llvm-capstone/mlir/docs/PassManagement.md
Mehdi Amini 7f00ba08aa Restore mlir-opt --run-reproducer option to opt-in running a reproducer
When tooling out there produces a reproducer that is archived, the first thing
a user is likely to expect is to process this as they do with any MLIR file.
However https://reviews.llvm.org/D126447 changed the behavior of mlir-opt to
eliminate the `--run-reproducer` option and instead automatically run it when
present in the input file. This creates a discrepancy in how mlir-opt behaves
when fed with an input file, and is surprising for users.
The explicit passing of `--run-reproducer` to express user-intent seems more
in line with what is expected from `mlir-opt`.

Reviewed By: rriddle, jpienaar

Differential Revision: https://reviews.llvm.org/D149820
2023-05-04 11:08:03 -07:00

53 KiB

Pass Infrastructure

[TOC]

Passes represent the basic infrastructure for transformation and optimization. This document provides an overview of the pass infrastructure in MLIR and how to use it.

See MLIR specification for more information about MLIR and its core aspects, such as the IR structure and operations.

See MLIR Rewrites for a quick start on graph rewriting in MLIR. If a transformation involves pattern matching operation DAGs, this is a great place to start.

Operation Pass

In MLIR, the main unit of abstraction and transformation is an operation. As such, the pass manager is designed to work on instances of operations at different levels of nesting. The structure of the pass manager, and the concept of nesting, is detailed further below. All passes in MLIR derive from OperationPass and adhere to the following restrictions; any noncompliance will lead to problematic behavior in multithreaded and other advanced scenarios:

  • Must not modify any state referenced or relied upon outside the current operation being operated on. This includes adding or removing operations from the parent block, changing the attributes(depending on the contract of the current operation)/operands/results/successors of the current operation.
  • Must not modify the state of another operation not nested within the current operation being operated on.
    • Other threads may be operating on these operations simultaneously.
  • Must not inspect the state of sibling operations.
    • Other threads may be modifying these operations in parallel.
    • Inspecting the state of ancestor/parent operations is permitted.
  • Must not maintain mutable pass state across invocations of runOnOperation. A pass may be run on many different operations with no guarantee of execution order.
    • When multithreading, a specific pass instance may not even execute on all operations within the IR. As such, a pass should not rely on running on all operations.
  • Must not maintain any global mutable state, e.g. static variables within the source file. All mutable state should be maintained by an instance of the pass.
  • Must be copy-constructible
    • Multiple instances of the pass may be created by the pass manager to process operations in parallel.

Op-Agnostic Operation Passes

By default, an operation pass is op-agnostic, meaning that it operates on the operation type of the pass manager that it is added to. This means a pass may operate on many different types of operations. Agnostic passes should be written such that they do not make assumptions on the operation they run on. Examples of this type of pass are Canonicalization Common Sub-Expression Elimination.

To create an agnostic operation pass, a derived class must adhere to the following:

  • Inherit from the CRTP class OperationPass.
  • Override the virtual void runOnOperation() method.

A simple pass may look like:

/// Here we utilize the CRTP `PassWrapper` utility class to provide some
/// necessary utility hooks. This is only necessary for passes defined directly
/// in C++. Passes defined declaratively use a cleaner mechanism for providing
/// these utilities.
struct MyOperationPass : public PassWrapper<MyOperationPass, OperationPass<>> {
  void runOnOperation() override {
    // Get the current operation being operated on.
    Operation *op = getOperation();
    ...
  }
};

Filtered Operation Pass

If a pass needs to constrain its execution to specific types or classes of operations, additional filtering may be applied on top. This transforms a once agnostic pass into one more specific to a certain context. There are various ways in which to filter the execution of a pass, and different contexts in which filtering may apply:

Operation Pass: Static Schedule Filtering

Static filtering allows for applying additional constraints on the operation types a pass may be scheduled on. This type of filtering generally allows for building more constrained passes that can only be scheduled on operations that satisfy the necessary constraints. For example, this allows for specifying passes that only run on operations of a certain, those that provide a certain interface, trait, or some other constraint that applies to all instances of that operation type. Below is an example of a pass that only permits scheduling on operations that implement FunctionOpInterface:

struct MyFunctionPass : ... {
  /// This method is used to provide additional static filtering, and returns if the
  /// pass may be scheduled on the given operation type.
  bool canScheduleOn(RegisteredOperationName opInfo) const override {
    return opInfo.hasInterface<FunctionOpInterface>();
  }

  void runOnOperation() {
    // Here we can freely cast to FunctionOpInterface, because our `canScheduleOn` ensures
    // that our pass is only executed on operations implementing that interface.
    FunctionOpInterface op = cast<FunctionOpInterface>(getOperation()); 
  }
};

When a pass with static filtering is added to an op-specific pass manager, it asserts that the operation type of the pass manager satisfies the static constraints of the pass. When added to an op-agnostic pass manager, that pass manager, and all passes contained within, inherits the static constraints of the pass. For example, if the pass filters on FunctionOpInterface, as in the MyFunctionPass example above, only operations that implement FunctionOpInterface will be considered when executing any passes within the pass manager. This invariant is important to keep in mind, as each pass added to an op-agnostic pass manager further constrains the operations that may be scheduled on it. Consider the following example:

func.func @foo() {
  // ...
  return
}

module @someModule {
  // ...
}

If we were to apply the op-agnostic pipeline, any(cse,my-function-pass), to the above MLIR snippet it would only run on the foo function operation. This is because the my-function-pass has a static filtering constraint to only schedule on operations implementing FunctionOpInterface. Remember that this constraint is inherited by the entire pass manager, so we never consider someModule for any of the passes, including cse which normally can be scheduled on any operation.

