FIR Language Reference

This page contains an overview of the Fortran IR operations, their syntax, and example usages.

fir.absent (::fir::AbsentOp)

create value to be passed for absent optional function argument

Syntax:

operation ::= `fir.absent` type($intype) attr-dict

Given the type of a function argument, create a value that will signal that an optional argument is absent in the call. On the caller side, fir.is_present can be used to query if the value of an optional argument was created with a fir.absent operation. It is undefined to use a value that was created by a fir.absent op in any other operation than fir.call and fir.is_present.

  %1 = fir.absent fir.box<fir.array<?xf32>>
  fir.call @_QPfoo(%1) : (fir.box<fir.array<?xf32>>) -> ()

Interfaces: NoSideEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Results:

Result Description
intype any reference or box like

fir.addc (::fir::AddcOp)

Syntax:

operation ::= `fir.addc` operands attr-dict `:` type($result)

Traits: Commutative, SameOperandsAndResultType

Interfaces: InferTypeOpInterface, NoSideEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Operands:

Operand Description
lhs Complex type
rhs Complex type

Results:

Result Description
result any type

fir.address_of (::fir::AddrOfOp)

convert a symbol to an SSA value

Syntax:

operation ::= `fir.address_of` `(` $symbol `)` attr-dict `:` type($resTy)

Convert a symbol (a function or global reference) to an SSA-value to be used in other Operations. References to Fortran symbols are distinguished via this operation from other arbitrary constant values.

  %p = fir.address_of(@symbol) : !fir.ref<f64>

Interfaces: NoSideEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Attributes:

Attribute MLIR Type Description
symbol ::mlir::SymbolRefAttr symbol reference attribute

Results:

Result Description
resTy any addressable

fir.allocmem (::fir::AllocMemOp)

allocate storage on the heap for an object of a given type

Creates a heap memory reference suitable for storing a value of the given type, T. The heap refernce returned has type !fir.heap<T>. The memory object is in an undefined state. allocmem operations must be paired with freemem operations to avoid memory leaks.

  %0 = fir.allocmem !fir.array<10 x f32>
  fir.freemem %0 : !fir.heap<!fir.array<10 x f32>>

Traits: AttrSizedOperandSegments

Interfaces: MemoryEffectOpInterface (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{MemoryEffects::Allocate on ::mlir::SideEffects::DefaultResource}

Attributes:

Attribute MLIR Type Description
in_type ::mlir::TypeAttr any type attribute
uniq_name ::mlir::StringAttr string attribute
bindc_name ::mlir::StringAttr string attribute

Operands:

Operand Description
typeparams any integer
shape any integer

Results:

Result Description
«unnamed» Reference to an ALLOCATABLE attribute type

fir.alloca (::fir::AllocaOp)

allocate storage for a temporary on the stack given a type

This primitive operation is used to allocate an object on the stack. A reference to the object of type !fir.ref<T> is returned. The returned object has an undefined/uninitialized state. The allocation can be given an optional name. The allocation may have a dynamic repetition count for allocating a sequence of locations for the specified type.

  %c = ... : i64
  %x = fir.alloca i32
  %y = fir.alloca !fir.array<8 x i64>
  %z = fir.alloca f32, %c

  %i = ... : i16
  %j = ... : i32
  %w = fir.alloca !fir.type<PT(len1:i16, len2:i32)> (%i, %j : i16, i32)

Note that in the case of %z, a contiguous block of memory is allocated and its size is a runtime multiple of a 32-bit REAL value.

In the case of %w, the arguments %i and %j are LEN parameters (len1, len2) to the type PT.

Finally, the operation is undefined if the ssa-value %c is negative.

Fortran Semantics: There is no language mechanism in Fortran to allocate space on the stack like C’s alloca() function. Therefore fir.alloca is not control-flow dependent. However, the lifetime of a stack allocation is often limited to a small region and a legal implementation may reuse stack storage in other regions when there is no conflict. For example, take the following code fragment.

  CALL foo(1)
  CALL foo(2)
  CALL foo(3)

A legal implementation can allocate a stack slot and initialize it with the constant 1, then pass that by reference to foo. Likewise for the second and third calls to foo, each stack slot being initialized accordingly. It is also a conforming implementation to reuse the same stack slot for all three calls, just initializing each in turn. This is possible as the lifetime of the copy of each constant need not exceed that of the CALL statement. Indeed, a user would likely expect a good Fortran compiler to perform such an optimization.

Until Fortran 2018, procedures defaulted to non-recursive. A legal implementation could therefore convert stack allocations to global allocations. Such a conversion effectively adds the SAVE attribute to all variables.

Some temporary entities (large arrays) probably should not be stack allocated as stack space can often be limited. A legal implementation can convert these large stack allocations to heap allocations regardless of whether the procedure is recursive or not.

The pinned attribute is used to flag fir.alloca operation in a specific region and avoid them being hoisted in an alloca hoisting pass.

Traits: AttrSizedOperandSegments

Interfaces: MemoryEffectOpInterface (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{MemoryEffects::Allocate on ::mlir::SideEffects::AutomaticAllocationScopeResource}

Attributes:

Attribute MLIR Type Description
in_type ::mlir::TypeAttr any type attribute
uniq_name ::mlir::StringAttr string attribute
bindc_name ::mlir::StringAttr string attribute
pinned ::mlir::UnitAttr unit attribute

Operands:

Operand Description
typeparams any integer
shape any integer

Results:

Result Description
«unnamed» Reference to an entity type

fir.array_access (::fir::ArrayAccessOp)

Fetch the reference of an element of an array value

Syntax:

operation ::= `fir.array_access` $sequence `,` $indices (`typeparams` $typeparams^)? attr-dict `:`
              functional-type(operands, results)

The array_access provides a reference to a single element from an array value. This is not a view in the immutable array, otherwise it couldn’t be stored to. It can be see as a logical copy of the element and its position in the array. This reference can be written to and modified without changing the original array.

The array_access operation is used to fetch the memory reference of an element in an array value.

  real :: a(n,m)
  ...
  ... a ...
  ... a(r,s+1) ...

One can use fir.array_access to recover the implied memory reference to the element a(i,j) in an array expression a as shown above. It can also be used to recover the reference element a(r,s+1) in the second expression.

  %s = fir.shape %n, %m : (index, index) -> !fir.shape<2>
  // load the entire array 'a'
  %v = fir.array_load %a(%s) : (!fir.ref<!fir.array<?x?xf32>>, !fir.shape<2>) -> !fir.array<?x?xf32>
  // fetch the value of one of the array value's elements
  %1 = fir.array_access %v, %i, %j : (!fir.array<?x?xf32>, index, index) -> !fir.ref<f32>

It is only possible to use array_access on an array_load result value or a value that can be trace back transitively to an array_load as the dominating source. Other array operation such as array_amend can be in between.

TODO: The above restriction is not enforced. The design of the operation might need to be revisited to avoid such restructions.

More information about array_access and other array operations can be found in flang/docs/FIRArrayOperations.md.

Traits: AttrSizedOperandSegments

Interfaces: NoSideEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Operands:

Operand Description
sequence FIR array type
indices coordinate type
typeparams any integer

Results:

Result Description
element Reference to an entity type

fir.array_amend (::fir::ArrayAmendOp)

Mark an array value as having been changed by reference.

Syntax:

operation ::= `fir.array_amend` $sequence `,` $memref attr-dict `:` functional-type(operands, results)

The array_amend operation marks an array value as having been changed via a reference obtained by an array_access. It acts as a logical transaction log that is used to merge the final result back with an array_merge_store operation.

  // fetch the value of one of the array value's elements
  %1 = fir.array_access %v, %i, %j : (!fir.array<?x?xT>, index, index) -> !fir.ref<T>
  // modify the element by storing data using %1 as a reference
  %2 = ... %1 ...
  // mark the array value
  %new_v = fir.array_amend %v, %2 : (!fir.array<?x?xT>, !fir.ref<T>) -> !fir.array<?x?xT>

More information about array_amend and other array operations can be found in flang/docs/FIRArrayOperations.md.

