Flang drivers

.. contents::

There are two main drivers in Flang:

  • the compiler driver, flang-new
  • the frontend driver, flang-new -fc1

NOTE: The diagrams in this document refer to flang as opposed to flang-new. Eventually, flang-new will be renamed as flang and the diagrams reflect the final design that we are still working towards.

The compiler driver will allow you to control all compilation phases (e.g. preprocessing, semantic checks, code-generation, code-optimisation, lowering and linking). For frontend specific tasks, the compiler driver creates a Fortran compilation job and delegates it to flang-new -fc1, the frontend driver. For linking, it creates a linker job and calls an external linker (e.g. LLVM‘s lld). It can also call other tools such as external assemblers (e.g. as). In Clang, the compiler driver can also link the generated binaries with LLVM’s static analysis/sanitizer libraries (e.g. MemorySanitizer). This is not yet available in Flang, but will be relatively easy to support once such libraries become available. Flang‘s compiler driver is intended for Flang’s end-users - its interface needs to remain stable. Otherwise, Flang's users will have to adjust their build scripts every time a compiler flag is changed.

Compiler Driver
Flang’s compiler driver and the tools that it runs

The frontend driver glues together and drives all of the Flang's frontend libraries. As such, it provides an easy-to-use and intuitive interface to the frontend. It uses MLIR and LLVM for code-generation and can be viewed as a driver for Flang, LLVM and MLIR libraries. Contrary to the compiler driver, it is not capable of calling any external tools (including linkers). It is aware of all the frontend internals that are “hidden” from the compiler driver. It accepts many frontend-specific options not available in flang-new and as such it provides a finer control over the frontend. Note that this tool is mostly intended for Flang developers. In particular, there are no guarantees about the stability of its interface and compiler developers can use it to experiment with new flags.

Frontend Driver
Flang's frontend driver and the libraries that it drives

Note that similarly to -Xclang in clang, you can use -Xflang to forward a frontend specific flag from the compiler directly to the frontend driver, e.g.:

flang-new -Xflang -fdebug-dump-parse-tree input.f95

In the invocation above, -fdebug-dump-parse-tree is forwarded to flang-new -fc1. Without the forwarding flag, -Xflang, you would see the following warning:

flang-new: warning: argument unused during compilation:

As -fdebug-dump-parse-tree is only supported by flang-new -fc1, flang-new will ignore it when used without Xflang.

Why Do We Need Two Drivers?

As hinted above, flang-new and flang-new -fc1 are two separate tools. The fact that these tools are accessed through one binary, flang-new, is just an implementation detail. Each tool has a separate list of options, albeit defined in the same file: clang/include/clang/Driver/Options.td.

The separation helps us split various tasks and allows us to implement more specialised tools. In particular, flang-new is not aware of various compilation phases within the frontend (e.g. scanning, parsing or semantic checks). It does not have to be. Conversely, the frontend driver, flang-new -fc1, needs not to be concerned with linkers or other external tools like assemblers. Nor does it need to know where to look for various systems libraries, which is usually OS and platform specific.

One helpful way of differentiating these tools is to keep in mind that:

  • the compiler driver is an end-user tool
  • frontend driver is a compiler developer tool with many additional options,

Also, Since the compiler driver can call external tools, e.g. linkers, it can be used to generate executables. The frontend driver cannot call external tools and hence can only generate object files. A similar model is implemented in Clang (clang vs clang -cc1 vs clang -cc1as), which is based on the architecture of GCC. In fact, Flang needs to adhere to this model in order to be able to re-use Clang's driver library. If you are more familiar with the architecture of GFortran than Clang, then flang-new corresponds to gfortran and flang-new -fc1 to f951.