Operation Pass: Static Filtering By Op Type

In the above section, we detailed a general mechanism for statically filtering the types of operations that a pass may be scheduled on. Sugar is provided on top of that mechanism to simplify the definition of passes that are restricted to scheduling on a single operation type. In these cases, a pass simply needs to provide the type of operation to the OperationPass base class. This will automatically instill filtering on that operation type:

/// Here we utilize the CRTP `PassWrapper` utility class to provide some
/// necessary utility hooks. This is only necessary for passes defined directly
/// in C++. Passes defined declaratively use a cleaner mechanism for providing
/// these utilities.
struct MyFunctionPass : public PassWrapper<MyOperationPass, OperationPass<func::FuncOp>> {
  void runOnOperation() {
    // Get the current operation being operated on.
    func::FuncOp op = getOperation();
  }
};

Operation Pass: Static Filtering By Interface

In the above section, we detailed a general mechanism for statically filtering the types of operations that a pass may be scheduled on. Sugar is provided on top of that mechanism to simplify the definition of passes that are restricted to scheduling on a specific operation interface. In these cases, a pass simply needs to inherit from the InterfacePass base class. This class is similar to OperationPass, but expects the type of interface to operate on. This will automatically instill filtering on that interface type:

/// Here we utilize the CRTP `PassWrapper` utility class to provide some
/// necessary utility hooks. This is only necessary for passes defined directly
/// in C++. Passes defined declaratively use a cleaner mechanism for providing
/// these utilities.
struct MyFunctionPass : public PassWrapper<MyOperationPass, InterfacePass<FunctionOpInterface>> {
  void runOnOperation() {
    // Get the current operation being operated on.
    FunctionOpInterface op = getOperation();
  }
};

Dependent Dialects

Dialects must be loaded in the MLIRContext before entities from these dialects (operations, types, attributes, ...) can be created. Dialects must also be loaded before starting the execution of a multi-threaded pass pipeline. To this end, a pass that may create an entity from a dialect that isn't guaranteed to already be loaded must express this by overriding the getDependentDialects() method and declare this list of Dialects explicitly. See also the dependentDialects field in the TableGen Specification.

Initialization

In certain situations, a Pass may contain state that is constructed dynamically, but is potentially expensive to recompute in successive runs of the Pass. One such example is when using PDL-based patterns, which are compiled into a bytecode during runtime. In these situations, a pass may override the following hook to initialize this heavy state:

  • LogicalResult initialize(MLIRContext *context)

This hook is executed once per run of a full pass pipeline, meaning that it does not have access to the state available during a runOnOperation call. More concretely, all necessary accesses to an MLIRContext should be driven via the provided context parameter, and methods that utilize "per-run" state such as getContext/getOperation/getAnalysis/etc. must not be used. In case of an error during initialization, the pass is expected to emit an error diagnostic and return a failure() which will abort the pass pipeline execution.

Analysis Management

An important concept, along with transformation passes, are analyses. These are conceptually similar to transformation passes, except that they compute information on a specific operation without modifying it. In MLIR, analyses are not passes but free-standing classes that are computed lazily on-demand and cached to avoid unnecessary recomputation. An analysis in MLIR must adhere to the following:

  • Provide a valid constructor taking either an Operation* or Operation* and AnalysisManager &.
    • The provided AnalysisManager & should be used to query any necessary analysis dependencies.
  • Must not modify the given operation.

An analysis may provide additional hooks to control various behavior:

  • bool isInvalidated(const AnalysisManager::PreservedAnalyses &)

Given a preserved analysis set, the analysis returns true if it should truly be invalidated. This allows for more fine-tuned invalidation in cases where an analysis wasn't explicitly marked preserved, but may be preserved (or invalidated) based upon other properties such as analyses sets. If the analysis uses any other analysis as a dependency, it must also check if the dependency was invalidated.

Querying Analyses

The base OperationPass class provides utilities for querying and preserving analyses for the current operation being processed.

  • OperationPass automatically provides the following utilities for querying analyses:
    • getAnalysis<>
      • Get an analysis for the current operation, constructing it if necessary.
    • getCachedAnalysis<>
      • Get an analysis for the current operation, if it already exists.
    • getCachedParentAnalysis<>
      • Get an analysis for a given parent operation, if it exists.
    • getCachedChildAnalysis<>
      • Get an analysis for a given child operation, if it exists.
    • getChildAnalysis<>
      • Get an analysis for a given child operation, constructing it if necessary.

Using the example passes defined above, let's see some examples:

/// An interesting analysis.
struct MyOperationAnalysis {
  // Compute this analysis with the provided operation.
  MyOperationAnalysis(Operation *op);
};

struct MyOperationAnalysisWithDependency {
  MyOperationAnalysisWithDependency(Operation *op, AnalysisManager &am) {
    // Request other analysis as dependency
    MyOperationAnalysis &otherAnalysis = am.getAnalysis<MyOperationAnalysis>();
    ...
  }

  bool isInvalidated(const AnalysisManager::PreservedAnalyses &pa) {
    // Check if analysis or its dependency were invalidated
    return !pa.isPreserved<MyOperationAnalysisWithDependency>() ||
           !pa.isPreserved<MyOperationAnalysis>();
  }
};

void MyOperationPass::runOnOperation() {
  // Query MyOperationAnalysis for the current operation.
  MyOperationAnalysis &myAnalysis = getAnalysis<MyOperationAnalysis>();

  // Query a cached instance of MyOperationAnalysis for the current operation.
  // It will not be computed if it doesn't exist.
  auto optionalAnalysis = getCachedAnalysis<MyOperationAnalysis>();
  if (optionalAnalysis)
    ...

  // Query a cached instance of MyOperationAnalysis for the parent operation of
  // the current operation. It will not be computed if it doesn't exist.
  auto optionalAnalysis = getCachedParentAnalysis<MyOperationAnalysis>();
  if (optionalAnalysis)
    ...
}

Preserving Analyses

Analyses that are constructed after being queried by a pass are cached to avoid unnecessary computation if they are requested again later. To avoid stale analyses, all analyses are assumed to be invalidated by a pass. To avoid invalidation, a pass must specifically mark analyses that are known to be preserved.

