MLIR

Multi-Level IR Compiler Framework

'emitc' Dialect

Dialect to generate C/C++ from MLIR. The EmitC dialect allows to convert operations from other MLIR dialects to EmitC ops. Those can be translated to C/C++ via the Cpp emitter.

The following convention is followed:

  • If template arguments are passed to an emitc.call_opaque operation, C++ is generated.
  • If tensors are used, C++ is generated.
  • If multiple return values are used within in a functions or an emitc.call_opaque operation, C++11 is required.
  • If floating-point type template arguments are passed to an emitc.call_opaque operation, C++20 is required.
  • Else the generated code is compatible with C99.

These restrictions are neither inherent to the EmitC dialect itself nor to the Cpp emitter and therefore need to be considered while implementing conversions.

After the conversion, C/C++ code can be emitted with mlir-translate. The tool supports translating MLIR to C/C++ by passing -mlir-to-cpp. Furthermore, code with variables declared at top can be generated by passing the additional argument -declare-variables-at-top.

Besides operations part of the EmitC dialect, the Cpp targets supports translating the following operations:

  • ‘cf’ Dialect
    • cf.br
    • cf.cond_br
  • ‘func’ Dialect
    • func.call
    • func.func
    • func.return
  • ‘arith’ Dialect
    • arith.constant

Operations 

source

emitc.add (emitc::AddOp) 

Addition operation

Syntax:

operation ::= `emitc.add` operands attr-dict `:` functional-type(operands, results)

With the add operation the arithmetic operator + (addition) can be applied.

Example:

// Custom form of the addition operation.
%0 = emitc.add %arg0, %arg1 : (i32, i32) -> i32
%1 = emitc.add %arg2, %arg3 : (!emitc.ptr<f32>, i32) -> !emitc.ptr<f32>
// Code emitted for the operations above.
int32_t v5 = v1 + v2;
float* v6 = v3 + v4;

Operands: 

OperandDescription
lhsany type
rhsany type

Results: 

ResultDescription
«unnamed»any type

emitc.apply (emitc::ApplyOp) 

Apply operation

Syntax:

operation ::= `emitc.apply` $applicableOperator `(` $operand `)` attr-dict `:` functional-type($operand, results)

With the apply operation the operators & (address of) and * (contents of) can be applied to a single operand.

Example:

// Custom form of applying the & operator.
%0 = emitc.apply "&"(%arg0) : (i32) -> !emitc.ptr<i32>

// Generic form of the same operation.
%0 = "emitc.apply"(%arg0) {applicableOperator = "&"}
    : (i32) -> !emitc.ptr<i32>

Attributes: 

AttributeMLIR TypeDescription
applicableOperator::mlir::StringAttrstring attribute

Operands: 

OperandDescription
operandany type

Results: 

ResultDescription
resultany type

emitc.assign (emitc::AssignOp) 

Assign operation

Syntax:

operation ::= `emitc.assign` $value `:` type($value) `to` $var `:` type($var) attr-dict

The assign operation stores an SSA value to the location designated by an EmitC variable. This operation doesn’t return any value. The assigned value must be of the same type as the variable being assigned. The operation is emitted as a C/C++ ‘=’ operator.

Example:

// Integer variable
%0 = "emitc.variable"(){value = 42 : i32} : () -> i32
%1 = emitc.call_opaque "foo"() : () -> (i32)

// Assign emitted as `... = ...;`
"emitc.assign"(%0, %1) : (i32, i32) -> ()

Operands: 

OperandDescription
varany type
valueany type

emitc.call (emitc::CallOp) 

Call operation

Syntax:

operation ::= `emitc.call` $callee `(` $operands `)` attr-dict `:` functional-type($operands, results)

The emitc.call operation represents a direct call to an emitc.func that is within the same symbol scope as the call. The operands and result type of the call must match the specified function type. The callee is encoded as a symbol reference attribute named “callee”.