Interfaces: NoSideEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Operands:

Operand Description
sequence FIR array type
memref Reference to an entity type

Results:

Result Description
«unnamed» FIR array type

fir.array_coor (::fir::ArrayCoorOp)

Find the coordinate of an element of an array

Syntax:

operation ::= `fir.array_coor` $memref (`(`$shape^`)`)? (`[`$slice^`]`)? $indices (`typeparams`
              $typeparams^)? attr-dict `:` functional-type(operands, results)

Compute the location of an element in an array when the shape of the array is only known at runtime.

This operation is intended to capture all the runtime values needed to compute the address of an array reference in a single high-level op. Given the following Fortran input:

  real :: a(n,m)
  ...
  ... a(i,j) ...

One can use fir.array_coor to determine the address of a(i,j).

  %s = fir.shape %n, %m : (index, index) -> !fir.shape<2>
  %1 = fir.array_coor %a(%s) %i, %j : (!fir.ref<!fir.array<?x?xf32>>, !fir.shape<2>, index, index) -> !fir.ref<f32>

Traits: AttrSizedOperandSegments

Interfaces: NoSideEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Operands:

Operand Description
memref any reference or box
shape any legal shape or shift type
slice FIR slice
indices coordinate type
typeparams any integer

Results:

Result Description
«unnamed» Reference to an entity type

fir.array_fetch (::fir::ArrayFetchOp)

Fetch the value of an element of an array value

Syntax:

operation ::= `fir.array_fetch` $sequence `,` $indices (`typeparams` $typeparams^)? attr-dict `:`
              functional-type(operands, results)

Fetch the value of an element in an array value.

  real :: a(n,m)
  ...
  ... a ...
  ... a(r,s+1) ...

One can use fir.array_fetch to fetch the (implied) value of a(i,j) in an array expression as shown above. It can also be used to extract the element a(r,s+1) in the second expression.

  %s = fir.shape %n, %m : (index, index) -> !fir.shape<2>
  // load the entire array 'a'
  %v = fir.array_load %a(%s) : (!fir.ref<!fir.array<?x?xf32>>, !fir.shape<2>) -> !fir.array<?x?xf32>
  // fetch the value of one of the array value's elements
  %1 = fir.array_fetch %v, %i, %j : (!fir.array<?x?xf32>, index, index) -> f32

It is only possible to use array_fetch on an array_load result value.

Traits: AttrSizedOperandSegments

Interfaces: NoSideEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Operands:

Operand Description
sequence FIR array type
indices coordinate type
typeparams any integer

Results:

Result Description
element any type

fir.array_load (::fir::ArrayLoadOp)

Load an array as a value.

Syntax:

operation ::= `fir.array_load` $memref (`(`$shape^`)`)? (`[`$slice^`]`)? (`typeparams` $typeparams^)?
              attr-dict `:` functional-type(operands, results)

This operation taken with array_merge_store captures Fortran’s copy-in/copy-out semantics. One way to think of this is that array_load creates a snapshot copy of the entire array. This copy can then be used as the “original value” of the array while the array’s new value is computed. The array_merge_store operation is the copy-out semantics, which merge the updates with the original array value to produce the final array result. This abstracts the copy operations as opposed to always creating copies or requiring dependence analysis be performed on the syntax trees and before lowering to the IR.

Load an entire array as a single SSA value.

  real :: a(o:n,p:m)
  ...
  ... = ... a ...

One can use fir.array_load to produce an ssa-value that captures an immutable value of the entire array a, as in the Fortran array expression shown above. Subsequent changes to the memory containing the array do not alter its composite value. This operation let’s one load an array as a value while applying a runtime shape, shift, or slice to the memory reference, and its semantics guarantee immutability.

  %s = fir.shape_shift %o, %n, %p, %m : (index, index, index, index) -> !fir.shape<2>
  // load the entire array 'a'
  %v = fir.array_load %a(%s) : (!fir.ref<!fir.array<?x?xf32>>, !fir.shape<2>) -> !fir.array<?x?xf32>
  // a fir.store here into array %a does not change %v

Traits: AttrSizedOperandSegments

Operands:

Operand Description
memref any reference or box
shape any legal shape or shift type
slice FIR slice
typeparams any integer

Results:

Result Description
«unnamed» FIR array type

fir.array_merge_store (::fir::ArrayMergeStoreOp)

Store merged array value to memory.

Syntax:

operation ::= `fir.array_merge_store` $original `,` $sequence `to` $memref (`[` $slice^ `]`)? (`typeparams`
              $typeparams^)? attr-dict `:` type(operands)

Store a merged array value to memory.

  real :: a(n,m)
  ...
  a = ...

One can use fir.array_merge_store to merge/copy the value of a in an array expression as shown above.

  %v = fir.array_load %a(%shape) : ...
  %r = fir.array_update %v, %f, %i, %j : (!fir.array<?x?xf32>, f32, index, index) -> !fir.array<?x?xf32>
  fir.array_merge_store %v, %r to %a : !fir.ref<!fir.array<?x?xf32>>

This operation merges the original loaded array value, %v, with the chained updates, %r, and stores the result to the array at address, %a.

Traits: AttrSizedOperandSegments

Operands:

Operand Description
original FIR array type
sequence FIR array type
memref any reference or box
slice FIR slice
typeparams any integer

fir.array_modify (::fir::ArrayModifyOp)

Get an address for an array value to modify it.

Syntax:

operation ::= `fir.array_modify` $sequence `,` $indices (`typeparams` $typeparams^)? attr-dict
              `:` functional-type(operands, results)

Modify the value of an element in an array value through actions done on the returned address. A new array value is also returned where all element values of the input array are identical except for the selected element which is the value after the modification done on the element address.

  real :: a(n)
  ...
  ! Elemental user defined assignment from type(SomeType) to real.
  a = value_of_some_type

One can use fir.array_modify to update the (implied) value of a(i) in an array expression as shown above.

  %s = fir.shape %n : (index) -> !fir.shape<1>
  // Load the entire array 'a'.
  %v = fir.array_load %a(%s) : (!fir.ref<!fir.array<?xf32>>, !fir.shape<1>)
      -> !fir.array<?xf32>
  // Update the value of one of the array value's elements with a user
  // defined assignment from %rhs.
  %new = fir.do_loop %i = ... (%inner = %v) {
    %rhs = ...
    %addr, %r = fir.array_modify %inner, %i, %j : (!fir.array<?xf32>,
 index) -> fir.ref<f32>, !fir.array<?xf32>
    fir.call @user_def_assign(%addr, %rhs) (fir.ref<f32>,
 fir.ref<!fir.type<SomeType>>) -> ()
    fir.result %r : !fir.ref<!fir.array<?xf32>>
  }
  fir.array_merge_store %v, %new to %a : !fir.ref<!fir.array<?xf32>>

An array value modification behaves as if a mapping function from the indices to the new value has been added, replacing the previous mapping. These mappings can be added to the ssa-value, but will not be materialized in memory until the fir.array_merge_store is performed.

Traits: AttrSizedOperandSegments

Interfaces: NoSideEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Operands:

Operand Description
sequence FIR array type
indices coordinate type
typeparams any integer

Results:

Result Description
«unnamed» Reference to an entity type
«unnamed» FIR array type

fir.array_update (::fir::ArrayUpdateOp)

Update the value of an element of an array value

Syntax:

operation ::= `fir.array_update` $sequence `,` $merge `,` $indices (`typeparams` $typeparams^)? attr-dict
              `:` functional-type(operands, results)

Updates the value of an element in an array value. A new array value is returned where all element values of the input array are identical except for the selected element which is the value passed in the update.

  real :: a(n,m)
  ...
  a = ...

One can use fir.array_update to update the (implied) value of a(i,j) in an array expression as shown above.

