docs/Extensions.md

<!--===- docs/Extensions.md

   Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
   See https://llvm.org/LICENSE.txt for license information.
   SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception

-->

# Fortran Extensions supported by Flang


As a general principle, this compiler will accept by default and
without complaint many legacy features, extensions to the standard
language, and features that have been deleted from the standard,
so long as the recognition of those features would not cause a
standard-conforming program to be rejected or misinterpreted.

Other non-standard features, which do conflict with the current
standard specification of the Fortran programming language, are
accepted if enabled by command-line options.

## Intentional violations of the standard

* Scalar `INTEGER` actual argument expressions (not variables!)
  are converted to the kinds of scalar `INTEGER` dummy arguments
  when the interface is explicit and the kinds differ.
  This conversion allows the results of the intrinsics like
  `SIZE` that (as mentioned below) may return non-default
  `INTEGER` results by default to be passed.  A warning is
  emitted when truncation is possible.  These conversions
  are not applied in calls to non-intrinsic generic procedures.
* We are not strict on the contents of `BLOCK DATA` subprograms
  so long as they contain no executable code, no internal subprograms,
  and allocate no storage outside a named `COMMON` block.  (C1415)
* Delimited list-directed (and NAMELIST) character output is required
  to emit contiguous doubled instances of the delimiter character
  when it appears in the output value.  When fixed-size records
  are being emitted, as is the case with internal output, this
  is not possible when the problematic character falls on the last
  position of a record.  No two other Fortran compilers do the same
  thing in this situation so there is no good precedent to follow.
  Because it seems least wrong, we emit one copy of the delimiter as
  the last character of the current record and another as the first
  character of the next record.  (The second-least-wrong alternative
  might be to flag a runtime error, but that seems harsh since it's
  not an explicit error in the standard, and the output may not have
  to be usable later as input anyway.)
  Consequently, the output is not suitable for use as list-directed or
  NAMELIST input.  If a later standard were to clarify this case, this
  behavior will change as needed to conform.
```
character(11) :: buffer(3)
character(10) :: quotes = '""""""""""'
write(buffer,*,delim="QUOTE") quotes
print "('>',a10,'<')", buffer
end
```
* The name of the control variable in an implied DO loop in an array
  constructor or DATA statement has a scope over the value-list only,
  not the bounds of the implied DO loop.  It is not advisable to use
  an object of the same name as the index variable in a bounds
  expression, but it will work, instead of being needlessly undefined.
* If both the `COUNT=` and the `COUNT_MAX=` optional arguments are
  present on the same call to the intrinsic subroutine `SYSTEM_CLOCK`,
  we require that their types have the same integer kind, since the
  kind of these arguments is used to select the clock rate.  In common
  with some other compilers, the clock rate varies from tenths of a
  second to nanoseconds depending on argument kind and platform support.
* If a dimension of a descriptor has zero extent in a call to
  `CFI_section`, `CFI_setpointer` or `CFI_allocate`, the lower
  bound on that dimension will be set to 1 for consistency with
  the `LBOUND()` intrinsic function.
* `-2147483648_4` is, strictly speaking, a non-conforming literal
  constant on a machine with 32-bit two's-complement integers as
  kind 4, because the grammar of Fortran expressions parses it as a
  negation of a literal constant, not a negative literal constant.
  This compiler accepts it with a portability warning.
* Construct names like `loop` in `loop: do j=1,n` are defined to
  be "local identifiers" and should be distinct in the "inclusive
  scope" -- i.e., not scoped by `BLOCK` constructs.
  As most (but not all) compilers implement `BLOCK` scoping of construct
  names, so does Flang, with a portability warning.
* 15.6.4 paragraph 2 prohibits an implicitly typed statement function
  from sharing the same name as a symbol in its scope's host, if it
  has one.
  We accept this usage with a portability warning.
* A module name from a `USE` statement can also be used as a
  non-global name in the same scope.  This is not conforming,
  but it is useful and unambiguous.
* The argument to `RANDOM_NUMBER` may not be an assumed-size array.
* `NULL()` without `MOLD=` is not allowed to be associated as an
  actual argument corresponding to an assumed-rank dummy argument;
  its rank in the called procedure would not be well-defined.
* When an index variable of a `FORALL` or `DO CONCURRENT` is present
  in the enclosing scope, and the construct does not have an explicit
  type specification for its index variables, some weird restrictions
  in F'2023 subclause 19.4 paragraphs 6 & 8 should apply.  Since this
  compiler properly scopes these names, violations of these restrictions
  elicit only portability warnings by default.
* The standard defines the intrinsic functions `MOD` and `MODULO`
  for real arguments using expressions in terms of `AINT` and `FLOOR`.
  These definitions yield fairly poor results due to floating-point
  cancellation, and every Fortran compiler (including this one)
  uses better algorithms.
* The rules for pairwise distinguishing the specific procedures of a
  generic interface are inadequate, as admitted in note C.11.6 of F'2023.
  Generic interfaces whose specific procedures can be easily proven by
  hand to be pairwise distinct (i.e., no ambiguous reference is possible)
  appear in real applications, but are still non-conforming under the
  incomplete tests in F'2023 15.4.3.4.5.
  These cases are compiled with optional portability warnings.
* `PROCEDURE(), BIND(C) :: PROC` is not conforming, as there is no
  procedure interface.  This compiler accepts it, since there is otherwise
  no way to declare an interoperable dummy procedure with an arbitrary
  interface like `void (*)()`.
* `PURE` functions are allowed to have dummy arguments that are
  neither `INTENT(IN)` nor `VALUE`, similar to `PURE` subroutines,
  with a warning.
  This enables atomic memory operations to be naturally represented
  as `PURE` functions, which allows their use in parallel constructs
  and `DO CONCURRENT`.
* A non-definable actual argument, including the case of a vector
  subscript, may be associated with an `ASYNCHRONOUS` or `VOLATILE`
  dummy argument, F'2023 15.5.2.5 p21 notwithstanding.
  The effects of these attributes are scoped over the lifetime of
  the procedure reference, and they can by added by internal subprograms
  and `BLOCK` constructs within the procedure.
  Further, a dummy argument can acquire the `ASYNCHRONOUS` attribute
  implicitly simply appearing in an asynchronous data transfer statement,
  without the attribute being visible in the procedure's explicit
  interface.
* When the name of an extended derived type's base type is the
  result of `USE` association with renaming, the name of the extended
  derived type's parent component is the new name by which the base
  is known in the scope of the extended derived type, not the original.
  This interpretation has usability advantages and is what six other
  Fortran compilers do, but is not conforming now that J3 approved an
  "interp" in June 2024 to the contrary.
* When an Arm processor raises an `ieee_overflow` or `ieee_underflow`
  exception, the `ieee_inexact` exception is also raised. This happens
  for a call to `ieee_set_flag` as well as for floating point expression
  evaluation.
* Arm has processors that allow a user to control what happens when an
  arithmetic exception is signaled, as well as processors that do not
  have this capability. An Arm executable will run on either type of
  processor, so it is effectively unknown at compile time whether or
  not this support will be available at runtime. The standard requires
  that a call to intrinsic module procedure `ieee_support_halting` with
  a constant argument has a compile time constant result in `constant
  expression` and `specification expression` contexts. In compilations
  where this information is not known at compile time, Flang generates code
  to determine the absence or presence of this capability at runtime.
  A call to `ieee_support_halting` in contexts that the standard requires
  to be constant will generate a compilation error. `ieee_support_standard`
  depends in part on `ieee_support_halting`, so this also applies to
  `ieee_support_standard` calls.
* An override for an inaccessible procedure binding works only within
  the same module; other apparent overrides of inaccessible bindings
  are actually new bindings of the same name.
  In the case of `DEFERRED` bindings in an `ABSTRACT` derived type,
  however, overrides are necessary, so they are permitted for inaccessible
  bindings with an optional warning.
* Main program name is allowed to be the same as the other symbols used
  in the main program, for example:
```
module m
end
program m
use m
end
```
  Note that internally the main program symbol name is all uppercase, unlike
  the names of all other symbols, which are usually all lowercase. This
  may make a difference in testing. It is not visible in the debug
  information, which spells the main program name in lowercase like every
  other name.
* A `PROCEDURE()` with no interface name or type may be called as an
  subroutine with an implicit interface, F'2023 15.4.3.6 paragraph 4 and
  C1525 notwithstanding.
  This is a universally portable feature, and it also applies to
  `PROCEDURE(), POINTER, NOPASS` derived type components.
  Such procedures may *not* be referenced as implicitly typed functions
  without first being associated with a function pointer.
* Some intrinsic functions (`COUNT`, `LBOUND`, `LCOBOUND`, `TRANSFER`,
  `UBOUND`, and `UCOBOUND`) have arguments (usually `DIM=`) that are documented
  as not allowing `OPTIONAL` dummy arguments to appear as their values.
  This prohibition appeared on the `DIM=` arguments of more
  intrinsic functions in earlier revisions of the ISO standard.
  (Perhaps these are meant to avoid misunderstanding these arguments,
  which appear in square brackets in the synopses, as if their dynamic
  presence at runtime could affect the semantics of the intrinsic in
  the same way as the static presence or absence of the argument does
  at compilation time, which would not be possible.)
  No compiler seems to enforce this requirement.
  We interpret it
  to mean that an `OPTIONAL` dummy argument may appear but must be present
  during execution, just as a pointer argument must be associated or an
  allocatable argument must be allocated.

