Commit
This commit does not belong to any branch on this repository, and may belong to a fork outside of the repository.
Rollup merge of rust-lang#36917 - nnethercote:speed-up-plug_leaks, r=…
…eddyb Speed up `plug_leaks` Profiling shows that `plug_leaks` and the functions it calls are hot on some benchmarks. It's very common that `skol_map` is empty in this function, and we can specialize `plug_leaks` in that case for some big speed-ups. The PR has two commits. I'm fairly confident that the first one is correct -- I traced through the code to confirm that the `fold_regions` and `pop_skolemized` calls are no-ops when `skol_map` is empty, and I also temporarily added an assertion to check that `result` ends up having the same value as `value` in that case. This commit is responsible for most of the improvement. I'm less confident about the second commit. The call to `resolve_type_vars_is_possible` can change `value` when `skol_map` is empty... but testing suggests that it doesn't matter if the call is omitted. So, please check both patches carefully, especially the second one! Here are the speed-ups for the first commit alone. stage1 compiler (built with old rustc, using glibc malloc), doing debug builds: ``` futures-rs-test 4.710s vs 4.538s --> 1.038x faster (variance: 1.009x, 1.005x) issue-32062-equ 0.415s vs 0.368s --> 1.129x faster (variance: 1.009x, 1.010x) issue-32278-big 1.884s vs 1.808s --> 1.042x faster (variance: 1.020x, 1.017x) jld-day15-parse 1.907s vs 1.668s --> 1.143x faster (variance: 1.011x, 1.007x) piston-image-0. 13.024s vs 12.421s --> 1.049x faster (variance: 1.004x, 1.012x) rust-encoding-0 3.335s vs 3.276s --> 1.018x faster (variance: 1.021x, 1.028x) ``` stage2 compiler (built with new rustc, using jemalloc), doing debug builds: ``` futures-rs-test 4.167s vs 4.065s --> 1.025x faster (variance: 1.006x, 1.018x) issue-32062-equ 0.383s vs 0.343s --> 1.118x faster (variance: 1.012x, 1.016x) issue-32278-big 1.680s vs 1.621s --> 1.036x faster (variance: 1.007x, 1.007x) jld-day15-parse 1.671s vs 1.478s --> 1.131x faster (variance: 1.016x, 1.004x) piston-image-0. 11.336s vs 10.852s --> 1.045x faster (variance: 1.003x, 1.006x) rust-encoding-0 3.036s vs 2.971s --> 1.022x faster (variance: 1.030x, 1.032x) ``` I've omitted the benchmarks for which the change was negligible. And here are the speed-ups for the first and second commit in combination. stage1 compiler (built with old rustc, using glibc malloc), doing debug builds: ``` futures-rs-test 4.684s vs 4.498s --> 1.041x faster (variance: 1.012x, 1.012x) issue-32062-equ 0.413s vs 0.355s --> 1.162x faster (variance: 1.019x, 1.006x) issue-32278-big 1.869s vs 1.763s --> 1.060x faster (variance: 1.013x, 1.018x) jld-day15-parse 1.900s vs 1.602s --> 1.186x faster (variance: 1.010x, 1.003x) piston-image-0. 12.907s vs 12.352s --> 1.045x faster (variance: 1.005x, 1.006x) rust-encoding-0 3.254s vs 3.248s --> 1.002x faster (variance: 1.063x, 1.045x) ``` stage2 compiler (built with new rustc, using jemalloc), doing debug builds: ``` futures-rs-test 4.183s vs 4.046s --> 1.034x faster (variance: 1.007x, 1.004x) issue-32062-equ 0.380s vs 0.340s --> 1.117x faster (variance: 1.020x, 1.003x) issue-32278-big 1.671s vs 1.616s --> 1.034x faster (variance: 1.031x, 1.012x) jld-day15-parse 1.661s vs 1.417s --> 1.172x faster (variance: 1.013x, 1.005x) piston-image-0. 11.347s vs 10.841s --> 1.047x faster (variance: 1.007x, 1.010x) rust-encoding-0 3.050s vs 3.000s --> 1.017x faster (variance: 1.016x, 1.012x) ``` @eddyb: `git blame` suggests that you should review this. Thanks!
- Loading branch information