  • All Pass classes automatically provide the following utilities for preserving analyses:
    • markAllAnalysesPreserved
    • markAnalysesPreserved<>
void MyOperationPass::runOnOperation() {
  // Mark all analyses as preserved. This is useful if a pass can guarantee
  // that no transformation was performed.
  markAllAnalysesPreserved();

  // Mark specific analyses as preserved. This is used if some transformation
  // was performed, but some analyses were either unaffected or explicitly
  // preserved.
  markAnalysesPreserved<MyAnalysis, MyAnalyses...>();
}

Pass Failure

Passes in MLIR are allowed to gracefully fail. This may happen if some invariant of the pass was broken, potentially leaving the IR in some invalid state. If such a situation occurs, the pass can directly signal a failure to the pass manager via the signalPassFailure method. If a pass signaled a failure when executing, no other passes in the pipeline will execute and the top-level call to PassManager::run will return failure.

void MyOperationPass::runOnOperation() {
  // Signal failure on a broken invariant.
  if (some_broken_invariant)
    return signalPassFailure();
}

Pass Manager

The above sections introduced the different types of passes and their invariants. This section introduces the concept of a PassManager, and how it can be used to configure and schedule a pass pipeline. There are two main classes related to pass management, the PassManager and the OpPassManager. The PassManager class acts as the top-level entry point, and contains various configurations used for the entire pass pipeline. The OpPassManager class is used to schedule passes to run at a specific level of nesting. The top-level PassManager also functions as an OpPassManager.

OpPassManager

An OpPassManager is essentially a collection of passes anchored to execute on operations at a given level of nesting. A pass manager may be op-specific (anchored on a specific operation type), or op-agnostic (not restricted to any specific operation, and executed on any viable operation type). Operation types that anchor pass managers must adhere to the following requirement:

  • Must be registered and marked IsolatedFromAbove.

    • Passes are expected not to modify operations at or above the current operation being processed. If the operation is not isolated, it may inadvertently modify or traverse the SSA use-list of an operation it is not supposed to.

Passes can be added to a pass manager via addPass.

An OpPassManager is generally created by explicitly nesting a pipeline within another existing OpPassManager via the nest<OpT> or nestAny methods. The former method takes the operation type that the nested pass manager will operate on. The latter method nests an op-agnostic pass manager, that may run on any viable operation type. Nesting in this sense, corresponds to the structural nesting within Regions of the IR.

For example, the following .mlir:

module {
  spirv.module "Logical" "GLSL450" {
    func @foo() {
      ...
    }
  }
}

Has the nesting structure of:

`builtin.module`
  `spirv.module`
    `spirv.func`

Below is an example of constructing a pipeline that operates on the above structure:

// Create a top-level `PassManager` class.
auto pm = PassManager::on<ModuleOp>(ctx);

// Add a pass on the top-level module operation.
pm.addPass(std::make_unique<MyModulePass>());

// Nest a pass manager that operates on `spirv.module` operations nested
// directly under the top-level module.
OpPassManager &nestedModulePM = pm.nest<spirv::ModuleOp>();
nestedModulePM.addPass(std::make_unique<MySPIRVModulePass>());

// Nest a pass manager that operates on functions within the nested SPIRV
// module.
OpPassManager &nestedFunctionPM = nestedModulePM.nest<func::FuncOp>();
nestedFunctionPM.addPass(std::make_unique<MyFunctionPass>());

// Nest an op-agnostic pass manager. This will operate on any viable
// operation, e.g. func.func, spirv.func, spirv.module, builtin.module, etc.
OpPassManager &nestedAnyPM = nestedModulePM.nestAny();
nestedAnyPM.addPass(createCanonicalizePass());
nestedAnyPM.addPass(createCSEPass());

// Run the pass manager on the top-level module.
ModuleOp m = ...;
if (failed(pm.run(m)))
    ... // One of the passes signaled a failure.

The above pass manager contains the following pipeline structure:

OpPassManager<ModuleOp>
  MyModulePass
  OpPassManager<spirv::ModuleOp>
    MySPIRVModulePass
    OpPassManager<func::FuncOp>
      MyFunctionPass
    OpPassManager<>
      Canonicalizer
      CSE

These pipelines are then run over a single operation at a time. This means that, for example, given a series of consecutive passes on func::FuncOp, it will execute all on the first function, then all on the second function, etc. until the entire program has been run through the passes. This provides several benefits:

  • This improves the cache behavior of the compiler, because it is only touching a single function at a time, instead of traversing the entire program.
  • This improves multi-threading performance by reducing the number of jobs that need to be scheduled, as well as increasing the efficiency of each job. An entire function pipeline can be run on each function asynchronously.

Dynamic Pass Pipelines

In some situations it may be useful to run a pass pipeline within another pass, to allow configuring or filtering based on some invariants of the current operation being operated on. For example, the Inliner Pass may want to run intraprocedural simplification passes while it is inlining to produce a better cost model, and provide more optimal inlining. To enable this, passes may run an arbitrary OpPassManager on the current operation being operated on or any operation nested within the current operation via the LogicalResult Pass::runPipeline(OpPassManager &, Operation *) method. This method returns whether the dynamic pipeline succeeded or failed, similarly to the result of the top-level PassManager::run method. A simple example is shown below:

void MyModulePass::runOnOperation() {
  ModuleOp module = getOperation();
  if (hasSomeSpecificProperty(module)) {
    OpPassManager dynamicPM("builtin.module");
    ...; // Build the dynamic pipeline.
    if (failed(runPipeline(dynamicPM, module)))
      return signalPassFailure();
  }
}

Note: though above the dynamic pipeline was constructed within the runOnOperation method, this is not necessary and pipelines should be cached when possible as the OpPassManager class can be safely copy constructed.