Example:

%2 = emitc.call @my_add(%0, %1) : (f32, f32) -> f32

Interfaces: CallOpInterface, SymbolUserOpInterface

Attributes: 

AttributeMLIR TypeDescription
callee::mlir::FlatSymbolRefAttrflat symbol reference attribute

Operands: 

OperandDescription
operandsvariadic of any type

Results: 

ResultDescription
«unnamed»variadic of any type

emitc.call_opaque (emitc::CallOpaqueOp) 

Opaque call operation

Syntax:

operation ::= `emitc.call_opaque` $callee `(` $operands `)` attr-dict `:` functional-type($operands, results)

The call_opaque operation represents a C++ function call. The callee can be an arbitrary non-empty string. The call allows specifying order of operands and attributes in the call as follows:

  • integer value of index type refers to an operand;
  • attribute which will get lowered to constant value in call;

Example:

// Custom form defining a call to `foo()`.
%0 = emitc.call_opaque "foo" () : () -> i32

// Generic form of the same operation.
%0 = "emitc.call_opaque"() {callee = "foo"} : () -> i32

Attributes: 

AttributeMLIR TypeDescription
callee::mlir::StringAttrstring attribute
args::mlir::ArrayAttrarray attribute
template_args::mlir::ArrayAttrarray attribute

Operands: 

OperandDescription
operandsvariadic of any type

Results: 

ResultDescription
«unnamed»variadic of any type

emitc.cast (emitc::CastOp) 

Cast operation

Syntax:

operation ::= `emitc.cast` $source attr-dict `:` type($source) `to` type($dest)

The cast operation performs an explicit type conversion and is emitted as a C-style cast expression. It can be applied to integer, float, index and EmitC types.

Example:

// Cast from `int32_t` to `float`
%0 = emitc.cast %arg0: i32 to f32

// Cast from `void` to `int32_t` pointer
%1 = emitc.cast %arg1 :
    !emitc.ptr<!emitc.opaque<"void">> to !emitc.ptr<i32>

Traits: SameOperandsAndResultShape

Interfaces: CastOpInterface

Operands: 

OperandDescription
sourceany type

Results: 

ResultDescription
destany type

emitc.cmp (emitc::CmpOp) 

Comparison operation

Syntax:

operation ::= `emitc.cmp` $predicate `,` operands attr-dict `:` functional-type(operands, results)

With the cmp operation the comparison operators ==, !=, <, <=, >, >=, <=> can be applied.

Its first argument is an attribute that defines the comparison operator:

  • equal to (mnemonic: "eq"; integer value: 0)
  • not equal to (mnemonic: "ne"; integer value: 1)
  • less than (mnemonic: "lt"; integer value: 2)
  • less than or equal to (mnemonic: "le"; integer value: 3)
  • greater than (mnemonic: "gt"; integer value: 4)
  • greater than or equal to (mnemonic: "ge"; integer value: 5)
  • three-way-comparison (mnemonic: "three_way"; integer value: 6)

Example:

// Custom form of the cmp operation.
%0 = emitc.cmp eq, %arg0, %arg1 : (i32, i32) -> i1
%1 = emitc.cmp lt, %arg2, %arg3 : 
    (
      !emitc.opaque<"std::valarray<float>">,
      !emitc.opaque<"std::valarray<float>">
    ) -> !emitc.opaque<"std::valarray<bool>">
// Code emitted for the operations above.
bool v5 = v1 == v2;
std::valarray<bool> v6 = v3 < v4;

Attributes: 

AttributeMLIR TypeDescription
predicate::mlir::emitc::CmpPredicateAttr
allowed 64-bit signless integer cases: 0, 1, 2, 3, 4, 5, 6

Enum cases:

  • eq (eq)
  • ne (ne)
  • lt (lt)
  • le (le)
  • gt (gt)
  • ge (ge)
  • three_way (three_way)

Operands: 

OperandDescription
lhsany type
rhsany type

Results: 

ResultDescription
«unnamed»any type

emitc.constant (emitc::ConstantOp) 

Constant operation

The constant operation produces an SSA value equal to some constant specified by an attribute. This can be used to form simple integer and floating point constants, as well as more exotic things like tensor constants. The constant operation also supports the EmitC opaque attribute and the EmitC opaque type. Since folding is supported, it should not be used with pointers.