Compiler Driver

The main entry point for Flang‘s compiler driver is implemented in flang/tools/flang-driver/driver.cpp. Flang’s compiler driver is implemented in terms of Clang's driver library, clangDriver. This approach allows us to:

  • benefit from Clang's support for various targets, platforms and operating systems
  • leverage Clang‘s ability to drive various backends available in LLVM, as well as linkers and assemblers. One implication of this dependency on Clang is that all of Flang’s compiler options are defined alongside Clang's options in clang/include/clang/Driver/Options.td. For options that are common for both Flang and Clang, the corresponding definitions are shared.

Internally, a clangDriver based compiler driver works by creating actions that correspond to various compilation phases, e.g. PreprocessJobClass, CompileJobClass, BackendJobClass or LinkJobClass from the clang::driver::Action::ActionClass enum. There are also other, more specialised actions, e.g. MigrateJobClass or InputClass, that do not map directly to common compilation steps. The actions to run are determined from the supplied compiler flags, e.g.

  • -E for PreprocessJobClass,
  • -c for CompileJobClass.

In most cases, the driver creates a chain of actions/jobs/phases where the output from one action is the input for the subsequent one. You can use the -ccc-print-phases flag to see the sequence of actions that the driver will create for your compiler invocation:

flang-new -ccc-print-phases -E file.f
+- 0: input, "file.f", f95-cpp-input
1: preprocessor, {0}, f95

As you can see, for -E the driver creates only two jobs and stops immediately after preprocessing. The first job simply prepares the input. For -c, the pipeline of the created jobs is more complex:

flang-new -ccc-print-phases -c file.f
         +- 0: input, "file.f", f95-cpp-input
      +- 1: preprocessor, {0}, f95
   +- 2: compiler, {1}, ir
+- 3: backend, {2}, assembler
4: assembler, {3}, object

Note that currently Flang does not support code-generation and flang-new will fail during the second step above with the following error:

error: code-generation is not available yet

The other phases are printed nonetheless when using -ccc-print-phases, as that reflects what clangDriver, the library, will try to create and run.

For actions specific to the frontend (e.g. preprocessing or code generation), a command to call the frontend driver is generated (more specifically, an instance of clang::driver::Command). Every command is bound to an instance of clang::driver::Tool. For Flang we introduced a specialisation of this class: clang::driver::Flang. This class implements the logic to either translate or forward compiler options to the frontend driver, flang-new -fc1.

You can read more on the design of clangDriver in Clang's Driver Design & Internals.

Frontend Driver

Flang‘s frontend driver is the main interface between compiler developers and the Flang frontend. The high-level design is similar to Clang’s frontend driver, clang -cc1 and consists of the following classes:

  • CompilerInstance, which is a helper class that encapsulates and manages various objects that are always required by the frontend (e.g. AllSources, AllCookedSources, Parsing, CompilerInvocation, etc.). In most cases CompilerInstance` owns these objects, but it also can share them with its clients when required. It also implements utility methods to construct and manipulate them.
  • CompilerInvocation encapsulates the configuration of the current invocation of the compiler as derived from the command-line options and the input files (in particular, file extensions). Among other things, it holds an instance of FrontendOptions. Like CompilerInstance, it owns the objects that it manages. It can share them with its clients that want to access them even after the corresponding CompilerInvocation has been destructed.
  • FrontendOptions holds options that control the behaviour of the frontend, as well as e.g. the list of the input files. These options come either directly from the users (through command-line flags) or are derived from e.g. the host system configuration.
  • FrontendAction and FrontendActions (the former being the base class for the latter) implement the actual actions to perform by the frontend. Usually there is one specialisation of FrontendActions for every compiler action flag (e.g. -E, -fdebug-unparse). These classes also contain various hooks that allow you to e.g. fine-tune the configuration of the frontend based on the input.

This list is not exhaustive and only covers the main classes that implement the driver. The main entry point for the frontend driver, fc1_main, is implemented in flang/tools/flang-driver/driver.cpp. It can be accessed by invoking the compiler driver, flang-new, with the -fc1 flag.

The frontend driver will only run one action at a time. If you specify multiple action flags, only the last one will be taken into account. The default action is ParseSyntaxOnlyAction, which corresponds to -fsyntax-only. In other words, flang-new -fc1 <input-file> is equivalent to flang-new -fc1 -fsyntax-only <input-file>.