The mechanism described in this section should be used whenever a pass pipeline should run in a nested fashion, i.e. when the nested pipeline cannot be scheduled statically along with the rest of the main pass pipeline. More specifically, a PassManager should generally never need to be constructed within a Pass. Using runPipeline also ensures that all analyses, instrumentations, and other pass manager related components are integrated with the dynamic pipeline being executed.

Instance Specific Pass Options

MLIR provides a builtin mechanism for passes to specify options that configure its behavior. These options are parsed at pass construction time independently for each instance of the pass. Options are defined using the Option<> and ListOption<> classes, and generally follow the LLVM command line flag definition rules. One major distinction from the LLVM command line functionality is that all ListOptions are comma-separated, and delimited sub-ranges within individual elements of the list may contain commas that are not treated as separators for the top-level list.

struct MyPass ... {
  /// Make sure that we have a valid default constructor and copy constructor to
  /// ensure that the options are initialized properly.
  MyPass() = default;
  MyPass(const MyPass& pass) {}

  /// Any parameters after the description are forwarded to llvm::cl::list and
  /// llvm::cl::opt respectively.
  Option<int> exampleOption{*this, "flag-name", llvm::cl::desc("...")};
  ListOption<int> exampleListOption{*this, "list-flag-name", llvm::cl::desc("...")};
};

For pass pipelines, the PassPipelineRegistration templates take an additional template parameter for an optional Option struct definition. This struct should inherit from mlir::PassPipelineOptions and contain the desired pipeline options. When using PassPipelineRegistration, the constructor now takes a function with the signature void (OpPassManager &pm, const MyPipelineOptions&) which should construct the passes from the options and pass them to the pm:

struct MyPipelineOptions : public PassPipelineOptions {
  // The structure of these options is the same as those for pass options.
  Option<int> exampleOption{*this, "flag-name", llvm::cl::desc("...")};
  ListOption<int> exampleListOption{*this, "list-flag-name",
                                    llvm::cl::desc("...")};
};

void registerMyPasses() {
  PassPipelineRegistration<MyPipelineOptions>(
    "example-pipeline", "Run an example pipeline.",
    [](OpPassManager &pm, const MyPipelineOptions &pipelineOptions) {
      // Initialize the pass manager.
    });
}

Pass Statistics

Statistics are a way to keep track of what the compiler is doing and how effective various transformations are. It is often useful to see what effect specific transformations have on a particular input, and how often they trigger. Pass statistics are specific to each pass instance, which allow for seeing the effect of placing a particular transformation at specific places within the pass pipeline. For example, they help answer questions like "What happens if I run CSE again here?".

Statistics can be added to a pass by using the 'Pass::Statistic' class. This class takes as a constructor arguments: the parent pass, a name, and a description. This class acts like an atomic unsigned integer, and may be incremented and updated accordingly. These statistics rely on the same infrastructure as llvm::Statistic and thus have similar usage constraints. Collected statistics can be dumped by the pass manager programmatically via PassManager::enableStatistics; or via -mlir-pass-statistics and -mlir-pass-statistics-display on the command line.

An example is shown below:

struct MyPass ... {
  /// Make sure that we have a valid default constructor and copy constructor to
  /// ensure that the options are initialized properly.
  MyPass() = default;
  MyPass(const MyPass& pass) {}
  StringRef getArgument() const final {
    // This is the argument used to refer to the pass in
    // the textual format (on the commandline for example).
    return "argument";
  }
  StringRef getDescription() const final {
    // This is a brief description of the pass.
    return  "description";
  }
  /// Define the statistic to track during the execution of MyPass.
  Statistic exampleStat{this, "exampleStat", "An example statistic"};

  void runOnOperation() {
    ...

    // Update the statistic after some invariant was hit.
    ++exampleStat;

    ...
  }
};

The collected statistics may be aggregated in two types of views:

A pipeline view that models the structure of the pass manager, this is the default view:

$ mlir-opt -pass-pipeline='any(func.func(my-pass,my-pass))' foo.mlir -mlir-pass-statistics

===-------------------------------------------------------------------------===
                         ... Pass statistics report ...
===-------------------------------------------------------------------------===
'func.func' Pipeline
  MyPass
    (S) 15 exampleStat - An example statistic
  VerifierPass
  MyPass
    (S)  6 exampleStat - An example statistic
  VerifierPass
VerifierPass

A list view that aggregates the statistics of all instances of a specific pass together:

$ mlir-opt -pass-pipeline='any(func.func(my-pass,my-pass))' foo.mlir -mlir-pass-statistics -mlir-pass-statistics-display=list

===-------------------------------------------------------------------------===
                         ... Pass statistics report ...
===-------------------------------------------------------------------------===
MyPass
  (S) 21 exampleStat - An example statistic

Pass Registration

Briefly shown in the example definitions of the various pass types is the PassRegistration class. This mechanism allows for registering pass classes so that they may be created within a textual pass pipeline description. An example registration is shown below:

void registerMyPass() {
  PassRegistration<MyPass>();
}
  • MyPass is the name of the derived pass class.
  • The pass getArgument() method is used to get the identifier that will be used to refer to the pass.
  • The pass getDescription() method provides a short summary describing the pass.

For passes that cannot be default-constructed, PassRegistration accepts an optional argument that takes a callback to create the pass:

void registerMyPass() {
  PassRegistration<MyParametricPass>(
    []() -> std::unique_ptr<Pass> {
      std::unique_ptr<Pass> p = std::make_unique<MyParametricPass>(/*options*/);
      /*... non-trivial-logic to configure the pass ...*/;
      return p;
    });
}

This variant of registration can be used, for example, to accept the configuration of a pass from command-line arguments and pass it to the pass constructor.

