Optimize dep-graph serialization and improve meta-data hashing #35406
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
The way we do HIR inlining introduces reads of the "Hir" into the graph, but this Hir in fact belongs to other crates, so when we try to load later, we ICE because the Hir nodes in question don't belond to the crate (and we haven't done inlining yet). This pass rewrites those HIR nodes to the metadata from which the inlined HIR was loaded.
The biggest problem, actually, is krate numbers being removed entirely, which can lead to array-index-out-of-bounds errors. cc rust-lang#35123 -- not a complete fix, since really we ought to "map" the old crate numbers to the new ones, not just detect changes.
The reads will occur naturally as the HIR/MIR is fetched from the tracked tables, and this winds up adding reads to the hir of foreign def-ids somehow.
When we hash the inputs to a MetaData node, we have to hash them in a consistent order. We achieve this by sorting the stringfied `DefPath` entries. Also, micro-optimie by cache more results across the saving process.
It's nice to get a rough idea of how much work we're saving.
We now detect inlined id's earlier (in the HIR map) and rewrite a read of them to be a read of the metadata for the associated item.
I cannot figure out how to write a test for this, but I observed incorrect edges as a result of not using memoized pattern here (e.g., LateLintCheck -> SizedConstraint).
this can actually be expensive!
|
|
| @@ -39,6 +39,14 @@ impl<'a, 'tcx> HashContext<'a, 'tcx> { | |||
| } | |||
| } | |||
|
|
|||
| pub fn is_hashable(dep_node: &DepNode<DefId>) -> bool { | |||
There was a problem hiding this comment.
I'd prefer if this was called is_input_node or is_graph_root to make it clearer what we are after.
|
Ok, I've reviewed everything after |
Per the discussion on rust-lang#34765, we make one `DepNode::Mir` variant and use it to represent both the MIR tracking map as well as passes that operate on MIR. We also track loads of cached MIR (which naturally comes from metadata). Note that the "HAIR" pass adds a read of TypeckItemBody because it uses a myriad of tables that are not individually tracked.
it now carries a def-id; supply a dummy
The new `Predecessors` type computes a set of interesting targets and their HIR predecessors, and discards everything in between.
Produces a deterministic hash, at least for a single platform / compiler-version.
This massively speeds up serialization. It also seems to produce deterministic metadata hashes (before I was seeing inconsistent results). Fixes rust-lang#35232.
|
@michaelwoerister OK, I addressed your nits. I will rebase atop of #35166, I think, and maybe try to land them together once you've revised the additional commits in there. |
nikomatsakis
force-pushed
the
incr-comp-35232
branch
from
698619a to
ecbcf1b
Compare
| // encoding, rather than having been retraced to a `DefId`. The | ||
| // reason for this is that this way we can include nodes that have | ||
| // been removed (which no longer have a `DefId` in the current | ||
| // compilation). |
|
LGTM. |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
This makes a number of related changes:
HIR -> WorkProductedges-Z incremental-info, which tells you how much re-use you are getting, and tells you why you are not getting re-useThe end result is that, at least in my tests, we are now able to get 100% re-use when doing
cargo clean; cargo buildtwice in a row (with an incremental directory, of course). Here are some rough performance numbers from thesyntex-syntaxcrate:Note that these numbers only reflect the first phase of incremental, where we are just skipping the trans/LLVM phases. They will get better as we incrementalize more of the compiler. Also, these numbers represent the simplest case, where no actual changes were made. There is still more work to do to handle the case some small number of changes were made correctly (in particular, verifying that the dep-graph is complete etc).
r? @michaelwoerister
Fixes #35167.