1 ==========================
2 Auto-Vectorization in LLVM
3 ==========================
5 LLVM has two vectorizers: The *Loop Vectorizer*, which operates on Loops,
6 and the *Basic Block Vectorizer*, which optimizes straight-line code. These
7 vectorizers focus on different optimization opportunities and use different
8 techniques. The BB vectorizer merges multiple scalars that are found in the
9 code into vectors while the Loop Vectorizer widens instructions in the
10 original loop to operate on multiple consecutive loop iterations.
15 LLVM’s Loop Vectorizer is now available and will be useful for many people.
16 It is not enabled by default, but can be enabled through clang using the
19 .. code-block:: console
21 $ clang -fvectorize -O3 file.c
23 If the ``-fvectorize`` flag is used then the loop vectorizer will be enabled
24 when running with ``-O3``, ``-O2``. When ``-Os`` is used, the loop vectorizer
25 will only vectorize loops that do not require a major increase in code size.
27 We plan to enable the Loop Vectorizer by default as part of the LLVM 3.3 release.
32 The LLVM Loop Vectorizer has a number of features that allow it to vectorize
35 Loops with unknown trip count
36 ------------------------------
38 The Loop Vectorizer supports loops with an unknown trip count.
39 In the loop below, the iteration ``start`` and ``finish`` points are unknown,
40 and the Loop Vectorizer has a mechanism to vectorize loops that do not start
41 at zero. In this example, ‘n’ may not be a multiple of the vector width, and
42 the vectorizer has to execute the last few iterations as scalar code. Keeping
43 a scalar copy of the loop increases the code size.
47 void bar(float *A, float* B, float K, int start, int end) {
48 for (int i = start; i < end; ++i)
52 Runtime Checks of Pointers
53 --------------------------
55 In the example below, if the pointers A and B point to consecutive addresses,
56 then it is illegal to vectorize the code because some elements of A will be
57 written before they are read from array B.
59 Some programmers use the 'restrict' keyword to notify the compiler that the
60 pointers are disjointed, but in our example, the Loop Vectorizer has no way of
61 knowing that the pointers A and B are unique. The Loop Vectorizer handles this
62 loop by placing code that checks, at runtime, if the arrays A and B point to
63 disjointed memory locations. If arrays A and B overlap, then the scalar version
64 of the loop is executed.
68 void bar(float *A, float* B, float K, int n) {
69 for (int i = 0; i < n; ++i)
75 --------------------------
77 In this example the ``sum`` variable is used by consecutive iterations of
78 the loop. Normally, this would prevent vectorization, but the vectorizer can
79 detect that ‘sum’ is a reduction variable. The variable ‘sum’ becomes a vector
80 of integers, and at the end of the loop the elements of the array are added
81 together to create the correct result. We support a number of different
82 reduction operations, such as addition, multiplication, XOR, AND and OR.
86 int foo(int *A, int *B, int n) {
88 for (int i = 0; i < n; ++i)
94 --------------------------
96 In this example the value of the induction variable ``i`` is saved into an
97 array. The Loop Vectorizer knows to vectorize induction variables.
101 void bar(float *A, float* B, float K, int n) {
102 for (int i = 0; i < n; ++i)
107 --------------------------
109 The Loop Vectorizer is able to "flatten" the IF statement in the code and
110 generate a single stream of instructions. The Loop Vectorizer supports any
111 control flow in the innermost loop. The innermost loop may contain complex
112 nesting of IFs, ELSEs and even GOTOs.
116 int foo(int *A, int *B, int n) {
118 for (int i = 0; i < n; ++i)
124 Pointer Induction Variables
125 ---------------------------
127 This example uses the "accumulate" function of the standard c++ library. This
128 loop uses C++ iterators, which are pointers, and not integer indices.
129 The Loop Vectorizer detects pointer induction variables and can vectorize
130 this loop. This feature is important because many C++ programs use iterators.
134 int baz(int *A, int n) {
135 return std::accumulate(A, A + n, 0);
139 --------------------------
141 The Loop Vectorizer can vectorize loops that count backwards.
145 int foo(int *A, int *B, int n) {
146 for (int i = n; i > 0; --i)
153 The Loop Vectorizer can vectorize code that becomes scatter/gather
158 int foo(int *A, int *B, int n, int k) {
159 for (int i = 0; i < n; ++i)
163 Vectorization of Mixed Types
164 --------------------------
166 The Loop Vectorizer can vectorize programs with mixed types. The Vectorizer
167 cost model can estimate the cost of the type conversion and decide if
168 vectorization is profitable.
172 int foo(int *A, char *B, int n, int k) {
173 for (int i = 0; i < n; ++i)
177 Vectorization of function calls
178 --------------------------
180 The Loop Vectorize can vectorize intrinsic math functions.
181 See the table below for a list of these functions.
183 +-----+-----+---------+
185 +-----+-----+---------+
187 +-----+-----+---------+
188 | log |log2 | log10 |
189 +-----+-----+---------+
191 +-----+-----+---------+
192 |fma |trunc|nearbyint|
193 +-----+-----+---------+
198 This section shows the the execution time of Clang on a simple benchmark:
199 `gcc-loops <http://llvm.org/viewvc/llvm-project/test-suite/trunk/SingleSource/UnitTests/Vectorizer/>`_.
200 This benchmarks is a collection of loops from the GCC autovectorization
201 `page <http://gcc.gnu.org/projects/tree-ssa/vectorization.html>`_ by Dorit Nuzman.
203 The chart below compares GCC-4.7, ICC-13, and Clang-SVN at -O3, running on a Sandybridge.
204 The Y-axis shows time in msec. Lower is better.
206 .. image:: gcc-loops.png
208 The Basic Block Vectorizer
209 ==========================
211 The Basic Block Vectorizer is not enabled by default, but it can be enabled
212 through clang using the command line flag:
214 .. code-block:: console
216 $ clang -fslp-vectorize file.c
218 The goal of basic-block vectorization (a.k.a. superword-level parallelism) is
219 to combine similar independent instructions within simple control-flow regions
220 into vector instructions. Memory accesses, arithemetic operations, comparison
221 operations and some math functions can all be vectorized using this technique
222 (subject to the capabilities of the target architecture).
224 For example, the following function performs very similar operations on its
225 inputs (a1, b1) and (a2, b2). The basic-block vectorizer may combine these
226 into vector operations.
230 int foo(int a1, int a2, int b1, int b2) {
231 int r1 = a1*(a1 + b1)/b1 + 50*b1/a1;
232 int r2 = a2*(a2 + b2)/b2 + 50*b2/a2;