Note: Make sure that the pass is copy-constructible in a way that does not share data as the pass manager may create copies of the pass to run in parallel.

Pass Pipeline Registration

Described above is the mechanism used for registering a specific derived pass class. On top of that, MLIR allows for registering custom pass pipelines in a similar fashion. This allows for custom pipelines to be available to tools like mlir-opt in the same way that passes are, which is useful for encapsulating common pipelines like the "-O1" series of passes. Pipelines are registered via a similar mechanism to passes in the form of PassPipelineRegistration. Compared to PassRegistration, this class takes an additional parameter in the form of a pipeline builder that modifies a provided OpPassManager.

void pipelineBuilder(OpPassManager &pm) {
  pm.addPass(std::make_unique<MyPass>());
  pm.addPass(std::make_unique<MyOtherPass>());
}

void registerMyPasses() {
  // Register an existing pipeline builder function.
  PassPipelineRegistration<>(
    "argument", "description", pipelineBuilder);

  // Register an inline pipeline builder.
  PassPipelineRegistration<>(
    "argument", "description", [](OpPassManager &pm) {
      pm.addPass(std::make_unique<MyPass>());
      pm.addPass(std::make_unique<MyOtherPass>());
    });
}

Textual Pass Pipeline Specification

The previous sections detailed how to register passes and pass pipelines with a specific argument and description. Once registered, these can be used to configure a pass manager from a string description. This is especially useful for tools like mlir-opt, that configure pass managers from the command line, or as options to passes that utilize dynamic pass pipelines.

To support the ability to describe the full structure of pass pipelines, MLIR supports a custom textual description of pass pipelines. The textual description includes the nesting structure, the arguments of the passes and pass pipelines to run, and any options for those passes and pipelines. A textual pipeline is defined as a series of names, each of which may in itself recursively contain a nested pipeline description. The syntax for this specification is as follows:

pipeline          ::= op-anchor `(` pipeline-element (`,` pipeline-element)* `)`
pipeline-element  ::= pipeline | (pass-name | pass-pipeline-name) options?
options           ::= '{' (key ('=' value)?)+ '}'
  • op-anchor
    • This corresponds to the mnemonic name that anchors the execution of the pass manager. This is either the name of an operation to run passes on, e.g. func.func or builtin.module, or any, for op-agnostic pass managers that execute on any viable operation (i.e. any operation that can be used to anchor a pass manager).
  • pass-name | pass-pipeline-name
    • This corresponds to the argument of a registered pass or pass pipeline, e.g. cse or canonicalize.
  • options
    • Options are specific key value pairs representing options defined by a pass or pass pipeline, as described in the "Instance Specific Pass Options" section. See this section for an example usage in a textual pipeline.

For example, the following pipeline:

$ mlir-opt foo.mlir -cse -canonicalize -convert-func-to-llvm='use-bare-ptr-memref-call-conv=1'

Can also be specified as (via the -pass-pipeline flag):

# Anchor the cse and canonicalize passes on the `func.func` operation.
$ mlir-opt foo.mlir -pass-pipeline='builtin.module(func.func(cse,canonicalize),convert-func-to-llvm{use-bare-ptr-memref-call-conv=1})'

# Anchor the cse and canonicalize passes on "any" viable root operation.
$ mlir-opt foo.mlir -pass-pipeline='builtin.module(any(cse,canonicalize),convert-func-to-llvm{use-bare-ptr-memref-call-conv=1})'

In order to support round-tripping a pass to the textual representation using OpPassManager::printAsTextualPipeline(raw_ostream&), override StringRef Pass::getArgument() to specify the argument used when registering a pass.

Declarative Pass Specification

Some aspects of a Pass may be specified declaratively, in a form similar to operations. This specification simplifies several mechanisms used when defining passes. It can be used for generating pass registration calls, defining boilerplate pass utilities, and generating pass documentation.

Consider the following pass specified in C++:

struct MyPass : PassWrapper<MyPass, OperationPass<ModuleOp>> {
  MyPass() = default;
  MyPass(const MyPass &) {}

  ...

  // Specify any options.
  Option<bool> option{
      *this, "example-option",
      llvm::cl::desc("An example option"), llvm::cl::init(true)};
  ListOption<int64_t> listOption{
      *this, "example-list",
      llvm::cl::desc("An example list option")};

  // Specify any statistics.
  Statistic statistic{this, "example-statistic", "An example statistic"};
};

/// Expose this pass to the outside world.
std::unique_ptr<Pass> foo::createMyPass() {
  return std::make_unique<MyPass>();
}

/// Register this pass.
void foo::registerMyPass() {
  PassRegistration<MyPass>();
}

This pass may be specified declaratively as so:

def MyPass : Pass<"my-pass", "ModuleOp"> {
  let summary = "My Pass Summary";
  let description = [{
    Here we can now give a much larger description of `MyPass`, including all of
    its various constraints and behavior.
  }];

  // A constructor must be provided to specify how to create a default instance
  // of MyPass. It can be skipped for this specific example, because both the
  // constructor and the registration methods live in the same namespace.
  let constructor = "foo::createMyPass()";

  // Specify any options.
  let options = [
    Option<"option", "example-option", "bool", /*default=*/"true",
           "An example option">,
    ListOption<"listOption", "example-list", "int64_t",
               "An example list option">
  ];

  // Specify any statistics.
  let statistics = [
    Statistic<"statistic", "example-statistic", "An example statistic">
  ];
}

Using the gen-pass-decls generator, we can generate most of the boilerplate above automatically. This generator takes as an input a -name parameter, that provides a tag for the group of passes that are being generated. This generator produces code with multiple purposes:

The first is to register the declared passes with the global registry. For each pass, the generator produces a registerPassName where PassName is the name of the definition specified in tablegen. It also generates a registerGroupPasses, where Group is the tag provided via the -name input parameter, that registers all of the passes present.