## Extensions, deletions, and legacy features supported by default

* Tabs in source
* A bare carriage return (CR, 0x0d) in the interior of a source line -- e.g.
  from a file with Windows line endings that has been mishandled -- is treated
  as a blank, except within a character or Hollerith literal where it is
  retained.
* `<>` as synonym for `.NE.` and `/=`
* `$` and `@` as legal characters in names
* Initialization in type declaration statements using `/values/`
* Saved variables without explicit or default initializers are zero initialized,
  except for scalar variables from the main program that are not explicitly
  initialized or marked with an explicit SAVE attribute (these variables may be
  placed on the stack by flang and not zero initialized). It is not advised to
  rely on this extension in new code.
* In a saved entity of a type with a default initializer, components without default
  values are zero initialized.
* Kind specification with `*`, e.g. `REAL*4`
* `DOUBLE COMPLEX` as a synonym for `COMPLEX(KIND(0.D0))` --
  but not when spelled `TYPE(DOUBLECOMPLEX)`.
* Signed complex literal constants
* DEC `STRUCTURE`, `RECORD`, with '%FILL'; but `UNION`, and `MAP`
  are not yet supported throughout compilation, and elicit a
  "not yet implemented" message.
* Structure field access with `.field`
* `BYTE` as synonym for `INTEGER(KIND=1)`; but not when spelled `TYPE(BYTE)`.
* When kind-param is used for REAL literals, allow a matching exponent letter
* Quad precision REAL literals with `Q`
* The `ISO_C_BINDING` module exports the named constants `c_float128` and
  `c_float128_complex` (both with the value 16), the kind parameter for
  128-bit (quad precision) real and complex C interoperable types. These
  extensions are gfortran-compatible.
* The `ISO_C_BINDING` module exports the named constants `c_float16` and
  `c_float16_complex` (both with the value 2), the kind parameter for
  16-bit (half precision) real and complex C interoperable types,
  corresponding to C `_Float16` (ISO/IEC TS 18661-3). `REAL(KIND=2)` and
  `COMPLEX(KIND=2)` are accordingly accepted as interoperable types. These
  extensions are gfortran-compatible.
* `X` prefix/suffix as synonym for `Z` on hexadecimal literals
* `B`, `O`, `Z`, and `X` accepted as suffixes as well as prefixes
* Support for using bare `L` in FORMAT statement
* Triplets allowed in array constructors
* `%LOC`, `%VAL`, and `%REF`
* Leading comma allowed before I/O item list
* Empty parentheses allowed in `PROGRAM P()`
* Missing parentheses allowed in `FUNCTION F`
* Cray based `POINTER(p,x)` and `LOC()` intrinsic (with `%LOC()` as
  an alias)
* Arithmetic `IF`.  (Which branch should NaN take? Fall through?)
* `ASSIGN` statement, assigned `GO TO`, and assigned format
* `PAUSE` statement
* Hollerith literals and edit descriptors
* Binary logical edit descriptor B (1/0 vs T/F)
* `NAMELIST` allowed in the execution part
* Omitted colons on type declaration statements with attributes
* COMPLEX constructor expression, e.g. `(x+y,z)`
* `+` and `-` before all primary expressions, e.g. `x*-y`
* `.NOT. .NOT.` accepted
* `NAME=` as synonym for `FILE=`
* Data edit descriptors without width or other details
* `D` lines in fixed form as comments or debug code
* `CARRIAGECONTROL=` on the OPEN and INQUIRE statements
* `CONVERT=` on the OPEN and INQUIRE statements
* `DISPOSE=` on the OPEN and INQUIRE statements
* Leading semicolons are ignored before any statement that
  could have a label
* The character `&` in column 1 in fixed form source is a variant form
  of continuation line.
* Character literals as elements of an array constructor without an explicit
  type specifier need not have the same length; the longest literal determines
  the length parameter of the implicit type, not the first.
* Outside a character literal, a comment after a continuation marker (&)
  need not begin with a comment marker (!).
* Classic C-style `/*comments*/` are skipped, so multi-language header
  files are easier to write and use. In fixed source form label fields, C
  comments are skipped only when preprocessing is enabled. Otherwise, valid
  Fortran programs could be rejected. For example:
```fortran
      x = x
     /* 2                           ! fixed-form continuation line
      print *, x, 'tail */ text'
```
* $ and \ edit descriptors are supported in FORMAT to suppress newline
  output on user prompts.
* Tabs in format strings (not `FORMAT` statements) are allowed on output.
* REAL and DOUBLE PRECISION variable and bounds in DO loops
* Integer literals without explicit kind specifiers that are out of range
  for the default kind of INTEGER are assumed to have the least larger kind
  that can hold them, if one exists.
* BOZ literals can be used as INTEGER values in contexts where the type is
  unambiguous: the right hand sides of assignments and initializations
  of INTEGER entities, as actual arguments to a few intrinsic functions
  (ACHAR, BTEST, CHAR), and as actual arguments of references to
  procedures with explicit interfaces whose corresponding dummy
  argument has a numeric type to which the BOZ literal may be
  converted.  BOZ literals are interpreted as default INTEGER only
  when they appear as the first items of array constructors with no
  explicit type.  Otherwise, they generally cannot be used if the type would
  not be known (e.g., `IAND(X'1',X'2')`, or as arguments of `DIM`, `MOD`,
  `MODULO`, and `SIGN`. Note that while other compilers may accept such usages,
  the type resolution of such BOZ literals usages is highly non portable).
  A warning is emitted when the BOZ literal is too large for the target.
* BOZ literals can also be used as REAL values in some contexts where the
  type is unambiguous, such as initializations of REAL parameters.
* `TRANSFER(boz, MOLD=integer or real scalar)` is accepted as an alternate
  spelling of `INT(boz, KIND=kind(mold))` or `REAL(boz, KIND=kind(mold))`.
* EQUIVALENCE of numeric and character sequences (a ubiquitous extension),
  as well as of sequences of non-default kinds of numeric types
  with each other.
* Values for whole anonymous parent components in structure constructors
  (e.g., `EXTENDEDTYPE(PARENTTYPE(1,2,3))` rather than `EXTENDEDTYPE(1,2,3)`
   or `EXTENDEDTYPE(PARENTTYPE=PARENTTYPE(1,2,3))`).
* Some intrinsic functions are specified in the standard as requiring the
  same type and kind for their arguments (viz., ATAN with two arguments,
  ATAN2, DIM, HYPOT, IAND, IEOR, IOR, MAX, MIN, MOD, and MODULO);
  we allow distinct types to be used, promoting
  the arguments as if they were operands to an intrinsic `+` operator,
  and defining the result type accordingly.
* DOUBLE COMPLEX intrinsics DREAL, DCMPLX, DCONJG, and DIMAG.
* The DFLOAT intrinsic function.
* INT_PTR_KIND intrinsic returns the kind of c_intptr_t.
* Restricted specific conversion intrinsics FLOAT, SNGL, IDINT, IFIX, DREAL,
  and DCMPLX accept arguments of any kind instead of only the default kind or
  double precision kind. Their result kinds remain as specified.
* Specific intrinsics AMAX0, AMAX1, AMIN0, AMIN1, DMAX1, DMIN1, MAX0, MAX1,
  MIN0, and MIN1 accept more argument types than specified. They are replaced by
  the related generics followed by conversions to the specified result types.
* When a scalar CHARACTER actual argument of the same kind is known to
  have a length shorter than the associated dummy argument, it is extended
  on the right with blanks, similar to assignment.
* When a dummy argument is `POINTER` or `ALLOCATABLE` and is `INTENT(IN)`, we
  relax enforcement of some requirements on actual arguments that must otherwise
  hold true for definable arguments.
* We allow a limited polymorphic `POINTER` or `ALLOCATABLE` actual argument
