FAQ

How to refer to MLIR in publications? Is there an accompanying paper? ¶

MLIR has been presented in the 2021 IEEE/ACM International Symposium on Code Generation and Optimization, the full text of the paper is available from IEEE. A pre-publication draft is available on arXiv but may be missing improvements and corrections. Please also note that MLIR keeps evolving and IR snippets presented in the paper may no longer use modern syntax, refer to the MLIR documentation for the new syntax.

To cite MLIR in academic or other publications, please use: Chris Lattner, Mehdi Amini, Uday Bondhugula, Albert Cohen, Andy Davis, Jacques Pienaar, River Riddle, Tatiana Shpeisman, Nicolas Vasilache, and Oleksandr Zinenko. “MLIR: Scaling compiler infrastructure for domain specific computation.” In 2021 IEEE/ACM International Symposium on Code Generation and Optimization (CGO), pp. 2-14. IEEE, 2021.

The BibTeX entry is as follows.

@inproceedings{mlir,
  author={Lattner, Chris and Amini, Mehdi and Bondhugula, Uday and Cohen, Albert and Davis, Andy and Pienaar, Jacques and Riddle, River and Shpeisman, Tatiana and Vasilache, Nicolas and Zinenko, Oleksandr},
  booktitle={2021 {{IEEE/ACM}} International Symposium on Code Generation and Optimization (CGO)},
  title={{{MLIR}}: Scaling Compiler Infrastructure for Domain Specific Computation},
  year={2021},
  volume={},
  number={},
  pages={2-14},
  doi={10.1109/CGO51591.2021.9370308}
}

Please do not cite the arXiv preprint as it is not a formal peer-reviewed publication.

Why is <small feature> not available in MLIR? ¶

On general basis, there is never a reason why a small feature is not available in MLIR other than nobody needed it enough to implement it. Consider submitting a patch. For larger features and dialects, follow the request-for-comments process.

MLIR is too heavy framework, should I just reimplement my own compiler from scratch? ¶

Maybe: it is hard to tell as it depends on your requirements, even C++ may already be too large for some micro-controllers. In our experience most projects ends up growing beyond what their original author intended, and reimplementing the features you would get from MLIR would also have a footprint. MLIR footprint is representative of the features it provides. More importantly we have a “you don’t pay for what you don’t use” approach: MLIR is very modular and you can link a binary with a very minimal set of libraries. If you use just the core IR, some pieces of the infrastructure, and a few dialects you should expect a few MBs. We have three examples in the repo showing some small possible configurations of MLIR, showing that the core of MLIR can take around 1MB.

What is the difference between the Tensor and Vector types? ¶

1. Conceptual: vectors are meant to and occur in lower level dialects - often where you expect hardware to have registers of that size. Tensors model higher-level “closer to the source” abstract representation. This is reflected in the abstraction modeled by the operations from the vector dialect, while Tensors would be more naturally present in the operations of the linalg dialect.
2. Tensors can be dynamically shaped, unranked, or have 0 dimensions ; but Vectors can’t be.
3. You can have a memref (a buffer in memory) containing Vectors but you can’t have a memref of a tensor type.
4. The set of allowed element types is different: the Tensor type isn’t limited while Vector is limited to float and integer types.
5. Tensors accept an optional “encoding” attribute, vector don’t at the moment.

Registered, loaded, dependent: what’s up with Dialects management? ¶

Before creating an Operation, a Type, or an Attribute, the associated Dialect must be already loaded in the MLIRContext. For example the Toy tutorial explicitly loads the Toy Dialect before emitting the Toy IR from the AST.

The process of loading a Dialect in the context is not thread-safe, which forces all involved Dialects to be loaded before the multi-threaded pass manager starts the execution. To keep the system modular and layered, invoking a pass pipeline should never require pre-loading dialects explicitly. This is achieved by requiring every pass to declare a list of dependent Dialects: these are Dialects for which an entity (Operation, Type, or Attribute) can be created by the pass, other than for Dialects that would already be in the input. For example, a convertLinalgToLoops pass would declare the SCF Dialect as dependent, but does not need to declare Linalg. See also dependent dialects in the pass infrastructure documentation.

Finally, dialects can be registered with the context. The sole purpose of the registration is to make these dialects available for the textual parser used by tools like mlir-opt or mlir-translate. A compiler frontend emitting the IR programmatically and invoking a pass pipeline would never need to register any dialects.

