AI基础架构Pass Infrastructure

AI基础架构Pass Infrastructure

• Operation Pass
  • OperationPass : Op-Specific
  • OperationPass : Op-Agnostic
  • Dependent Dialects
  • Initialization
• Analysis Management
  • Querying Analyses
  • Preserving Analyses
• Pass Failure
• Pass Manager
  • OpPassManager
• Dynamic Pass Pipelines
• Instance Specific Pass Options
• Pass Statistics
• Pass Registration
  • Pass Pipeline Registration
  • Textual Pass Pipeline Specification
• Declarative Pass Specification
  • Tablegen Specification
• Pass Instrumentation
  • Standard Instrumentations
• Crash and Failure Reproduction
  • Local Reproducer Generation

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.

Passes代表了转换和优化的基本基础架构。本文概述了MLIR中的Passes基础结构以及如何使用。

有关MLIR 及其核心方面（如IR结构和算子）的更多信息，请参见 MLIR规范。

有关 在MLIR中进行图形重写的快速入门，请参见 MLIR Rewrites。如果转换涉及模式匹配算子作DAG，那么这是一个很好的起点。

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:

• Modify any state referenced or relied upon outside the current 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.
• Modify the state of another operation not nested within the current operation being operated on.
  • Other threads may be operating on these operations simultaneously.
• Inspect the state of sibling operations.
  • Other threads may be modifying these operations in parallel.
  • Inspecting the state of ancestor/parent operations is permitted.
• 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.
• 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.

When creating an operation pass, there are two different types to choose from depending on the usage scenario:

在MLIR中，抽象和转换的主要单元是一个 算子 。这样，通过管理器被设计为在不同嵌套级别上的算子实例上工作。Passes管理器的结构 和嵌套的概念将详细介绍。MLIR中的所有Passes均源自OperationPass并遵守限制；在多线程和其它高级方案中，任何不符合都会导致有问题的行为：

• 修改正在算子的电流之外引用或依赖的任何状态。这包括在父块中添加或删除算子，更改属性（取决于当前算子的协议）/当前算子的算子数/结果/后继对象。
• 修改另一个未嵌套在当前算子中的算子的状态。
  • 其它线程可能同时在这些算子上运行。
• 检查同级算子的状态。
  • 其它线程可能正在并行修改这些算子。
  • 允许检查祖先/父项算子的状态。
• 跨调用保持可变的通过状态runOnOperation。传递可能在许多不同的算子上运行，而不能保证执行顺序。
  • 在进行多线程处理时，特定的传递实例甚至可能不会在IR内的所有算子上执行。因此，传递不应该依赖于所有算子上的运行。
• 保持任何全局可变状态，例如源文件中的静态变量。所有可变状态都应通过传递实例来维护。
• 必须是可复制的
  • 流程管理器可以创建流程的多个实例，以并行处理算子。

创建算子传递时，根据使用情况，有两种不同的类型可供选择：

OperationPass : Op-Specific 

An op-specific operation pass operates explicitly on a given operation type. This operation type must adhere to the restrictions set by the pass manager for pass execution.

To define an op-specific operation pass, a derived class must adhere to the following:

• Inherit from the CRTP class OperationPass and provide the operation type as an additional template parameter.
• Override the virtual void runOnOperation() method.

A simple pass may look like:

在op-specific算子上传递给定类型明确的算子。此类型必须遵守pass管理器为pass执行设置的限制。

要定义特定于算子的通道，派生类必须遵守以下规定：

• 继承自CRTP类，OperationPass并提供算子类型作为附加模板参数。
• 覆盖虚拟void runOnOperation()方法。

一个简单的pass可能看起来像：

namespace {

/// 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<OperationPass<FuncOp>,

                                           MyFunctionPass> {

  void runOnOperation() override {

    // Get the current FuncOp operation being operated on.

    FuncOp f = getOperation();

    // Walk the operations within the function.

    f.walk([](Operation *inst) {

      ....

    });

  }

};

} // end anonymous namespace

/// Register this pass so that it can be built via from a textual pass pipeline.

