Missed code elimination with std::mem::replace/std::mem::swap #44701
Description
Edit: As pointed out by @oyvindln, the problem appears to be a regression introduced in 1.20.
Demonstration: https://godbolt.org/g/5uuzVL
Version: rustc 1.22.0-nightly (277476c 2017-09-16) , -C opt-level=3 -C target-cpu=native
Code:
use std::rc::Rc; use std::cell::RefCell; pub struct Buffer { buf: Vec<u8>, pool: Rc<RefCell<Vec<Vec<u8>>>>, } impl Drop for Buffer { fn drop(&mut self) { self.pool.borrow_mut() .push(std::mem::replace(&mut self.buf, vec![])); } }Expected result: An optimal code would move self.buf directly into self.pool and then reset self.buf in-place. An acceptable code would moveself.buf into a temporary on the stack, move the temporary into self.pool and reset self.buf in-place.
Observed result:
- Space for two
std::Vec<u8>(each 24 bytes) is allocated on the stack,-48(%rbp)(A) and-96(%rbp)(B). self.bufis copied to A.self.bufis reset in-place.- A is copied to B.
- B is copied to A.
- A is inserted into
self.pool.
Steps 4 and 5 is a completely unnecessary copying of 48 bytes and could be safely removed. Replacing std::mem::replace with an equivalent std::mem::swap call produces a slightly different code with the same basic problem.
Compiler output:
<example::Buffer as core::ops::drop::Drop>::drop: pushq %rbp movq %rsp, %rbp pushq %rbx subq $88, %rsp movq %rdi, %rax movq 24(%rax), %rbx cmpq $0, 16(%rbx) jne .LBB6_6 leaq 16(%rbx), %rcx movq $-1, 16(%rbx) leaq 24(%rbx), %rdi movq %rdi, -64(%rbp) movq %rcx, -56(%rbp) ;2 movq 16(%rax), %rcx movq %rcx, -32(%rbp) vmovups (%rax), %xmm0 vmovaps %xmm0, -48(%rbp) ;3 movq $1, (%rax) vxorps %xmm0, %xmm0, %xmm0 vmovups %xmm0, 8(%rax) ;4 movq -32(%rbp), %rax movq %rax, -80(%rbp) vmovaps -48(%rbp), %xmm0 vmovaps %xmm0, -96(%rbp) ;5 movq -80(%rbp), %rax movq %rax, -32(%rbp) vmovaps -96(%rbp), %xmm0 vmovaps %xmm0, -48(%rbp) ;6 movq 40(%rbx), %rax cmpq 32(%rbx), %rax jne .LBB6_4 callq <alloc::raw_vec::RawVec<T, A>>::double movq 40(%rbx), %rax .LBB6_4: movq 24(%rbx), %rcx leaq (%rax,%rax,2), %rax movq -32(%rbp), %rdx movq %rdx, 16(%rcx,%rax,8) vmovaps -48(%rbp), %xmm0 vmovups %xmm0, (%rcx,%rax,8) incq 40(%rbx) movq $0, 16(%rbx) addq $88, %rsp popq %rbx popq %rbp retq .LBB6_6: callq core::result::unwrap_failed movq %rax, %rbx leaq -48(%rbp), %rdi callq core::ptr::drop_in_place leaq -64(%rbp), %rdi callq core::ptr::drop_in_place movq %rbx, %rdi callq _Unwind_Resume@PLT