llvm-project/flang/documentation/Calls.md

<!--
Copyright (c) 2019, NVIDIA CORPORATION.  All rights reserved.
-->

## Procedure reference implementation protocol

Fortran function and subroutine references are complicated.
This document attempts to collect the requirements imposed by the 2018
standard (and legacy extensions) on programs and implementations, work
through the implications of the various features, and propose both a
runtime model and a compiler design.

All section, requirement, and constraint numbers herein pertain to
the Fortran 2018 standard.

This note does not consider calls to intrinsic procedures, statement
functions, or calls to internal runtime support library routines.

## Interfaces

Referenced procedures may or may not have declared interfaces
available to their call sites.

Calls to procedures with some post-Fortran '77 features require an
explicit interface (15.4.2.2):

* keyword arguments
* procedures that are `ELEMENTAL` or `BIND(C)`
* procedures that are required to be `PURE` due to the context of the call
  (specification expression, `DO CONCURRENT`, `FORALL`)
* dummy arguments with these attributes: `ALLOCATABLE`, `POINTER`,
  `VALUE`, `TARGET`, `OPTIONAL`, `ASYNCHRONOUS`, `VOLATILE`
  (but *not* `CONTIGUOUS`, `INTENT()`)
* dummy arguments that are coarrays, have assumed-shape/-rank,
  have parameterized derived types, &/or are polymorphic
* function results that are arrays, `ALLOCATABLE`, `POINTER`,
  or have derived types with any length parameters that are
  neither constant nor assumed

Module procedures, internal procedures, procedure pointers,
type-bound procedures, and recursive references by a procedure to itself
always have explicit interfaces.

Other uses of procedures besides calls may also require explicit interfaces,
such as procedure pointer assignment, type-bound procedure bindings, &c.

Note that non-parameterized monomorphic derived type arguments do
not by themselves require the use of an explicit interface; we should perhaps emit a warning
when the derived type of an actual argument to an implicit interface is
not a `SEQUENCE` type (7.5.2.3).

A procedure that can be called with an implicit interface can
have a dummy procedure that requires an explicit interface.

### Implicit interfaces

In the absence of any characteristic or context that requires an
explicit interface (see above), an external function or subroutine (R503)
or `ENTRY` (R1541) can be called directly or indirectly via its implicit interface.
Each of the arguments can be passed as a simple address, including
dummy procedures.
Procedures that *can* be called via an implicit interface can
enjoy more thorough checking
by semantics when they do have a visible external interface, but must be
compiled as if all calls to them were through the implicit interface.

Internal and module subprograms that are ever passed as arguments &/or
assigned as targets of procedure pointers are also potentially at risk of
being called indirectly through implicit interfaces.

Every procedure callable via an implicit interface
can and must be distiguished at compilation time.
Such procedures should respect the external naming conventions (when external)
and any legacy ABI used for Fortran '77 programs on the target architecture,
so that portable libraries can be compiled
and used by distinct implementations (and their versions)
of Fortran.

Note that functions callable via implicit interfaces still have known result
types, possibly by means of implicit typing of their names.
They can also be `CHARACTER(*)` assumed-length character functions.

In other words: procedures that can be referenced with implicit interfaces
have argument lists that comprise only addresses of actual arguments,
and possibly the length of an assumed-length `CHARACTER(*)` result,
and they can return only scalar values of intrinsic types.
None of their arguments or results need be (or can be) implemented
with descriptors,
and any internal procedures passed to them as arguments must be
addresses of trampolines.

Note that `INTENT` and `CONTIGUOUS` attributes do not, by themselves,
require the use of explicit interface; neither do dummy procedures.
Analyses of calls to procedures with implicit interfaces must make
allowances or impose restrictions capable of dealing with any of the
invisible `INTENT` and `CONTIGUOUS` attributes of dummy arguments
in the called procedure.

## Protocol overview

Here is a summary script of all of the actions that may need to be taken
by the calling procedure and its referenced procedure to effect
the call, entry, exit, and return steps of the procedure reference
protocol.
The order of these steps is not particularly strict, and we have
some design alternatives that are explored further below.