  %s = fir.shape %n, %m : (index, index) -> !fir.shape<2>
  // load the entire array 'a'
  %v = fir.array_load %a(%s) : (!fir.ref<!fir.array<?x?xf32>>, !fir.shape<2>) -> !fir.array<?x?xf32>
  // update the value of one of the array value's elements
  // %r_{ij} = %f  if (i,j) = (%i,%j),   %v_{ij} otherwise
  %r = fir.array_update %v, %f, %i, %j : (!fir.array<?x?xf32>, f32, index, index) -> !fir.array<?x?xf32>
  fir.array_merge_store %v, %r to %a : !fir.ref<!fir.array<?x?xf32>>

An array value update behaves as if a mapping function from the indices to the new value has been added, replacing the previous mapping. These mappings can be added to the ssa-value, but will not be materialized in memory until the fir.array_merge_store is performed.

Traits: AttrSizedOperandSegments

Interfaces: NoSideEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Operands:

Operand Description
sequence FIR array type
merge any type
indices coordinate type
typeparams any integer

Results:

Result Description
«unnamed» FIR array type

fir.box_addr (::fir::BoxAddrOp)

return a memory reference to the boxed value

Syntax:

operation ::= `fir.box_addr` operands attr-dict `:` functional-type(operands, results)

This operator is overloaded to work with values of type box, boxchar, and boxproc. The result for each of these cases, respectively, is the address of the data, the address of the CHARACTER data, and the address of the procedure.

  %51 = fir.box_addr %box : (!fir.box<f64>) -> !fir.ref<f64>
  %52 = fir.box_addr %boxchar : (!fir.boxchar<1>) -> !fir.ref<!fir.char<1>>
  %53 = fir.box_addr %boxproc : (!fir.boxproc<!P>) -> !P

Interfaces: NoSideEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Operands:

Operand Description
val any box

Results:

Result Description
«unnamed» any code or data reference

fir.boxchar_len (::fir::BoxCharLenOp)

return the LEN type parameter from a boxchar value

Syntax:

operation ::= `fir.boxchar_len` operands attr-dict `:` functional-type(operands, results)

Extracts the LEN type parameter from a boxchar value.

  %45 = ... : !boxchar<1>  // CHARACTER(20)
  %59 = fir.boxchar_len %45 : (!fir.boxchar<1>) -> i64  // len=20

Interfaces: NoSideEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Operands:

Operand Description
val CHARACTER type descriptor.

Results:

Result Description
«unnamed» any integer

fir.box_dims (::fir::BoxDimsOp)

return the dynamic dimension information for the boxed value

Syntax:

operation ::= `fir.box_dims` $val `,` $dim attr-dict `:` functional-type(operands, results)

Returns the triple of lower bound, extent, and stride for dim dimension of val, which must have a box type. The dimensions are enumerated from left to right from 0 to rank-1. This operation has undefined behavior if dim is out of bounds.

  %c1   = arith.constant 0 : i32
  %52:3 = fir.box_dims %40, %c1 : (!fir.box<!fir.array<*:f64>>, i32) -> (index, index, index)

The above is a request to return the left most row (at index 0) triple from the box. The triple will be the lower bound, extent, and byte-stride, which are the values encoded in a standard descriptor.

Interfaces: InferTypeOpInterface, NoSideEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Operands:

Operand Description
val The type of a Fortran descriptor
dim any integer

Results:

Result Description
«unnamed» index
«unnamed» index
«unnamed» index

fir.box_elesize (::fir::BoxEleSizeOp)

return the size of an element of the boxed value

Syntax:

operation ::= `fir.box_elesize` operands attr-dict `:` functional-type(operands, results)

Returns the size of an element in an entity of box type. This size may not be known until runtime.

  %53 = fir.box_elesize %40 : (!fir.box<f32>, i32) -> i32  // size=4
  %54 = fir.box_elesize %40 : (!fir.box<!fir.array<*:f32>>, i32) -> i32

In the above example, %53 may box an array of REAL values while %54 must box an array of REAL values (with dynamic rank and extent).

Interfaces: NoSideEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Operands:

Operand Description
val The type of a Fortran descriptor

Results:

Result Description
«unnamed» any integer

fir.box_isalloc (::fir::BoxIsAllocOp)

is the boxed value an ALLOCATABLE?

Syntax:

operation ::= `fir.box_isalloc` operands attr-dict `:` functional-type(operands, results)

Determine if the boxed value was from an ALLOCATABLE entity. This will return true if the originating box value was from a fir.embox op with a mem-ref value that had the type !fir.heap.

  %r = ... : !fir.heap<i64>
  %b = fir.embox %r : (!fir.heap<i64>) -> !fir.box<i64>
  %a = fir.box_isalloc %b : (!fir.box<i64>) -> i1  // true

The canonical descriptor implementation will carry a flag to record if the variable is an ALLOCATABLE.

Interfaces: NoSideEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Operands:

Operand Description
val The type of a Fortran descriptor

Results:

Result Description
«unnamed» bool-like

fir.box_isarray (::fir::BoxIsArrayOp)

is the boxed value an array?

Syntax:

operation ::= `fir.box_isarray` operands attr-dict `:` functional-type(operands, results)

Determine if the boxed value has a positive (> 0) rank. This will return true if the originating box value was from a fir.embox with a memory reference value that had the type !fir.array and/or a shape argument.

  %r = ... : !fir.ref<i64>
  %c_100 = arith.constant 100 : index
  %d = fir.shape %c_100 : (index) -> !fir.shape<1>
  %b = fir.embox %r(%d) : (!fir.ref<i64>, !fir.shape<1>) -> !fir.box<i64>
  %a = fir.box_isarray %b : (!fir.box<i64>) -> i1  // true

Interfaces: NoSideEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Operands:

Operand Description
val The type of a Fortran descriptor

Results:

Result Description
«unnamed» bool-like

fir.box_isptr (::fir::BoxIsPtrOp)

is the boxed value a POINTER?

Syntax:

operation ::= `fir.box_isptr` operands attr-dict `:` functional-type(operands, results)

Determine if the boxed value was from a POINTER entity.

  %p = ... : !fir.ptr<i64>
  %b = fir.embox %p : (!fir.ptr<i64>) -> !fir.box<i64>
  %a = fir.box_isptr %b : (!fir.box<i64>) -> i1  // true

Interfaces: NoSideEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Operands:

Operand Description
val The type of a Fortran descriptor

Results:

Result Description
«unnamed» bool-like

fir.boxproc_host (::fir::BoxProcHostOp)

returns the host instance pointer (or null)

Syntax:

operation ::= `fir.boxproc_host` operands attr-dict `:` functional-type(operands, results)

Extract the host context pointer from a boxproc value.

  %8 = ... : !fir.boxproc<(!fir.ref<!fir.type<T>>) -> i32>
  %9 = fir.boxproc_host %8 : (!fir.boxproc<(!fir.ref<!fir.type<T>>) -> i32>) -> !fir.ref<tuple<i32, i32>>

In the example, the reference to the closure over the host procedure’s variables is returned. This allows an internal procedure to access the host’s variables. It is up to lowering to determine the contract between the host and the internal procedure.

Interfaces: NoSideEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Operands:

Operand Description
val

Results:

Result Description
«unnamed» Reference to an entity type

fir.box_rank (::fir::BoxRankOp)

return the number of dimensions for the boxed value

Syntax:

operation ::= `fir.box_rank` operands attr-dict `:` functional-type(operands, results)

Return the rank of a value of box type. If the value is scalar, the rank is 0.

  %57 = fir.box_rank %40 : (!fir.box<!fir.array<*:f64>>) -> i32
  %58 = fir.box_rank %41 : (!fir.box<f64>) -> i32

The example %57 shows how one would determine the rank of an array that has deferred rank at runtime. This rank should be at least 1. In %58, the descriptor may be either an array or a scalar, so the value is nonnegative.

Interfaces: NoSideEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Operands:

Operand Description
val The type of a Fortran descriptor

Results:

Result Description
«unnamed» any integer

fir.box_tdesc (::fir::BoxTypeDescOp)

return the type descriptor for the boxed value

Syntax:

operation ::= `fir.box_tdesc` operands attr-dict `:` functional-type(operands, results)

Return the opaque type descriptor of a value of box type. A type descriptor is an implementation defined value that fully describes a type to the Fortran runtime.