/// (Pass registration is discussed more below)

void registerMyPass() {

  PassRegistration<MyFunctionPass>(

    "flag-name-to-invoke-pass-via-mlir-opt", "Pass description here");

}

OperationPass : Op-Agnostic 

An op-agnostic pass operates on the operation type of the pass manager that it is added to. This means that passes of this type may operate on several different operation types. Passes of this type are generally written generically using operation interfaces and traits . Examples of this type of pass are Common Sub-Expression Elimination and Inlining .

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

• Inherit from the CRTP class OperationPass.

op-agnostic上，被添加到管理器的算子类型进行运算。这意味着这种类型的通道可以在几种不同的算子类型上进行运算。通常使用算子接口 和 特征来写这种类型的pass 。此类传递的示例包括“ 常见子表达式消除” 和“ 内联” 。

要创建算子传递，派生类必须遵守以下内容：

• 继承自CRTP类OperationPass。

struct MyOperationPass : public PassWrapper<OperationPass<>, MyOperationPass> {

    // Get the current operation being operated on.

    Operation *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 ne loaded must express this by overriding the getDependentDialects() method and declare this list of Dialects explicitly.

必须先在MLIRContext中加载语言对话，然后才能从这些语言对话创建实体（算子，类型，属性等）。在开始执行多线程传递pass之前，还必须加载语言对话。可能无法保证已从其加载的语言对话创建实体的pass，必须通过重写

getDependentDialects()

 方法并明确声明此语言对话列表来表达。

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.

在某些情况下，通过可能包含动态构造的状态，但是在连续运行过程中重新计算可能会很费时。一个这样的例子就是使用 PDL基于 模式的模式 ，该模式在运行时被编译成字节码。在这些情况下，通过可能会覆盖以下hook，初始化此重载状态：

此hook在一次完整pass管道的每次运行中执行一次，意味着无法访问runOnOperation调用期间的可用状态。更具体地，所有必要的访问的MLIRContext应通过提供驱动context参数，利用 “per-run”状态，诸如 getContext/ getOperation/ getAnalysis/等，不得使用。如果初始化期间发生错误，则该传递将发出错误诊断并返回a failure()，中止pass管道的执行。

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 an Operation*.
• 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.

一个重要的概念，以及转换过程，都是分析。这些在概念上与转换过程相似，除了在不修改计算信息的特定算子情况下。在MLIR中，分析不是pass，而是pass独立式类，这些类是按需延迟计算并缓存，以避免不必要的重新计算。MLIR中的分析必须遵循以下条件：

• 提供有效构造函数Operation*。
• 一定不能修改给定的算子。

分析可能会提供其它hook来控制各种行为：

给定一个保留的分析集，如果该分析真无效，则分析返回true。如果没有明确标记保留分析，但可以根据其它属性（例如分析集）保留（或使之无效），则可以进行更精细的无效化。

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:

OperationPass基类提供用于查询和保留当前正在处理算子和分析的实用程序。

• OperationPass自动提供以下实用程序来查询分析：
  • • 获得当前算子的分析，并在必要时进行构建。
  • • 获取当前算子的分析（如果已经存在）。
  • • 获取给定父算子的分析（如果存在）。
  • • 对给定的子算子（如果存在）进行分析。
  • • 对给定的子算子进行分析，并在必要时进行构造。

使用上面定义的示例传递，看一些示例：

/// An interesting analysis.

struct MyOperationAnalysis {

  // Compute this analysis with the provided operation.

  MyOperationAnalysis(Operation *op);

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>();

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<>

通过查询后构造的分析将被缓存，以避免不必要的计算（如果稍后再次请求的话）。为了避免过时的分析，假定所有分析都通过了一次pass验证就无效了。为避免无效，pass必须特别标记已知保留的分析。

• 所有Pass类都会自动提供以下用于保留分析的实用程序：

  // 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.

允许MLIR中的传递正常失败。如果pass的某些不变式被破坏，可能会使IR处于某种无效状态，则可能会发生这种情况。如果发生这种情况，则pass可以该

signalPassFailure

方法直接向pass管理器发出故障信号。如果在执行过程中pass指示失败，则流水线中将不会执行其它任何传递，并且对

PassManager::run

顶级调用，返回

failure

。

  // 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.