  to be associated with a compatible monomorphic dummy argument, as
  our implementation, like others, supports a reallocation that would
  change the dynamic type
* Assignment of `LOGICAL` to `INTEGER` and vice versa (but not other types) is
  allowed.  The values are normalized to canonical `.TRUE.`/`.FALSE.`.
  The values are also normalized for assignments of `LOGICAL(KIND=K1)` to
  `LOGICAL(KIND=K2)`, when `K1 != K2`.
* Static initialization of `LOGICAL` with `INTEGER` is allowed in `DATA` statements
  and object initializers.
  The results are *not* normalized to canonical `.TRUE.`/`.FALSE.`.
  Static initialization of `INTEGER` with `LOGICAL` is also permitted.
* An effectively empty source file (no program unit) is accepted and
  produces an empty relocatable output file.
* A `RETURN` statement may appear in a main program.
* DATA statement initialization is allowed for procedure pointers outside
  structure constructors.
* Nonstandard intrinsic functions: ISNAN, SIZEOF
* A forward reference to an INTEGER dummy argument is permitted to appear
  in a specification expression, such as an array bound, in a scope with
  IMPLICIT NONE(TYPE).
* A forward reference to a default INTEGER scalar
  `COMMON` block variable is permitted to appear in a specification
  expression, such as an array bound, in a scope with IMPLICIT NONE(TYPE)
  if the name of the variable would have caused it to be implicitly typed
  as default INTEGER if IMPLICIT NONE(TYPE) were absent.
* A named constant defined by a `PARAMETER` statement is permitted to appear
  before its explicit type declaration in a scope with IMPLICIT NONE(TYPE);
  it acquires the type it would have had under implicit typing rules (F2023 8.7),
  and a later explicit declaration must specify that same type (F2023 8.6.11 p2).
* An inquiry function whose result depends only on the declared type, kind
  type parameters, or rank of its argument, such as `KIND` or `DIGITS`, can
  appear in a constant expression when that argument is a conditional
  argument, as in `integer, parameter :: k = kind((flag ? a : b))`.
  F2023 10.1.12 item (6) admits only a constant expression or a variable as
  the argument of such a specification inquiry, and a conditional argument
  is neither. All consequent-args have the same declared type and kind type
  parameters (C1538) and, unless all are assumed-rank, the same rank
  (C1539), so the result doesn't depend on the condition. Flang accepts
  `kind(a + b)` in a constant expression on the same terms.
* OPEN(ACCESS='APPEND') is interpreted as OPEN(POSITION='APPEND')
  to ease porting from Sun Fortran.
* Intrinsic subroutines EXIT([status]) and ABORT()
* The definition of simple contiguity in 9.5.4 applies only to arrays;
  we also treat scalars as being trivially contiguous, so that they
  can be used in contexts like data targets in pointer assignments
  with bounds remapping.
* The `CONTIGUOUS` attribute can be redundantly applied to simply
  contiguous objects, including scalars, with a portability warning.
* `IS_CONTIGUOUS` is always true when its argument is a named constant
  (`PARAMETER`) or a subobject of a named constant, even when the standard's
  rules in F2023 8.5.7 would otherwise classify the subobject as non-contiguous
  (for example, a strided section like `a(::2)`).  This happens because Flang
  constant-folds an array section of a named constant into a fresh constant array
  containing only the selected elements, and that fresh array is trivially
  contiguous.  When `-pedantic` is enabled, Flang emits the
  `-Wconstant-is-contiguous` warning at each such call.  For example:
```fortran
integer, parameter :: a(5) = [1,2,3,4,5]
print *, is_contiguous(a(::2))                   ! prints T in Flang
```
  Other compilers may report `a(::2)` as non-contiguous.
* We support some combinations of specific procedures in generic
  interfaces that a strict reading of the standard would preclude
  when their calls must nonetheless be distinguishable.
  Specifically, `ALLOCATABLE` dummy arguments are distinguishing
  if an actual argument acceptable to one could not be passed to
  the other & vice versa because exactly one is polymorphic or
  exactly one is unlimited polymorphic).
* External unit 0 is predefined and connected to the standard error output,
  and defined as `ERROR_UNIT` in the intrinsic `ISO_FORTRAN_ENV` module.
* Objects in blank COMMON may be initialized.
* Initialization of COMMON blocks outside of BLOCK DATA subprograms.
* A named COMMON block may be redundantly initialized (via `DATA`
  statements or declaration initializers) in more than one program
  unit, with a portability warning, provided that every appearance
  that initializes the block does so identically: the same members
  are initialized to the same values everywhere the block appears. A
  first initialized appearance that leaves some members uninitialized
  while a later appearance initializes them (or vice versa) is a
  conflict, not a duplicate, and remains a hard error, as does any
  appearance that initializes a shared member to a different value.
  A member that is only indirectly initialized via an object
  equivalenced with it, rather than directly by a `DATA` statement or
  declaration initializer, is conservatively treated as a conflict at
  every appearance, since the equivalenced objects are not compared
  for agreement.
* Multiple specifications of the SAVE attribute on the same object
  are allowed, with a warning.
* Specific intrinsic functions BABS, IIABS, JIABS, KIABS, ZABS, and CDABS.
* A `POINTER` component's type need not be a sequence type when
  the component appears in a derived type with `SEQUENCE`.
  (This case should probably be an exception to constraint C740 in
  the standard.)
* Format expressions that have type but are not character and not
  integer scalars are accepted so long as they are simply contiguous.
  This legacy extension supports pre-Fortran'77 usage in which
  variables initialized in DATA statements with Hollerith literals
  as modifiable formats.
* At runtime, `NAMELIST` input will skip over `NAMELIST` groups
  with other names, and will treat text before and between groups
  as if they were comment lines, even if not begun with `!`.
* Commas are required in FORMAT statements and character variables
  only when they prevent ambiguity.
* Legacy names `AND`, `OR`, and `XOR` are accepted as aliases for
  the standard intrinsic functions `IAND`, `IOR`, and `IEOR`
  respectively.
* A digit count of d=0 is accepted in Ew.0, Dw.0, and Gw.0 output
  editing if no nonzero scale factor (kP) is in effect.
* The name `IMAG` is accepted as an alias for the generic intrinsic
  function `AIMAG`.
* The legacy extension intrinsic functions `IZEXT` and `JZEXT`
  are supported; `ZEXT` has different behavior with various older
  compilers, so it is not supported.
* Flang doesn't impose a limit on the number of continuation lines
  allowed for a single statement.
* When a type-bound procedure declaration statement has neither interface
  nor attributes, the "::" before the bindings is optional, even
  if a binding has renaming with "=> proc".
  The colons are not necessary for an unambiguous parse, C768
  notwithstanding.
* A type-bound procedure binding can be passed as an actual
  argument corresponding to a dummy procedure and can be used as
  the target of a procedure pointer assignment statement.
* An explicit `INTERFACE` can declare the interface of a
  procedure pointer even if it is not a dummy argument.
* A `NOPASS` type-bound procedure binding is required by C1529
  to apply only to a scalar data-ref, but most compilers don't
  enforce it and the constraint is not necessary for a correct
  implementation.