Example:

// Integer constant
%0 = "emitc.constant"(){value = 42 : i32} : () -> i32

// Constant emitted as `char = CHAR_MIN;`
%1 = "emitc.constant"()
    {value = #emitc.opaque<"CHAR_MIN"> : !emitc.opaque<"char">}
    : () -> !emitc.opaque<"char">

Traits: ConstantLike

Attributes: 

AttributeMLIR TypeDescription
value::mlir::AttributeAn opaque attribute or TypedAttr instance

Results: 

ResultDescription
«unnamed»any type

emitc.declare_func (emitc::DeclareFuncOp) 

An operation to declare a function

Syntax:

operation ::= `emitc.declare_func` $sym_name attr-dict

The declare_func operation allows to insert a function declaration for an emitc.func at a specific position. The operation only requires the callee of the emitc.func to be specified as an attribute.

Example:

emitc.declare_func @bar
emitc.func @foo(%arg0: i32) -> i32 {
  %0 = emitc.call @bar(%arg0) : (i32) -> (i32)
  emitc.return %0 : i32
}

emitc.func @bar(%arg0: i32) -> i32 {
  emitc.return %arg0 : i32
}
// Code emitted for the operations above.
int32_t bar(int32_t v1);
int32_t foo(int32_t v1) {
  int32_t v2 = bar(v1);
  return v2;
}

int32_t bar(int32_t v1) {
  return v1;
}

Interfaces: SymbolUserOpInterface

Attributes: 

AttributeMLIR TypeDescription
sym_name::mlir::FlatSymbolRefAttrflat symbol reference attribute

emitc.div (emitc::DivOp) 

Division operation

Syntax:

operation ::= `emitc.div` operands attr-dict `:` functional-type(operands, results)

With the div operation the arithmetic operator / (division) can be applied.

Example:

// Custom form of the division operation.
%0 = emitc.div %arg0, %arg1 : (i32, i32) -> i32
%1 = emitc.div %arg2, %arg3 : (f32, f32) -> f32
// Code emitted for the operations above.
int32_t v5 = v1 / v2;
float v6 = v3 / v4;

Operands: 

OperandDescription
«unnamed»floating-point or integer or index or EmitC opaque type
«unnamed»floating-point or integer or index or EmitC opaque type

Results: 

ResultDescription
«unnamed»floating-point or integer or index or EmitC opaque type

emitc.expression (emitc::ExpressionOp) 

Expression operation

Syntax:

operation ::= `emitc.expression` attr-dict (`noinline` $do_not_inline^)? `:` type($result) $region

The expression operation returns a single SSA value which is yielded by its single-basic-block region. The operation doesn’t take any arguments.

As the operation is to be emitted as a C expression, the operations within its body must form a single Def-Use tree of emitc ops whose result is yielded by a terminating yield.

Example:

%r = emitc.expression : () -> i32 {
  %0 = emitc.add %a, %b : (i32, i32) -> i32
  %1 = emitc.call "foo"(%0) : () -> i32
  %2 = emitc.add %c, %d : (i32, i32) -> i32
  %3 = emitc.mul %1, %2 : (i32, i32) -> i32
  yield %3
}

May be emitted as

int32_t v7 = foo(v1 + v2) * (v3 + v4);

The operations allowed within expression body are emitc.add, emitc.apply, emitc.call, emitc.cast, emitc.cmp, emitc.div, emitc.mul, emitc.rem and emitc.sub.

When specified, the optional do_not_inline indicates that the expression is to be emitted as seen above, i.e. as the rhs of an EmitC SSA value definition. Otherwise, the expression may be emitted inline, i.e. directly at its use.

Traits: HasOnlyGraphRegion, NoRegionArguments, SingleBlockImplicitTerminator<emitc::YieldOp>, SingleBlock

Attributes: 

AttributeMLIR TypeDescription
do_not_inline::mlir::UnitAttrunit attribute

Results: 

ResultDescription
resultany type

emitc.for (emitc::ForOp) 

For operation

The emitc.for operation represents a C loop of the following form:

for (T i = lb; i < ub; i += step) { /* ... */ } // where T is typeof(lb)

The operation takes 3 SSA values as operands that represent the lower bound, upper bound and step respectively, and defines an SSA value for its induction variable. It has one region capturing the loop body. The induction variable is represented as an argument of this region. This SSA value is a signless integer or index. The step is a value of same type.

