Replace `(Body, DefId)` with `Body` where possible
Follow-up to #77430.
I `grep`-ed for parameter lists in which a `Body` appeared within a few lines of a `DefId`, so it's possible that I missed some cases, but this should be pretty complete. Most of these changes were mechanical, but there's a few places where I started calling things "caller" and "callee" when multiple `DefId`s were in-scope at once. Also, we should probably have a helper function on `Body` that returns a `LocalDefId`. I can do that in this PR or in a follow-up.
Allow a unique name to be assigned to dataflow graphviz output
Previously, if the same analysis were invoked multiple times in a single compilation session, the graphviz output for later runs would overwrite that of previous runs. Allow callers to add a unique identifier to each run so this can be avoided.
Don't compile regex at every function call.
Use `SyncOnceCell` to only compile it once.
I believe this still adds some kind of locking mechanism?
Related issue: https://github.com/rust-lang/rust/issues/76817
Implement a generic Destination Propagation optimization on MIR
This takes the work that was originally started by `@eddyb` in https://github.com/rust-lang/rust/pull/47954, and then explored by me in https://github.com/rust-lang/rust/pull/71003, and implements it in a general (ie. not limited to acyclic CFGs) and dataflow-driven way (so that no additional infrastructure in rustc is needed).
The pass is configured to run at `mir-opt-level=2` and higher only. To enable it by default, some followup work on it is still needed:
* Performance needs to be evaluated. I did some light optimization work and tested against `tuple-stress`, which caused trouble in my last attempt, but didn't go much in depth here.
* We can also enable the pass only at `opt-level=2` and higher, if it is too slow to run in debug mode, but fine when optimizations run anyways.
* Debuginfo needs to be fixed after locals are merged. I did not look into what is required for this.
* Live ranges of locals (aka `StorageLive` and `StorageDead`) are currently deleted. We either need to decide that this is fine, or if not, merge the variable's live ranges (or remove these statements entirely – https://github.com/rust-lang/rust/issues/68622).
Some benchmarks of the pass were done in https://github.com/rust-lang/rust/pull/72635.
Make graphviz font configurable
Alternative to PR #76776.
To change the graphviz output to use an alternative `fontname` value,
add a command line option like: `rustc --graphviz-font=monospace`.
r? @ecstatic-morse
Alternative to PR ##76776.
To change the graphviz output to use an alternative `fontname` value,
add a command line option like: `rustc --graphviz-font=monospace`.
Many developers use a dark theme with editors and IDEs, but this
typically doesn't extend to graphviz output.
When I bring up a MIR graphviz document, the white background is
strikingly bright. This new option changes the colors used for graphviz
output to work better in dark-themed UIs.
Support dataflow problems on arbitrary lattices
This PR implements last of the proposed extensions I mentioned in the design meeting for the original dataflow refactor. It extends the current dataflow framework to work with arbitrary lattices, not just `BitSet`s. This is a prerequisite for dataflow-enabled MIR const-propagation. Personally, I am skeptical of the usefulness of doing const-propagation pre-monomorphization, since many useful constants only become known after monomorphization (e.g. `size_of::<T>()`) and users have a natural tendency to hand-optimize the rest. It's probably worth exprimenting with, however, and others have shown interest cc `@rust-lang/wg-mir-opt.`
The `Idx` associated type is moved from `AnalysisDomain` to `GenKillAnalysis` and replaced with an associated `Domain` type that must implement `JoinSemiLattice`. Like before, each `Analysis` defines the "bottom value" for its domain, but can no longer override the dataflow join operator. Analyses that want to use set intersection must now use the `lattice::Dual` newtype. `GenKillAnalysis` impls have an additional requirement that `Self::Domain: BorrowMut<BitSet<Self::Idx>>`, which effectively means that they must use `BitSet<Self::Idx>` or `lattice::Dual<BitSet<Self::Idx>>` as their domain.
Most of these changes were mechanical. However, because a `Domain` is no longer always a powerset of some index type, we can no longer use an `IndexVec<BasicBlock, GenKillSet<A::Idx>>>` to store cached block transfer functions. Instead, we use a boxed `dyn Fn` trait object. I discuss a few alternatives to the current approach in a commit message.
The majority of new lines of code are to preserve existing Graphviz diagrams for those unlucky enough to have to debug dataflow analyses. I find these diagrams incredibly useful when things are going wrong and considered regressing them unacceptable, especially the pretty-printing of `MovePathIndex`s, which are used in many dataflow analyses. This required a parallel `fmt` trait used only for printing dataflow domains, as well as a refactoring of the `graphviz` module now that we cannot expect the domain to be a `BitSet`. Some features did have to be removed, such as the gen/kill display mode (which I didn't use but existed to mirror the output of the old dataflow framework) and line wrapping. Since I had to rewrite much of it anyway, I took the opportunity to switch to a `Visitor` for printing dataflow state diffs instead of using cursors, which are error prone for code that must be generic over both forward and backward analyses. As a side-effect of this change, we no longer have quadratic behavior when writing graphviz diagrams for backward dataflow analyses.
r? `@pnkfelix`
I've tried a few ways of implementing this, but each fell short.
Adding an auxiliary `_Idx` associated type to `Analysis` that defaults
to `!` but is overridden in the blanket impl of `Analysis` for `A:
GenKillAnalysis` to `A::Idx` seems promising, but the trait solver is
unable to prove equivalence between `A::Idx` and `A::_Idx` within the
overridden version of `into_engine`. Without full-featured
specialization, removing `into_engine` or splitting it into a different
trait would have a significant ergonomic penalty.
Alternatively, we could erase the index type and store a
`GenKillSet<u32>` as well as a function pointer for transmuting between
`&mut A::Domain` and `&mut BitSet<u32>` in the hopes that LLVM can
devirtualize a simple function pointer better than the boxed closure.
However, this is brittle, requires `unsafe` code, and doesn't work for
index types that aren't the same size as a `u32` (e.g. `usize`) since
`GenKillSet` stores a `HybridBitSet`, which may be a `Vec<I>`. Perhaps
safe transmute could help here?