### Before the call:

1. Compute &/or copy into temporary storage the values of
   some actual argument expressions and designators (see below).
1. Create and populate descriptors for assumed-shape/-rank arrays,
   parameterized derived types with length, polymorphic types,
   coarrays, and non-`POINTER` actual arguments (that are`TARGET`
   or procedures) associated with `INTENT(IN) POINTER`
   dummy arrays (15.5.2.7, C.10.4).
1. When passing internal procedures, procedure pointers, or dummy
   procedure descriptors as arguments to target procedures
   that can be called via implicit interfaces, package them as
   trampolines; pass or forward other possible internal procedures
   as descriptors containing host variable links.
1. Possibly allocate function result storage,
   when its size can be known by all callers; function results that are
   neither `POINTER` nor `ALLOCATABLE` must have explicit shapes (C816).
1. Create and populate a descriptor for the function result, if it
   needs one (deferred-shape/-length `POINTER`, any `ALLOCATABLE`,
   parameterized derived type with non-constant length parameters, &c.).
1. Capture the values of host-escaping local objects in memory;
   package them into single address (for calls to internal procedures &
   for calls that pass internal procedures as arguments).
1. Resolve the target procedure's polymorphic binding, if any.
1. Marshal actual argument addresses/values into registers.
1. Marshal an extra argument for the assumed-length `CHARACTER` result
   length or the function result descriptor.
1. Set the host variable link register when calling an internal procedure
   from its host or another internal procedure, a procedure pointer,
   or dummy procedure (when it has a descriptor).
1. Jump.

### On entry:
1. Shuffle `ENTRY` dummy arguments & jump to common entry point.
1. Complete `VALUE` copying if this step will not always be done
   by the caller (as I think it should be).
1. Finalize &/or re-initialize `INTENT(OUT)` non-pointer
   actual arguments (see below).
1. Optionally compact assumed-shape arguments for contiguity on one
   or more leading dimensions to improve SIMD vectorization, if not
   `TARGET` and not already sufficiently contiguous.
   (PGI does this in the caller, whether the callee needs it or not.)
1. Complete allocation of function result storage, if that has
   not been done by the caller.
1. Initialize components of derived type local variables,
   including the function result.

Execute the callee, populating the function result or selecting
the subroutine's alternate return.

### On exit:
1. Clean up local scope (finalization, deallocation)
1. Deallocate `VALUE` argument temporaries.
   (But don't finalize them; see 7.5.6.3(3)).
1. Replace any assumed-shape argument data that were compacted on
   entry for partial contiguity for SIMD vectorization, if possibly
   modified across the call (never when `INTENT(IN)` or `VALUE`).
1. Identify alternate `RETURN` to caller.
1. Marshal results.
1. Jump

### On return to the caller:
1. Save the result registers, if any.
1. Copy actual argument array designator data that was copied into
   a temporary back into its original storage (see below).
1. Complete deallocation of actual argument temporaries (not `VALUE`).
1. Reload definable host-escaping local objects from memory, if they
   were saved to memory by the host before the call.
1. `GO TO` alternate return, if any.
1. Use the function result in an expression.
1. Eventually, finalize &/or deallocate the function result.

## The messy details

### Copying actual argument values into temporary storage

There are several conditions that require the compiler to generate
code that allocates and populates temporary storage for an actual
argument.

First, actual arguments that are expressions, not designators, obviously
need to be computed and captured into memory in order to be passed
by reference.
This includes parenthesized designators like `(X)` as an important
special case, which are expressions in Fortran.
(This case also technically includes constants, but those are better
implemented by passing addresses in read-only memory when the interface
is explicit.)
The dummy argument cannot be known to have `INTENT(OUT)` or
`INTENT(IN OUT)`.

Small scalar or elemental `VALUE` arguments may be passed in registers.
Multiple elemental `VALUE` arguments might be packed into SIMD registers.

Actual arguments that are designators, not expressions, must also
be copied into temporaries in the following situations.

1. Coindexed objects need to be copied into the local image.
   This can get very involved if they contain `ALLOCATABLE`
   components, which also need to be copied, along with their
   `ALLOCATABLE` components, and may be best implemented with a runtime
   library routine working off a description of the type.
1. Actual arguments associated with dummies with the `VALUE`
   attribute need to copied; this could be done on either
   side of the call, but there are optimization opportunities
   on the caller's side.
1. In non-elemental calls, the values of array sections with
   vector-valued subscripts need to be compacted into temporaries.
   These actual arguments are not definable, and they are not allowed to
   be associated with non-`VALUE` dummy arguments with the attributes
   `INTENT(IN)`, `INTENT(IN OUT)`, `ASYNCHRONOUS`, or `VOLATILE`
   (15.4.2.4(21)); `INTENT()` can't always be checked.
1. Non-simply-contiguous (9.5.4) arrays being passed to non-`POINTER`
   dummy arguments that must be contiguous due to a `CONTIGUOUS`
   attribute, implicit interface, or not being assumed-shape/-rank.
   This should be a runtime decision, so that actual arguments
   that turn out to be contiguous can be passed cheaply.
   This rule does not apply to coarray dummies, whose actual arguments
   are required to be simply contiguous when this rule would otherwise
   force the use of a temporary (15.5.2.8); neither does it apply
   to `ASYNCHRONOUS` and `VOLATILE` actual arguments, which are
   disallowed when it matters (C1539, C1540).
   *Only temporaries created by this contiguity requirement are
   subject to being copied back to the original variable after
   the call* (see below).