This operation has no result. The body region must contain exactly one block that terminates with emitc.yield. Calling ForOp::build will create such a region and insert the terminator implicitly if none is defined, so will the parsing even in cases when it is absent from the custom format. For example:

// Index case.
emitc.for %iv = %lb to %ub step %step {
  ... // body
}
...
// Integer case.
emitc.for %iv_32 = %lb_32 to %ub_32 step %step_32 : i32 {
  ... // body
}

Traits: RecursiveMemoryEffects, SingleBlockImplicitTerminator<emitc::YieldOp>, SingleBlock

Operands: 

OperandDescription
lowerBoundinteger or index or EmitC opaque type
upperBoundinteger or index or EmitC opaque type
stepinteger or index or EmitC opaque type

emitc.func (emitc::FuncOp) 

An operation with a name containing a single SSACFG region

Operations within the function cannot implicitly capture values defined outside of the function, i.e. Functions are IsolatedFromAbove. All external references must use function arguments or attributes that establish a symbolic connection (e.g. symbols referenced by name via a string attribute like SymbolRefAttr). While the MLIR textual form provides a nice inline syntax for function arguments, they are internally represented as “block arguments” to the first block in the region.

Only dialect attribute names may be specified in the attribute dictionaries for function arguments, results, or the function itself.

Example:

// A function with no results:
emitc.func @foo(%arg0 : i32) {
  emitc.call_opaque "bar" (%arg0) : (i32) -> ()
  emitc.return
}

// A function with its argument as single result:
emitc.func @foo(%arg0 : i32) -> i32 {
  emitc.return %arg0 : i32
}

// A function with specifiers attribute:
emitc.func @example_specifiers_fn_attr() -> i32
            attributes {specifiers = ["static","inline"]} {
  %0 = emitc.call_opaque "foo" (): () -> i32
  emitc.return %0 : i32
}

// An external function definition:
emitc.func private @extern_func(i32)
                    attributes {specifiers = ["extern"]}

Traits: AutomaticAllocationScope, IsolatedFromAbove

Interfaces: CallableOpInterface, FunctionOpInterface, Symbol

Attributes: 

AttributeMLIR TypeDescription
sym_name::mlir::StringAttrstring attribute
function_type::mlir::TypeAttrtype attribute of function type
specifiers::mlir::ArrayAttrstring array attribute
arg_attrs::mlir::ArrayAttrArray of dictionary attributes
res_attrs::mlir::ArrayAttrArray of dictionary attributes

emitc.if (emitc::IfOp) 

If-then-else operation

The if operation represents an if-then-else construct for conditionally executing two regions of code. The operand to an if operation is a boolean value. For example:

emitc.if %b  {
  ...
} else {
  ...
}

The “then” region has exactly 1 block. The “else” region may have 0 or 1 blocks. The blocks are always terminated with emitc.yield, which can be left out to be inserted implicitly. This operation doesn’t produce any results.

Traits: NoRegionArguments, RecursiveMemoryEffects, SingleBlockImplicitTerminator<emitc::YieldOp>, SingleBlock

Interfaces: RegionBranchOpInterface

Operands: 

OperandDescription
condition1-bit signless integer

emitc.include (emitc::IncludeOp) 

Include operation

The include operation allows to define a source file inclusion via the #include directive.

Example:

// Custom form defining the inclusion of `<myheader>`.
emitc.include <"myheader.h">

// Generic form of the same operation.
"emitc.include" (){include = "myheader.h", is_standard_include} : () -> ()

// Custom form defining the inclusion of `"myheader"`.
emitc.include "myheader.h"

// Generic form of the same operation.
"emitc.include" (){include = "myheader.h"} : () -> ()

Traits: HasParent<ModuleOp>

Attributes: 

AttributeMLIR TypeDescription
include::mlir::StringAttrstring attribute
is_standard_include::mlir::UnitAttrunit attribute

emitc.literal (emitc::LiteralOp) 

Literal operation

Syntax:

operation ::= `emitc.literal` $value attr-dict `:` type($result)

The literal operation produces an SSA value equal to some constant specified by an attribute.