  %7 = fir.box_tdesc %41 : (!fir.box<f64>) -> !fir.tdesc<f64>

Interfaces: NoSideEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Operands:

Operand Description
val The type of a Fortran descriptor

Results:

Result Description
«unnamed» FIR Type descriptor type

fir.call (::fir::CallOp)

call a procedure

Call the specified function or function reference.

Provides a custom parser and pretty printer to allow a more readable syntax in the FIR dialect, e.g. fir.call @sub(%12) or fir.call %20(%22,%23).

  %a = fir.call %funcref(%arg0) : (!fir.ref<f32>) -> f32
  %b = fir.call @function(%arg1, %arg2) : (!fir.ref<f32>, !fir.ref<f32>) -> f32

Interfaces: CallOpInterface

Attributes:

Attribute MLIR Type Description
callee ::mlir::SymbolRefAttr symbol reference attribute

Operands:

Operand Description
args any type

Results:

Result Description
«unnamed» any type

fir.char_convert (::fir::CharConvertOp)

Primitive to convert an entity of type CHARACTER from one KIND to a
different KIND.

Syntax:

operation ::= `fir.char_convert` $from `for` $count `to` $to attr-dict `:` type(operands)

Copy a CHARACTER (must be in memory) of KIND k1 to a CHARACTER (also must be in memory) of KIND k2 where k1 != k2 and the buffers do not overlap. This latter restriction is unchecked, as the Fortran language definition eliminates the overlapping in memory case.

The number of code points copied is specified explicitly as the second argument. The length of the !fir.char type is ignored.

  fir.char_convert %1 for %2 to %3 : !fir.ref<!fir.char<1,?>>, i32,
      !fir.ref<!fir.char<2,20>>

Should future support for encodings other than ASCII be supported, codegen can generate a call to a runtime helper routine which will map the code points from UTF-8 to UCS-2, for example. Such remappings may not always be possible as they may involve the creation of more code points than the count limit. These details are left as future to-dos.

Operands:

Operand Description
from any reference
count any integer
to any reference

fir.cmpc (::fir::CmpcOp)

complex floating-point comparison operator

A complex comparison to handle complex types found in FIR.

Traits: SameOperandsAndResultShape, SameTypeOperands

Interfaces: NoSideEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Operands:

Operand Description
lhs Complex type
rhs Complex type

Results:

Result Description
«unnamed» any logical

fir.constc (::fir::ConstcOp)

create a complex constant

A complex constant. Similar to the standard dialect complex type, but this extension allows constants with APFloat values that are not supported in the standard dialect.

Interfaces: NoSideEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Results:

Result Description
«unnamed» Complex type

fir.convert (::fir::ConvertOp)

encapsulates all Fortran scalar type conversions

Syntax:

operation ::= `fir.convert` $value attr-dict `:` functional-type($value, results)

Generalized type conversion. Convert the ssa value from type T to type U. Not all pairs of types have conversions. When types T and U are the same type, this instruction is a NOP and may be folded away.

  %v = ... : i64
  %w = fir.convert %v : (i64) -> i32

The example truncates the value %v from an i64 to an i32.

Interfaces: NoSideEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Operands:

Operand Description
value any type

Results:

Result Description
res FIR dialect type

fir.coordinate_of (::fir::CoordinateOp)

Finds the coordinate (location) of a value in memory

Compute the internal coordinate address starting from a boxed value or unboxed memory reference. Returns a memory reference. When computing the coordinate of an array element, the rank of the array must be known and the number of indexing expressions must not exceed the rank of the array.

This operation will apply the access map from a boxed value implicitly.

Unlike LLVM’s GEP instruction, one cannot stride over the outermost reference; therefore, the leading 0 index must be omitted.

  %i = ... : index
  %h = ... : !fir.heap<!fir.array<100 x f32>>
  %p = fir.coordinate_of %h, %i : (!fir.heap<!fir.array<100 x f32>>, index) -> !fir.ref<f32>

In the example, %p will be a pointer to the %i-th f32 value in the array %h.

Interfaces: NoSideEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Attributes:

Attribute MLIR Type Description
baseType ::mlir::TypeAttr any type attribute

Operands:

Operand Description
ref any reference or box
coor coordinate type

Results:

Result Description
«unnamed» fir.ref or fir.llvm_ptr

fir.dt_entry (::fir::DTEntryOp)

map entry in a dispatch table

An entry in a dispatch table. Allows a function symbol to be bound to a specifier method identifier. A dispatch operation uses the dynamic type of a distinguished argument to determine an exact dispatch table and uses the method identifier to select the type-bound procedure to be called.

  fir.dt_entry method_name, @uniquedProcedure

Traits: HasParent

Attributes:

Attribute MLIR Type Description
method ::mlir::StringAttr string attribute
proc ::mlir::SymbolRefAttr symbol reference attribute

fir.dispatch (::fir::DispatchOp)

call a type-bound procedure

Perform a dynamic dispatch on the method name via the dispatch table associated with the first argument. The attribute ‘pass_arg_pos’ can be used to select a dispatch argument other than the first one.

  %r = fir.dispatch methodA(%o) : (!fir.box<none>) -> i32

Attributes:

Attribute MLIR Type Description
method ::mlir::StringAttr string attribute

Operands:

Operand Description
object The type of a Fortran descriptor
args any type

Results:

Result Description
«unnamed» any type

fir.dispatch_table (::fir::DispatchTableOp)

Dispatch table definition

Define a dispatch table for a derived type with type-bound procedures.

A dispatch table is an untyped symbol that contains a list of associations between method identifiers and corresponding FuncOp symbols.

The ordering of associations in the map is determined by the front end.

  fir.dispatch_table @_QDTMquuzTfoo {
    fir.dt_entry method1, @_QFNMquuzTfooPmethod1AfooR
    fir.dt_entry method2, @_QFNMquuzTfooPmethod2AfooII
  }

Traits: ImplicitFirTerminator, IsolatedFromAbove

Interfaces: Symbol

fir.divc (::fir::DivcOp)

Syntax:

operation ::= `fir.divc` operands attr-dict `:` type($result)

Traits: SameOperandsAndResultType

Interfaces: InferTypeOpInterface, NoSideEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Operands:

Operand Description
lhs Complex type
rhs Complex type

Results:

Result Description
result any type

fir.do_loop (::fir::DoLoopOp)

generalized loop operation

Generalized high-level looping construct. This operation is similar to MLIR’s scf.for.

  %l = arith.constant 0 : index
  %u = arith.constant 9 : index
  %s = arith.constant 1 : index
  fir.do_loop %i = %l to %u step %s unordered {
    %x = fir.convert %i : (index) -> i32
    %v = fir.call @compute(%x) : (i32) -> f32
    %p = fir.coordinate_of %A, %i : (!fir.ref<!fir.array<?xf32>>, index) -> !fir.ref<f32>
    fir.store %v to %p : !fir.ref<f32>
  }

The above example iterates over the interval [%l, %u]. The unordered keyword indicates that the iterations can be executed in any order.

Traits: FirRegionTerminator, RecursiveSideEffects

Interfaces: LoopLikeOpInterface

Attributes:

Attribute MLIR Type Description
unordered ::mlir::UnitAttr unit attribute
finalValue ::mlir::UnitAttr unit attribute

Operands:

Operand Description
lowerBound index
upperBound index
step index
initArgs any type

Results:

Result Description
results any type

fir.emboxchar (::fir::EmboxCharOp)

boxes a given CHARACTER reference and its LEN parameter

Syntax:

operation ::= `fir.emboxchar` $memref `,` $len attr-dict `:` functional-type(operands, results)

Create a boxed CHARACTER value. The CHARACTER type has the LEN type parameter, the value of which may only be known at runtime. Therefore, a variable of type CHARACTER has both its data reference as well as a LEN type parameter.

