Compiler directives supported by Flang

A list of non-standard directives supported by Flang

    subroutine clear(arr,bytes)
!dir$ ignore_tkr arr
      integer(1), intent(out) :: arr(bytes)
  end interface

Directive Details


Directives are commonly used in Fortran programs to specify additional actions to be performed by the compiler. The directives are always specified with the !dir$ or cdir$ prefix.

Loop Directives

Some directives are associated with the following construct, for example loop directives. Directives on loops are used to specify additional transformation to be performed by the compiler like enabling vectorisation, unrolling, interchange etc.

Currently loop directives are not accepted in the presence of OpenMP or OpenACC constructs on the loop. This should be implemented as it is used in some applications.

Array Expressions

It is to be decided whether loop directives should also be able to be associated with array expressions.


Directives that are associated with constructs must appear in the same section as the construct they are associated with, for example loop directives must appear in the executable section as the loops appear there. To facilitate this the parse tree is corrected to move such directives that appear in the specification part into the execution part.

When a directive that must be associated with a construct appears, a search forward from that directive to the next non-directive construct is performed to check that that construct matches the expected construct for the directive. Skipping other intermediate directives allows multiple directives to appear on the same construct.


Evaluation is extended with a new field called dirs for representing directives associated with that Evaluation. When lowering loop directives, the associated Do Loop’s evaluation is found and the directive is added to it. This information is used only during the lowering of the loop.

Representation in LLVM

The llvm.loop metadata is used in LLVM to provide information to the optimizer about the loop. For example, the llvm.loop.vectorize.enable metadata informs the optimizer that a loop can be vectorized without considering its cost-model. This attribute is added to the loop condition branch.

Representation in MLIR

The MLIR LLVM dialect models this by an attribute called LoopAnnotation Attribute. The attribute can be added to the latch of the loop in the cf dialect and is then carried through lowering to the LLVM dialect.


Since directives must maintain a flow from source to LLVM IR, an integration test is provided that tests the vector always directive, as well as individual lit tests for each of the parsing, semantics and lowering stages.