Traits: AlwaysSpeculatableImplTrait

Interfaces: ConditionallySpeculatable, NoMemoryEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Attributes: 

AttributeMLIR TypeDescription
value::mlir::StringAttrstring attribute

Results: 

ResultDescription
resultany type

emitc.mul (emitc::MulOp) 

Multiplication operation

Syntax:

operation ::= `emitc.mul` operands attr-dict `:` functional-type(operands, results)

With the mul operation the arithmetic operator * (multiplication) can be applied.

Example:

// Custom form of the multiplication operation.
%0 = emitc.mul %arg0, %arg1 : (i32, i32) -> i32
%1 = emitc.mul %arg2, %arg3 : (f32, f32) -> f32
// Code emitted for the operations above.
int32_t v5 = v1 * v2;
float v6 = v3 * v4;

Operands: 

OperandDescription
«unnamed»floating-point or integer or index or EmitC opaque type
«unnamed»floating-point or integer or index or EmitC opaque type

Results: 

ResultDescription
«unnamed»floating-point or integer or index or EmitC opaque type

emitc.rem (emitc::RemOp) 

Remainder operation

Syntax:

operation ::= `emitc.rem` operands attr-dict `:` functional-type(operands, results)

With the rem operation the arithmetic operator % (remainder) can be applied.

Example:

// Custom form of the remainder operation.
%0 = emitc.rem %arg0, %arg1 : (i32, i32) -> i32
// Code emitted for the operation above.
int32_t v5 = v1 % v2;

Operands: 

OperandDescription
«unnamed»integer or index or EmitC opaque type
«unnamed»integer or index or EmitC opaque type

Results: 

ResultDescription
«unnamed»integer or index or EmitC opaque type

emitc.return (emitc::ReturnOp) 

Function return operation

Syntax:

operation ::= `emitc.return` attr-dict ($operand^ `:` type($operand))?

The emitc.return operation represents a return operation within a function. The operation takes zero or exactly one operand and produces no results. The operand number and type must match the signature of the function that contains the operation.

Example:

emitc.func @foo() : (i32) {
  ...
  emitc.return %0 : i32
}

Traits: AlwaysSpeculatableImplTrait, HasParent<FuncOp>, ReturnLike, Terminator

Interfaces: ConditionallySpeculatable, NoMemoryEffect (MemoryEffectOpInterface), RegionBranchTerminatorOpInterface

Effects: MemoryEffects::Effect{}

Operands: 

OperandDescription
operandany type

emitc.sub (emitc::SubOp) 

Subtraction operation

Syntax:

operation ::= `emitc.sub` operands attr-dict `:` functional-type(operands, results)

With the sub operation the arithmetic operator - (subtraction) can be applied.

Example:

// Custom form of the substraction operation.
%0 = emitc.sub %arg0, %arg1 : (i32, i32) -> i32
%1 = emitc.sub %arg2, %arg3 : (!emitc.ptr<f32>, i32) -> !emitc.ptr<f32>
%2 = emitc.sub %arg4, %arg5 : (!emitc.ptr<i32>, !emitc.ptr<i32>)
    -> !emitc.opaque<"ptrdiff_t">
// Code emitted for the operations above.
int32_t v7 = v1 - v2;
float* v8 = v3 - v4;
ptrdiff_t v9 = v5 - v6;

Operands: 

OperandDescription
lhsany type
rhsany type

Results: 

ResultDescription
«unnamed»any type

emitc.variable (emitc::VariableOp) 

Variable operation

The variable operation produces an SSA value equal to some value specified by an attribute. This can be used to form simple integer and floating point variables, as well as more exotic things like tensor variables. The variable operation also supports the EmitC opaque attribute and the EmitC opaque type. If further supports the EmitC pointer type, whereas folding is not supported. The variable is emitted as a C/C++ local variable.

