Implement ISO/IEC JTC1 SC22 WG14 N1169 fixed-point types and arithmetic

[bevin-hansson]

Bugzilla Link	43896
Version	trunk
OS	All
CC	@bevin-hansson,@bjope,@PiJoules,@zygoloid

Extended Description

This is a ticket to track the remaining work required for supporting the ISO/IEC JTC1 SC22 WG14 N1169 fixed-point types and arithmetic.

Although this ticket is filed under Clang Frontend, the remaining work also includes patches to LLVM and other parts of Clang.

The currently estimated work that remains is:

LLVM:

• Intrinsic for fixed-point [saturating] division
• Possibly intrinsics for floating point/fixed-point conversion
• Possibly intrinsic for saturating left shift

Clang:

• Precision/limits macros (stdfix.h)
• Constant evaluation
  ** Subtraction
  ** Multiplication
  ** Division
  ** Shift
  ** Floating point conversion
• Compound assignment
• Codegen
  ** Multiplication
  ** Division
  ** Shift
  ** Unary operations
  ** Floating point conversion

[bjope]

Looking at the earlier "remaining work" list several things should have been implemented by now.

I sent out a status report to cfe-dev/llvm-dev on March 15, 2021:

Status regarding support for fixed point numbers (and Embedded-C) in Clang/LLVM

During the years there have been a number of questions on cfe-dev and llvm-dev
about support for fixed point arithmetic in LLVM.

In 2018 Leonard Chan wrote a proposal for adding fixed point number support in
clang [1], accompanied by discussions about how to support the types in LLVM
[2]. Since then a number of patches has been submitted and nowadays
Clang/LLVM has lots of support for fixed point arithmetic, based on the
Embedded-C draft [3].

A major part of chapter 4 in [3] should have been implemented by now.
Here is a list of features implemented:

• Fixed point types can be enabled using the -ffixed-point clang option.

• Variables can be declared using the _Fract and _Accum type specifiers (see
  for example clang/test/Frontend/fixed_point_declarations.c for examples).

• Saturated fixed point types are available by using the _Sat specifier
  for the _Fract and _Accum types.

• Support for conversion between signed/unsigned fixed-point types and
  integral or floating point types.

• Support for literals, using the following fixed-point suffixes:
  hr: short _Fract
  uhr: unsigned short _Fract
  r: _Fract
  ur: unsigned _Fract
  lr: long _Fract
  ulr: unsigned long _Fract
  hk: short _Accum
  uhk: unsigned short _Accum
  k: _Accum
  uk: unsigned _Accum
  lk: long _Accum
  ulk: unsigned long _Accum

• Support for operations involving fixed-point types:
  Unary (++, --, +, -, !)
  Binary (+, -, *, /, <<, >>)
  Comparison (<, <=, >=, >, ==, !=)
  Assignment / Compound Assignment (=, +=, -=, *=, /=, <<=, >>=)

Future improvements:

• Improve user documentation (the comment that _Fract types aren't supported
  has been removed from the UsersManual, but we haven't really added anything
  describing that we now support fixed point types).

• Support for format specifiers (r/R/k/K) for fixed-point arguments in I/O
  functions.

• Support for pragmas such as FX_FRACT_OVERFLOW and FX_ACCUM_OVERFLOW.
  (This might be a difference compared to what they have implemented in
  gcc, see [4].)

• The Embedded-C draft [3] mentions a stdfix.h library.
  (This has not been implemented. Also lacking in gcc according to [4].)

• Implement sanitizer checks for fixed point shifts. The LLVM IR for shift has
  UB semantics when shift count is too large. So either we need to clamp shift
  count when lowering shift operators (if there are special rules for
  embedded-c), or if it is UB on c level it would be nice if the sanitizer
  catch such faults. There are some TODO:s left in the code (see for example
  ScalarExprEmitter::EmitShl) indicating that the fixed point lowering
  currently bypass the code for sanitizing the shift count.

• Add some C++ compatibility. One thing that limits the fixed point support
  to C is lack of name mangling rules for fixed point types (at least according
  to the Itanium C++ ABI). One solution is to create a wrapper class, overriding
  various operators. Another idea is to add name mangling (there is a proposal
  for the Itanium C++ ABI here [5]).

References:

[1] - https://lists.llvm.org/pipermail/cfe-dev/2018-April/057756.html
[2] - https://lists.llvm.org/pipermail/llvm-dev/2018-August/125433.html
[3] - http://www.open-std.org/jtc1/sc22/wg14/www/docs/n1169.pdf
[4] - https://gcc.gnu.org/wiki/FixedPointArithmetic
[5] - itanium-cxx-abi/cxx-abi#56