr [rip + std::io::stdio::_print@GOTPCREL]
lea rsi, [rsp + 40]
lea rdi, [rsp + 24]
call <pstd::collections::btree_map::Tree<i32, ()>>::dealloc::<pstd::collections::btree_map::CustomTuning>
cmp qword ptr [rsp + 24], 0
jne .LBB10_75
mov rdi, qword ptr [rsp + 32]
mov esi, 32
mov edx, 8
call qword ptr [rip + __rustc::__rust_dealloc@GOTPCREL]
.LBB10_75:
add rsp, 1480
.cfi_def_cfa_offset 56
pop rbx
.cfi_def_cfa_offset 48
pop r12
.cfi_def_cfa_offset 40
pop r13
.cfi_def_cfa_offset 32
pop r14
.cfi_def_cfa_offset 24
pop r15
.cfi_def_cfa_offset 16
pop rbp
.cfi_def_cfa_offset 8
ret
.LBB10_60:
.cfi_def_cfa_offset 1536
lea rdi, [rip + .Lanon.4e23a766b1161c58f315fd4481a40bbc.26]
lea rdx, [rip + .Lanon.4e23a766b1161c58f315fd4481a40bbc.36]
mov esi, 32
call qword ptr [rip + core::panicking::panic@GOTPCREL]
jmp .LBB10_4
.LBB10_3:
mov edi, 8
mov esi, 32
call qword ptr [rip + alloc::alloc::handle_alloc_error@GOTPCREL]
jmp .LBB10_4
.LBB10_13:
lea rdi, [rip + .Lanon.4e23a766b1161c58f315fd4481a40bbc.38]
lea rdx, [rip + .Lanon.4e23a766b1161c58f315fd4481a40bbc.39]
mov esi, 39
call qword ptr [rip + core::panicking::panic@GOTPCREL]
jmp .LBB10_4
.LBB10_16:
lea rdi, [rip + .Lanon.4e23a766b1161c58f315fd4481a40bbc.38]
lea rdx, [rip + .Lanon.4e23a766b1161c58f315fd4481a40bbc.42]
mov esi, 39
call qword ptr [rip + core::panicking::panic@GOTPCREL]
jmp .LBB10_4
.LBB10_21:
lea rdi, [rip + .Lanon.4e23a766b1161c58f315fd4481a40bbc.38]
lea rdx, [rip + .Lanon.4e23a766b1161c58f315fd4481a40bbc.42]
mov esi, 39
call qword ptr [rip + core::panicking::panic@GOTPCREL]
.LBB10_4:
ud2
jmp .LBB10_65
mov r15, rax
cmp qword ptr [rsp + 24], 0
jne .LBB10_66
mov rdi, qword ptr [rsp + 32]
mov esi, 32
mov edx, 8
call qword ptr [rip + __rustc::__rust_dealloc@GOTPCREL]
mov rdi, r15
call _Unwind_Resume@PLT
jmp .LBB10_25
mov r15, rax
cmp qword ptr [rsp + 768], 0
je .LBB10_28
cmp qword ptr [rsp + 784], 0
jne .LBB10_28
mov r14, qword ptr [rsp + 792]
jmp .LBB10_8
jmp .LBB10_65
mov r15, rax
test r12, r12
jne .LBB10_8
mov esi, 32
mov edx, 8
mov rdi, rbp
call qword ptr [rip + __rustc::__rust_dealloc@GOTPCREL]
jmp .LBB10_8
mov r15, rax
test rbx, rbx
jne .LBB10_18
mov esi, 32
mov edx, 8
mov rdi, r13
call qword ptr [rip + __rustc::__rust_dealloc@GOTPCREL]
.LBB10_18:
mov r13, rbp
mov rbp, r12
jmp .LBB10_26
.LBB10_25:
mov r15, rax
.LBB10_26:
mov esi, 32
mov edx, 8
mov rdi, r14
call qword ptr [rip + __rustc::__rust_dealloc@GOTPCREL]
jmp .LBB10_27
mov r15, rax
.LBB10_27:
mov r14, r13
test rbp, rbp
jne .LBB10_28
.LBB10_8:
mov esi, 32
mov edx, 8
mov rdi, r14
call qword ptr [rip + __rustc::__rust_dealloc@GOTPCREL]
.LBB10_28:
lea rdi, [rsp + 56]
call core::ptr::drop_in_place::<pstd::collections::btree_set::BTreeSet<i32>>
jmp .LBB10_66
call qword ptr [rip + core::panicking::panic_in_cleanup@GOTPCREL]
.LBB10_65:
mov r15, rax
lea rdi, [rsp + 16]
call core::ptr::drop_in_place::<pstd::collections::btree_set::BTreeSet<i32>>
.LBB10_66:
mov rdi, r15
call _Unwind_Resume@PLT
call qword ptr [rip + core::panicking::panic_in_cleanup@GOTPCREL]
Maybe I am mis-interpreting this code:
.LBB10_55:
mov edx, 704
mov rdi, r14
xor esi, esi
call qword ptr [rip + memset@GOTPCREL]
Now I look at it more carefully, it seems to be zeroing 704 bytes, but I don’t know why. Hmm, maybe it needs to zero various parts of the struct ( Iterator ) so it zeroes all of it rather than fiddling around. Except if it is big, that can be slow. I think…
[ There are 2 fixed-length vecs in the struct and some other stuff, I think the only thing that needs to be initialised is the “other stuff” and the two vec lengths, but instead of doing 2 word-stores, it is zeroing it all… Yes, I think that is it. ]
Ok, I finally managed to make a simple example:
This code:
```
movl $8016, %edx
movq %rbx, %rdi
xorl %esi, %esi
callq *%r13
is clearing the whole struct, 8016 bytes, when only a couple of words (16 bytes) need to be set.
