llvm-project/mlir/g3doc/ConversionToLLVMDialect.md

Conversion to the LLVM IR Dialect

Conversion to the LLVM IR Dialect can be performed by the specialized dialect conversion pass by running

mlir-opt -lower-to-llvm <filename.mlir>

It performs type and operation conversions for a subset of operations from standard, built-in and super-vector dialects as described in this document. We use the terminology defined by the LLVM IR Dialect description throughout this document.

[TOC]

Type Conversion

Scalar Types

Scalar types are converted to their LLVM counterparts if they exist. The following conversions are currently implemented.

• i* converts to !llvm.type<"i*">
• f16 converts to !llvm.type<"half">
• f32 converts to !llvm.type<"float">
• f64 converts to !llvm.type<"double">

Note: bf16 type is not supported by LLVM IR and cannot be converted.

Index Type

Index type is converted to a wrapped LLVM IR integer with bitwidth equal to the bitwidth of the pointer size as specified by the data layout of the LLVM module contained in the LLVM Dialect object. For example, on x86-64 CPUs it converts to !llvm.type<"i64">.

Vector Types

LLVM IR only supports one-dimensional vectors, unlike MLIR where vectors can be multi-dimensional. Vector types cannot be nested in either IR. In the one-dimensional case, MLIR vectors are converted to LLVM IR vectors of the same size with element type converted using these conversion rules. In the n-dimensional case, MLIR vectors are converted to (n-1)-dimensional array types of one-dimensional vectors.

For example, vector<4 x f32> converts to !llvm.type<"<4 x float>"> and vector<4 x 8 x 16 f32> converts to !llvm<"[4 x [8 x <16 x float>]]">.

Memref Types

Memref types in MLIR have both static and dynamic information associated with them. The dynamic information comprises the buffer pointer as well as sizes and strides of any dynamically sized dimensions. Memref types are normalized and converted to a descriptor that is only dependent on the rank of the memref. The descriptor contains:

1. the pointer to the data buffer, followed by
2. a lowered index-type integer containing the distance between the beginning of the buffer and the first element to be accessed through the memref, followed by
3. an array containing as many index-type integers as the rank of the memref: the array represents the size, in number of elements, of the memref along the given dimension. For constant MemRef dimensions, the corresponding size entry is a constant whose runtime value must match the static value, followed by
4. a second array containing as many 64-bit integers as the rank of the MemRef: the second array represents the "stride" (in tensor abstraction sense), i.e. the number of consecutive elements of the underlying buffer.

For constant memref dimensions, the corresponding size entry is a constant whose runtime value matches the static value. This normalization serves as an ABI for the memref type to interoperate with externally linked functions. In the particular case of rank 0 memrefs, the size and stride arrays are omitted, resulting in a struct containing a pointer + offset.

Examples:

memref<f32> -> !llvm.type<"{ float*, i64 }">
memref<1 x f32> -> !llvm.type<"{ float*, i64, [1 x i64], [1 x i64] }">
memref<? x f32> -> !llvm.type<"{ float*, i64, [1 x i64], [1 x i64] }">
memref<10x42x42x43x123 x f32> -> !llvm.type<"{ float*, i64, [5 x i64], [5 x i64] }">
memref<10x?x42x?x123 x f32> -> !llvm.type<"{ float*, i64, [5 x i64], [5 x i64]  }">

// Memref types can have vectors as element types
memref<1x? x vector<4xf32>> -> !llvm.type<"{ <4 x float>*, i64, [1 x i64], [1 x i64] }">

Function Types

Function types get converted to LLVM function types. The arguments are converted individually according to these rules. The result types need to accommodate the fact that LLVM IR functions always have a return type, which may be a Void type. The converted function always has a single result type. If the original function type had no results, the converted function will have one result of the wrapped void type. If the original function type had one result, the converted function will have one result converted using these rules. Otherwise, the result type will be a wrapped LLVM IR structure type where each element of the structure corresponds to one of the results of the original function, converted using these rules. In high-order functions, function-typed arguments and results are converted to a wrapped LLVM IR function pointer type (since LLVM IR does not allow passing functions to functions without indirection) with the pointee type converted using these rules.