以上各节介绍了pass的不同类型及其不变性。本节介绍PassManager的概念，以及如何配置和计划pass管道。与pass管理相关的主要类别有两种：

PassManager

和

OpPassManager

PassManager

类作为顶层的入口点，包含用于整个pass管道的各种配置。该

OpPassManager

用于调度类会将以嵌套的一个特定的水平上运行。顶级 

PassManager

还用作

OpPassManager

OpPassManager 

An OpPassManager is essentially a collection of passes to execute on an operation of a specific type. This operation type must adhere to the following requirement:

• Must be registered and marked IsolatedFromAbove .
  • Passes are expected to not 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. The pass must either be an op-specific pass operating on the same operation type as OpPassManager, or an op-agnostic pass.

An OpPassManager is generally created by explicitly nesting a pipeline within another existing OpPassManager via the nest<> method. This method takes the operation type that the nested pass manager will operate on. At the top-level, a PassManager acts as an OpPassManager. Nesting in this sense, corresponds to the structural nesting within Regions of the IR.

For example, the following .mlir:

AnOpPassManager本质上是要在特定类型的算子上执行的pass的集合。算子类型必须符合以下要求：

• 必须注册并标记 IsolatedFromAbove 。
  • 预期pass不会修改正在处理的当前操作或更高的操作。如果操作不是孤立的，则可能会无意间修改或遍历不应执行的操作的SSA使用列表。

可以通过将pass添加到pass管理器addPass。该pass必须是采用 op-specific与相同的算子类型进行操作OpPassManager的op-agnosticpass，或者是pass。

OpPassManager通常OpPassManager通过将该nest<>方法显式嵌套在另一个现有pass中来创建An 。此方法采用嵌套pass管理器，将对其进行操作的操作类型。在顶层，a PassManager充当OpPassManager。从这个意义上讲，嵌套对应 于 IR区域内的 结构嵌套 。

例如，以下内容.mlir：

module {

  spv.module "Logical" "GLSL450" {

    func @foo() {

      ...

    }

Has the nesting structure of:

`module`

  `spv.module`

    `function`

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

// Create a top-level `PassManager` class. If an operation type is not

// explicitly specific, the default is the builtin `module` operation.

PassManager pm(ctx);

// Note: We could also create the above `PassManager` this way.

PassManager pm(ctx, /*operationName=*/"module");

// 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<FuncOp>();

nestedFunctionPM.addPass(std::make_unique<MyFunctionPass>());

// 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<FuncOp>

      MyFunctionPass

These pipelines are then run over a single operation at a time. This means that, for example, given a series of consecutive passes on 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("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 follow the LLVM command line flag definition rules. See below for a few examples:

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.

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 -pass-statistics and -pass-statistics-display on the command line.

An example is shown below:

  /// 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='func(my-pass,my-pass)' foo.mlir -pass-statistics

===-------------------------------------------------------------------------===

                         ... Pass statistics report ...

'func' Pipeline

  MyPass

    (S) 15 exampleStat - An example statistic

  VerifierPass

    (S)  6 exampleStat - An example statistic

VerifierPass

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

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

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:

  PassRegistration<MyPass>("argument", "description");

• MyPass is the name of the derived pass class.
• “argument” is the argument used to refer to the pass in the textual format.
• “description” is a brief description of the pass.

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

  PassRegistration<MyParametricPass>(

    "argument", "description",

    []() -> 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>());

  // Register an existing pipeline builder function.

  PassPipelineRegistration<>(

    "argument", "description", pipelineBuilder);

  // Register an inline pipeline builder.