// Tablegen options: -gen-pass-decls -name="Example"

// Passes.h

namespace foo {
#define GEN_PASS_REGISTRATION
#include "Passes.h.inc"
} // namespace foo

void registerMyPasses() {
  // Register all of the passes.
  foo::registerExamplePasses();
  
  // Or

  // Register `MyPass` specifically.
  foo::registerMyPass();
}

The second is to provide a way to configure the pass options. These classes are named in the form of MyPassOptions, where MyPass is the name of the pass definition in tablegen. The configurable parameters reflect the options declared in the tablegen file. These declarations can be enabled for the whole group of passes by defining the GEN_PASS_DECL macro, or on a per-pass basis by defining GEN_PASS_DECL_PASSNAME where PASSNAME is the uppercase version of the name specified in tablegen.

// .h.inc

#ifdef GEN_PASS_DECL_MYPASS

struct MyPassOptions {
    bool option = true;
    ::llvm::ArrayRef<int64_t> listOption;
};

#undef GEN_PASS_DECL_MYPASS
#endif // GEN_PASS_DECL_MYPASS

If the constructor field has not been specified in the tablegen declaration, then autogenerated file will also contain the declarations of the default constructors.

// .h.inc

#ifdef GEN_PASS_DECL_MYPASS
...

std::unique_ptr<::mlir::Pass> createMyPass();
std::unique_ptr<::mlir::Pass> createMyPass(const MyPassOptions &options);

#undef GEN_PASS_DECL_MYPASS
#endif // GEN_PASS_DECL_MYPASS

The last purpose of this generator is to emit a base class for each of the passes, containing most of the boiler plate related to pass definitions. These classes are named in the form of MyPassBase and are declared inside the impl namespace, where MyPass is the name of the pass definition in tablegen. We can update the original C++ pass definition as so:

// MyPass.cpp

/// Include the generated base pass class definitions.
namespace foo {
#define GEN_PASS_DEF_MYPASS
#include "Passes.h.inc"
}

/// Define the main class as deriving from the generated base class.
struct MyPass : foo::impl::MyPassBase<MyPass> {
  using MyPassBase::MyPassBase;

  /// The definitions of the options and statistics are now generated within
  /// the base class, but are accessible in the same way.
};

These definitions can be enabled on a per-pass basis by defining the appropriate preprocessor GEN_PASS_DEF_PASSNAME macro, with PASSNAME equal to the uppercase version of the name of the pass definition in tablegen. If the constructor field has not been specified in tablegen, then the default constructors are also defined and expect the name of the actual pass class to be equal to the name defined in tablegen.

Using the gen-pass-doc generator, markdown documentation for each of the passes can be generated. See Passes.md for example output of real MLIR passes.

Tablegen Specification

The Pass class is used to begin a new pass definition. This class takes as an argument the registry argument to attribute to the pass, as well as an optional string corresponding to the operation type that the pass operates on. The class contains the following fields:

  • summary
    • A short one-line summary of the pass, used as the description when registering the pass.
  • description
    • A longer, more detailed description of the pass. This is used when generating pass documentation.
  • dependentDialects
    • A list of strings representing the Dialect classes this pass may introduce entities, Attributes/Operations/Types/etc., of.
  • constructor
    • A code block used to create a default instance of the pass.
  • options
    • A list of pass options used by the pass.
  • statistics
    • A list of pass statistics used by the pass.

Options

Options may be specified via the Option and ListOption classes. The Option class takes the following template parameters:

  • C++ variable name
    • A name to use for the generated option variable.
  • argument
    • The argument name of the option.
  • type
    • The C++ type of the option.
  • default value
    • The default option value.
  • description
    • A one-line description of the option.
  • additional option flags
    • A string containing any additional options necessary to construct the option.
def MyPass : Pass<"my-pass"> {
  let options = [
    Option<"option", "example-option", "bool", /*default=*/"true",
           "An example option">,
  ];
}

The ListOption class takes the following fields:

  • C++ variable name
    • A name to use for the generated option variable.
  • argument
    • The argument name of the option.
  • element type
    • The C++ type of the list element.
  • description
    • A one-line description of the option.
  • additional option flags
    • A string containing any additional options necessary to construct the option.
def MyPass : Pass<"my-pass"> {
  let options = [
    ListOption<"listOption", "example-list", "int64_t",
               "An example list option">
  ];
}

Statistic

Statistics may be specified via the Statistic, which takes the following template parameters:

  • C++ variable name
    • A name to use for the generated statistic variable.
  • display name
    • The name used when displaying the statistic.
  • description
    • A one-line description of the statistic.
def MyPass : Pass<"my-pass"> {
  let statistics = [
    Statistic<"statistic", "example-statistic", "An example statistic">
  ];
}

Pass Instrumentation

MLIR provides a customizable framework to instrument pass execution and analysis computation, via the PassInstrumentation class. This class provides hooks into the PassManager that observe various events:

  • runBeforePipeline
    • This callback is run just before a pass pipeline, i.e. pass manager, is executed.
  • runAfterPipeline
    • This callback is run right after a pass pipeline has been executed, successfully or not.
  • runBeforePass
    • This callback is run just before a pass is executed.
  • runAfterPass
    • This callback is run right after a pass has been successfully executed. If this hook is executed, runAfterPassFailed will not be.
  • runAfterPassFailed
    • This callback is run right after a pass execution fails. If this hook is executed, runAfterPass will not be.
  • runBeforeAnalysis
    • This callback is run just before an analysis is computed.
    • If the analysis requested another analysis as a dependency, the runBeforeAnalysis/runAfterAnalysis pair for the dependency can be called from inside of the current runBeforeAnalysis/runAfterAnalysis pair.
  • runAfterAnalysis
    • This callback is run right after an analysis is computed.