  CHARACTER(LEN=10) :: var
  %4 = ...         : !fir.ref<!fir.array<10 x !fir.char<1>>>
  %5 = arith.constant 10 : i32
  %6 = fir.emboxchar %4, %5 : (!fir.ref<!fir.array<10 x !fir.char<1>>>, i32) -> !fir.boxchar<1>

In the above %4 is a memory reference to a buffer of 10 CHARACTER units. This buffer and its LEN value (10) are wrapped into a pair in %6.

Interfaces: NoSideEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Operands:

Operand Description
memref any reference
len any integer

Results:

Result Description
«unnamed» CHARACTER type descriptor.

fir.embox (::fir::EmboxOp)

boxes a given reference and (optional) dimension information

Syntax:

operation ::= `fir.embox` $memref (`(` $shape^ `)`)? (`[` $slice^ `]`)? (`typeparams` $typeparams^)?
              (`map` $accessMap^)? attr-dict `:` functional-type(operands, results)

Create a boxed reference value. In Fortran, the implementation can require extra information about an entity, such as its type, rank, etc. This auxilliary information is packaged and abstracted as a value with box type by the calling routine. (In Fortran, these are called descriptors.)

  %c1 = arith.constant 1 : index
  %c10 = arith.constant 10 : index
  %5 = ... : !fir.ref<!fir.array<10 x i32>>
  %6 = fir.embox %5 : (!fir.ref<!fir.array<10 x i32>>) -> !fir.box<!fir.array<10 x i32>>

The descriptor tuple may contain additional implementation-specific information through the use of additional attributes. Specifically, - shape: emboxing an array may require shape information (an array’s lower bounds and extents may not be known until runtime), - slice: an array section can be described with a slice triple, - typeparams: for emboxing a derived type with LEN type parameters, - accessMap: unused/experimental.

Traits: AttrSizedOperandSegments

Interfaces: NoSideEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Attributes:

Attribute MLIR Type Description
accessMap ::mlir::AffineMapAttr AffineMap attribute

Operands:

Operand Description
memref any reference
shape any legal shape type
slice FIR slice
typeparams any integer

Results:

Result Description
«unnamed» The type of a Fortran descriptor

fir.emboxproc (::fir::EmboxProcOp)

boxes a given procedure and optional host context

Syntax:

operation ::= `fir.emboxproc` $func (`,` $host^)? attr-dict `:` functional-type(operands, results)

Creates an abstract encapsulation of a PROCEDURE POINTER along with an optional pointer to a host instance context. If the pointer is not to an internal procedure or the internal procedure does not need a host context then the form takes only the procedure’s symbol.

  %f = ... : (i32) -> i32
  %0 = fir.emboxproc %f : ((i32) -> i32) -> !fir.boxproc<(i32) -> i32>

An internal procedure requiring a host instance for correct execution uses the second form. The closure of the host procedure’s state is passed as a reference to a tuple. It is the responsibility of the host to manage the context’s values accordingly, up to and including inhibiting register promotion of local values.

  %4 = ... : !fir.ref<tuple<!fir.ref<i32>, !fir.ref<i32>>>
  %g = ... : (i32) -> i32
  %5 = fir.emboxproc %g, %4 : ((i32) -> i32, !fir.ref<tuple<!fir.ref<i32>, !fir.ref<i32>>>) -> !fir.boxproc<(i32) -> i32>

Interfaces: NoSideEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Operands:

Operand Description
func function type
host Reference to an entity type

Results:

Result Description
«unnamed»

fir.extract_value (::fir::ExtractValueOp)

Extract a value from an aggregate SSA-value

Syntax:

operation ::= `fir.extract_value` $adt `,` $coor attr-dict `:` functional-type(operands, results)

Extract a value from an entity with a type composed of tuples, arrays, and/or derived types. Returns the value from entity with the type of the specified component. Cannot be used on values of !fir.box type. It can also be used to access complex parts and elements of a character string.

Note that the entity ssa-value must be of compile-time known size in order to use this operation.

  %f = fir.field_index field, !fir.type<X{field:i32}>
  %s = ... : !fir.type<X>
  %v = fir.extract_value %s, %f : (!fir.type<X>, !fir.field) -> i32

Interfaces: NoSideEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Attributes:

Attribute MLIR Type Description
coor ::mlir::ArrayAttr array attribute

Operands:

Operand Description
adt any composite

Results:

Result Description
res FIR dialect type

fir.field_index (::fir::FieldIndexOp)

create a field index value from a field identifier

Generate a field (offset) value from an identifier. Field values may be lowered into exact offsets when the layout of a Fortran derived type is known at compile-time. The type of a field value is !fir.field and these values can be used with the fir.coordinate_of, fir.extract_value, or fir.insert_value instructions to compute (abstract) addresses of subobjects.

  %f = fir.field_index field, !fir.type<X{field:i32}>

Interfaces: NoSideEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Attributes:

Attribute MLIR Type Description
field_id ::mlir::StringAttr string attribute
on_type ::mlir::TypeAttr any type attribute

Operands:

Operand Description
typeparams any integer

Results:

Result Description
res FIR dialect type

fir.end (::fir::FirEndOp)

the end instruction

The end terminator is a special terminator used inside various FIR operations that have regions. End is thus the custom invisible terminator for these operations. It is implicit and need not appear in the textual representation.

Traits: Terminator

fir.freemem (::fir::FreeMemOp)

free a heap object

Syntax:

operation ::= `fir.freemem` $heapref attr-dict `:` qualified(type($heapref))

Deallocates a heap memory reference that was allocated by an allocmem. The memory object that is deallocated is placed in an undefined state after fir.freemem. Optimizations may treat the loading of an object in the undefined state as undefined behavior. This includes aliasing references, such as the result of an fir.embox.

  %21 = fir.allocmem !fir.type<ZT(p:i32){field:i32}>
  ...
  fir.freemem %21 : !fir.heap<!fir.type<ZT>>

Interfaces: MemoryEffectOpInterface (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{MemoryEffects::Free on ::mlir::SideEffects::DefaultResource}

Operands:

Operand Description
heapref Reference to an ALLOCATABLE attribute type

fir.gentypedesc (::fir::GenTypeDescOp)

generate a type descriptor for a given type

Generates a constant object that is an abstract type descriptor of the specified type. The meta-type of a type descriptor for the type T is !fir.tdesc<T>.

  !T = !fir.type<T{...}>
  %t = fir.gentypedesc !T  // returns value of !fir.tdesc<!T>

Interfaces: NoSideEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Attributes:

Attribute MLIR Type Description
in_type ::mlir::TypeAttr Fortran surface type

Results:

Result Description
res FIR dialect type

fir.global_len (::fir::GlobalLenOp)

map a LEN parameter to a global

A global entity (that is not an automatic data object) can have extra LEN parameter (compile-time) constants associated with the instance’s type. These values can be bound to the global instance used fir.global_len.

  global @g : !fir.type<t(len1:i32)> {
    fir.global_len len1, 10 : i32
    %1 = fir.undefined !fir.type<t(len1:i32)>
    fir.has_value %1 : !fir.type<t(len1:i32)>
  }

Attributes:

Attribute MLIR Type Description
lenparam ::mlir::StringAttr string attribute
intval ::mlir::IntegerAttr arbitrary integer attribute

fir.global (::fir::GlobalOp)

Global data

A global variable or constant with initial values.