    "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-name `(` pipeline-element (`,` pipeline-element)* `)`

pipeline-element  ::= pipeline | (pass-name | pass-pipeline-name) options?

options           ::= '{' (key ('=' value)?)+ '}'

• op-name
  • This corresponds to the mnemonic name of an operation to run passes on, e.g. func or module.
• 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-std-to-llvm='use-bare-ptr-memref-call-conv=1'

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

$ mlir-opt foo.mlir -pass-pipeline='func(cse,canonicalize),convert-std-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(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"), llvm::cl::ZeroOrMore,

      llvm::cl::MiscFlags::CommaSeparated};

  // 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>("my-pass", "My pass summary");

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.

  let constructor = "foo::createMyPass()";

  let options = [

    Option<"option", "example-option", "bool", /*default=*/"true",

           "An example option">,

    ListOption<"listOption", "example-list", "int64_t",

               "An example list option",

               "llvm::cl::ZeroOrMore, llvm::cl::MiscFlags::CommaSeparated">

  ];

  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 two chunks of output:

The first is a code block for registering the declarative passes with the global registry. For each pass, the generator produces a registerFooPass where Foo 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.

// gen-pass-decls -name="Example"

#define GEN_PASS_REGISTRATION

#include "Passes.h.inc"

  // Register all of the passes.

  registerExamplePasses();

  // Register `MyPass` specifically.

  registerMyPassPass();

The second is 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, where MyPass is the name of the pass definition in tablegen. We can update the original C++ pass definition as so:

/// Include the generated base pass class definitions.

#define GEN_PASS_CLASSES

/// Define the main class as deriving from the generated base class.

struct MyPass : MyPassBase<MyPass> {

  /// The explicit constructor is no longer explicitly necessary when defining

  /// pass options and statistics, the base class takes care of that

  /// automatically.

  /// The definitions of the options and statistics are now generated within

  /// the base class, but are accessible in the same way.

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.
• • 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.
• • 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"> {

The ListOption class takes the following fields:

• element type
  • The C++ type of the list element.

Statistic 

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

• • A name to use for the generated statistic variable.
• display name
  • The name used when displaying the statistic.
• • A one line description of the 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.
• 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 = ...;

// Add the instrumentation to the pass manager.

unsigned domInfoCount;

pm.addInstrumentation(

    std::make_unique<DominanceCounterInstrumentation>(domInfoCount));

// Run the pass manager on a module operation.

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 -pass-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 -pass-timing-display=list.

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

Pipeline 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='func(cse,canonicalize)' -convert-std-to-llvm -pass-timing

  Total Execution Time: 0.0249 seconds

   0.0058 ( 70.8%)  '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='func(cse,canonicalize)' -convert-std-to-llvm -pass-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' 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:

• print-ir-before=(comma-separated-pass-list)
  • Print the IR before each of the passes provided within the pass list.
• print-ir-before-all
  • Print the IR before every pass in the pipeline.

$ mlir-opt foo.mlir -pass-pipeline='func(cse)' -print-ir-before=cse

*** IR Dump Before CSE ***

func @simple_constant() -> (i32, i32) {

  %c1_i32 = constant 1 : i32

  %c1_i32_0 = constant 1 : i32

  return %c1_i32, %c1_i32_0 : i32, i32

• print-ir-after=(comma-separated-pass-list)
  • Print the IR after each of the passes provided within the pass list.
• print-ir-after-all
  • Print the IR after every pass in the pipeline.

$ mlir-opt foo.mlir -pass-pipeline='func(cse)' -print-ir-after=cse

*** IR Dump After CSE ***

  return %c1_i32, %c1_i32 : i32, i32

• 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 ‘print-ir-after’ options above, as this option alone does not enable printing.

$ mlir-opt foo.mlir -pass-pipeline='func(cse,cse)' -print-ir-after=cse -print-ir-after-change

• 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(cse)' -print-ir-after=cse -print-ir-module-scope

*** IR Dump After CSE ***  ('func' operation: @bar)

func @bar(%arg0: f32, %arg1: f32) -> f32 {

*** IR Dump After CSE ***  ('func' operation: @simple_constant)

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 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. A potential reproducible may have the form:

// configuration: -pass-pipeline='func(cse,canonicalize),inline' -verify-each

  func @foo() {

The configuration dumped can be passed to mlir-opt by specifying -run-reproducer flag. This will result in parsing the first line configuration of the reproducer and adding those to the command line options.

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 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.

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

// configuration: -pass-pipeline='func(canonicalize)' -verify-each

人工智能芯片与自动驾驶