Avoid stack -> Box copy when data can be put in Box directly

[jrmuizel]

In https://bugzilla.mozilla.org/show_bug.cgi?id=1787715 (currently not public for security reasons) we ran into a stack overflow when flate2 calls Box::<CompressorOxide>>::default() in DeflateBackend::make() here:
https://github.com/rust-lang/flate2-rs/blob/cc5ed7f817cc5e5712b2bb924ed19cab4f389a47/src/ffi/rust.rs#L131

DeflateBackend::make() is getting compiled with a call to chkstk(65664)

A reduced version of this is here:
https://rust.godbolt.org/z/hTE3785xY

pub const LZ_CODE_BUF_SIZE: usize = 64 * 1024;

struct LZOxide {
    pub codes: [u8; LZ_CODE_BUF_SIZE],
}

impl LZOxide {
    #[inline(never)]
    const fn new() -> Self {
        LZOxide {
            codes: [0; LZ_CODE_BUF_SIZE],
        }
    }
}

 pub struct CompressorOxide {
    lz: LZOxide,
    huff: Box<u32>,
    pub codes: [u8; LZ_CODE_BUF_SIZE],
}

impl Default for CompressorOxide {
    #[inline(always)]
    fn default() -> Self {
        CompressorOxide {
            lz: LZOxide::new(),
            huff: Box::default(),
            codes: [0; LZ_CODE_BUF_SIZE],

        }
    }
}

pub fn f() -> Box<CompressorOxide> {
    Box::default()
}

A further reduced version that doesn't require any of the special box syntax in Default() is here:
https://rust.godbolt.org/z/8e5EPafzb

pub const LZ_CODE_BUF_SIZE: usize = 64 * 1024;

struct LZOxide {
    pub codes: [u8; LZ_CODE_BUF_SIZE],
}

impl LZOxide {
    #[inline(never)]
    const fn new() -> Self {
        LZOxide {
            codes: [0; LZ_CODE_BUF_SIZE],
        }
    }
}

 pub struct CompressorOxide {
    lz: LZOxide,
    huff: Box<u32>,
    //pub codes: [u8; LZ_CODE_BUF_SIZE],
}

impl Default for CompressorOxide {
    #[inline(always)]
    fn default() -> Self {
        CompressorOxide {
            lz: LZOxide::new(),
            huff: Box::default(),
            //codes: [0; LZ_CODE_BUF_SIZE],

        }
    }
}

use std::ptr;
use std::alloc::{Layout, alloc};

pub fn g() -> Box<CompressorOxide> {
    
    let layout = Layout::new::<CompressorOxide>();
    let ptr = 
            unsafe { alloc(layout) as *mut CompressorOxide };
            assert!(!ptr.is_null());
            unsafe { ptr::write(ptr, CompressorOxide {
            lz: LZOxide::new(),
            huff: Box::default(),

        });
    Box::from_raw(ptr)}

}

This seems to have to do with LZOxide::new() not being inlined. If it's inlined the problem goes away. If the field initialization order of CompressorOxide is swapped
so that we call Box::<u32>::default() before LZOxide::new() the problem also goes away.

@nikic

[Enselic]

Triage: Can you explain a bit more what you want rustc to change here? If you have particular inlining needs, using #[inline(always)] does not seem unreasonable?

[safinaskar]

@rustbot label A-box

[Enselic]

Triage: Thanks for adding the label, but at this point I think we can go ahead and close this as obsolete since the reporter did not get back with a clarification.

[lqd]

Why close, is the codegen issue fixed? What the OP wants is to avoid the large copy to the stack to do a heap allocation. That being said, it’s surely far from the only open issue about this.

[Enselic]

I think there is an issue about that already which would make this a duplicate. I'll look for it later today. Thanks for clarifying.

[lqd]

Im sure there are a ton, yes.

[jrmuizel]

Why close, is the codegen issue fixed? What the OP wants is to avoid the large copy to the stack to do a heap allocation. That being said, it’s surely far from the only open issue about this.

Yep.

[Enselic]

I only found #53827 which I don't think is a perfect duplicate. So I

• reopened this issue
• retitled it to be more generic (and IMHO descriptive)
• inlined the code in case godbolt goes down (temporarily or permanently)

Sorry for any inconvenience and thank you for the cooperation. Of course feel free to make whatever further adjustments you like.