Examples:

// zero-ary function type with no results.
() -> ()
// is converted to a zero-ary function with `void` result
!llvm.type<"void ()">

// unary function with one result
(i32) -> (i64)
// has its argument and result type converted, before creating the LLVM IR function type
!llvm.type<"i64 (i32)">

// binary function with one result
(i32, f32) -> (i64)
// has its arguments handled separately
!llvm.type<"i64 (i32, float)">

// binary function with two results
(i32, f32) -> (i64, f64)
// has its result aggregated into a structure type
!llvm.type<"{i64, double} (i32, f32)">

// function-typed arguments or results in higher-order functions
(() -> ()) -> (() -> ())
// are converted into pointers to functions
!llvm.type<"void ()* (void ()*)">

Calling Convention

Function Signature Conversion

MLIR function type is built into the representation, even the functions in dialects including a first-class function type must have the built-in MLIR function type. During the conversion to LLVM IR, function signatures are converted as follows:

• the outer type remains the built-in MLIR function;
• function arguments are converted individually following these rules;
• function results:
  • zero-result functions remain zero-result;
  • single-result functions have their result type converted according to these rules;
  • multi-result functions have a single result type of the wrapped LLVM IR structure type with elements corresponding to the converted original results.

Rationale: function definitions remain analyzable within MLIR without having to abstract away the function type. In order to remain consistent with the regular MLIR functions, we do not introduce a void result type since we cannot create a value of void type that MLIR passes might expect to be returned from a function.

Examples:

// zero-ary function type with no results.
func @foo() -> ()
// remains as is
func @foo() -> ()

// unary function with one result
func @bar(i32) -> (i64)
// has its argument and result type converted
func @bar(!llvm.type<"i32">) -> !llvm.type<"i64">

// binary function with one result
func @baz(i32, f32) -> (i64)
// has its arguments handled separately
func @baz(!llvm.type<"i32">, !llvm.type<"float">) -> !llvm.type<"i64">

// binary function with two results
func @qux(i32, f32) -> (i64, f64)
// has its result aggregated into a structure type
func @qux(!llvm.type<"i32">, !llvm.type<"float">) -> !llvm.type<"{i64, double}">

// function-typed arguments or results in higher-order functions
func @quux(() -> ()) -> (() -> ())
// are converted into pointers to functions
func @quux(!llvm.type<"void ()*">) -> !llvm.type<"void ()*">
// the call flow is handled by the LLVM dialect `call` operation supporting both
// direct and indirect calls

Result Packing

In case of multi-result functions, the returned values are inserted into a structure-typed value before being returned and extracted from it at the call site. This transformation is a part of the conversion and is transparent to the defines and uses of the values being returned.

Example:

func @foo(%arg0: i32, %arg1: i64) -> (i32, i64) {
  return %arg0, %arg1 : i32, i64
}
func @bar() {
  %0 = constant 42 : i32
  %1 = constant 17 : i64
  %2:2 = call @foo(%0, %1) : (i32, i64) -> (i32, i64)
  "use_i32"(%2#0) : (i32) -> ()
  "use_i64"(%2#1) : (i64) -> ()
}

// is transformed into

func @foo(%arg0: !llvm.type<"i32">, %arg1: !llvm.type<"i64">) -> !llvm.type<"{i32, i64}"> {
  // insert the vales into a structure
  %0 = llvm.mlir.undef :  !llvm.type<"{i32, i64}">
  %1 = llvm.insertvalue %arg0, %0[0] : !llvm.type<"{i32, i64}">
  %2 = llvm.insertvalue %arg1, %1[1] : !llvm.type<"{i32, i64}">

  // return the structure value
  llvm.return %2 : !llvm.type<"{i32, i64}">
}
func @bar() {
  %0 = llvm.mlir.constant(42 : i32) : !llvm.type<"i32">
  %1 = llvm.mlir.constant(17) : !llvm.type<"i64">

  // call and extract the values from the structure
  %2 = llvm.call @bar(%0, %1) : (%arg0: !llvm.type<"i32">, %arg1: !llvm.type<"i64">) -> !llvm.type<"{i32, i64}">
  %3 = llvm.extractvalue %2[0] : !llvm.type<"{i32, i64}">
  %4 = llvm.extractvalue %2[1] : !llvm.type<"{i32, i64}">

  // use as before
  "use_i32"(%3) : (!llvm.type<"i32">) -> ()
  "use_i64"(%4) : (!llvm.type<"i64">) -> ()
}

Repeated Successor Removal

Since the goal of the LLVM IR dialect is to reflect LLVM IR in MLIR, the dialect and the conversion procedure must account for the differences between block arguments and LLVM IR PHI nodes. In particular, LLVM IR disallows PHI nodes with different values coming from the same source. Therefore, the LLVM IR dialect disallows operations that have identical successors accepting arguments, which would lead to invalid PHI nodes. The conversion process resolves the potential PHI source ambiguity by injecting dummy blocks if the same block is used more than once as a successor in an instruction. These dummy blocks branch unconditionally to the original successors, pass them the original operands (available in the dummy block because it is dominated by the original block) and are used instead of them in the original terminator operation.

