steven_blocks: could not compiler under windows? memory allocation of 15728640 bytes failed, STATUS_STACK_BUFFER_OVERRUN

[woodgear]

       Fresh mime_guess v2.0.0-alpha.6
       Fresh rsa_public_encrypt_pkcs1 v0.2.0
       Fresh reqwest v0.9.18
memory allocation of 15728640 bytes failed
error: Could not compile `steven_blocks`.

Caused by:
  process didn't exit successfully: `C:\Users\wucong\.cargo\bin\sccache.exe rustc --edition=2018 --crate-name steven_blocks blocks\src\lib.rs --color always --crate-type lib --emit=dep-info,link -C opt-level=1 -C debuginfo=2 -C debug-assertions=on -C metadata=fff764ec044c94f8 -C extra-filename=-fff764ec044c94f8 --out-dir C:\Users\wucong\work\lab\stevenarella\target\debug\deps -C incremental=C:\Users\wucong\work\lab\stevenarella\target\debug\incremental -L dependency=C:\Users\wucong\work\lab\stevenarella\target\debug\deps --extern cgmath=C:\Users\wucong\work\lab\stevenarella\target\debug\deps\libcgmath-df2dce11f9bb8253.rlib --extern collision=C:\Users\wucong\work\lab\stevenarella\target\debug\deps\libcollision-7b0ee92f28b81d67.rlib --extern lazy_static=C:\Users\wucong\work\lab\stevenarella\target\debug\deps\liblazy_static-7fa98efeb9d42cb9.rlib --extern steven_shared=C:\Users\wucong\work\lab\stevenarella\target\debug\deps\libsteven_shared-fe5e6bf3fccae278.rlib` (exit code: 0xc0000409, STATUS_STACK_BUFFER_OVERRUN)

what should i do?

[iceiix]

This looks similar to the issue I was hitting in #174, unfortunately steven_blocks requires a lot of memory to compile (I've seen it peak at 3.3 GB) because of the recursive macro expansion.

@woodgear How much memory do you have? I test building on a 32 GB macOS system, and build on SourceHut (Linux) and AppVeyor (Windows), but not sure of their memory limits.

But I agree this excessive memory requirement is kind of ridiculous (steven_blocks also takes the longest time to build by far), would like to fix it but have yet to find how to improve it, may require rewriting the blocks macro system.

[iceiix]

15728640 bytes = 15 megabytes. I'm now hitting a similar problem on the same machine with Rust 1.40 (#253), not a memory allocation failure, but an unreasonably long build time (2.5+ hr, 25+ GB).

This high memory usage in steven_blocks is a longstanding problem, certainly should be addressed. May have to rewrite the whole blocks system to use fewer macros, or is there a way it could be optimized to build reasonably quickly? This will only get worse once more blocks are added, and we are behind. But steven_blocks has became quite complex, I'm not sure what to do here, may need assistance from someone with more expertise with Rust macros…

[iceiix]

Updating to Rust version 1.40 #253 (comment), time cargo build --release takes 4 hours 12 minutes (!).

The bottleneck seems to be in release optimization. Testing to isolate the problem, on nightly with -Z time passes, completes in about 2 minutes. No optimization, stable, also about less than 2 minutes. Returning to nightly and effectively using --release with -Z time-passes: blocks $ cargo clean ; time cargo rustc -- -C 'opt-level=3' -Z time-passes — still building, 25+ minutes, 4.5GB+ and counting…

time: 0.002	codegen passes [4gqa868n29dtlfbb]
time: 0.652	LTO passes
time: 0.075	codegen passes [473etbgnzq2x8af8]
time: 3.476	LTO passes
time: 4.218	LTO passes
time: 4.255	codegen passes [22fyvf1uiha0mrri]
time: 4.234	codegen passes [6ffwyc85r8r5jgh]
time: 63.064	LTO passes
Building [======================================================>  ] 48/49: steven_blocks   

[iceiix]

https://internals.rust-lang.org/t/rust-staticlibs-and-optimizing-for-size/5746/4 mentions https://bugzilla.mozilla.org/show_bug.cgi?id=1386371
Disable Rust LTO by default in local builds, in Firefox.

Rust LTO is very slow and cause significant build time regressions, but Rust LTO's run-time performance speedup is small. Michael Woerister says disabling Rust LTO "would basically eliminate the build time for gkrust" (the Stylo crate).

Is this still true, or how much would steven_blocks benefit form LTO, is it causing the 125x slowdown or another optimization?

https://doc.rust-lang.org/nightly/rustc/codegen-options/index.html#opt-level

0: no optimizations, also turns on cfg(debug_assertions).
1: basic optimizations
2: some optimizations
3: all optimizations
s: optimize for binary size
z: optimize for binary size, but also turn off loop vectorization.

Worth noting, steven itself has this optimization profile in Config.toml:

[profile.dev]
# Steven runs horrendously slow with no optimizations, and often freezes.
# However, building with full -O3 optimizations takes too long for a debug build.
# Use an -O1 optimization level strikes a good compromise between build and program performance.
opt-level = 1

should steven_blocks have its Cargo.toml specify profile.release opt-level=1 (or 2)?

Still waiting on the clean time cargo rustc -- -C 'opt-level=3' -Z time-passes build, in blocks/, going on 37 minutes and 6.5GB+, if it takes as long as last time expect 4+ hours.