* A label may follow a semicolon in fixed form source.
* A logical dummy argument to a `BIND(C)` procedure, or a logical
  component to a `BIND(C)` derived type does not have to have
  `KIND=C_BOOL` since it can be converted to/from `_Bool` without
  loss of information.
* The character length of the `SOURCE=` or `MOLD=` in `ALLOCATE`
  may be distinct from the constant character length, if any,
  of an allocated object.
* When a name is brought into a scope by multiple ways,
  such as USE-association as well as an `IMPORT` from its host,
  it's an error only if the resolution is ambiguous.
* When USE association brings an equivalent external procedure interface
  and a legacy BLAS generic from the intrinsic `cublas` or `cublas_v2`
  module into the same scope under the same local name, Flang selects
  the intrinsic module's generic, regardless of the order of the two USE
  statements. A generic already merged from both `cublas` and `cublas_v2`
  is not covered, including when it is re-exported by another module.
  The generic must contain a same-named host specific with characteristics
  equal to those of the external interface, and at least one specific
  with a CUDA dummy data attribute.
  Generic names beginning with `cublas` are excluded.
  A warning is emitted by default, including with `-pedantic`.
  Use `-Wno-prefer-intrinsic-module-use-association` to suppress the warning,
  or `-fno-prefer-intrinsic-module-use-association` to disable the extension.
  This extension is deprecated and may be removed at any time; it was added
  to support BerkeleyGW.
* An entity may appear in a `DATA` statement before its explicit
  type declaration under `IMPLICIT NONE(TYPE)`.
* `INCLUDE` lines can start in any column, can be preceded in
  fixed form source by a '0' in column 6, can contain spaces
  between the letters of the word INCLUDE, and can have a
  numeric character literal kind prefix on the file name.
* Intrinsic procedures SIND, COSD, TAND and ATAND. Constant folding
  is currently not supported for these procedures but this is planned.
* When a pair of quotation marks in a character literal are split
  by a line continuation in free form, the second quotation mark
  may appear at the beginning of the continuation line without an
  ampersand, althought one is required by the standard.
* Unrestricted `INTRINSIC` functions are accepted for use in
  `PROCEDURE` statements in generic interfaces, as in some other
  compilers.
* A `NULL()` pointer is treated as an unallocated allocatable
  when associated with an `INTENT(IN)` allocatable dummy argument.
* `READ(..., SIZE=n)` is accepted with `NML=` and `FMT=*` with
  a portability warning.
  The Fortran standard doesn't allow `SIZE=` with formatted input
  modes that might require look-ahead, perhaps to ease implementations.
* When a file included via an `INCLUDE` line or `#include` directive
  has a continuation marker at the end of its last line in free form,
  Fortran line continuation works.
* A `NAMELIST` input group may omit its trailing `/` character if
  it is followed by another `NAMELIST` input group.
* A `NAMELIST` input group may begin with either `&` or `$`.
* In `NAMELIST` input, an assignment to a scalar item may omit its
  value (e.g. `l=`, immediately followed by the next name-value pair,
  the group terminator, or end-of-record).  F2023 13.11.2 p1 requires
  one or more values to follow the `=`, but classic nvfortran and
  gfortran accept the empty form and leave the item's current value
  unchanged.  Flang follows the same convention.  For example, given
  a namelist group `nml` with a `LOGICAL` scalar `l`, an `INTEGER`
  scalar `i_count`, and a `REAL` scalar `r_value`, the input record
  ```
  &nml l= i_count=7 r_value=2.72/
  ```
  leaves `l` unchanged and assigns `7` and `2.72` to `i_count` and
  `r_value` respectively.  Without this extension, the runtime would
  abort with `Bad character 'i' in LOGICAL input field`.
* In `NAMELIST` input, a `!` character is accepted as terminating the
  current value and introducing a comment even when it is not preceded
  by a value separator.  For example, `name=0.01!comment` is accepted
  as if it had been written `name=0.01 !comment`.  F2023 13.11.3.6 p.1
  requires a value separator (blank, comma, slash, or end of record)
  before a `!` comment introducer in namelist input, but gfortran,
  ifx, and classic nvfortran all accept this form as a widely-used
  extension.  Inside a character literal constant the `!` is taken
  literally as required by the standard.
* A comma (or semicolon in `DECIMAL='COMMA'` or `DC` mode) in a
  fixed-width numeric input field terminates the field rather than
  signaling an invalid character error.
* Arguments to the intrinsic functions `MAX` and `MIN` are converted
  when necessary to the type of the result.
  An `OPTIONAL`, `POINTER`, or `ALLOCATABLE` argument after
  the first two cannot be converted, as it may not be present.
* A derived type that meets (most of) the requirements of an interoperable
  derived type can be used as such where an interoperable type is
  required, with warnings, even if it lacks the BIND(C) attribute.
* A "mult-operand" in an expression can be preceded by a unary
  `+` or `-` operator.
* `BIND(C, NAME="...", CDEFINED)` signifies that the storage for an
  interoperable variable will be allocated outside of Fortran,
  probably by a C or C++ external definition.
* An automatic data object may be declared in the specification part
  of the main program.
* A local data object may appear in a specification expression, even
  when it is not a dummy argument or in COMMON, so long as it is
  has the SAVE attribute and was initialized.
* `PRINT namelistname` is accepted and interpreted as
  `WRITE(*,NML=namelistname)`, a near-universal extension.
* A character length specifier in a component or entity declaration
  is accepted before an array specification (`ch*3(2)`) as well
  as afterwards.
* A zero field width is allowed for logical formatted output (`L0`).
* `OPEN(..., FORM='BINARY')` is accepted as a legacy synonym for
  the standard `OPEN(..., FORM='UNFORMATTED', ACCESS='STREAM')`.
* A character string edit descriptor is allowed in an input format
  with an optional compilation-time warning.  When executed, it
  is treated as an 'nX' positioning control descriptor that skips
  over the same number of characters, without comparison.
* A passed-object dummy argument for a procedure binding is allowed
  to be a pointer so long as it is `INTENT(IN)`.
  (This extension is not yet supported for procedure pointer component
  interfaces.)
* A data object can be initialized multiple times by `DATA` statements
  and default component initialization, but only when all initializations
  are to the same value.  Distinct initializations remain errors.
* A pointer component that has no default initialization or explicit value
  in a structure constructor is defaulted to `NULL()`.
* Multiple specifications of a prefix-spec on the same procedure are allowed,
  with a warning.  C1552 (F2023) specifies that at most one of each shall be
  present.
* An assumed-rank entity is an acceptable `NAMELIST` group item.
* A named constant (`PARAMETER`) may appear as a `namelist-group-object` in a
  `NAMELIST` statement.  The Fortran standard requires namelist group objects
  to be variables, but this usage is accepted by Flang as an extension.
  When `-pedantic` is enabled, Flang emits a warning for this case.
  For example:
```
program p
  implicit none
  integer, parameter :: k = 3
  namelist /g/ k
end program
```
* When the argument to intrinsic `ALLOCATED(p)` is actually a pointer
  rather than an allocatable, it is interpreted as `ASSOCIATED(p)` with a
  stern warning.

