[AMDGPU] Make v8i16/v8f16 legal

[llvm-project.git] / flang / 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

```eval_rst
.. contents::
   :local:
```

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 is in milliseconds
  for kinds <= 4 and nanoseconds otherwise where the target system
  supports these rates.

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

* Tabs in source
* `<>` as synonym for `.NE.` and `/=`
* `$` and `@` as legal characters in names
* Initialization in type declaration statements using `/values/`
* Kind specification with `*`, e.g. `REAL*4`
* `DOUBLE COMPLEX`
* 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)`
* Quad precision REAL literals with `Q`
* `X` prefix/suffix as synonym for `Z` on hexadecimal literals
* `B`, `O`, `Z`, and `X` accepted as suffixes as well as prefixes
* 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
* `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.
* $ 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 assigments 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')`).
* BOZ literals can also be used as REAL values in some contexts where the
  type is unambiguous, such as initializations of REAL parameters.
* EQUIVALENCE of numeric and character sequences (a ubiquitous extension)
* 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, 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.
* Assignment of `LOGICAL` to `INTEGER` and vice versa (but not other types) is
  allowed.  The values are normalized.
* 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 a default INTEGER scalar dummy argument 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 dummy argument would have caused it to be implicitly typed
  as default INTEGER if IMPLICIT NONE(TYPE) were absent.
* 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.
* 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.
* 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 `!`.

### Extensions supported when enabled by options

* C-style backslash escape sequences in quoted CHARACTER literals
  (but not Hollerith) [-fbackslash]
* 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]
* Old-style `PARAMETER pi=3.14` statement without parentheses
  [-falternative-parameter-statement]

### 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`
* 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.
* IF (integer expression) THEN ... END IF  (PGI/Intel)
* Comparsion 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.
* 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.

## 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.

## 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.  f18 will complain when a
  generic intrinsic function's inferred result type does not
  match an explicit declaration.  This message is a warning.

## Standard features that might as well not be

* f18 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.

## 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 variabless, and
  load uninitialized values.

  In f18, 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.