[iceiix]

Completed in 5 hours 9 minutes this time:

   time: 0.002	codegen passes [4gqa868n29dtlfbb]
    time: 0.652	LTO passes
    time: 0.075	codegen passes [473etbgnzq2x8af8]
    time: 3.476	LTO passes
    time: 4.218	LTO passes
    time: 4.255	codegen passes [22fyvf1uiha0mrri]
    time: 4.234	codegen passes [6ffwyc85r8r5jgh]
    time: 63.064	LTO passes
    time: 15574.417	codegen passes [2w3r8ppd44eh9kz5]
  time: 15707.329	LLVM passes
  time: 0.003	serialize work products
  time: 0.746	linking
time: 15771.815		total
    Finished dev [unoptimized + debuginfo] target(s) in 263m 19s

real	309m45.295s
user	198m44.124s
sys	14m54.290s

15574.417 is spent on "codegen passes [2w3r8ppd44eh9kz5]". Other hits for 2w3r8ppd44eh9kz5:

    time: 3.984	llvm function passes [2w3r8ppd44eh9kz5]
      time: 2.733	persist query result cache
          time: 0.020	getting serialized graph
          time: 0.024	encoding serialized graph
        time: 0.044	encode dep-graph
      time: 0.070	persist dep-graph
    time: 2.803	serialize dep graph
  time: 5.799	codegen
    time: 63.088	llvm module passes [2w3r8ppd44eh9kz5]

15771.815 seconds = 4.38 hours on "codegen passes". This again points to codegen options.

[iceiix]

cargo clean ; time cargo rustc -- -C 'opt-level=0' -Z time-passes

    Finished dev [unoptimized + debuginfo] target(s) in 1m 41s

cargo clean ; time cargo rustc -- -C 'opt-level=1' -Z time-passes

  time: 29.178	LLVM passes
  time: 0.000	serialize work products
  time: 0.164	linking
time: 97.822		total
    Finished dev [unoptimized + debuginfo] target(s) in 2m 05s

cargo clean ; time cargo rustc -- -C 'opt-level=2' -Z time-passes

waiting 12+ minutes, 1.48 GB+ and growing. Killed. update: rerunning, 3.7+ hours, 20+ GB, before I killed it again. Whatever optimization pass is slow, it occurs during opt-level=2.

Try this: https://users.rust-lang.org/t/improve-compile-time-and-executable-size-by-counting-lines-of-llvm-ir/14203 https://github.com/dtolnay/cargo-llvm-lines

[iceiix]

blocks $ cargo llvm-lines
warning: file found to be present in multiple build targets: steven/src/main.rs
   Compiling steven_shared v0.0.1 (steven/shared)
   Compiling steven_blocks v0.0.1 (steven/blocks)
    Finished dev [unoptimized + debuginfo] target(s) in 1m 19s
 151828    2  core::ops::function::FnOnce::call_once
  43954  488  core::option::Option<T>::map
  24252  896  core::ptr::real_drop_in_place
  19669    1  steven_blocks::Block::get_model
  15834   39  alloc::raw_vec::RawVec<T,A>::reserve_internal
  14674    1  steven_blocks::Block::get_flat_offset
  12548    1  steven_blocks::Block::get_hierarchical_data
  12362    1  steven_blocks::Block::get_collision_boxes
  11195    1  steven_blocks::Block::get_model_variant
   9678    1  <steven_blocks::Block as core::fmt::Debug>::fmt
   6394   46  <<steven_blocks::VANILLA_ID_MAP as core::ops::deref::Deref>::deref::__static_ref_initialize::CombinationIter<facing> as core::iter::traits::iterator::Iterator>::next
   6305    1  steven_blocks::Block::get_material
   5976    1  steven_blocks::Block::match_multipart
   5642   14  <<steven_blocks::VANILLA_ID_MAP as core::ops::deref::Deref>::deref::__static_ref_initialize::CombinationIter<facing,half,shape,waterlogged> as core::iter::traits::iterator::Iterator>::next
   5075    1  steven_blocks::Block::get_tint
   5004   36  core::slice::<impl [T]>::copy_from_slice
   4990    1  steven_blocks::Block::get_modid
   4736   37  alloc::raw_vec::RawVec<T,A>::allocate_in
   4698    9  <<steven_blocks::VANILLA_ID_MAP as core::ops::deref::Deref>::deref::__static_ref_initialize::CombinationIter<north,south,west,east,waterlogged> as core::iter::traits::iterator::Iterator>::next
   4680   39  core::result::Result<T,E>::map_err
   4679    1  steven_blocks::Block::get_internal_id
   4600   40  core::result::Result<T,E>::map
   3885   37  core::iter::traits::exact_size::ExactSizeIterator::len
   3838   38  core::result::Result<T,E>::unwrap_or_else
   3591    7  <<steven_blocks::VANILLA_ID_MAP as core::ops::deref::Deref>::deref::__static_ref_initialize::CombinationIter<facing,half,hinge,open,powered> as core::iter::traits::iterator::Iterator>::next
...

get_model, get_flat_offset, get_hierarchical_data, get_collision_boxes, get_model_variant are among the largest in terms of IR lines. All these functions match *self, example:

            #[allow(unused_variables)]
            pub fn get_model(&self) -> (String, String) {
                match *self {
                    $(
                        Block::$name {
                            $($fname,)?
                        } => {
                            let parts = $model;
                            (String::from(parts.0), String::from(parts.1))
                        }
                    )+
                }
            }

Can attributes be used to control optimization? Maybe not, only code generation attributes are: inline, cold, no_builtins, target_feature. Leaning towards reducing the optimization level in the release profile. At least opt-level=1 is reasonably fast, only about 24 seconds slower than opt-level=0.

[iceiix]

steven_blocks builds fast in cd blocks && cargo build --release, but not at the top-level (going on 20+ minutes, 3.12 GB). Why is blocks/Cargo.toml profile.release (and even [dependencies], given #256) not used from the top-level Cargo.toml with [dependencies.steven_gl] path = "./gl"?