### `SIMPLE` classification of extensions

The following extensions are `SIMPLE`: `IARGC`, `INT2`, `INT8`, `INT_PTR_KIND`, `ISNAN`,
`IZEXT`, `JZEXT`, `LNBLNK`, `LOC`, `SELECTED_UNSIGNED_KIND`, `SIZEOF`, `UINT`, `UMASKL`,
and `UMASKR`. All specific intrinsic extensions are also `SIMPLE`.

The following extensions are not classified as `SIMPLE`: `CHDIR`, `DSECNDS`, `ETIME`,
`FSEEK`, `FTELL`, `GETCWD`, `GETGID`, `GETPID`, `GETUID`, `HOSTNM`, `IRAND`, `MALLOC`,
`PUTENV`, `RAND`, `RENAME`, `RTC`, `SECNDS`, `SECOND`, `SYSTEM`, `TIME`, `TIMEF`, and `UNLINK`.

### Extensions supported when enabled by options

* C-style backslash escape sequences in quoted CHARACTER literals
  (but not Hollerith) [-fbackslash], including Unicode escapes
  with `\U`.
* Logical abbreviations `.T.`, `.F.`, `.N.`, `.A.`, `.O.`, and `.X.`
  [-flogical-abbreviations]
* `.XOR.` as a synonym for `.NEQV.` [-fxor-operator]
* The default `INTEGER` type is required by the standard to occupy
  the same amount of storage as the default `REAL` type.  Default
  `REAL` is of course 32-bit IEEE-754 floating-point today.  This legacy
  rule imposes an artificially small constraint in some cases
  where Fortran mandates that something have the default `INTEGER`
  type: specifically, the results of references to the intrinsic functions
  `SIZE`, `STORAGE_SIZE`,`LBOUND`, `UBOUND`, `SHAPE`, and the location reductions
  `FINDLOC`, `MAXLOC`, and `MINLOC` in the absence of an explicit
  `KIND=` actual argument.  We return `INTEGER(KIND=8)` by default in
  these cases when the `-flarge-sizes` option is enabled.
  `SIZEOF` and `C_SIZEOF` always return `INTEGER(KIND=8)`.
* Treat each specification-part like is has `IMPLICIT NONE`
  [-fimplicit-none-type-always]
* Ignore occurrences of `IMPLICIT NONE` and `IMPLICIT NONE(TYPE)`
  [-fimplicit-none-type-never]
* Treat a subprogram in a submodule as if it had a missing `MODULE` prefix
  when its name matches a separate module procedure interface in an ancestor
  module [-fimplicit-module-prefix]. This extension is disabled by default
  because the unprefixed subprogram can instead be a conforming local
  procedure. Without this extension, `-pedantic` or `-Wportability` diagnoses
  a likely missing prefix without changing the program. When the extension
  is enabled, `-Wimplicit-module-prefix` or `-pedantic` reports each repaired
  prefix. Since the extension cannot distinguish a missing prefix from an
  intentionally local procedure with the same name as an ancestor interface,
  it can reject a conforming program when that interface is implemented in a
  different submodule. This behavior is compatible with gfortran. Only
  definitions in the current source are repaired; a module file keeps the
  interpretation chosen when it was compiled.
* Old-style `PARAMETER pi=3.14` statement without parentheses
  [-falternative-parameter-statement]
* `UNSIGNED` type (-funsigned)
* Default exponent of zero, e.g. `3.14159E`, on a READ from a
  fixed-width input field.  Includes the case with only an
  exponent letter for compatibility with other compilers.
* Allow a data object or function pointer as the `C_LOC()`
  argument (not just pointers/targets). The compiler will not
  reason about aliases created through non-target non-pointer
  arguments and code generated using such aliases may exhibit
  unexpected behavior. This is for compatibility with
  legacy code; legacy code should be updated to be correct.
  This could be removed at any time.
  Use `-Wrelaxed-c-loc-checks` (alongside `-frelaxed-c-loc-checks`) to
  enable a diagnostic warning for affected call sites.
  [-frelaxed-c-loc-checks]

