Split the type context into a global and a local (inference-only) one. #33425
Conversation
|
(rust_highfive has picked a reviewer for you, use r? to override) |
|
Nice memory savings, but 3.5x time regression in item-bodies checking? |
|
@durka Just noticed that actually, will need to investigate. We might have to make lifting cheaper. |
|
About 25% of the time appears to be spend in |
|
What are the callers? |
|
@arielb1 I believe I'd need |
|
I've found a simple way of turning the split off while keeping all of the infrastructural changes, if we want to merge this without any significant impact on performance. But it turns out that making So either the cost of checking |
|
@nikomatsakis I'm pretty sure the problem here is the global selection/evaluation caches which used to allow Should I just add a separate flag for freshened types, and keep them in the global context, just for the caches? |
Hmm. Seems worth a try! I'm a bit wary of removing |
|
@nikomatsakis I'd keep the current behavior of |
|
☔ The latest upstream changes (presumably #33138) made this pull request unmergeable. Please resolve the merge conflicts. |
|
☔ The latest upstream changes (presumably #33267) made this pull request unmergeable. Please resolve the merge conflicts. |
eddyb
force-pushed
the
rift
branch
2 times, most recently
from
e2a30af
to
3316c7a
Compare
|
☔ The latest upstream changes (presumably #33091) made this pull request unmergeable. Please resolve the merge conflicts. |
|
☔ The latest upstream changes (presumably #33523) made this pull request unmergeable. Please resolve the merge conflicts. |
|
@eddyb r=me you are my hero 👑 |
|
📌 Commit 42eb703 has been approved by |
Split the type context into a global and a local (inference-only) one. After this change, each `InferCtxt` creates its own local type interner for types with inference by-products. Most of the code which handles both a global and a local interner uses `'gcx` and `'tcx` for them. A reference to the type context in that situation (e.g. `infcx.tcx`) is `TyCtxt<'a, 'gcx, 'tcx>`. The global type context which used to be `&'a TyCtxt<'tcx>` is now `TyCtxt<'a, 'tcx, 'tcx>`. In order to minimize the number of extra lifetime parameters, many functions became methods. Where possible (some inherent impls), lifetime parameters were added on the impl, not each method. As inference by-products no longer escape their inference contexts, memory usage is lower. Example of `-Z time-passes` excerpt for `librustc`, stage1 (~100MB gains): Before "rustc: Split local type contexts interners from the global one.": ``` time: 0.395; rss: 335MB item-types checking time: 15.392; rss: 472MB item-bodies checking time: 0.000; rss: 472MB drop-impl checking time: 1.140; rss: 478MB const checking time: 0.139; rss: 478MB privacy checking time: 0.024; rss: 478MB stability index time: 0.072; rss: 478MB intrinsic checking time: 0.038; rss: 478MB effect checking time: 0.255; rss: 478MB match checking time: 0.128; rss: 484MB liveness checking time: 1.372; rss: 484MB rvalue checking time: 1.404; rss: 597MB MIR dump time: 0.809; rss: 599MB MIR passes ``` After: ``` time: 0.467; rss: 337MB item-types checking time: 17.443; rss: 395MB item-bodies checking time: 0.000; rss: 395MB drop-impl checking time: 1.423; rss: 401MB const checking time: 0.141; rss: 401MB privacy checking time: 0.024; rss: 401MB stability index time: 0.116; rss: 401MB intrinsic checking time: 0.038; rss: 401MB effect checking time: 0.382; rss: 401MB match checking time: 0.132; rss: 407MB liveness checking time: 1.678; rss: 407MB rvalue checking time: 1.614; rss: 503MB MIR dump time: 0.957; rss: 512MB MIR passes ``` **NOTE**: Functions changed to methods weren't re-indented to keep this PR easier to review. Once approved, the changes will be mechanically performed. However, indentation changes of function arguments are there - and I believe there's a way to hide whitespace-only changes in diffs on GitHub.
It turns out that `ast_ty_to_ty` is supposed to be updating the `def` after it finishes, but at some point in the past it stopped doing so. This was never noticed because of the `ast_ty_to_ty_cache`, but that cache was recently removed. This PR fixes the code to update the def properly, but apparently that is not quite enough to make the operation idempotent, so for now we reintroduce the cache too. Fixes rust-lang#33425.
…-type-path, r=eddyb re-introduce a cache for ast-ty-to-ty It turns out that `ast_ty_to_ty` is supposed to be updating the `def` after it finishes, but at some point in the past it stopped doing so. This was never noticed because of the `ast_ty_to_ty_cache`, but that cache was recently removed. This PR fixes the code to update the def properly, but apparently that is not quite enough to make the operation idempotent, so for now we reintroduce the cache too. Fixes rust-lang#33425. r? @eddyb
bluss
added
the
relnotes
After this change, each
InferCtxtcreates its own local type interner for types with inference by-products.Most of the code which handles both a global and a local interner uses
'gcxand'tcxfor them.A reference to the type context in that situation (e.g.
infcx.tcx) isTyCtxt<'a, 'gcx, 'tcx>.The global type context which used to be
&'a TyCtxt<'tcx>is nowTyCtxt<'a, 'tcx, 'tcx>.In order to minimize the number of extra lifetime parameters, many functions became methods.
Where possible (some inherent impls), lifetime parameters were added on the impl, not each method.
As inference by-products no longer escape their inference contexts, memory usage is lower.
Example of
-Z time-passesexcerpt forlibrustc, stage1 (~100MB gains):Before "rustc: Split local type contexts interners from the global one.":
After:
NOTE: Functions changed to methods weren't re-indented to keep this PR easier to review.
Once approved, the changes will be mechanically performed.
However, indentation changes of function arguments are there - and I believe there's a way to hide whitespace-only changes in diffs on GitHub.