missed optimization: coalesce PathBuf reallocation

[lolbinarycat]

compare the following disassemblies:

#![allow(dead_code)]

use std::path::{Path, PathBuf};

#[no_mangle]
pub fn boxed_path() -> Box<Path> {
    Path::new("aaa").into()
}

boxed_path:
        lea     rdi, [rip + .L__unnamed_1]
        mov     esi, 3
        jmp     qword ptr [rip + std::path::<impl core::convert::From<&std::path::Path> for alloc::boxed::Box<std::path::Path>>::from::hfd2d6fbd02ea1491@GOTPCREL]

.L__unnamed_1:
        .zero   3,97

#![allow(dead_code)]

use std::path::{Path, PathBuf};

#[no_mangle]
pub fn boxed_path() -> Box<Path> {
    PathBuf::from("aaa").into()
}

boxed_path:
        push    rbx
        sub     rsp, 32
        lea     rsi, [rip + .L__unnamed_1]
        lea     rbx, [rsp + 8]
        mov     edx, 3
        mov     rdi, rbx
        call    qword ptr [rip + std::sys::os_str::bytes::Slice::to_owned::hda0e6bb365f52f1f@GOTPCREL]
        mov     rdi, rbx
        call    qword ptr [rip + std::ffi::os_str::OsString::into_boxed_os_str::haa190229882bb506@GOTPCREL]
        add     rsp, 32
        pop     rbx
        ret

.L__unnamed_1:
        .zero   3,97

[workingjubilee]

@lolbinarycat hmm, you should probably either provide a godbolt link or specify things like the optimization level?

[ptrca]

I guess that's because of inlining differences.

[ryanavella]

I guess that's because of inlining differences.

Even if more could be inlined in the From<PathBuf> for Box<Path> case, it eventually delegates to Vec::shrink_to_fit. This reminds me of similar issues where the optimizer struggles to reason about alloc+realloc pairs that can be coalesced down to a single alloc.

[veera-sivarajan]

@rustbot label -needs-triage +C-optimization +T-compiler +A-io +E-needs-mcve

[saethlin]

This reminds me of similar issues where the optimizer struggles to reason about alloc+realloc pairs that can be coalesced down to a single alloc.

While this problem does exist, it's not relevant here. This example is just inlining.