The flang-to-external-fc script

The flang-to-external-fc wrapper script for flang-new was introduced as a development tool and to facilitate testing. The flang-to-external-fc wrapper script will:

  • use flang-new to unparse the input source file (i.e. it will run flang-new -fc1 -fdebug-unparse <input-file>), and then
  • call a host Fortran compiler, e.g. gfortran, to compile the unparsed file.

Here's a basic breakdown of what happens inside flang-to-external-fc when you run flang-to-external-fc file.f90:

flang-new -fc1 -fdebug-unparse file.f90 -o file-unparsed.f90
gfortran file-unparsed.f90

This is a simplified version for illustration purposes only. In practice, flang-to-external-fc adds a few more frontend options and it also supports various other use cases (e.g. compiling C files, linking existing object files). gfortran is the default host compiler used by flang-to-external-fc. You can change it by setting the FLANG_FC environment variable.

Adding new Compiler Options

Adding a new compiler option in Flang consists of two steps:

  • define the new option in a dedicated TableGen file,
  • parse and implement the option in the relevant drivers that support it.

Option Definition

All of Flang's compiler and frontend driver options are defined in clang/include/clang/Driver/Options.td in Clang. When adding a new option to Flang, you will either:

  • extend the existing definition for an option that is already available in one of Clang's drivers (e.g. clang), but not yet available in Flang, or
  • add a completely new definition if the option that you are adding has not been defined yet.

There are many predefined TableGen classes and records that you can use to fine tune your new option. The list of available configurations can be overwhelming at times. Sometimes the easiest approach is to find an existing option that has similar semantics to your new option and start by copying that.

For every new option, you will also have to define the visibility of the new option. This is controlled through the Flags field. You can use the following Flang specific option flags to control this:

  • FlangOption - this option will be available in the flang-new compiler driver,
  • FC1Option - this option will be available in the flang-new -fc1 frontend driver,
  • FlangOnlyOption - this option will not be visible in Clang drivers.

Please make sure that options that you add are only visible in drivers that can support it. For example, options that only make sense for Fortran input files (e.g. -ffree-form) should not be visible in Clang and be marked as FlangOnlyOption.

When deciding what OptionGroup to use when defining a new option in the Options.td file, many new options fall into one of the following two categories:

  • Action_Group - options that define an action to run (e.g. -fsyntax-only, -E)
  • f_Group - target independent compiler flags (e.g. -ffixed-form, -fopenmp) There are also other groups and occasionally you will use them instead of the groups listed above.

Option Implementation

First, every option needs to be parsed. Flang compiler options are parsed in two different places, depending on which driver they belong to:

  • frontend driver: flang/lib/Frontend/CompilerInvocation.cpp,
  • compiler driver: clang/lib/Driver/ToolChains/Flang.cpp.

The parsing will depend on the semantics encoded in the TableGen definition.

When adding a compiler driver option (i.e. an option that contains FlangOption among its Flags) that you also intend to be understood by the frontend, make sure that it is either forwarded to flang-new -fc1 or translated into some other option that is accepted by the frontend driver. In the case of options that contain both FlangOption and FC1Option among its flags, we usually just forward from flang-new to flang-new -fc1. This is then tested in flang/test/Driver/frontend-forward.F90.

What follows is usually very dependant on the meaning of the corresponding option. In general, regular compiler flags (e.g. -ffree-form) are mapped to some state within the driver. A lot of this state is stored within an instance of FrontendOptions, but there are other more specialised classes too. Action flags (e.g. -fsyntax-only) are usually more complex overall, but also more structured in terms of the implementation.