PassInstrumentation instances may be registered directly with a PassManager instance via the addInstrumentation method. Instrumentations added to the PassManager are run in a stack like fashion, i.e. the last instrumentation to execute a runBefore* hook will be the first to execute the respective runAfter* hook. The hooks of a PassInstrumentation class are guaranteed to be executed in a thread-safe fashion, so additional synchronization is not necessary. Below in an example instrumentation that counts the number of times the DominanceInfo analysis is computed:

struct DominanceCounterInstrumentation : public PassInstrumentation {
  /// The cumulative count of how many times dominance has been calculated.
  unsigned &count;

  DominanceCounterInstrumentation(unsigned &count) : count(count) {}
  void runAfterAnalysis(llvm::StringRef, TypeID id, Operation *) override {
    if (id == TypeID::get<DominanceInfo>())
      ++count;
  }
};

MLIRContext *ctx = ...;
PassManager pm(ctx);

// Add the instrumentation to the pass manager.
unsigned domInfoCount;
pm.addInstrumentation(
    std::make_unique<DominanceCounterInstrumentation>(domInfoCount));

// Run the pass manager on a module operation.
ModuleOp m = ...;
if (failed(pm.run(m)))
    ...

llvm::errs() << "DominanceInfo was computed " << domInfoCount << " times!\n";

Standard Instrumentations

MLIR utilizes the pass instrumentation framework to provide a few useful developer tools and utilities. Each of these instrumentations are directly available to all users of the MLIR pass framework.

Pass Timing

The PassTiming instrumentation provides timing information about the execution of passes and computation of analyses. This provides a quick glimpse into what passes are taking the most time to execute, as well as how much of an effect a pass has on the total execution time of the pipeline. Users can enable this instrumentation directly on the PassManager via enableTiming. This instrumentation is also made available in mlir-opt via the -mlir-timing flag. The PassTiming instrumentation provides several different display modes for the timing results, each of which is described below:

List Display Mode

In this mode, the results are displayed in a list sorted by total time with each pass/analysis instance aggregated into one unique result. This view is useful for getting an overview of what analyses/passes are taking the most time in a pipeline. This display mode is available in mlir-opt via -mlir-timing-display=list.

$ mlir-opt foo.mlir -mlir-disable-threading -pass-pipeline='builtin.module(func.func(cse,canonicalize),convert-func-to-llvm)' -mlir-timing -mlir-timing-display=list

===-------------------------------------------------------------------------===
                      ... Pass execution timing report ...
===-------------------------------------------------------------------------===
  Total Execution Time: 0.0203 seconds

   ---Wall Time---  --- Name ---
   0.0047 ( 55.9%)  Canonicalizer
   0.0019 ( 22.2%)  VerifierPass
   0.0016 ( 18.5%)  LLVMLoweringPass
   0.0003 (  3.4%)  CSE
   0.0002 (  1.9%)  (A) DominanceInfo
   0.0084 (100.0%)  Total
Tree Display Mode

In this mode, the results are displayed in a nested pipeline view that mirrors the internal pass pipeline that is being executed in the pass manager. This view is useful for understanding specifically which parts of the pipeline are taking the most time, and can also be used to identify when analyses are being invalidated and recomputed. This is the default display mode.

$ mlir-opt foo.mlir -mlir-disable-threading -pass-pipeline='builtin.module(func.func(cse,canonicalize),convert-func-to-llvm)' -mlir-timing

===-------------------------------------------------------------------------===
                      ... Pass execution timing report ...
===-------------------------------------------------------------------------===
  Total Execution Time: 0.0249 seconds

   ---Wall Time---  --- Name ---
   0.0058 ( 70.8%)  'func.func' Pipeline
   0.0004 (  4.3%)    CSE
   0.0002 (  2.6%)      (A) DominanceInfo
   0.0004 (  4.8%)    VerifierPass
   0.0046 ( 55.4%)    Canonicalizer
   0.0005 (  6.2%)    VerifierPass
   0.0005 (  5.8%)  VerifierPass
   0.0014 ( 17.2%)  LLVMLoweringPass
   0.0005 (  6.2%)  VerifierPass
   0.0082 (100.0%)  Total
Multi-threaded Pass Timing

When multi-threading is enabled in the pass manager the meaning of the display slightly changes. First, a new timing column is added, User Time, that displays the total time spent across all threads. Secondly, the Wall Time column displays the longest individual time spent amongst all of the threads. This means that the Wall Time column will continue to give an indicator on the perceived time, or clock time, whereas the User Time will display the total cpu time.

$ mlir-opt foo.mlir -pass-pipeline='builtin.module(func.func(cse,canonicalize),convert-func-to-llvm)'  -mlir-timing

===-------------------------------------------------------------------------===
                      ... Pass execution timing report ...
===-------------------------------------------------------------------------===
  Total Execution Time: 0.0078 seconds

   ---User Time---   ---Wall Time---  --- Name ---
   0.0177 ( 88.5%)     0.0057 ( 71.3%)  'func.func' Pipeline
   0.0044 ( 22.0%)     0.0015 ( 18.9%)    CSE
   0.0029 ( 14.5%)     0.0012 ( 15.2%)      (A) DominanceInfo
   0.0038 ( 18.9%)     0.0015 ( 18.7%)    VerifierPass
   0.0089 ( 44.6%)     0.0025 ( 31.1%)    Canonicalizer
   0.0006 (  3.0%)     0.0002 (  2.6%)    VerifierPass
   0.0004 (  2.2%)     0.0004 (  5.4%)  VerifierPass
   0.0013 (  6.5%)     0.0013 ( 16.3%)  LLVMLoweringPass
   0.0006 (  2.8%)     0.0006 (  7.0%)  VerifierPass
   0.0200 (100.0%)     0.0081 (100.0%)  Total