In dialect conversion, I want an operation to be removed after its users get converted, how do I do that? ¶

This operation can be marked “illegal” and you can just do speculatively rewriter.eraseOp(op);. The operation won’t be actually removed right now, instead when mark something as erased you are basically saying to the driver “I expect all uses of this to go away by the time everything is over”. The conversion will fail if the operation you marked as erased doesn’t actually get erased at the end.

Why is dialect X missing feature Y? ¶

Most likely, nobody has had a need for it yet. Many MLIR components, dialects even more than others, grew out of specific needs and are extended by volunteers sending patches to add the missing bits. Everybody is welcome to contribute!

In some specfic cases, the dialect design might have explicitly decided against implementing a feature or chose an alternative modeling that provides a similar functionality. Such design decisions are usually noted in the dialect or rationale documents.

Many dialects define a constant operation, how do I get a constant value generically? ¶

#include "mlir/IR/Matchers.h"

// Return the constant attribute, or null if the Operation isn't a constant.
Attribute getConstantAttr(Operation *constantOp) {
  Attribute constant;
  matchPattern(value.getDefiningOp(), m_Constant());
  return constant;
}

What is the difference between traits and interfaces? ¶

Both traits and interfaces can be used to inject common behavior into operations, types and attributes without introducing duplication. However, conceptually these are quite different.

Traits inject static behavior into operations/types/attributes whereas interfaces dynamically dispatch behavior based on their runtime type. For instance, since ModuleOp implements the SymbolTable trait, mlir::ModuleOp exposes lookupSymbol as a member function. However, there is no type-erased way to access this functionality – it is available only via mlir::ModuleOp. On the other hand, if an operation implements CallOpInterface, its implementation of getCallableForCallee can be invoked in a type-erased manner by dyn_casting the operation to a CallOpInterface. The caller does not need to know the concrete type of the operation to do this.

There is one similarity between interfaces and traits: both their presence can be checked dynamically (i.e. without access to the concrete type). Specifically, presence of traits can be checked using Operation::hasTrait and presence of interfaces can be checked using isa<>. However, this similarity does not run deep, and was only added for practical ergonomic reasons.

How to convert a memref to a pointer? ¶

It is impossible in the general case. Structured memory reference (memref) type is not (only) a pointer. This type supports multi-dimensional indexing and customizable data layouts to support advanced yet analyzable addressing modes. Implementing address computation requires understanding the layout and storing additional information such as sizes and layout parameters that would be impossible with a plain, single-typed pointer to a contiguous block of data. Even the single-dimensional memref<?xi8> with the default layout is not a pointer as it must store at the very least the size of the data (think C++ std::string vs. C NULL-terminated const char *).

It is, however, possible to define operations that create pointer-like types out of a memref as well as operations that, conversely, create memref out of pointers combined with additional information. Before implementing such operations, dialect authors are advised to carefully consider the implication of such operations on aliasing properties of the resulting IR.

Interoperability with C is often cited to motivate an opaque cast from memrefs to pointers. The LLVM IR target provides an interface compatible with C for a well-defined subset of memrefs with strided layout. At the function boundary, it even provides a minimalist support for passing memrefs as bare pointers provided their sizes are known statically and their layout is trivially identity.

What’s with “op symbol declaration cannot have public visibility”? ¶

A common mistake is to try to provide a function declaration (that is a function without a body) but leaving it “public”. Declaration must be private, only definitions can be public in the MLIR symbol system. See the symbol visibility documentation.

I’m confused about iterating on getUsers() vs getUses(): what’s the difference? ¶

The “users” of an SSA value are instances of Operation, while the “uses” refer to the operands of these operations. For example considering test.op(%0, %0) : ..., when iterating on the “uses” of %0 you would see two instances of OpOperand (one for each use in test.op), whereas iterating on the “users” of %0 would yield directly two Operation * corresponding to test.op. Note that you see test.op twice as it is twice a user of %0, it’s up to the call site to use a set to unique these if needed. The tutorial on use-def chains may help understand the details as well.

How to programmatically obtain the “name” of the SSA value (%foo)? ¶

The values names are not part of the IR and are only there to make textual representation of the IR easier for humans to read. They are generated by the IR printer on-the-fly and may differ depending on the printer configuration. While it is technically possible to configure the printer to produce predictable names, in particular names with specific prefixes via the OpAsmOpInterface, one is strongly discouraged from relying on the textual names. Therefore there is intentionally no support for obtaining these names easily.