Action Options

For options that correspond to an action (i.e. marked as Action_Group), you will have to define a dedicated instance of FrontendActions in flang/include/flang/Frontend/FrontendOptions.h. For example, for -fsyntax-only we defined:

class ParseSyntaxOnlyAction : public PrescanAndSemaAction {
  void ExecuteAction() override;

Command line options are mapped to frontend actions through the Fortran::frontend::ActionKind enum. For every new action option that you add, you will have to add a dedicated entry in that enum (e.g. ParseSyntaxOnly for -fsyntax-only) and a corresponding case in ParseFrontendArgs function in the CompilerInvocation.cpp file, e.g.:

    case clang::driver::options::OPT_fsyntax_only:
      opts.programAction = ParseSyntaxOnly;

Note that this simply sets the program/frontend action within the frontend driver. You still have make sure that the corresponding frontend action class is instantiated when your new action option is used. The relevant switch statement is implemented in Fortran::frontend::CreatedFrontendBaseAction in the ExecuteCompilerInvocation.cpp file. Here's an example for -fsyntax-only:

  case ParseSyntaxOnly:
    return std::make_unique<ParseSyntaxOnlyAction>();

At this point you should be able to trigger that frontend action that you have just added using your new frontend option.

CMake Support

As of #7246 (and soon to be released CMake 3.24.0), cmake can detect flang-new as a supported Fortran compiler. You can configure your CMake projects to use flang-new as follows:

cmake -DCMAKE_Fortran_FLAGS="-flang-experimental-exec" -DCMAKE_Fortran_COMPILER=<path/to/flang-new> <src/dir>

You should see the following in the output:

-- The Fortran compiler identification is LLVMFlang <version>

where <version> corresponds to the LLVM Flang version. Note that while generating executables remains experimental, you will need to inform CMake to use the -flang-experimental-exec flag when invoking flang-new as in the example above.


In LIT, we define two variables that you can use to invoke Flang's drivers:

  • %flang is expanded as flang-new (i.e. the compiler driver)
  • %flang_fc1 is expanded as flang-new -fc1 (i.e. the frontend driver)

For most regression tests for the frontend, you will want to use %flang_fc1. In some cases, the observable behaviour will be identical regardless of whether %flang or %flang_fc1 is used. However, when you are using %flang instead of %flang_fc1, the compiler driver will add extra flags to the frontend driver invocation (i.e. flang-new -fc1 -<extra-flags>). In some cases that might be exactly what you want to test. In fact, you can check these additional flags by using the -### compiler driver command line option.

Lastly, you can use ! REQUIRES: <feature> for tests that will only work when <feature> is available. For example, you can use! REQUIRES: shell to mark a test as only available on Unix-like systems (i.e. systems that contain a Unix shell). In practice this means that the corresponding test is skipped on Windows.

Frontend Driver Plugins

Plugins are an extension to the frontend driver that make it possible to run extra user defined frontend actions, in the form of a specialization of a PluginParseTreeAction. These actions are run during compilation, after semantic checks. Similarly to Clang, Flang leverages LoadLibraryPermanently from LLVM's llvm::sys::DynamicLibrary to load dynamic objects that implement plugins. The process for using plugins includes:

Flang plugins are limited to flang-new -fc1 and are currently only available / been tested on Linux.

Creating a Plugin

There are three parts required for plugins to work:

  1. PluginParseTreeAction subclass
  2. Implementation of ExecuteAction
  3. Plugin registration

There is an example plugin located in flang/example/PrintFlangFunctionNames that demonstrates these points by using the ParseTree API to print out function and subroutine names declared in the input file.

A PluginParseTreeAction Subclass

This subclass will wrap everything together and represent the FrontendAction corresponding to your plugin. It will need to inherit from PluginParseTreeAction (defined in flang/include/flang/FrontendActions.h), in order to have access to the parse tree post semantic checks, and also so that it can be registered, e.g.

class PrintFunctionNamesAction : public PluginParseTreeAction

Implementation of ExecuteAction

Like in other frontend actions, the driver looks for an ExecuteAction function to run, so in order for your plugin to do something, you will need to implement the ExecuteAction method in your plugin class. This method will contain the implementation of what the plugin actually does, for example:

// Forward declaration
struct ParseTreeVisitor;

void ExecuteAction() override {
  ParseTreeVisitor visitor;
  Fortran::parser::Walk(getParsing().parseTree(), visitor);

In the example plugin, the ExecuteAction method first creates an instance of visitor struct, before passing it together with the parse tree to the Fortran::parser::Walk function that will traverse the parse tree. The parse tree will normally be generated by the frontend driver and can be retrieved in your plugin through the getParsing() member method. Implementation and details of the Walk function can be found in flang/include/flang/Parser/parse-tree-visitor.h.