IR Printing

When debugging it is often useful to dump the IR at various stages of a pass pipeline. This is where the IR printing instrumentation comes into play. This instrumentation allows for conditionally printing the IR before and after pass execution by optionally filtering on the pass being executed. This instrumentation can be added directly to the PassManager via the enableIRPrinting method. mlir-opt provides a few useful flags for utilizing this instrumentation:

  • mlir-print-ir-before=(comma-separated-pass-list)
    • Print the IR before each of the passes provided within the pass list.
  • mlir-print-ir-before-all
    • Print the IR before every pass in the pipeline.
$ mlir-opt foo.mlir -pass-pipeline='func.func(cse)' -mlir-print-ir-before=cse

*** IR Dump Before CSE ***
func.func @simple_constant() -> (i32, i32) {
  %c1_i32 = arith.constant 1 : i32
  %c1_i32_0 = arith.constant 1 : i32
  return %c1_i32, %c1_i32_0 : i32, i32
}
  • mlir-print-ir-after=(comma-separated-pass-list)
    • Print the IR after each of the passes provided within the pass list.
  • mlir-print-ir-after-all
    • Print the IR after every pass in the pipeline.
$ mlir-opt foo.mlir -pass-pipeline='func.func(cse)' -mlir-print-ir-after=cse

*** IR Dump After CSE ***
func.func @simple_constant() -> (i32, i32) {
  %c1_i32 = arith.constant 1 : i32
  return %c1_i32, %c1_i32 : i32, i32
}
  • mlir-print-ir-after-change
    • Only print the IR after a pass if the pass mutated the IR. This helps to reduce the number of IR dumps for "uninteresting" passes.
    • Note: Changes are detected by comparing a hash of the operation before and after the pass. This adds additional run-time to compute the hash of the IR, and in some rare cases may result in false-positives depending on the collision rate of the hash algorithm used.
    • Note: This option should be used in unison with one of the other 'mlir-print-ir-after' options above, as this option alone does not enable printing.
$ mlir-opt foo.mlir -pass-pipeline='func.func(cse,cse)' -mlir-print-ir-after=cse -mlir-print-ir-after-change

*** IR Dump After CSE ***
func.func @simple_constant() -> (i32, i32) {
  %c1_i32 = arith.constant 1 : i32
  return %c1_i32, %c1_i32 : i32, i32
}
  • mlir-print-ir-after-failure
    • Only print IR after a pass failure.
    • This option should not be used with the other mlir-print-ir-after flags above.
$ mlir-opt foo.mlir -pass-pipeline='func.func(cse,bad-pass)' -mlir-print-ir-after-failure

*** IR Dump After BadPass Failed ***
func.func @simple_constant() -> (i32, i32) {
  %c1_i32 = arith.constant 1 : i32
  return %c1_i32, %c1_i32 : i32, i32
}
  • mlir-print-ir-module-scope
    • Always print the top-level module operation, regardless of pass type or operation nesting level.
    • Note: Printing at module scope should only be used when multi-threading is disabled(-mlir-disable-threading)
$ mlir-opt foo.mlir -mlir-disable-threading -pass-pipeline='func.func(cse)' -mlir-print-ir-after=cse -mlir-print-ir-module-scope

*** IR Dump After CSE ***  ('func.func' operation: @bar)
func.func @bar(%arg0: f32, %arg1: f32) -> f32 {
  ...
}

func.func @simple_constant() -> (i32, i32) {
  %c1_i32 = arith.constant 1 : i32
  %c1_i32_0 = arith.constant 1 : i32
  return %c1_i32, %c1_i32_0 : i32, i32
}

*** IR Dump After CSE ***  ('func.func' operation: @simple_constant)
func.func @bar(%arg0: f32, %arg1: f32) -> f32 {
  ...
}

func.func @simple_constant() -> (i32, i32) {
  %c1_i32 = arith.constant 1 : i32
  return %c1_i32, %c1_i32 : i32, i32
}

Crash and Failure Reproduction

The pass manager in MLIR contains a builtin mechanism to generate reproducibles in the event of a crash, or a pass failure. This functionality can be enabled via PassManager::enableCrashReproducerGeneration or via the command line flag mlir-pass-pipeline-crash-reproducer. In either case, an argument is provided that corresponds to the output .mlir file name that the reproducible should be written to. The reproducible contains the configuration of the pass manager that was executing, as well as the initial IR before any passes were run. The reproducer is stored within the assembly format as an external resource. A potential reproducible may have the form:

module {
  func.func @foo() {
    ...
  }
}

{-#
  external_resources: {
    mlir_reproducer: {
      pipeline: "builtin.module(func.func(cse,canonicalize),inline)",
      disable_threading: true,
      verify_each: true
    }
  }
#-}

The configuration dumped can be passed to mlir-opt by specifying -run-reproducer flag. This will result in parsing the configuration of the reproducer and adjusting the necessary opt state, e.g. configuring the pass manager, context, etc.

Beyond specifying a filename, one can also register a ReproducerStreamFactory function that would be invoked in the case of a crash and the reproducer written to its stream.

Local Reproducer Generation

An additional flag may be passed to PassManager::enableCrashReproducerGeneration, and specified via mlir-pass-pipeline-local-reproducer on the command line, that signals that the pass manager should attempt to generate a "local" reproducer. This will attempt to generate a reproducer containing IR right before the pass that fails. This is useful for situations where the crash is known to be within a specific pass, or when the original input relies on components (like dialects or passes) that may not always be available.

Note: Local reproducer generation requires that multi-threading is disabled(-mlir-disable-threading)

For example, if the failure in the previous example came from the canonicalize pass, the following reproducer would be generated:

module {
  func.func @foo() {
    ...
  }
}

{-#
  external_resources: {
    mlir_reproducer: {
      pipeline: "builtin.module(func.func(canonicalize))",
      disable_threading: true,
      verify_each: true
    }
  }
#-}