Incremental Compilation In Detail

The incremental compilation scheme is, in essence, a surprisingly simple extension to the overall query system. It relies on the fact that:

  1. queries are pure functions -- given the same inputs, a query will always yield the same result, and
  2. the query model structures compilation in an acyclic graph that makes dependencies between individual computations explicit.

This chapter will explain how we can use these properties for making things incremental and then goes on to discuss version implementation issues.

A Basic Algorithm For Incremental Query Evaluation

As explained in the query evaluation model primer, query invocations form a directed-acyclic graph. Here's the example from the previous chapter again:

  list_of_all_hir_items <----------------------------- type_check_crate()
                                                               |
                                                               |
  Hir(foo) <--- type_of(foo) <--- type_check_item(foo) <-------+
                                      |                        |
                    +-----------------+                        |
                    |                                          |
                    v                                          |
  Hir(bar) <--- type_of(bar) <--- type_check_item(bar) <-------+

Since every access from one query to another has to go through the query context, we can record these accesses and thus actually build this dependency graph in memory. With dependency tracking enabled, when compilation is done, we know which queries were invoked (the nodes of the graph) and for each invocation, which other queries or input has gone into computing the query's result (the edges of the graph).

Now suppose we change the source code of our program so that HIR of bar looks different than before. Our goal is to only recompute those queries that are actually affected by the change while re-using the cached results of all the other queries. Given the dependency graph we can do exactly that. For a given query invocation, the graph tells us exactly what data has gone into computing its results, we just have to follow the edges until we reach something that has changed. If we don't encounter anything that has changed, we know that the query still would evaluate to the same result we already have in our cache.

Taking the type_of(foo) invocation from above as an example, we can check whether the cached result is still valid by following the edges to its inputs. The only edge leads to Hir(foo), an input that has not been affected by the change. So we know that the cached result for type_of(foo) is still valid.

The story is a bit different for type_check_item(foo): We again walk the edges and already know that type_of(foo) is fine. Then we get to type_of(bar) which we have not checked yet, so we walk the edges of type_of(bar) and encounter Hir(bar) which has changed. Consequently the result of type_of(bar) might yield a different result than what we have in the cache and, transitively, the result of type_check_item(foo) might have changed too. We thus re-run type_check_item(foo), which in turn will re-run type_of(bar), which will yield an up-to-date result because it reads the up-to-date version of Hir(bar). Also, we re-run type_check_item(bar) because result of type_of(bar) might have changed.

The Problem With The Basic Algorithm: False Positives

If you read the previous paragraph carefully you'll notice that it says that type_of(bar) might have changed because one of its inputs has changed. There's also the possibility that it might still yield exactly the same result even though its input has changed. Consider an example with a simple query that just computes the sign of an integer:

  IntValue(x) <---- sign_of(x) <--- some_other_query(x)

Let's say that IntValue(x) starts out as 1000 and then is set to 2000. Even though IntValue(x) is different in the two cases, sign_of(x) yields the result + in both cases.

If we follow the basic algorithm, however, some_other_query(x) would have to (unnecessarily) be re-evaluated because it transitively depends on a changed input. Change detection yields a "false positive" in this case because it has to conservatively assume that some_other_query(x) might be affected by that changed input.

Unfortunately it turns out that the actual queries in the compiler are full of examples like this and small changes to the input often potentially affect very large parts of the output binaries. As a consequence, we had to make the change detection system smarter and more accurate.

Improving Accuracy: The red-green Algorithm

The "false positives" problem can be solved by interleaving change detection and query re-evaluation. Instead of walking the graph all the way to the inputs when trying to find out if some cached result is still valid, we can check if a result has actually changed after we were forced to re-evaluate it.

We call this algorithm the red-green algorithm because nodes in the dependency graph are assigned the color green if we were able to prove that its cached result is still valid and the color red if the result has turned out to be different after re-evaluating it.