You will have to define your own visitor struct. It should define different Pre and Post functions that take the type of a specific ParseTree node as an argument. When the Walk function is traversing the parse tree, these functions will be run before/after a node of that type is visited. Template functions for Pre/Post are defined so that when a node is visited that you have not defined a function for, it will still be able to continue. Pre returns a bool indicating whether to visit that node's children or not. For example:

struct ParseTreeVisitor {
  template <typename A> bool Pre(const A&) { return true; }
  template <typename A> void Post(const A&) {}
  void Post(const Fortran::parser::FunctionStmt &f) {
    llvm::outs() << std::get<Fortran::parser::Name>(f.t).ToString() << "\n" ;

The different types of nodes and also what each node structure contains are defined in flang/include/flang/Parser/parse-tree.h. In the example, there is a Post function, with a line that gets the Name element from a tuple t in the FunctionStmt struct and prints it. This function will be run after every FunctionStmt node is visited in the parse tree.

Plugin Registration

A plugin registry is used to store names and descriptions of a collection of plugins. The Flang plugin registry, defined in flang/include/flang/Frontend/FrontendPluginRegistry.h, is an alias of llvm::Registry of type PluginParseTreeAction.

The plugin will need to be registered, which will add the Plugin to the registry and allow it to be used. The format is as follows, with print-fns being the plugin name that is used later to call the plugin and Print Function names being the description:

static FrontendPluginRegistry::Add<PrintFunctionNamesAction> X(
    "print-fns", "Print Function names");

Loading and Running a Plugin

In order to use plugins, there are 2 command line options made available to the frontend driver, flang-new -fc1:

Invocation of the example plugin is done through:

flang-new -fc1 -load flangPrintFunctionNames.so -plugin print-fns file.f90

Both these options are parsed in flang/lib/Frontend/CompilerInvocation.cpp and fulfil their actions in flang/lib/FrontendTool/ExecuteCompilerInvocation.cpp

The -load <dsopath> option

This loads the plugin shared object library, with the path given at <dsopath>, using LoadLibraryPermantly from LLVM's llvm::sys::DynamicLibrary, which itself uses dlopen. During this stage, the plugin is registered with the registration line from the plugin, storing the name and description.

The -plugin <name> option

This sets frontend::ActionKind programAction in FrontendOptions to PluginAction, through which it searches the plugin registry for the plugin name from <name>. If found, it returns the instantiated plugin, otherwise it reports an error diagnostic and returns nullptr.

Enabling In-Tree Plugins

For in-tree plugins, there is the CMake flag FLANG_PLUGIN_SUPPORT, enabled by default, that controls the exporting of executable symbols from flang-new, which plugins need access to. Additionally, there is the CMake flag FLANG_BUILD_EXAMPLES, turned off by default, that is used to control if the example programs are built. This includes plugins that are in the flang/example directory and added as a sub_directory to the flang/examples/CMakeLists.txt, for example, the PrintFlangFunctionNames plugin. It is also possible to develop plugins out-of-tree.


Note that the traversal API presented here is under active development and might change in the future. We expect it to evolve as support for new language features are added. This document and the examples will be updated accordingly.

The current ParseTree structure is not suitable for modifications. The copy constructors are not available and hence duplicating code might not be trivial. Please take this into consideration when designing your plugin. In particular, creating a transformation plugin will be noticeably harder than analysis plugins that just consume (rather than edit) ParseTree.