The layout is 8 bytes for the length + 4000 MaybeUninit bytes, times two.
Well I think it is fairly clear what is going on, the compiler decided to use memset to clear (initialise) an entire struct with large uninit sections, which is somewhat inefficient.
Yes, this looks to be behavior that was designed for cases where uninitialized data is put into something like small padding sections appearing in a struct, and memset-ting the whole thing would be faster than trying to leaving the uninitialized padding untouched. I could not immediately find a fitting issue on Github yet, perhaps it might be worth just opening one.
Actually testing a bit on godbolt.org, it might seem like this is actually a new / regression behavior in Rust 1.93?
Edit: Bisection points to
as the relevant change.
I’ve made an issue ^^
This is usually an LLVM behaviour. It will prefer const-folding to "real" values rather than to poison/undef, for example, even if poison would be correct.
Demo: https://llvm.godbolt.org/z/sbrvb5vq3
It would be legal for that to return poison, but that's not what LLVM does.
So typically the answer is that uninit in Rust gives undef in LLVM, but then LLVM turns that into something else.
Note that the PR linked earlier (Fix MaybeUninit codegen using GVN by saethlin · Pull Request #147827 · rust-lang/rust · GitHub) was fixing a different version of this where MaybeUninit::uninitialized() got codegen'd as copying a value from rodata, which is obviously also bad.
If you look at what we give LLVM: https://rust.godbolt.org/z/8WcjT98dY
@anon.29d017c02cc7c950eacbf58ef03f904f.0 = private unnamed_addr constant [4000 x i8] undef, align 4
define internal void @example[50e99ae1af2fd79d]::test(ptr dead_on_unwind noalias noundef writable sret([8016 x i8]) align 8 captures(none) dereferenceable(8016) %_0, i32 noundef %i) unnamed_addr {
start:
%_3 = alloca [4008 x i8], align 8
%result = alloca [8016 x i8], align 8
call void @llvm.lifetime.start.p0(ptr %result)
%0 = getelementptr inbounds i8, ptr %_3, i64 4000
store i64 0, ptr %0, align 8
call void @llvm.memcpy.p0.p0.i64(ptr align 8 %_3, ptr align 4 @anon.29d017c02cc7c950eacbf58ef03f904f.0, i64 4000, i1 false)
call void @llvm.memcpy.p0.p0.i64(ptr align 8 %result, ptr align 8 %_3, i64 4008, i1 false)
%1 = getelementptr inbounds i8, ptr %result, i64 4008
call void @llvm.memcpy.p0.p0.i64(ptr align 8 %1, ptr align 8 %_3, i64 4008, i1 false)
call void @example[50e99ae1af2fd79d]::dosomething(ptr noalias noundef align 8 dereferenceable(8016) %result, i32 noundef %i)
call void @llvm.memcpy.p0.p0.i64(ptr align 8 %_0, ptr align 8 %result, i64 8016, i1 false)
call void @llvm.lifetime.end.p0(ptr %result)
ret void
}
It's copying from a fully-undef private constant.
And later, MemCpyOptPass replaces that sequence of memcpys with a memset: https://rust.godbolt.org/z/46bjEn446