The meat of red-green change tracking is implemented in the try-mark-green algorithm, that, you've guessed it, tries to mark a given node as green:

fn try_mark_green(tcx, current_node) -> bool {

    // Fetch the inputs to `current_node`, i.e. get the nodes that the direct
    // edges from `node` lead to.
    let dependencies = tcx.dep_graph.get_dependencies_of(current_node);

    // Now check all the inputs for changes
    for dependency in dependencies {

        match tcx.dep_graph.get_node_color(dependency) {
            Green => {
                // This input has already been checked before and it has not
                // changed; so we can go on to check the next one
            }
            Red => {
                // We found an input that has changed. We cannot mark
                // `current_node` as green without re-running the
                // corresponding query.
                return false
            }
            Unknown => {
                // This is the first time we look at this node. Let's try
                // to mark it green by calling try_mark_green() recursively.
                if try_mark_green(tcx, dependency) {
                    // We successfully marked the input as green, on to the
                    // next.
                } else {
                    // We could *not* mark the input as green. This means we
                    // don't know if its value has changed. In order to find
                    // out, we re-run the corresponding query now!
                    tcx.run_query_for(dependency);

                    // Fetch and check the node color again. Running the query
                    // has forced it to either red (if it yielded a different
                    // result than we have in the cache) or green (if it
                    // yielded the same result).
                    match tcx.dep_graph.get_node_color(dependency) {
                        Red => {
                            // The input turned out to be red, so we cannot
                            // mark `current_node` as green.
                            return false
                        }
                        Green => {
                            // Re-running the query paid off! The result is the
                            // same as before, so this particular input does
                            // not invalidate `current_node`.
                        }
                        Unknown => {
                            // There is no way a node has no color after
                            // re-running the query.
                            panic!("unreachable")
                        }
                    }
                }
            }
        }
    }

    // If we have gotten through the entire loop, it means that all inputs
    // have turned out to be green. If all inputs are unchanged, it means
    // that the query result corresponding to `current_node` cannot have
    // changed either.
    tcx.dep_graph.mark_green(current_node);

    true
}

NOTE: The actual implementation can be found in compiler/rustc_query_system/src/dep_graph/graph.rs

By using red-green marking we can avoid the devastating cumulative effect of having false positives during change detection. Whenever a query is executed in incremental mode, we first check if its already green. If not, we run try_mark_green() on it. If it still isn't green after that, then we actually invoke the query provider to re-compute the result. Re-computing the query might then itself involve recursively invoking more queries, which can mean we come back to the try_mark_green() algorithm for the dependencies recursively.

The Real World: How Persistence Makes Everything Complicated

The sections above described the underlying algorithm for incremental compilation but because the compiler process exits after being finished and takes the query context with its result cache with it into oblivion, we have to persist data to disk, so the next compilation session can make use of it. This comes with a whole new set of implementation challenges:

  • The query result cache is stored to disk, so they are not readily available for change comparison.
  • A subsequent compilation session will start off with new version of the code that has arbitrary changes applied to it. All kinds of IDs and indices that are generated from a global, sequential counter (e.g. NodeId, DefId, etc) might have shifted, making the persisted results on disk not immediately usable anymore because the same numeric IDs and indices might refer to completely new things in the new compilation session.
  • Persisting things to disk comes at a cost, so not every tiny piece of information should be actually cached in between compilation sessions. Fixed-sized, plain-old-data is preferred to complex things that need to run through an expensive (de-)serialization step.

The following sections describe how the compiler solves these issues.

A Question Of Stability: Bridging The Gap Between Compilation Sessions

As noted before, various IDs (like DefId) are generated by the compiler in a way that depends on the contents of the source code being compiled. ID assignment is usually deterministic, that is, if the exact same code is compiled twice, the same things will end up with the same IDs. However, if something changes, e.g. a function is added in the middle of a file, there is no guarantee that anything will have the same ID as it had before.

