LanguageExtensions.rst

  #pragma clang optimize off
  // This function will be decorated with optnone.
  int foo() {
    // ... code
  }

  // optnone conflicts with always_inline, so bar() will not be decorated.
  __attribute__((always_inline)) int bar() {
    // ... code
  }
  #pragma clang optimize on

If no ``on`` is found to close an ``off`` region, the end of the region is the
end of the compilation unit.

Note that a stray ``#pragma clang optimize on`` does not selectively enable
additional optimizations when compiling at low optimization levels. This feature
can only be used to selectively disable optimizations.

The pragma has an effect on functions only at the point of their definition; for
function templates, this means that the state of the pragma at the point of an
instantiation is not necessarily relevant. Consider the following example:

.. code-block:: c++

  template<typename T> T twice(T t) {
    return 2 * t;
  }

  #pragma clang optimize off
  template<typename T> T thrice(T t) {
    return 3 * t;
  }

  int container(int a, int b) {
    return twice(a) + thrice(b);
  }
  #pragma clang optimize on

In this example, the definition of the template function ``twice`` is outside
the pragma region, whereas the definition of ``thrice`` is inside the region.
The ``container`` function is also in the region and will not be optimized, but
it causes the instantiation of ``twice`` and ``thrice`` with an ``int`` type; of
these two instantiations, ``twice`` will be optimized (because its definition
was outside the region) and ``thrice`` will not be optimized.

Extensions for loop hint optimizations
======================================

The ``#pragma clang loop`` directive is used to specify hints for optimizing the
subsequent for, while, do-while, or c++11 range-based for loop. The directive
provides options for vectorization, interleaving, unrolling and
distribution. Loop hints can be specified before any loop and will be ignored if
the optimization is not safe to apply.

Vectorization and Interleaving
------------------------------

A vectorized loop performs multiple iterations of the original loop
in parallel using vector instructions. The instruction set of the target
processor determines which vector instructions are available and their vector
widths. This restricts the types of loops that can be vectorized. The vectorizer
automatically determines if the loop is safe and profitable to vectorize. A
vector instruction cost model is used to select the vector width.

Interleaving multiple loop iterations allows modern processors to further
improve instruction-level parallelism (ILP) using advanced hardware features,
such as multiple execution units and out-of-order execution. The vectorizer uses
a cost model that depends on the register pressure and generated code size to
select the interleaving count.

Vectorization is enabled by ``vectorize(enable)`` and interleaving is enabled
by ``interleave(enable)``. This is useful when compiling with ``-Os`` to
manually enable vectorization or interleaving.

.. code-block:: c++

  #pragma clang loop vectorize(enable)
  #pragma clang loop interleave(enable)
  for(...) {
    ...
  }

The vector width is specified by ``vectorize_width(_value_)`` and the interleave
count is specified by ``interleave_count(_value_)``, where
_value_ is a positive integer. This is useful for specifying the optimal
width/count of the set of target architectures supported by your application.

.. code-block:: c++

  #pragma clang loop vectorize_width(2)
  #pragma clang loop interleave_count(2)
  for(...) {
    ...
  }

Specifying a width/count of 1 disables the optimization, and is equivalent to
``vectorize(disable)`` or ``interleave(disable)``.

Loop Unrolling
--------------

Unrolling a loop reduces the loop control overhead and exposes more
opportunities for ILP. Loops can be fully or partially unrolled. Full unrolling
eliminates the loop and replaces it with an enumerated sequence of loop
iterations. Full unrolling is only possible if the loop trip count is known at
compile time. Partial unrolling replicates the loop body within the loop and
reduces the trip count.

If ``unroll(enable)`` is specified the unroller will attempt to fully unroll the
loop if the trip count is known at compile time. If the fully unrolled code size
is greater than an internal limit the loop will be partially unrolled up to this
limit. If the trip count is not known at compile time the loop will be partially
unrolled with a heuristically chosen unroll factor.

.. code-block:: c++

  #pragma clang loop unroll(enable)
  for(...) {
    ...
  }

If ``unroll(full)`` is specified the unroller will attempt to fully unroll the
loop if the trip count is known at compile time identically to
``unroll(enable)``. However, with ``unroll(full)`` the loop will not be unrolled
if the loop count is not known at compile time.

.. code-block:: c++

  #pragma clang loop unroll(full)
  for(...) {
    ...
  }

The unroll count can be specified explicitly with ``unroll_count(_value_)`` where
_value_ is a positive integer. If this value is greater than the trip count the
loop will be fully unrolled. Otherwise the loop is partially unrolled subject
to the same code size limit as with ``unroll(enable)``.

.. code-block:: c++

  #pragma clang loop unroll_count(8)
  for(...) {
    ...
  }

Unrolling of a loop can be prevented by specifying ``unroll(disable)``.

Loop Distribution
-----------------

Loop Distribution allows splitting a loop into multiple loops.  This is
beneficial for example when the entire loop cannot be vectorized but some of the
resulting loops can.

If ``distribute(enable))'' is specified and the loop has memory dependencies
that inhibit vectorization, the compiler will attempt to isolate the offending
operations into a new loop.  This optimization is not enabled by default, only
loops marked with the pragma are considered.

.. code-block:: c++

  #pragma clang loop distribute(enable)
  for (i = 0; i < N; ++i) {
    S1: A[i + 1] = A[i] + B[i];
    S2: C[i] = D[i] * E[i];
  }

This loop will be split into two loops between statements S1 and S2.  The
second loop containing S2 will be vectorized.

Loop Distribution is currently not enabled by default in the optimizer because
it can hurt performance in some cases.  For example, instruction-level
parallelism could be reduced by sequentializing the execution of the
statements S1 and S2 above.

If Loop Distribution is turned on globally with
``-mllvm -enable-loop-distribution``, specifying ``distribute(disable)`` can
be used the disable it on a per-loop basis.

Additional Information
----------------------

For convenience multiple loop hints can be specified on a single line.

.. code-block:: c++

  #pragma clang loop vectorize_width(4) interleave_count(8)
  for(...) {
    ...
  }

If an optimization cannot be applied any hints that apply to it will be ignored.
For example, the hint ``vectorize_width(4)`` is ignored if the loop is not
proven safe to vectorize. To identify and diagnose optimization issues use
`-Rpass`, `-Rpass-missed`, and `-Rpass-analysis` command line options. See the
user guide for details.