Lastly, if ParseTree modifications are performed, then it might be necessary to re-analyze expressions and modify scope or symbols. You can check Semantics.md for more details on how ParseTree is edited e.g. during the semantic checks.

LLVM Pass Plugins

Pass plugins are dynamic shared objects that consist of one or more LLVM IR passes. The -fpass-plugin option enables these passes to be passed to the middle-end where they are added to the optimization pass pipeline and run after lowering to LLVM IR.The exact position of the pass in the pipeline will depend on how it has been registered with the llvm::PassBuilder. See the documentation for llvm::PassBuilder for details.

The framework to enable pass plugins in flang-new uses the exact same machinery as that used by clang and thus has the same capabilities and limitations.

In order to use a pass plugin, the pass(es) must be compiled into a dynamic shared object which is then loaded using the -fpass-plugin option.

flang-new -fpass-plugin=/path/to/plugin.so <file.f90>

This option is available in both the compiler driver and the frontend driver. Note that LLVM plugins are not officially supported on Windows.

LLVM Pass Extensions

Pass extensions are similar to plugins, except that they can also be linked statically. Setting -DLLVM_${NAME}_LINK_INTO_TOOLS to ON in the cmake command turns the project into a statically linked extension. An example would be Polly, e.g., using -DLLVM_POLLY_LINK_INTO_TOOLS=ON would link Polly passes into flang-new as built-in middle-end passes.

See the WritingAnLLVMNewPMPass documentation for more details.

Ofast and Fast Math

-Ofast in Flang means -O3 -ffast-math. -fstack-arrays will be added to -Ofast in the future (https://github.com/llvm/llvm-project/issues/59231).

-ffast-math means the following:

  • -fno-honor-infinities
  • -fno-honor-nans
  • -fassociative-math
  • -freciprocal-math
  • -fapprox-func
  • -fno-signed-zeros
  • -ffp-contract=fast

These correspond to LLVM IR Fast Math attributes: https://llvm.org/docs/LangRef.html#fast-math-flags

When -ffast-math is specified, any linker steps generated by the compiler driver will also link to crtfastmath.o, which adds a static constructor that sets the FTZ/DAZ bits in MXCSR, affecting not only the current only the current compilation unit but all static and shared libraries included in the program. Setting these bits causes denormal floating point numbers to be flushed to zero.

Comparison with GCC/GFortran

GCC/GFortran translate -Ofast to -O3 -ffast-math -fstack-arrays -fno-semantic-interposition. -fstack-arrays is TODO for Flang. -fno-semantic-interposition is not used because clang does not enable this as part of -Ofast as the default behaviour is similar.

GCC/GFortran has a wider definition of -ffast-math: also including -fno-trapping-math, -fno-rounding-math, and -fsignaling-nans; these aren't included in Flang because Flang currently has no support for strict floating point and so always acts as though these flags were specified.

GCC/GFortran will also set flush-to-zero mode: linking crtfastmath.o, the same as Flang.

Comparison with nvfortran

nvfortran defines -fast as -O2 -Munroll=c:1 -Mnoframe -Mlre -Mpre -Mvect=simd -Mcache_align -Mflushz -Mvect.

  • -O2 -Munroll=c:1 -Mlre -Mautoinline -Mpre -Mvect-simd affect code optimization. flang -O3 should enable all optimizations for execution time, similarly to clang -O3. The -O3 pipeline has passes that perform transformations like inlining, vectorisation, unrolling, etc. Additionally, the GVN and LICM passes perform redundancy elimination like Mpre and Mlre
  • -Mnoframe: the equivalent flag would be -fomit-frame-pointer. This flag is not yet supported in Flang and so Flang follows GFortran in not including this in -Ofast. There is no plan to include this flag as part of -Ofast.
  • -Mcache_align: there is no equivalent flag in Flang or Clang.
  • -Mflushz: flush-to-zero mode - when -ffast-math is specified, Flang will link to crtfastmath.o to ensure denormal numbers are flushed to zero.