As a consequence we cannot represent the data in our on-disk cache the same way it is represented in memory. For example, if we just stored a piece of type information like TyKind::FnDef(DefId, &'tcx Substs<'tcx>) (as we do in memory) and then the contained DefId points to a different function in a new compilation session we'd be in trouble.

The solution to this problem is to find "stable" forms for IDs which remain valid in between compilation sessions. For the most important case, DefIds, these are the so-called DefPaths. Each DefId has a corresponding DefPath but in place of a numeric ID, a DefPath is based on the path to the identified item, e.g. std::collections::HashMap. The advantage of an ID like this is that it is not affected by unrelated changes. For example, one can add a new function to std::collections but std::collections::HashMap would still be std::collections::HashMap. A DefPath is "stable" across changes made to the source code while a DefId isn't.

There is also the DefPathHash which is just a 128-bit hash value of the DefPath. The two contain the same information and we mostly use the DefPathHash because it simpler to handle, being Copy and self-contained.

This principle of stable identifiers is used to make the data in the on-disk cache resilient to source code changes. Instead of storing a DefId, we store the DefPathHash and when we deserialize something from the cache, we map the DefPathHash to the corresponding DefId in the current compilation session (which is just a simple hash table lookup).

The HirId, used for identifying HIR components that don't have their own DefId, is another such stable ID. It is (conceptually) a pair of a DefPath and a LocalId, where the LocalId identifies something (e.g. a hir::Expr) locally within its "owner" (e.g. a hir::Item). If the owner is moved around, the LocalIds within it are still the same.

Checking Query Results For Changes: HashStable And Fingerprints

In order to do red-green-marking we often need to check if the result of a query has changed compared to the result it had during the previous compilation session. There are two performance problems with this though:

  • We'd like to avoid having to load the previous result from disk just for doing the comparison. We already computed the new result and will use that. Also loading a result from disk will "pollute" the interners with data that is unlikely to ever be used.
  • We don't want to store each and every result in the on-disk cache. For example, it would be wasted effort to persist things to disk that are already available in upstream crates.

The compiler avoids these problems by using so-called Fingerprints. Each time a new query result is computed, the query engine will compute a 128 bit hash value of the result. We call this hash value "the Fingerprint of the query result". The hashing is (and has to be) done "in a stable way". This means that whenever something is hashed that might change in between compilation sessions (e.g. a DefId), we instead hash its stable equivalent (e.g. the corresponding DefPath). That's what the whole HashStable infrastructure is for. This way Fingerprints computed in two different compilation sessions are still comparable.

The next step is to store these fingerprints along with the dependency graph. This is cheap since fingerprints are just bytes to be copied. It's also cheap to load the entire set of fingerprints together with the dependency graph.

Now, when red-green-marking reaches the point where it needs to check if a result has changed, it can just compare the (already loaded) previous fingerprint to the fingerprint of the new result.

This approach works rather well but it's not without flaws:

  • There is a small possibility of hash collisions. That is, two different results could have the same fingerprint and the system would erroneously assume that the result hasn't changed, leading to a missed update.

    We mitigate this risk by using a high-quality hash function and a 128 bit wide hash value. Due to these measures the practical risk of a hash collision is negligible.

  • Computing fingerprints is quite costly. It is the main reason why incremental compilation can be slower than non-incremental compilation. We are forced to use a good and thus expensive hash function, and we have to map things to their stable equivalents while doing the hashing.

A Tale Of Two DepGraphs: The Old And The New

The initial description of dependency tracking glosses over a few details that quickly become a head scratcher when actually trying to implement things. In particular it's easy to overlook that we are actually dealing with two dependency graphs: The one we built during the previous compilation session and the one that we are building for the current compilation session.

When a compilation session starts, the compiler loads the previous dependency graph into memory as an immutable piece of data. Then, when a query is invoked, it will first try to mark the corresponding node in the graph as green. This means really that we are trying to mark the node in the previous dep-graph as green that corresponds to the query key in the current session. How do we do this mapping between current query key and previous DepNode? The answer is again Fingerprints: Nodes in the dependency graph are identified by a fingerprint of the query key. Since fingerprints are stable across compilation sessions, computing one in the current session allows us to find a node in the dependency graph from the previous session. If we don't find a node with the given fingerprint, it means that the query key refers to something that did not yet exist in the previous session.

