Compile time regression with the new trait solver and diesel

[weiznich]

I tried this code: diesel-rs/diesel@381be19 with both RUSTFLAGS=-Znext-solver=globally set and not set

I expected to see this happen: Code compiles similar fast with both combinations

Instead, this happened: RUSTFLAGS=-Znext-solver=globally results in the code compiling takes over 10 times as long. For me the compiler takes ~10s to compile diesel from scratch with the old trait solver. The same operation takes 1:53 min, so over 10 times of what the old trait solver needed.

Meta

rustc --version --verbose:

rustc 1.83.0-nightly (9ff5fc4ff 2024-10-03)
binary: rustc
commit-hash: 9ff5fc4ffbbe1e911527aa054e789b05ae55ffcc
commit-date: 2024-10-03
host: x86_64-unknown-linux-gnu
release: 1.83.0-nightly
LLVM version: 19.1.0

@rustbot label +I-compiletime +WG-trait-system-refactor

[lcnr]

does #130821 impact the compile times here, currently travelling to EuroRust and can't get to compile diesel on the dev desktops

[weiznich]

I've tried it again with today's nightly and I fear the results are at least not better.

Version:

rustc 1.83.0-nightly (3ae715c8c 2024-10-07)
binary: rustc
commit-hash: 3ae715c8c63f9aeac47cbf7d8d9dadb3fa32c638
commit-date: 2024-10-07
host: x86_64-unknown-linux-gnu
release: 1.83.0-nightly
LLVM version: 19.1.1

Times:
Old Solver: 19s
New Solver: 3:47min

(Differences in the old solver times are likely due to running on battery right now, due to traveling to Vienna as well)

[weiznich]

I checked this again today with the following nightly version:

rustc 1.84.0-nightly (143ce0920 2024-11-10)
binary: rustc
commit-hash: 143ce0920a2307b19831160a01f06f107610f1b2
commit-date: 2024-11-10
host: x86_64-unknown-linux-gnu
release: 1.84.0-nightly
LLVM version: 19.1.3

Times for cargo check -p diesel:
Old Solver: 20s
New Solver: 3:33min

cc @compiler-errors @lcnr

[lqd]

Is performance testing the new solver expected or useful now compared to when it's more complete (and the work to improve its performance actually starts)?

[lcnr]

I think knowing about the regression in diesel is useful, even if it isn't a priority right now. It will definitely be useful once bootstrap is working and we're looking into our benchmarks in more detail

[weiznich]

For reference I just redid the measurements with the current rustc version:

Benchmark 1: cargo +nightly check -p diesel
  Time (mean ± σ):      4.136 s ±  0.064 s    [User: 3.900 s, System: 0.214 s]
  Range (min … max):    4.061 s …  4.225 s    10 runs
 
Benchmark 2: RUSTFLAGS=-Znext-solver=globally cargo +nightly check -p diesel
  Time (mean ± σ):      8.383 s ±  1.814 s    [User: 8.316 s, System: 0.393 s]
  Range (min … max):    7.729 s … 13.544 s    10 runs

rustc 1.90.0-nightly (9982d6462 2025-07-20)
binary: rustc
commit-hash: 9982d6462bedf1e793f7b2dbd655a4e57cdf67d4
commit-date: 2025-07-20
host: x86_64-unknown-linux-gnu
release: 1.90.0-nightly
LLVM version: 20.1.8

These numbers seem to indicate that the new solver got quite a bit faster since the last tests, it's now only 2 times slower than the old solver. Great to see the new solver progressing in the right direction 👍

(The actual baseline times differ, due to that measurements are taken on a different system)