While we are unlikely to want to _needlessly_ use a temporary for
an actual argument that does not require one for any of these
reasons above, we are specifically disallowed from doing so
by the standard in cases where pointers to the original target
data are required to be valid across the call (15.5.2.4(9-10)).
In particular, compaction of assumed-shape arrays for contiguity
on the leading dimension to ease SIMD vectorization cannot be
done safely for `TARGET` dummies.

Actual arguments associated with known `INTENT(OUT)` dummies that
require allocation of a temporary -- and this can only be for reasons of
contiguity -- don't have to populate it, but they do have to perform
minimal initialization of any `ALLOCATABLE` components so that
the runtime doesn't crash when the callee finalizes and deallocates
them.
Note that calls to implicit interfaces must conservatively allow
for the use of `INTENT(OUT)` by the callee.

Except for `VALUE` and known `INTENT(IN)` dummy arguments, the original
contents of local designators that have been compacted into temporaries
could optionally have their `ALLOCATABLE` components invalidated
across the call as an aid to debugging.

Except for `VALUE` and known `INTENT(IN)` dummy arguments, the contents of
the temporary storage will be copied back into the actual argument
designator after control returns from the procedure, and it may be necessary
to preserve addresses (or the values of subscripts and cosubscripts
needed to recalculate them) of the actual argument designator, or its
elements, in additional temporary storage if they can't be safely or
quickly recomputed after the call.

### `INTENT(OUT)` preparation

Actual arguments that are associated with `INTENT(OUT)`
dummy arguments are required to be definable.
This cannot always be checked, as the use of `INTENT(OUT)`
does not by itself mandate the use of an explicit interface.

`INTENT(OUT)` arguments are finalized (as if) on entry to the called
procedure.  In particular, in calls to elemental procedures,
the elements of an array are finalized by a scalar or elemental
`FINAL` procedure (7.5.6.3(7)).

Derived type components that are `ALLOCATABLE` are finalized
and deallocated.
Components with initializers are (re)initialized.

The preparation of actual arguments for `INTENT(OUT)` could be
done on either side of the call.  If the preparation is
done by the caller, there is an optimization opportunity
in situations where unmodified incoming `INTENT(OUT)` dummy
arguments are being passed onward as outgoing `INTENT(OUT)`
arguments.

### Copying temporary storage back into actual argument designators

Except for `VALUE` and known `INTENT(IN)` dummy arguments and array sections
with vector-valued subscripts (15.5.2.4(21)), temporary storage into
which actual argument data were compacted for contiguity before the call
must be redistributed back to its original storage by the caller after
the return.

In conjunction with saved cosubscript values, a standard descriptor
suffices to represent a pointer to the original storage into which the
temporary data should be redistributed.

Note that coindexed objects with `ALLOCATABLE` ultimate components
are required to be associated only with dummy arguments with the
`VALUE` &/or `INTENT(IN)` attributes (15.6.2.4(6)), so there is no
requirement that the local image somehow reallocate remote storage
when copying the data back.

### Polymorphic bindings

Calls to the type-bound procedures of monomorphic types are
resolved at compilation time, as are calls to `NON_OVERRIDABLE`
type-bound procedures.
The resolution of calls to overridable type-bound procedures of
polymorphic types must be completed at execution (generic resolution
of type-bound procedure bindings from actual argument types, kinds,
and ranks is always a compilation-time task (C.10.6)).

Each derived type that declares or inherits any overridable
type-bound procedure bindings must correspond to a static constant
table of code addresses (or, more likely, a static constant type
description containing or pointing to such a table, along with
information used by the runtime support library for initialization,
copying, finalization, and I/O of type instances).  Each overridable
type-bound procedure in the type corresponds to an index into this table.

### Host association linkage

Calls to dummy procedures and procedure pointers that resolve to
internal procedures need to pass an additional argument that
addresses on block of storage in the stack frame of the their
host subprogram that was active at the time they were passed as an
actual argument or associated with a procedure pointer.
This is similar to a static link in implementations of programming
languages with nested subprograms, although Fortran only allows
one level of nesting.
The 64-bit x86 and little-endian OpenPower ABIs reserve registers
for this purpose (`%r10` & `R11`); 64-bit ARM has a reserved register
that can be used (`x18`).