So, having found the dep-node in the previous dependency graph, we can look up its dependencies (i.e. also dep-nodes in the previous graph) and continue with the rest of the try-mark-green algorithm. The next interesting thing happens when we successfully marked the node as green. At that point we copy the node and the edges to its dependencies from the old graph into the new graph. We have to do this because the new dep-graph cannot acquire the node and edges via the regular dependency tracking. The tracking system can only record edges while actually running a query -- but running the query, although we have the result already cached, is exactly what we want to avoid.

Once the compilation session has finished, all the unchanged parts have been copied over from the old into the new dependency graph, while the changed parts have been added to the new graph by the tracking system. At this point, the new graph is serialized out to disk, alongside the query result cache, and can act as the previous dep-graph in a subsequent compilation session.

Didn't You Forget Something?: Cache Promotion

The system described so far has a somewhat subtle property: If all inputs of a dep-node are green then the dep-node itself can be marked as green without computing or loading the corresponding query result. Applying this property transitively often leads to the situation that some intermediate results are never actually loaded from disk, as in the following example:

   input(A) <-- intermediate_query(B) <-- leaf_query(C)

The compiler might need the value of leaf_query(C) in order to generate some output artifact. If it can mark leaf_query(C) as green, it will load the result from the on-disk cache. The result of intermediate_query(B) is never loaded though. As a consequence, when the compiler persists the new result cache by writing all in-memory query results to disk, intermediate_query(B) will not be in memory and thus will be missing from the new result cache.

If there subsequently is another compilation session that actually needs the result of intermediate_query(B) it will have to be re-computed even though we had a perfectly valid result for it in the cache just before.

In order to prevent this from happening, the compiler does something called "cache promotion": Before emitting the new result cache it will walk all green dep-nodes and make sure that their query result is loaded into memory. That way the result cache doesn't unnecessarily shrink again.

Incremental Compilation and the Compiler Backend

The compiler backend, the part involving LLVM, is using the query system but it is not implemented in terms of queries itself. As a consequence it does not automatically partake in dependency tracking. However, the manual integration with the tracking system is pretty straight-forward. The compiler simply tracks what queries get invoked when generating the initial LLVM version of each codegen unit (CGU), which results in a dep-node for each CGU. In subsequent compilation sessions it then tries to mark the dep-node for a CGU as green. If it succeeds, it knows that the corresponding object and bitcode files on disk are still valid. If it doesn't succeed, the entire CGU has to be recompiled.

This is the same approach that is used for regular queries. The main differences are:

  • that we cannot easily compute a fingerprint for LLVM modules (because they are opaque C++ objects),

  • that the logic for dealing with cached values is rather different from regular queries because here we have bitcode and object files instead of serialized Rust values in the common result cache file, and

  • the operations around LLVM are so expensive in terms of computation time and memory consumption that we need to have tight control over what is executed when and what stays in memory for how long.

The query system could probably be extended with general purpose mechanisms to deal with all of the above but so far that seemed like more trouble than it would save.

Query Modifiers

The query system allows for applying modifiers to queries. These modifiers affect certain aspects of how the system treats the query with respect to incremental compilation:

  • eval_always - A query with the eval_always attribute is re-executed unconditionally during incremental compilation. I.e. the system will not even try to mark the query's dep-node as green. This attribute has two use cases:

    • eval_always queries can read inputs (from files, global state, etc). They can also produce side effects like writing to files and changing global state.

    • Some queries are very likely to be re-evaluated because their result depends on the entire source code. In this case eval_always can be used as an optimization because the system can skip recording dependencies in the first place.