The example creates a global variable (writable) named @_QV_Mquark_Vvarble with some initial values. The initializer should conform to the variable’s type.

  fir.global @_QV_Mquark_Vvarble : tuple<i32, f32> {
    %1 = arith.constant 1 : i32
    %2 = arith.constant 2.0 : f32
    %3 = fir.undefined tuple<i32, f32>
    %z = arith.constant 0 : index
    %o = arith.constant 1 : index
    %4 = fir.insert_value %3, %1, %z : (tuple<i32, f32>, i32, index) -> tuple<i32, f32>
    %5 = fir.insert_value %4, %2, %o : (tuple<i32, f32>, f32, index) -> tuple<i32, f32>
    fir.has_value %5 : tuple<i32, f32>
  }

Traits: IsolatedFromAbove

Interfaces: Symbol

Attributes:

Attribute MLIR Type Description
sym_name ::mlir::StringAttr string attribute
symref ::mlir::SymbolRefAttr symbol reference attribute
type ::mlir::TypeAttr any type attribute
initVal ::mlir::Attribute any attribute
constant ::mlir::UnitAttr unit attribute
linkName ::mlir::StringAttr string attribute

fir.has_value (::fir::HasValueOp)

terminator for GlobalOp

Syntax:

operation ::= `fir.has_value` $resval attr-dict `:` type($resval)

The terminator for a GlobalOp with a body.

  global @variable : tuple<i32, f32> {
    %0 = arith.constant 45 : i32
    %1 = arith.constant 100.0 : f32
    %2 = fir.undefined tuple<i32, f32>
    %3 = arith.constant 0 : index
    %4 = fir.insert_value %2, %0, %3 : (tuple<i32, f32>, i32, index) -> tuple<i32, f32>
    %5 = arith.constant 1 : index
    %6 = fir.insert_value %4, %1, %5 : (tuple<i32, f32>, f32, index) -> tuple<i32, f32>
    fir.has_value %6 : tuple<i32, f32>
  }

Traits: HasParent, Terminator

Operands:

Operand Description
resval any type

fir.if (::fir::IfOp)

if-then-else conditional operation

Used to conditionally execute operations. This operation is the FIR dialect’s version of loop.if.

  %56 = ... : i1
  %78 = ... : !fir.ref<!T>
  fir.if %56 {
    fir.store %76 to %78 : !fir.ref<!T>
  } else {
    fir.store %77 to %78 : !fir.ref<!T>
  }

Traits: FirRegionTerminator, NoRegionArguments, RecursiveSideEffects

Operands:

Operand Description
condition 1-bit signless integer

Results:

Result Description
results any type

fir.insert_on_range (::fir::InsertOnRangeOp)

insert sub-value into a range on an existing sequence

Syntax:

operation ::= `fir.insert_on_range` $seq `,` $val custom<CustomRangeSubscript>($coor) attr-dict `:` functional-type(operands, results)

Insert copies of a value into an entity with an array type of constant shape and size. Returns a new ssa value with the same type as the original entity. The values are inserted at a contiguous range of indices in Fortran row-to-column element order as specified by lower and upper bound coordinates.

  %a = fir.undefined !fir.array<10x10xf32>
  %c = arith.constant 3.0 : f32
  %1 = fir.insert_on_range %a, %c from (0, 0) to (7, 2) : (!fir.array<10x10xf32>, f32) -> !fir.array<10x10xf32>

The first 28 elements of %1, with coordinates from (0,0) to (7,2), have the value 3.0.

Interfaces: NoSideEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Attributes:

Attribute MLIR Type Description
coor ::mlir::DenseIntElementsAttr index elements attribute

Operands:

Operand Description
seq FIR array type
val any type

Results:

Result Description
«unnamed» FIR array type

fir.insert_value (::fir::InsertValueOp)

insert a new sub-value into a copy of an existing aggregate

Syntax:

operation ::= `fir.insert_value` $adt `,` $val `,` $coor attr-dict `:` functional-type(operands, results)

Insert a value into an entity with a type composed of tuples, arrays, and/or derived types. Returns a new ssa value with the same type as the original entity. Cannot be used on values of !fir.box type. It can also be used to set complex parts and elements of a character string.

Note that the entity ssa-value must be of compile-time known size in order to use this operation.

  %a = ... : !fir.array<10xtuple<i32, f32>>
  %f = ... : f32
  %o = ... : i32
  %c = arith.constant 1 : i32
  %b = fir.insert_value %a, %f, %o, %c : (!fir.array<10x20xtuple<i32, f32>>, f32, i32, i32) -> !fir.array<10x20xtuple<i32, f32>>

Interfaces: NoSideEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Attributes:

Attribute MLIR Type Description
coor ::mlir::ArrayAttr array attribute

Operands:

Operand Description
adt any composite
val any type

Results:

Result Description
«unnamed» any composite

fir.is_present (::fir::IsPresentOp)

is this optional function argument present?

Syntax:

operation ::= `fir.is_present` operands attr-dict `:` functional-type(operands, results)

Determine if an optional function argument is PRESENT (i.e. that it was not created by a fir.absent op on the caller side).

  func @_QPfoo(%arg0: !fir.box<!fir.array<?xf32>>) {
    %0 = fir.is_present %arg0 : (!fir.box<!fir.array<?xf32>>) -> i1
    ...

Interfaces: NoSideEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Operands:

Operand Description
val any reference or box like

Results:

Result Description
«unnamed» bool-like

fir.iterate_while (::fir::IterWhileOp)

DO loop with early exit condition

This single-entry, single-exit looping construct is useful for lowering counted loops that can exit early such as, for instance, implied-DO loops. It is very similar to fir::DoLoopOp with the addition that it requires a single loop-carried bool value that signals an early exit condition to the operation. A true disposition means the next loop iteration should proceed. A false indicates that the fir.iterate_while operation should terminate and return its iteration arguments. This is a degenerate counted loop in that the loop is not guaranteed to execute all iterations.

An example iterate_while that returns the counter value, the early termination condition, and an extra loop-carried value is shown here. This loop counts from %lo to %up (inclusive), stepping by %c1, so long as the early exit (%ok) is true. The iter_args %sh value is also carried by the loop. The result triple is the values of %i=phi(%lo,%i+%c1), %ok=phi(%okIn,%okNew), and %sh=phi(%shIn,%shNew) from the last executed iteration.

  %v:3 = fir.iterate_while (%i = %lo to %up step %c1) and (%ok = %okIn) iter_args(%sh = %shIn) -> (index, i1, i16) {
    %shNew = fir.call @bar(%sh) : (i16) -> i16
    %okNew = fir.call @foo(%sh) : (i16) -> i1
    fir.result %i, %okNew, %shNew : index, i1, i16
  }

Traits: FirRegionTerminator, RecursiveSideEffects

Interfaces: LoopLikeOpInterface

Attributes:

Attribute MLIR Type Description
finalValue ::mlir::UnitAttr unit attribute

Operands:

Operand Description
lowerBound index
upperBound index
step index
iterateIn 1-bit signless integer
initArgs any type

Results:

Result Description
results any type

fir.len_param_index (::fir::LenParamIndexOp)

create a field index value from a LEN type parameter identifier

Generate a LEN parameter (offset) value from an LEN parameter identifier. The type of a LEN parameter value is !fir.len and these values can be used with the fir.coordinate_of instructions to compute (abstract) addresses of LEN parameters.

  %e = fir.len_param_index len1, !fir.type<X(len1:i32)>
  %p = ... : !fir.box<!fir.type<X>>
  %q = fir.coordinate_of %p, %e : (!fir.box<!fir.type<X>>, !fir.len) -> !fir.ref<i32>

Interfaces: NoSideEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Attributes:

Attribute MLIR Type Description
field_id ::mlir::StringAttr string attribute
on_type ::mlir::TypeAttr any type attribute

Operands:

Operand Description
typeparams any integer

Results:

Result Description
res FIR dialect type

fir.load (::fir::LoadOp)

load a value from a memory reference

Load a value from a memory reference into an ssa-value (virtual register). Produces an immutable ssa-value of the referent type. A memory reference has type !fir.ref<T>, !fir.heap<T>, or !fir.ptr<T>.

  %a = fir.alloca i32
  %l = fir.load %a : !fir.ref<i32>

The ssa-value has an undefined value if the memory reference is undefined or null.

Interfaces: MemoryEffectOpInterface (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{MemoryEffects::Read on ::mlir::SideEffects::DefaultResource}

Operands:

Operand Description
memref any reference

Results:

Result Description
res FIR dialect type

fir.mulc (::fir::MulcOp)

Syntax:

operation ::= `fir.mulc` operands attr-dict `:` type($result)

Traits: Commutative, SameOperandsAndResultType

Interfaces: InferTypeOpInterface, NoSideEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Operands:

Operand Description
lhs Complex type
rhs Complex type

Results:

Result Description
result any type

fir.negc (::fir::NegcOp)

Syntax:

operation ::= `fir.negc` operands attr-dict `:` type($result)

Traits: SameOperandsAndResultType

Interfaces: InferTypeOpInterface, NoSideEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Operands:

Operand Description
operand Complex type

Results:

Result Description
result any type

fir.no_reassoc (::fir::NoReassocOp)

synthetic op to prevent reassociation

Syntax:

operation ::= `fir.no_reassoc` $val attr-dict `:` type($val)

Primitive operation meant to intrusively prevent operator reassociation. The operation is otherwise a nop and the value returned is the same as the argument.

