MIR optimizations are optimizations run on the MIR to produce better MIR before codegen. This is important for two reasons: first, it makes the final generated executable code better, and second, it means that LLVM has less work to do, so compilation is faster. Note that since MIR is generic (not monomorphized yet), these optimizations are particularly effective; we can optimize the generic version, so all of the monomorphizations are cheaper!
MIR optimizations run after borrow checking. We run a series of optimization
passes over the MIR to improve it. Some passes are required to run on all code,
some passes don't actually do optimizations but only check stuff, and some
passes are only turned on in
optimized_mir query is called to produce the optimized MIR
for a given
DefId. This query makes sure that the borrow checker has
run and that some validation has occurred. Then, it steals the MIR,
optimizes it, and returns the improved MIR.
Make a Rust source file in
src/test/mir-optthat shows the code you want to optimize. This should be kept simple, so avoid
println!or other formatting code if it's not necessary for the optimization. The reason for this is that
format!, etc. generate a lot of MIR that can make it harder to understand what the optimization does to the test.
./x.py test --bless src/test/mir-opt/<your-test>.rsto generate a MIR dump. Read this README for instructions on how to dump things.
Commit the current working directory state. The reason you should commit the test output before you implement the optimization is so that you (and your reviewers) can see a before/after diff of what the optimization changed.
Implement a new optimization in
compiler/rustc_mir_transform/src. The fastest and easiest way to do this is to
./x.py test --bless src/test/mir-opt/<your-test>.rsto regenerate the MIR dumps. Look at the diffs to see if they are what you expect.
./x.py test src/test/uito see if your optimization broke anything.
If there are issues with your optimization, experiment with it a bit and repeat steps 5 and 6.
Commit and open a PR. You can do this at any point, even if things aren't working yet, so that you can ask for feedback on the PR. Open a "WIP" PR (just prefix your PR title with
[WIP]or otherwise note that it is a work in progress) in that case.
Make sure to commit the blessed test output as well! It's necessary for CI to pass and it's very helpful to reviewers.
If you have any questions along the way, feel free to ask in
#t-compiler/wg-mir-opt on Zulip.
The list of passes run and the order in which they are run is defined by the
run_optimization_passes function. It contains an array of passes to
run. Each pass in the array is a struct that implements the
The array is an array of
&dyn MirPass trait objects. Typically, a pass is
implemented in its own module of the
Some examples of passes are:
CleanupNonCodegenStatements: remove some of the info that is only needed for analyses, rather than codegen.
ConstProp: Does constant propagation
You can see the "Implementors" section of the
MirPass rustdocs for more examples.
MIR optimizations can come in various levels of readiness. Experimental
optimizations may cause miscompilations, or slow down compile times.
These passes are still included in nightly builds to gather feedback and make it easier to modify
the pass. To enable working with slow or otherwise experimental optimization passes,
you can specify the
-Z mir-opt-level debug flag. You can find the
definitions of the levels in the compiler MCP. If you are developing a MIR pass and
want to query whether your optimization pass should run, you can check the
current level using
Optimization fuel is a compiler option (
-Z fuel=<crate>=<value>) that allows for fine grained
control over which optimizations can be applied during compilation: each optimization reduces
fuel by 1, and when fuel reaches 0 no more optimizations are applied. The primary use of fuel
is debugging optimizations that may be incorrect or misapplied. By changing the fuel
value, you can bisect a compilation session down to the exact incorrect optimization
(this behaves like a kind of binary search through the optimizations).
MIR optimizations respect fuel, and in general each pass should check fuel by calling
tcx.consider_optimizing and skipping the optimization if fuel
is empty. There are a few considerations:
- If the pass is considered "guaranteed" (for example, it should always be run because it is
needed for correctness), then fuel should not be used. An example of this is
- In some cases, an initial pass is performed to gather candidates, which are then iterated to
perform optimizations. In these situations, we should allow for the initial gathering pass
and then check fuel as close to the mutation as possible. This allows for the best
debugging experience, because we can determine where in the list of candidates an optimization
may have been misapplied. Examples of this are