### Extensions and legacy features deliberately not supported

* `.LG.` as synonym for `.NE.`
* `REDIMENSION`
* Allocatable `COMMON`
* Expressions in formats
* `ACCEPT` as synonym for `READ *`
* `TYPE` as synonym for `PRINT`
* `ARRAY` as synonym for `DIMENSION`
* `VIRTUAL` as synonym for `DIMENSION`
* `ENCODE` and `DECODE` as synonyms for internal I/O
* `IMPLICIT AUTOMATIC`, `IMPLICIT STATIC`
* Default exponent of zero, e.g. `3.14159E`, on a literal constant
* Characters in defined operators that are neither letters nor digits
* `B` suffix on unquoted octal constants
* `Z` prefix on unquoted hexadecimal constants (dangerous)
* `T` and `F` as abbreviations for `.TRUE.` and `.FALSE.` in DATA (PGI/XLF)
* Use of host FORMAT labels in internal subprograms (PGI-only feature)
* ALLOCATE(TYPE(derived)::...) as variant of correct ALLOCATE(derived::...) (PGI only)
* Defining an explicit interface for a subprogram within itself (PGI only)
* USE association of a procedure interface within that same procedure's definition
* NULL() as a structure constructor expression for an ALLOCATABLE component (PGI).
* Conversion of LOGICAL to INTEGER in expressions.
* Use of INTEGER data with the intrinsic logical operators `.NOT.`, `.AND.`, `.OR.`,
  and `.XOR.`.
* IF (integer expression) THEN ... END IF  (PGI/Intel)
* Comparison of LOGICAL with ==/.EQ. rather than .EQV. (also .NEQV.) (PGI/Intel)
* Procedure pointers in COMMON blocks (PGI/Intel)
* Underindexing multi-dimensional arrays (e.g., A(1) rather than A(1,1)) (PGI only)
* Legacy PGI `NCHARACTER` type and `NC` Kanji character literals
* Using non-integer expressions for array bounds (e.g., REAL A(3.14159)) (PGI/Intel)
* Mixing INTEGER types as operands to bit intrinsics (e.g., IAND); only two
  compilers support it, and they disagree on sign extension.
* Module & program names that conflict with an object inside the unit (PGI only).
* When the same name is brought into scope via USE association from
  multiple modules, the name must refer to a generic interface; PGI
  allows a name to be a procedure from one module and a generic interface
  from another.
  Flang supports the limited intrinsic CUBLAS exception described above.
* Type parameter declarations must come first in a derived type definition;
  some compilers allow them to follow `PRIVATE`, or be intermixed with the
  component declarations.
* Wrong argument types in calls to specific intrinsics that have different names than the
  related generics. Some accepted exceptions are listed above in the allowed extensions.
  PGI, Intel, and XLF support this in ways that are not numerically equivalent.
  PGI converts the arguments while Intel and XLF replace the specific by the related generic.
* VMS listing control directives (`%LIST`, `%NOLIST`, `%EJECT`)
* Continuation lines on `INCLUDE` lines
* `NULL()` actual argument corresponding to an `ALLOCATABLE` dummy data object
* User (non-intrinsic) `ELEMENTAL` procedures may not be passed as actual
  arguments, in accordance with the standard; some Fortran compilers
  permit such usage.
* Constraint C1406, which prohibits the same module name from being used
  in a scope for both an intrinsic and a non-intrinsic module, is implemented
  as a portability warning only, not a hard error.
* IBM @PROCESS directive is accepted but ignored.

## Preprocessing behavior

* The preprocessor is always run, whatever the filename extension may be.
* We respect Fortran comments in macro actual arguments (like GNU, Intel, NAG;
  unlike PGI and XLF) on the principle that macro calls should be treated
  like function references.  Fortran's line continuation methods also work.
* We implement the `__COUNTER__` preprocessing extension,
  see [Non-standard Extensions](Preprocessing.md#non-standard-extensions)

## Standard features not silently accepted

* Fortran explicitly ignores type declaration statements when they
  attempt to type the name of a generic intrinsic function (8.2 p3).
  One can declare `CHARACTER::COS` and still get a real result
  from `COS(3.14159)`, for example.  Flang will complain when a
  generic intrinsic function's inferred result type does not
  match an explicit declaration.  This message is a warning.

* There is no restriction in the standard against assigning
  to a whole polymorphic allocatable under control of a `WHERE`
  construct or statement, but there is no good portable
  behavior to implement and the standard isn't entirely clear
  what it should mean.
  (Other compilers allow it, but the results are never meaningful;
  some never change the type, some change the type according to
  the value of the last mask element, some treat these
  assignment statements as no-ops, and the rest crash during compilation.)
  The compiler flags this case as an error.

* F2023 12.6.3 restricts enumeration types in I/O only for list-directed
  transfers (prohibited) and formatted transfers (which must use an `I`, `B`,
  `O`, or `Z` edit descriptor); it places no restriction on unformatted I/O.
  Flang is currently stricter than the standard here and rejects an
  enumeration type -- whether a bare item or reached as a component of a
  derived type not processed by defined I/O -- in unformatted I/O with an
  error.  This can be a temporary flang limitation while enumeration-type
  support is incomplete, not a standard requirement.

## Standard features that might as well not be

* Flang supports designators with constant expressions, properly
  constrained, as initial data targets for data pointers in
  initializers of variable and component declarations and in
  `DATA` statements; e.g., `REAL, POINTER :: P => T(1:10:2)`.
  This Fortran 2008 feature might as well be viewed like an
  extension; no other compiler that we've tested can handle
  it yet.
* According to 11.1.3.3p1, if a selector of an `ASSOCIATE` or
  related construct is defined by a variable, it has the `TARGET`
  attribute if the variable was a `POINTER` or `TARGET`.
  We read this to include the case of the variable being a
  pointer-valued function reference.
  No other Fortran compiler seems to handle this correctly for
  `ASSOCIATE`, though NAG gets it right for `SELECT TYPE`.
* The standard doesn't explicitly require that a named constant that
  appears as part of a complex-literal-constant be a scalar, but
  most compilers emit an error when an array appears.
  Flang supports them with a portability warning.
* Flang does not enforce a blanket prohibition against generic
  interfaces containing a mixture of functions and subroutines.
  We allow both to appear, unlike several other Fortran compilers.
  This is especially desirable when two generics of the same
  name are combined due to USE association and the mixture may
  be inadvertent.
* Since Fortran 90, `INCLUDE` lines have been allowed to have
  a numeric kind parameter prefix on the file name.  No other
  Fortran compiler supports them that I can find.
* A `SEQUENCE` derived type is required (F'2023 C745) to have
  at least one component.  No compiler enforces this constraint;
  this compiler emits a warning.
* Many compilers disallow a `VALUE` assumed-length character dummy
  argument, which has been standard since F'2008.
  We accept this usage with an optional portability warning.
* The `ASYNCHRONOUS` attribute can be implied by usage in data
  transfer I/O statements.  Only one other compiler supports this
  correctly.  This compiler does, apart from objects in asynchronous
  NAMELIST I/O, for which an actual asynchronous runtime implementation
  seems unlikely.

## Behavior in cases where the standard is ambiguous or indefinite

* When an inner procedure of a subprogram uses the value or an attribute
  of an undeclared name in a specification expression and that name does
  not appear in the host, it is not clear in the standard whether that
  name is an implicitly typed local variable of the inner procedure or a
  host association with an implicitly typed local variable of the host.
  For example:
```
module module
 contains
  subroutine host(j)
    ! Although "m" never appears in the specification or executable
    ! parts of this subroutine, both of its contained subroutines
    ! might be accessing it via host association.
    integer, intent(in out) :: j
    call inner1(j)
    call inner2(j)
   contains
    subroutine inner1(n)
      integer(kind(m)), intent(in) :: n
      m = n + 1
    end subroutine
    subroutine inner2(n)
      integer(kind(m)), intent(out) :: n
      n = m + 2
    end subroutine
  end subroutine
end module

program demo
  use module
  integer :: k
  k = 0
  call host(k)
  print *, k, " should be 3"
end

```

  Other Fortran compilers disagree in their interpretations of this example;
  some seem to treat the references to `m` as if they were host associations
  to an implicitly typed variable (and print `3`), while others seem to
  treat them as references to implicitly typed local variables, and
  load uninitialized values.

  In Flang, we chose to emit an error message for this case since the standard
  is unclear, the usage is not portable, and the issue can be easily resolved
  by adding a declaration.

* In subclause 7.5.6.2 of Fortran 2018 the standard defines a partial ordering
  of the final subroutine calls for finalizable objects, their non-parent
  components, and then their parent components.
  (The object is finalized, then the non-parent components of each element,
  and then the parent component.)
  Some have argued that the standard permits an implementation
  to finalize the parent component before finalizing an allocatable component in
  the context of deallocation, and the next revision of the language may codify
  this option.
  In the interest of avoiding needless confusion, this compiler implements what
  we believe to be the least surprising order of finalization.
  Specifically: all non-parent components are finalized before
  the parent, allocatable or not;
  all finalization takes place before any deallocation;
  and no object or subobject will be finalized more than once.