The presence of this operation prevents any local optimizations. In the example below, this would prevent possibly replacing the multiply and add operations with a single FMA operation.

  %98 = arith.mulf %96, %97 : f32
  %99 = fir.no_reassoc %98 : f32
  %a0 = arith.addf %99, %95 : f32

Traits: SameOperandsAndResultType

Interfaces: InferTypeOpInterface, NoSideEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Operands:

Operand Description
val any type

Results:

Result Description
res FIR dialect type

fir.rebox (::fir::ReboxOp)

create a box given another box and (optional) dimension information

Syntax:

operation ::= `fir.rebox` $box (`(` $shape^ `)`)? (`[` $slice^ `]`)? attr-dict `:` functional-type(operands, results)

Create a new boxed reference value from another box. This is meant to be used when the taking a reference to part of a boxed value, or to an entire boxed value with new shape or type information.

The new extra information can be:

  • new shape information (new lower bounds, new rank, or new extents. New rank/extents can only be provided if the original fir.box is contiguous in all dimension but maybe the first row). The shape operand must be provided to set new shape information.
  • new type (only for derived types). It is possible to set the dynamic type of the new box to one of the parent types of the input box dynamic type. Type parameters cannot be changed. This change is reflected in the requested result type of the new box.

A slice argument can be provided to build a reference to part of a boxed value. In this case, the shape operand must be absent or be a fir.shift that can be used to provide a non default origin for the slice.

The following example illustrates creating a fir.box for x(10:33:2) where x is described by a fir.box and has non default lower bounds, and then applying a new 2-dimension shape to this fir.box.

  %0 = fir.slice %c10, %c33, %c2 : (index, index, index) -> !fir.slice<1>
  %1 = fir.shift %c0 : (index) -> !fir.shift<1>
  %2 = fir.rebox %x(%1) [%0] : (!fir.box<!fir.array<?xf32>>, !fir.shift<1>, !fir.slice<1>) -> !fir.box<!fir.array<?xf32>>
  %3 = fir.shape %c3, %c4 : (index, index) -> !fir.shape<2>
  %4 = fir.rebox %2(%3) : (!fir.box<!fir.array<?xf32>>, !fir.shape<2>) -> !fir.box<!fir.array<?x?xf32>>

Traits: AttrSizedOperandSegments

Interfaces: NoSideEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Operands:

Operand Description
box The type of a Fortran descriptor
shape any legal shape or shift type
slice FIR slice

Results:

Result Description
«unnamed» The type of a Fortran descriptor

fir.result (::fir::ResultOp)

special terminator for use in fir region operations

Syntax:

operation ::= `fir.result` ($results^ `:` type($results))? attr-dict

Result takes a list of ssa-values produced in the block and forwards them as a result to the operation that owns the region of the block. The operation can retain the values or return them to its parent block depending upon its semantics.

Traits: HasParent<IfOp, DoLoopOp, IterWhileOp>, ReturnLike, Terminator

Interfaces: NoSideEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Operands:

Operand Description
results any type

fir.save_result (::fir::SaveResultOp)

save an array, box, or record function result SSA-value to a memory location

Syntax:

operation ::= `fir.save_result` $value `to` $memref (`(` $shape^ `)`)? (`typeparams` $typeparams^)?
              attr-dict `:` type(operands)

Save the result of a function returning an array, box, or record type value into a memory location given the shape and LEN parameters of the result.

Function results of type fir.box, fir.array, or fir.rec are abstract values that require a storage to be manipulated on the caller side. This operation allows associating such abstract result to a storage. In later lowering of the function interfaces, this storage might be used to pass the result in memory.

For arrays, result, it is required to provide the shape of the result. For character arrays and derived types with LEN parameters, the LEN parameter values must be provided.

The fir.save_result associated to a function call must immediately follow the call and be in the same block.

  %buffer = fir.alloca fir.array<?xf32>, %c100
  %shape = fir.shape %c100
  %array_result = fir.call @foo() : () -> fir.array<?xf32>
  fir.save_result %array_result to %buffer(%shape)
  %coor = fir.array_coor %buffer%(%shape), %c5
  %fifth_element = fir.load %coor : f32

The above fir.save_result allows saving a fir.array function result into a buffer to later access its 5th element.

Traits: AttrSizedOperandSegments

Operands:

Operand Description
value fir.box, fir.array or fir.type
memref any reference
shape any legal shape type
typeparams any integer

fir.select_case (::fir::SelectCaseOp)

Fortran’s SELECT CASE statement

Similar to select, select_case provides a way to express Fortran’s SELECT CASE construct. In this case, the selector value is matched against variables (not just constants) and ranges. The structure is the same as select, but select_case allows for the expression of more complex match conditions.

  fir.select_case %arg : i32 [
        #fir.point, %0, ^bb1(%0 : i32),
        #fir.lower, %1, ^bb2(%2,%arg,%arg2,%1 : i32,i32,i32,i32),
        #fir.interval, %2, %3, ^bb3(%2,%arg2 : i32,i32),
        #fir.upper, %arg, ^bb4(%1 : i32),
        unit, ^bb5]

Traits: AttrSizedOperandSegments, Terminator

Interfaces: BranchOpInterface

Operands:

Operand Description
selector any type
compareArgs any type
targetArgs any type

Successors:

Successor Description
targets any successor

fir.select (::fir::SelectOp)

a multiway branch

A multiway branch terminator with similar semantics to C’s switch statement. A selector value is matched against a list of constants of the same type for a match. When a match is found, control is transferred to the corresponding basic block. A select must have at least one basic block with a corresponding unit match, and that block will be selected when all other conditions fail to match.

  fir.select %arg:i32 [1, ^bb1(%0 : i32),
                       2, ^bb2(%2,%arg,%arg2 : i32,i32,i32),
                      -3, ^bb3(%arg2,%2 : i32,i32),
                       4, ^bb4(%1 : i32),
                    unit, ^bb5]

Traits: AttrSizedOperandSegments, Terminator

Interfaces: BranchOpInterface

Operands:

Operand Description
selector any type
compareArgs any type
targetArgs any type

Successors:

Successor Description
targets any successor

fir.select_rank (::fir::SelectRankOp)

Fortran’s SELECT RANK statement

Similar to select, select_rank provides a way to express Fortran’s SELECT RANK construct. In this case, the rank of the selector value is matched against constants of integer type. The structure is the same as select, but select_rank determines the rank of the selector variable at runtime to determine the best match.

  fir.select_rank %arg:i32 [1, ^bb1(%0 : i32),
                            2, ^bb2(%2,%arg,%arg2 : i32,i32,i32),
                            3, ^bb3(%arg2,%2 : i32,i32),
                           -1, ^bb4(%1 : i32),
                         unit, ^bb5]

Traits: AttrSizedOperandSegments, Terminator

Interfaces: BranchOpInterface

Operands:

Operand Description
selector any type
compareArgs any type
targetArgs any type

Successors:

Successor Description
targets any successor

fir.select_type (::fir::SelectTypeOp)

Fortran’s SELECT TYPE statement

Similar to select, select_type provides a way to express Fortran’s SELECT TYPE construct. In this case, the type of the selector value is matched against a list of type descriptors. The structure is the same as select, but select_type determines the type of the selector variable at runtime to determine the best match.

  fir.select_type %arg : !fir.box<()> [
      #fir.instance<!fir.type<type1>>, ^bb1(%0 : i32),
      #fir.instance<!fir.type<type2>>, ^bb2(%2 : i32),
      #fir.subsumed<!fir.type<type3>>, ^bb3(%2 : i32),
      #fir.instance<!fir.type<type4>>, ^bb4(%1,%3 : i32,f32),
      unit, ^bb5]

Traits: AttrSizedOperandSegments, Terminator

Interfaces: BranchOpInterface

Operands:

Operand Description
selector any type
compareArgs any type
targetArgs any type

Successors:

Successor Description
targets any successor

fir.shape (::fir::ShapeOp)

generate an abstract shape vector of type !fir.shape

Syntax:

operation ::= `fir.shape` operands attr-dict `:` functional-type(operands, results)

The arguments are an ordered list of integral type values that define the runtime extent of each dimension of an array. The shape information is given in the same row-to-column order as Fortran. This abstract shape value must be applied to a reified object, so all shape information must be specified. The extent must be nonnegative.