Example:

// Integer variable
%0 = "emitc.variable"(){value = 42 : i32} : () -> i32

// Variable emitted as `int32_t* = NULL;`
%1 = "emitc.variable"()
    {value = #emitc.opaque<"NULL"> : !emitc.opaque<"int32_t*">}
    : () -> !emitc.opaque<"int32_t*">

Since folding is not supported, it can be used with pointers. As an example, it is valid to create pointers to variable operations by using apply operations and pass these to a call operation.

%0 = "emitc.variable"() {value = 0 : i32} : () -> i32
%1 = "emitc.variable"() {value = 0 : i32} : () -> i32
%2 = emitc.apply "&"(%0) : (i32) -> !emitc.ptr<i32>
%3 = emitc.apply "&"(%1) : (i32) -> !emitc.ptr<i32>
emitc.call_opaque "write"(%2, %3)
  : (!emitc.ptr<i32>, !emitc.ptr<i32>) -> ()

Attributes: 

AttributeMLIR TypeDescription
value::mlir::AttributeAn opaque attribute or TypedAttr instance

Results: 

ResultDescription
«unnamed»any type

emitc.verbatim (emitc::VerbatimOp) 

Verbatim operation

Syntax:

operation ::= `emitc.verbatim` $value attr-dict

The verbatim operation produces no results and the value is emitted as is followed by a line break (’\n’ character) during translation.

Note: Use with caution. This operation can have arbitrary effects on the semantics of the emitted code. Use semantically more meaningful operations whenever possible. Additionally this op is NOT intended to be used to inject large snippets of code.

This operation can be used in situations where a more suitable operation is not yet implemented in the dialect or where preprocessor directives interfere with the structure of the code. One example of this is to declare the linkage of external symbols to make the generated code usable in both C and C++ contexts:

#ifdef __cplusplus
extern "C" {
#endif

...

#ifdef __cplusplus
}
#endif

Attributes: 

AttributeMLIR TypeDescription
value::mlir::StringAttrstring attribute

emitc.yield (emitc::YieldOp) 

Block termination operation

Syntax:

operation ::= `emitc.yield` attr-dict ($result^ `:` type($result))?

“yield” terminates its parent EmitC op’s region, optionally yielding an SSA value. The semantics of how the values are yielded is defined by the parent operation. If “yield” has an operand, the operand must match the parent operation’s result. If the parent operation defines no values, then the “emitc.yield” may be left out in the custom syntax and the builders will insert one implicitly. Otherwise, it has to be present in the syntax to indicate which value is yielded.

Traits: AlwaysSpeculatableImplTrait, HasParent<ExpressionOp, IfOp, ForOp>, Terminator

Interfaces: ConditionallySpeculatable, NoMemoryEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Operands: 

OperandDescription
resultany type

Attributes 

OpaqueAttr 

An opaque attribute

Syntax:

#emitc.opaque<
  ::llvm::StringRef   # value
>

An opaque attribute of which the value gets emitted as is.

Example:

#emitc.opaque<"">
#emitc.opaque<"NULL">
#emitc.opaque<"nullptr">

Parameters: 

ParameterC++ typeDescription
value::llvm::StringRefthe opaque value

Types 

OpaqueType 

EmitC opaque type

Syntax:

!emitc.opaque<
  ::llvm::StringRef   # value
>

An opaque data type of which the value gets emitted as is.

Example:

!emitc.opaque<"int">
!emitc.opaque<"mytype">
!emitc.opaque<"std::vector<std::string>">

Parameters: 

ParameterC++ typeDescription
value::llvm::StringRefthe opaque value

PointerType 

EmitC pointer type

Syntax:

!emitc.ptr<
  Type   # pointee
>

A pointer data type.

Example:

// Pointer emitted as `int32_t*`
!emitc.ptr<i32>
// Pointer emitted as `float*`
!emitc.ptr<f32>
// Pointer emitted as `int*`
!emitc.ptr<!emitc.opaque<"int">>

Parameters: 

ParameterC++ typeDescription
pointeeType