* When `RECL=` is set via the `OPEN` statement for a sequential formatted input
  file, it functions as an effective maximum record length.
  Longer records, if any, will appear as if they had been truncated to
  the value of `RECL=`.
  (Other compilers ignore `RECL=`, signal an error, or apply effective truncation
  to some forms of input in this situation.)
  For sequential formatted output, RECL= serves as a limit on record lengths
  that raises an error when it is exceeded.

* When a `DATA` statement in a `BLOCK` construct could be construed as
  either initializing a host-associated object or declaring a new local
  initialized object, Flang interprets the standard's classification of
  a `DATA` statement as being a "declaration" rather than a "specification"
  construct, and notes that the `BLOCK` construct is defined as localizing
  names that have specifications in the `BLOCK` construct.
  So this example will elicit an error about multiple initialization:
```
subroutine subr
  integer n = 1
  block
    data n/2/
  end block
end subroutine
```

  Other Fortran compilers disagree with each other in their interpretations
  of this example.
  The precedent among the most commonly used compilers
  agrees with Flang's interpretation: a `DATA` statement without any other
  specification of the name refers to the host-associated object.

* Many Fortran compilers allow a non-generic procedure to be `USE`-associated
  into a scope that also contains a generic interface of the same name
  but does not have the `USE`-associated non-generic procedure as a
  specific procedure.
```
module m1
 contains
  subroutine foo(n)
    integer, intent(in) :: n
  end subroutine
end module

module m2
  use m1, only: foo
  interface foo
    module procedure noargs
  end interface
 contains
  subroutine noargs
  end subroutine
end module
```

  This case elicits a warning from Flang, as it should not be treated
  any differently than the same case with the non-generic procedure of
  the same name being defined in the same scope rather than being
  `USE`-associated into it, which is explicitly non-conforming in the
  standard and not allowed by most other compilers.
  If the `USE`-associated entity of the same name is not a procedure,
  most compilers disallow it as well.

* Fortran 2018 19.3.4p1: "A component name has the scope of its derived-type
  definition.  Outside the type definition, it may also appear ..." which
  seems to imply that within its derived-type definition, a component
  name is in its scope, and at least shadows any entity of the same name
  in the enclosing scope and might be read, thanks to the "also", to mean
  that a "bare" reference to the name could be used in a specification inquiry.
  However, most other compilers do not allow a component to shadow exterior
  symbols, much less appear in specification inquiries, and there are
  application codes that expect exterior symbols whose names match
  components to be visible in a derived-type definition's default initialization
  expressions, and so Flang follows that precedent.

* 19.3.1p1 "Within its scope, a local identifier of an entity of class (1)
  or class (4) shall not be the same as a global identifier used in that scope..."
  is read so as to allow the name of a module, submodule, main program,
  or `BLOCK DATA` subprogram to also be the name of an local entity in its
  scope, with a portability warning, since that global name is not actually
  capable of being "used" in its scope.

* In the definition of the `ASSOCIATED` intrinsic function (16.9.16), its optional
  second argument `TARGET=` is required to be "allowable as the data-target or
  proc-target in a pointer assignment statement (10.2.2) in which POINTER is
  data-pointer-object or proc-pointer-object."  Some Fortran compilers
  interpret this to require that the first argument (`POINTER=`) be a valid
  left-hand side for a pointer assignment statement -- in particular, it
  cannot be `NULL()`, but also it is required to be modifiable.
  As there is  no good reason to disallow (say) an `INTENT(IN)` pointer here,
  or even `NULL()` as a well-defined case that is always `.FALSE.`,
  this compiler doesn't require the `POINTER=` argument to be a valid
  left-hand side for a pointer assignment statement, and we emit a
  portability warning when it is not.

* Flang allows a `USE` statement to reference a module that is defined later
  in the same compilation unit, so long as mutual dependencies do not form
  a cycle.
  This feature forestalls any risk of such a `USE` statement reading an
  obsolete module file from a previous compilation and then overwriting
  that file later.

* Flang allows `OPTIONAL` dummy arguments to interoperable procedures
  unless they are `VALUE` (C865).

* Flang processes the `NAMELIST` group declarations in a scope after it
  has resolved all of the names in that scope.  This means that names
  that appear before their local declarations do not resolve to host
  associated objects and do not elicit errors about improper redeclarations
  of implicitly typed entities.

* Standard Fortran allows forward references to derived types, which
  can lead to ambiguity when combined with host association.
  Some Fortran compilers resolve the type name to the host type,
  others to the forward-referenced local type; this compiler diagnoses
  an error.
```
module m
  type ambiguous; integer n; end type
 contains
  subroutine s
    type(ambiguous), pointer :: ptr
    type ambiguous; real a; end type
  end
end
```

* When an intrinsic procedure appears in the specification part of a module
  only in function references, but not an explicit `INTRINSIC` statement,
  its name is not brought into other scopes by a `USE` statement.

* The subclause on rounding in formatted I/O (13.7.2.3.8 in F'2023)
  only discusses rounding for decimal-to/from-binary conversions,
  omitting any mention of rounding for hexadecimal conversions.
  As other compilers do apply rounding, so does this one.

* For real `MAXVAL`, `MINVAL`, `MAXLOC`, and `MINLOC`, NaN values are
  essentially ignored unless there are some unmasked array entries and
  *all* of them are NaNs.

* When `INDEX` is used as an unrestricted specific intrinsic function
  in the context of an actual procedure, as the explicit interface in
  a `PROCEDURE` declaration statement, or as the target of a procedure
  pointer assignment, its interface has exactly two dummy arguments
  (`STRING=` and `SUBSTRING=`), and includes neither `BACK=` nor
  `KIND=`.
  This is how `INDEX` as an unrestricted specific intrinsic function was
  documented in FORTRAN '77 and Fortran '90; later revisions of the
  standard deleted the argument information from the section on
  unrestricted specific intrinsic functions.
  At least one other compiler (XLF) seems to expect that the interface for
  `INDEX` include an optional `BACK=` argument, but it doesn't actually
  work.

* Allocatable components of array and structure constructors are deallocated
  after use without calling final subroutines.
  The standard does not specify when and how deallocation of array and structure
  constructors allocatable components should happen. All compilers free the
  memory after use, but the behavior when the allocatable component is a derived
  type with finalization differ, especially when dealing with nested array and
  structure constructors expressions. Some compilers call final routine for the
  allocatable components of each constructor sub-expressions, some call it only
  for the allocatable component of the top level constructor, and some only
  deallocate the memory. Deallocating only the memory offers the most
  flexibility when lowering such expressions, and it is not clear finalization
  is desirable in such context (Fortran interop 1.6.2 in F2018 standards require
  array and structure constructors not to be finalized, so it also makes sense
  not to finalize their allocatable components when releasing their storage).

* F'2023 19.4 paragraph 5: "If integer-type-spec appears in data-implied-do or
  ac-implied-do-control it has the specified type and type parameters; otherwise
  it has the type and type parameters that it would have if it were the name of
  a variable in the innermost executable construct or scoping unit that includes
  the DATA statement or array constructor, and this type shall be integer type."
  Reading "would have if it were" as being the subjunctive, this would mean that
  an untyped implied DO index variable should be implicitly typed according to
  the rules active in the enclosing scope.  But all other Fortran compilers interpret
  the "would have if it were" as meaning "has if it is" -- i.e., if the name
  is visible in the enclosing scope, the type of that name is used as the
  type of the implied DO index.  So this is an error, not a simple application
  of the default implicit typing rule:
```
character j
print *, [(j,j=1,10)]
```

* The Fortran standard doesn't mention integer overflow explicitly. In many cases,
  however, integer overflow makes programs non-conforming.
  Flang follows other widely-used Fortran compilers. Specifically, Flang assumes
  integer overflow never occurs in address calculations and increment of
  do-variable unless the option `-fwrapv` is enabled.