The host subprogram objects that are visible to any of their internal
subprograms need to be resident in memory across any calls to them
(direct or not).  Any host subprogram object that might be defined
during a call to an internal subprogram needs to be reloaded after
a call or reside permanently in memory.
A simple conservative analysis of the internal subprograms can
identify all of these escaping objects and their definable subset.

The address of the host subprogram storage used to hold the escaping
objects needs to be saved alongside the code address(es) that
represent a procedure pointer.
It also needs to be conveyed alongside the text address for a
dummy procedure.

For procedures that can be called with an implicit interface,
we cannot use a "procedure pointer descriptor" to pass them an
a procedure argument -- a Fortran '77 routine
with an `EXTERNAL` dummy argument expects to receive a single
address argument.  Instead, when passing an actual procedure
to a procedure that can be called with an implicit interface,
we will need to package the host link in a trampoline.

GNU Fortran and Intel Fortran construct trampolines by writing
a sequence of machine instructions to a block of storage in the
host's stack frame, which requires the stack to be executable,
which seems inadvisable for security reasons;
XLF manages trampolines in its runtime support library, which adds some overhead
to their construction and a reclamation obligation;
NAG Fortran manages a static fixed-sized stack of trampolines
per call site, imposing a hidden limit on recursion and foregoing
reentrancy;
PGI passes host variable links in descriptors in additional arguments
that are not always successfully forwarded across implicit interfaces,
sometimes leading to crashes when they turn out to be needed.

F18 will manage a pool of trampolines in its runtime support library
that can be used to pass internal procedures as actual arguments
to targets that might be called with an implicit interface, so that
a bare code address is used to represent the actual argument.
But targets that can only be called with an explicit interface
have the option of using a "fat pointer" (or additional argument)
to represent a dummy procedure so as
to avoid the overhead of constructing and reclaiming a trampoline.

### Naming

External subroutines and functions (R503) and `ENTRY` points (R1541)
with `BIND(C)` (R808) have linker-visible names that are either explicitly
specified in the program or determined by straightforward rules.
The names of other external procedures that might be called
via an implicit interface should respect the conventions of the target architecture
for legacy Fortran '77 programs; this is typically something like `foo_`.

In other cases, however, we have fewer constraints on external naming,
as well as some additional requirements and goals.

Module procedures need to be distinguished by the name of their module
and (when they have one) the submodule where their interface was
defined.
Note that submodule names are distinct in their modules, not hierarchical,
so at most two levels of qualification are needed.

Pure `ELEMENTAL` functions (15.8) must use distinct names for any alternate
entry points used for packed SIMD arguments of various widths if we support
calls to these functions in SIMD parallel contexts.
There are already conventions for these names in `libpgmath`.

It would be great to have the option of distinguishing the names of
external procedures that cannot be called via an implicit interface
as a means for catching attempts to do so and causing them to fail with
link errors.
But this turns out to not be possible due to the possibility of passing
an `EXTERNAL` procedure without an interface as an argument to
another procedure with an implicit interface.

Last, there must be reasonably permanent naming conventions used
by the F18 runtime library for those unrestricted specific intrinsic
functions (table 16.2 in 16.8) and extensions that can be passed as
arguments.

In these cases where external naming is at the discretion
of the implementation, we should use names that are not in the C language
user namespace, begin with something that identifies
the current incompatible version of F18, the module, the submodule, and
elemental SIMD width, and are followed by the external name.
The parts of the external name can be separated by some character that
is acceptable for use in LLVM IR and assembly language but not in user
Fortran or C code, or by switching case
(so long as there's a way to cope with extension names that don't begin
with letters).

In particular, the period (`.`) seems safe to use as a separator character,
so a `Fa.` prefix can serve to isolate these discretionary names from
other uses and to identify the earliest link-compatible version.
For examples: `Fa.mod.foo`, `Fa.mod.submod.foo`.

## Further topics to document

### Arguments
* Alternate return specifiers
* `%VAL()` and `%REF()`
* Unrestricted specific intrinsic functions as actual arguments
* Check definability of known `INTENT(OUT)` and `INTENT(IN OUT)` actuals.
* Whether lower bounds in argument descriptors should be
  initialized (they shouldn't be used)

### Other
* SIMD variants of `ELEMENTAL` procedures (& unrestricted specific intrinsics)
* Interoperable procedures
* Multiple code addresses for dummy procedures
* Elemental calls with array arguments