  %d = fir.shape %row_sz, %col_sz : (index, index) -> !fir.shape<2>

Interfaces: NoSideEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Operands:

Operand Description
extents any integer

Results:

Result Description
«unnamed» shape of a multidimensional array object

fir.shape_shift (::fir::ShapeShiftOp)

generate an abstract shape and shift vector of type `!fir.shapeshift`

Syntax:

operation ::= `fir.shape_shift` operands attr-dict `:` functional-type(operands, results)

The arguments are an ordered list of integral type values that is a multiple of 2 in length. Each such pair is defined as: the lower bound and the extent for that dimension. The shifted shape information is given in the same row-to-column order as Fortran. This abstract shifted shape value must be applied to a reified object, so all shifted shape information must be specified. The extent must be nonnegative.

  %d = fir.shape_shift %lo, %extent : (index, index) -> !fir.shapeshift<1>

Interfaces: NoSideEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Operands:

Operand Description
pairs any integer

Results:

Result Description
«unnamed» shape and origin of a multidimensional array object

fir.shift (::fir::ShiftOp)

generate an abstract shift vector of type !fir.shift

Syntax:

operation ::= `fir.shift` operands attr-dict `:` functional-type(operands, results)

The arguments are an ordered list of integral type values that define the runtime lower bound of each dimension of an array. The shape information is given in the same row-to-column order as Fortran. This abstract shift value must be applied to a reified object, so all shift information must be specified.

  %d = fir.shift %row_lb, %col_lb : (index, index) -> !fir.shift<2>

Interfaces: NoSideEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Operands:

Operand Description
origins any integer

Results:

Result Description
«unnamed» lower bounds of a multidimensional array object

fir.slice (::fir::SliceOp)

generate an abstract slice vector of type !fir.slice

Syntax:

operation ::= `fir.slice` $triples (`path` $fields^)? (`substr` $substr^)? attr-dict `:`
              functional-type(operands, results)

The array slicing arguments are an ordered list of integral type values that must be a multiple of 3 in length. Each such triple is defined as: the lower bound, the upper bound, and the stride for that dimension, as in Fortran syntax. Both bounds are inclusive. The array slice information is given in the same row-to-column order as Fortran. This abstract slice value must be applied to a reified object, so all slice information must be specified. The extent must be nonnegative and the stride must not be zero.

  %d = fir.slice %lo, %hi, %step : (index, index, index) -> !fir.slice<1>

To support generalized slicing of Fortran’s dynamic derived types, a slice op can be given a component path (narrowing from the product type of the original array to the specific elemental type of the sliced projection).

  %fld = fir.field_index component, !fir.type<t{...component:ct...}>
  %d = fir.slice %lo, %hi, %step path %fld :
      (index, index, index, !fir.field) -> !fir.slice<1>

Projections of !fir.char type can be further narrowed to invariant substrings.

  %d = fir.slice %lo, %hi, %step substr %offset, %width :
      (index, index, index, index, index) -> !fir.slice<1>

Traits: AttrSizedOperandSegments

Interfaces: NoSideEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Operands:

Operand Description
triples any integer
fields coordinate type
substr any integer

Results:

Result Description
«unnamed» FIR slice

fir.store (::fir::StoreOp)

store an SSA-value to a memory location

Store an ssa-value (virtual register) to a memory reference. The stored value must be of the same type as the referent type of the memory reference.

  %v = ... : f64
  %p = ... : !fir.ptr<f64>
  fir.store %v to %p : !fir.ptr<f64>

The above store changes the value to which the pointer is pointing and not the pointer itself. The operation is undefined if the memory reference, %p, is undefined or null.

Interfaces: MemoryEffectOpInterface (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{MemoryEffects::Write on ::mlir::SideEffects::DefaultResource}

Operands:

Operand Description
value any type
memref any reference

fir.string_lit (::fir::StringLitOp)

create a string literal constant

An FIR constant that represents a sequence of characters that correspond to Fortran’s CHARACTER type, including a LEN. We support CHARACTER values of different KINDs (different constant sizes).

  %1 = fir.string_lit "Hello, World!"(13) : !fir.char<1> // ASCII
  %2 = fir.string_lit [158, 2345](2) : !fir.char<2>      // Wide chars

Interfaces: NoSideEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Results:

Result Description
«unnamed» FIR character type

fir.subc (::fir::SubcOp)

Syntax:

operation ::= `fir.subc` operands attr-dict `:` type($result)

Traits: SameOperandsAndResultType

Interfaces: InferTypeOpInterface, NoSideEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Operands:

Operand Description
lhs Complex type
rhs Complex type

Results:

Result Description
result any type

fir.unboxchar (::fir::UnboxCharOp)

unbox a boxchar value into a pair value

Syntax:

operation ::= `fir.unboxchar` operands attr-dict `:` functional-type(operands, results)

Unboxes a value of boxchar type into a pair consisting of a memory reference to the CHARACTER data and the LEN type parameter.

  %45   = ... : !fir.boxchar<1>
  %46:2 = fir.unboxchar %45 : (!fir.boxchar<1>) -> (!fir.ref<!fir.char<1>>, i32)

Interfaces: NoSideEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Operands:

Operand Description
boxchar CHARACTER type descriptor.

Results:

Result Description
«unnamed» Reference to an entity type
«unnamed» any integer

fir.unboxproc (::fir::UnboxProcOp)

unbox a boxproc value into a pair value

Syntax:

operation ::= `fir.unboxproc` operands attr-dict `:` functional-type(operands, results)

Unboxes a value of boxproc type into a pair consisting of a procedure pointer and a pointer to a host context.

  %47   = ... : !fir.boxproc<() -> i32>
  %48:2 = fir.unboxproc %47 : (!fir.ref<() -> i32>, !fir.ref<tuple<f32, i32>>)

Interfaces: NoSideEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Operands:

Operand Description
boxproc

Results:

Result Description
«unnamed» function type
refTuple Reference to an entity type

fir.undefined (::fir::UndefOp)

explicit undefined value of some type

Syntax:

operation ::= `fir.undefined` type($intype) attr-dict

Constructs an ssa-value of the specified type with an undefined value. This operation is typically created internally by the mem2reg conversion pass. An undefined value can be of any type except !fir.ref<T>.

  %a = fir.undefined !fir.array<10 x !fir.type<T>>

The example creates an array shaped ssa value. The array is rank 1, extent 10, and each element has type !fir.type<T>.

Interfaces: NoSideEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Results:

Result Description
intype any type

fir.unreachable (::fir::UnreachableOp)

the unreachable instruction

Syntax:

operation ::= `fir.unreachable` attr-dict

Terminates a basic block with the assertion that the end of the block will never be reached at runtime. This instruction can be used immediately after a call to the Fortran runtime to terminate the program, for example. This instruction corresponds to the LLVM IR instruction unreachable.

  fir.unreachable

Traits: Terminator

fir.zero_bits (::fir::ZeroOp)

explicit polymorphic zero value of some type

Syntax:

operation ::= `fir.zero_bits` type($intype) attr-dict

Constructs an ssa-value of the specified type with a value of zero for all bits.

  %a = fir.zero_bits !fir.box<!fir.array<10 x !fir.type<T>>>

The example creates a value of type box where all bits are zero.

Interfaces: NoSideEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Results:

Result Description
intype any type