* Two new ieee_round_type values were added in Flang beyond the four values
  defined in f03 and f08: ieee_away and ieee_other. Contemporary hardware
  typically does not have support for these rounding modes;
  ieee_support_rounding calls for these values return false.
  ieee_set_rounding_mode calls that attempt to set the rounding mode to one
  of these values in violation of the restriction in f23 clause 17.11.42 set
  the mode to ieee_nearest.

* Some compilers allow an `INTENT(OUT)` dummy argument's value to appear
  via host association in a specification expression.  A non-host-associated
  use is an error because an `INTENT(OUT)` dummy argument's value is not
  defined.  The argument put forth to accept this usage in a `BLOCK` construct
  or inner procedure is that the language in 10.1.11 (specification expressions)
  allows any host-associated object to appear, but that's unconvincing
  because it would also allow a host-associated `OPTIONAL` dummy argument to
  be used in a nested scope, and that doesn't make sense.  This compiler
  accepts an `INTENT(OUT)` non-`OPTIONAL` host-associated value to appear
  in a specification expression via host association with a portability
  warning since such values may have become defined by the time the nested
  expression's value is required.

* Intrinsic assignment of arrays is defined elementally, and intrinsic
  assignment of derived type components is defined componentwise.
  However, when intrinsic assignment takes place for an array of derived
  type, the order of the loop nesting is not defined.
  Some compilers will loop over the elements, assigning all of the components
  of each element before proceeding to the next element.
  This compiler loops over all of the components, and assigns all of
  the elements for each component before proceeding to the next component.
  A program using defined assignment might be able to detect the difference.

* The standard forbids instances of derived types with defined unformatted
  WRITE subroutines from appearing in the I/O list of an `INQUIRE(IOLENGTH=...)`
  statement.  It then also says that these defined I/O procedures should be
  ignored for that statement.  So we allow them to appear (like most
  compilers) and don't use any defined unformatted WRITE that might have been
  defined.

* Forward references to target objects are allowed to appear
  in the initializers of data pointer declarationss.
  Forward references to target objects are not accepted in the default
  initializers of derived type component declarations, however,
  since these default values need to be available to process incomplete
  structure constructors.

* When an `ALLOCATE` or `DEALLOCATE` statement with multiple variables
  has a `STAT=` specifier that allows the program to continue execution
  after an error, the variables after the one with the error are left
  deallocated (or allocated).  This interpretation allows the program to
  identify the variable that encountered the problem while avoiding any
  ambiguity in the case of multiple errors with distinct status codes.
  Some compilers work differently; for maximum portability, avoid
  `ALLOCATE` and `DEALLOCATE` statements with error recovery for
  multiple variables.

* When a "null" value is encountered in list-directed input, the
  corresponding effective item in the data list is left unchanged,
  even when it has a derived type with a defined `READ(FORMATTED)`
  subroutine.  This is the most literal reading of F'2023 13.10.3.2p2
  and the portable interpretation across the most common Fortran
  compilers.

* `NAMELIST` child input statements are allowed to advance to further
  input records.
  Further, advancement is allowed when the parent input statement is
  a non-child (top level) list-directed input statement, or, recursively,
  an intermediate child list-directed input statement that can advance.
  This means that non-`NAMELIST` list-directed child input statements are
  not allowed to advance when they have an ancestor formatted input statement
  that is not list-directed and there is no intervening `NAMELIST`.
  This design allows format-driven input with `DT` editing to retain
  control over advancement in child input, while otherwise allowing it.

* When output takes place to a file under `ACCESS="STREAM"` after
  repositioning it to an earlier position, some compilers will
  truncate the file; this behavior is similar to the implicit
  `ENDFILE` that takes place under sequential output after a
  `BACKSPACE` or `REWIND` statement.
  Truncation of streams is not specified in the standard, however,
  and it does not take place with all compilers.
  In this one, truncation is optional; it occurs by default,
  but it can be disabled via `FORT_TRUNCATE_STREAM=0` in the
  environment at execution time.

* Some expression errors, like out-of-range known subscript values,
  are noted only as warnings when they appear in code known to be
  dead anyway at compilation time.

* A reference with a constant subscript that is out of range is accepted
  with a warning rather than rejected with an error.  A subscript value is
  required to be within its bounds only when the reference is executed
  (F'2023 9.5.3.1 paragraph 2), so a reference that never runs does not
  render a program nonconforming; that case cannot be recognized in general
  -- consider a procedure whose only call site is in dead code, or one that
  is never called at all.  Note that the warning, not an error, is also what
  appears when the reference *is* executed.  The endpoints of array sections
  get the same treatment.  Cosubscripts do not: their requirement is F'2023
  9.6 paragraph 2 and a cosubscript list determines an image index, so an
  out-of-cobounds constant cosubscript remains an error.  Neither do an
  out-of-range subscript in a reference to a named constant array, an
  out-of-range `DATA` statement designator, or an out-of-range substring;
  those remain errors as well.
  Use `-fno-out-of-bounds-subscripts` to make these references errors again,
  or `-Wno-out-of-bounds-subscripts` to silence the warning entirely.
  Note that a module file compiled with the warning may produce errors in a
  dependent compilation that uses `-fno-out-of-bounds-subscripts`, since the
  interface is re-analyzed there; those errors point into the module file.

## Behavior in cases where the standard is clear but disputed

* Unless one uses `-ffast-math`, directly or by implication,
  the parentheses in real expressions like `A+(B*C)` will prevent
  fusing the multiplication with the addition using an "FMA"
  operation that only rounds the final result.
  F'2023 10.1.2.4 paragraph 2 reads "Once the interpretation of a
  numeric intrinsic operation is established, the processor may
  evaluate any mathematically equivalent expression, provided that
  the integrity of parentheses is not violated."
  This compiler honors that long-standing guarantee of the language,
  because subclause 10.1.8 ("Integrity of parentheses") requires that
  "any expression in parentheses shall be treated as a data entity".
  When the calculation of the value of that data entity yields
  `REAL` or `COMPLEX`, its kind is known, which implies rounding.
  So one needn't worry about default compilation modes breaking
  the parentheses in cases like `(X*X)-(Y*Y)` when they're meant
  to prevent inconsistent rounding from the use of FMA.
  (Parentheses prevent fusion in only 4 out of 7 other compilers,
  so this interpretation is not portable.)

## De Facto Standard Features

* `EXTENDS_TYPE_OF()` returns `.TRUE.` if both of its arguments have the
  same type, a case that is technically implementation-defined.

* `ENCODING=` is not in the list of changeable modes on an I/O unit,
  but every Fortran compiler allows the encoding to be changed on an
  open unit.

* A `NAMELIST` input item that references a scalar element of a vector
  or contiguous array can be used as the initial element of a storage
  sequence.  For example, "&GRP A(1)=1. 2. 3./" is treated as if had been
  "&GRP A(1:)=1. 2. 3./".
  This extension is necessarily disabled when the type of the array
  has an accessible defined formatted READ subroutine.