Example:

  cond_br %0, ^bb1(%1 : i32), ^bb1(%2 : i32)
^bb1(%3 : i32)
  "use"(%3) : (i32) -> ()

leads to a new basic block being inserted,

  cond_br %0, ^bb1(%1 : i32), ^dummy
^bb1(%3 : i32):
  "use"(%3) : (i32) -> ()
^dummy:
  br ^bb1(%4 : i32)

before the conversion to the LLVM IR dialect:

  llvm.cond_br  %0, ^bb1(%1 : !llvm.type<"i32">), ^dummy
^bb1(%3 : !llvm.type<"i32">):
  "use"(%3) : (!llvm.type<"i32">) -> ()
^dummy:
  llvm.br ^bb1(%2 : !llvm.type<"i32">)

Memref Model

Memref Descriptor

Within a converted function, a memref-typed value is represented by a memref descriptor, the type of which is the structure type obtained by converting from the memref type. This descriptor holds a pointer to a linear buffer storing the data, and dynamic sizes of the memref value. It is created by the allocation operation and is updated by the conversion operations that may change static dimensions into dynamic and vice versa.

Note: LLVM IR conversion does not support memrefs in non-default memory spaces or memrefs with non-identity layouts.

Index Linearization

Accesses to a memref element are transformed into an access to an element of the buffer pointed to by the descriptor. The position of the element in the buffer is calculated by linearizing memref indices in row-major order (lexically first index is the slowest varying, similar to C). The computation of the linear address is emitted as arithmetic operation in the LLVM IR dialect. Static sizes are introduced as constants. Dynamic sizes are extracted from the memref descriptor.

Accesses to zero-dimensional memref (that are interpreted as pointers to the elemental type) are directly converted into llvm.load or llvm.store without any pointer manipulations.

Examples:

An access to a zero-dimensional memref is converted into a plain load:

// before
%0 = load %m[] : memref<f32>

// after
%0 = llvm.load %m : !llvm.type<"float*">

An access to a memref with indices:

%0 = load %m[1,2,3,4] : memref<10x?x13x?xf32>

is transformed into the equivalent of the following code:

// obtain the buffer pointer
%b = llvm.extractvalue %m[0] : !llvm.type<"{float*, i64, i64}">

// obtain the components for the index
%sub1 = llvm.mlir.constant(1) : !llvm.type<"i64">  // first subscript
%sz2 = llvm.extractvalue %m[1]
    : !llvm.type<"{float*, i64, i64}"> // second size (dynamic, second descriptor element)
%sub2 = llvm.mlir.constant(2) : !llvm.type<"i64">  // second subscript
%sz3 = llvm.mlir.constant(13) : !llvm.type<"i64">  // third size (static)
%sub3 = llvm.mlir.constant(3) : !llvm.type<"i64">  // third subscript
%sz4 = llvm.extractvalue %m[1]
    : !llvm.type<"{float*, i64, i64}"> // fourth size (dynamic, third descriptor element)
%sub4 = llvm.mlir.constant(4) : !llvm.type<"i64">  // fourth subscript

// compute the linearized index
// %sub4 + %sub3 * %sz4 + %sub2 * (%sz3 * %sz4) + %sub1 * (%sz2 * %sz3 * %sz4) =
// = ((%sub1 * %sz2 + %sub2) * %sz3 + %sub3) * %sz4 + %sub4
%idx0 = llvm.mul %sub1, %sz2 : !llvm.type<"i64">
%idx1 = llvm.add %idx0, %sub : !llvm.type<"i64">
%idx2 = llvm.mul %idx1, %sz3 : !llvm.type<"i64">
%idx3 = llvm.add %idx2, %sub3 : !llvm.type<"i64">
%idx4 = llvm.mul %idx3, %sz4 : !llvm.type<"i64">
%idx5 = llvm.add %idx4, %sub4 : !llvm.type<"i64">

// obtain the element address
%a = llvm.getelementptr %b[%idx5] : (!llvm.type<"float*">, !llvm.type<"i64">) -> !llvm.type<"float*">

// perform the actual load
%0 = llvm.load %a : !llvm.type<"float*">

In practice, the subscript and size extraction will be interleaved with the linear index computation. For stores, the address computation code is identical and only the actual store operation is different.

Note: the conversion does not perform any sort of common subexpression elimination when emitting memref accesses.