  • no_hash - Applying no_hash to a query tells the system to not compute the fingerprint of the query's result. This has two consequences:

    • Not computing the fingerprint can save quite a bit of time because fingerprinting is expensive, especially for large, complex values.

    • Without the fingerprint, the system has to unconditionally assume that the result of the query has changed. As a consequence anything depending on a no_hash query will always be re-executed.

    Using no_hash for a query can make sense in two circumstances:

    • If the result of the query is very likely to change whenever one of its inputs changes, e.g. a function like |a, b, c| -> (a * b * c). In such a case recomputing the query will always yield a red node if one of the inputs is red so we can spare us the trouble and default to red immediately. A counter example would be a function like |a| -> (a == 42) where the result does not change for most changes of a.

    • If the result of a query is a big, monolithic collection (e.g. index_hir) and there are "projection queries" reading from that collection (e.g. hir_owner). In such a case the big collection will likely fulfill the condition above (any changed input means recomputing the whole collection) and the results of the projection queries will be hashed anyway. If we also hashed the collection query it would mean that we effectively hash the same data twice: once when hashing the collection and another time when hashing all the projection query results. no_hash allows us to avoid that redundancy and the projection queries act as a "firewall", shielding their dependents from the unconditionally red no_hash node.

  • cache_on_disk_if - This attribute is what determines which query results are persisted in the incremental compilation query result cache. The attribute takes an expression that allows per query invocation decisions. For example, it makes no sense to store values from upstream crates in the cache because they are already available in the upstream crate's metadata.

  • anon - This attribute makes the system use "anonymous" dep-nodes for the given query. An anonymous dep-node is not identified by the corresponding query key, instead its ID is computed from the IDs of its dependencies. This allows the red-green system to do its change detection even if there is no query key available for a given dep-node -- something which is needed for handling trait selection because it is not based on queries.

The Projection Query Pattern

It's interesting to note that eval_always and no_hash can be used together in the so-called "projection query" pattern. It is often the case that there is one query that depends on the entirety of the compiler's input (e.g. the indexed HIR) and another query that projects individual values out of this monolithic value (e.g. a HIR item with a certain DefId). These projection queries allow for building change propagation "firewalls" because even if the result of the monolithic query changes (which it is very likely to do) the small projections can still mostly be marked as green.

  +------------+
  |            |           +---------------+           +--------+
  |            | <---------| projection(x) | <---------| foo(a) |
  |            |           +---------------+           +--------+
  |            |
  | monolithic |           +---------------+           +--------+
  |   query    | <---------| projection(y) | <---------| bar(b) |
  |            |           +---------------+           +--------+
  |            |
  |            |           +---------------+           +--------+
  |            | <---------| projection(z) | <---------| baz(c) |
  |            |           +---------------+           +--------+
  +------------+

Let's assume that the result monolithic_query changes so that also the result of projection(x) has changed, i.e. both their dep-nodes are being marked as red. As a consequence foo(a) needs to be re-executed; but bar(b) and baz(c) can be marked as green. However, if foo, bar, and baz would have directly depended on monolithic_query then all of them would have had to be re-evaluated.

This pattern works even without eval_always and no_hash but the two modifiers can be used to avoid unnecessary overhead. If the monolithic query is likely to change at any minor modification of the compiler's input it makes sense to mark it as eval_always, thus getting rid of its dependency tracking cost. And it always makes sense to mark the monolithic query as no_hash because we have the projections to take care of keeping things green as much as possible.

Shortcomings of the Current System

There are many things that still can be improved.

Incrementality of on-disk data structures

The current system is not able to update on-disk caches and the dependency graph in-place. Instead it has to rewrite each file entirely in each compilation session. The overhead of doing so is a few percent of total compilation time.

Unnecessary data dependencies

Data structures used as query results could be factored in a way that removes edges from the dependency graph. Especially "span" information is very volatile, so including it in query result will increase the chance that the result won't be reusable. See https://github.com/rust-lang/rust/issues/47389 for more information.