11 *ThinLTO* compilation is a new type of LTO that is both scalable and
12 incremental. *LTO* (Link Time Optimization) achieves better
13 runtime performance through whole-program analysis and cross-module
14 optimization. However, monolithic LTO implements this by merging all
15 input into a single module, which is not scalable
16 in time or memory, and also prevents fast incremental compiles.
18 In ThinLTO mode, as with regular LTO, clang emits LLVM bitcode after the
19 compile phase. The ThinLTO bitcode is augmented with a compact summary
20 of the module. During the link step, only the summaries are read and
21 merged into a combined summary index, which includes an index of function
22 locations for later cross-module function importing. Fast and efficient
23 whole-program analysis is then performed on the combined summary index.
25 However, all transformations, including function importing, occur
26 later when the modules are optimized in fully parallel backends.
27 By default, linkers_ that support ThinLTO are set up to launch
28 the ThinLTO backends in threads. So the usage model is not affected
29 as the distinction between the fast serial thin link step and the backends
30 is transparent to the user.
32 For more information on the ThinLTO design and current performance,
33 see the LLVM blog post `ThinLTO: Scalable and Incremental LTO
34 <http://blog.llvm.org/2016/06/thinlto-scalable-and-incremental-lto.html>`_.
35 While tuning is still in progress, results in the blog post show that
36 ThinLTO already performs well compared to LTO, in many cases matching
37 the performance improvement.
46 The 3.9 release of clang includes ThinLTO support. However, ThinLTO
47 is under active development, and new features, improvements and bugfixes
48 are being added for the next release. For the latest ThinLTO support,
49 `build a recent version of clang and LLVM
50 <https://llvm.org/docs/CMake.html>`_.
57 ThinLTO is currently supported for the following linkers:
59 - **gold (via the gold-plugin)**:
60 Similar to monolithic LTO, this requires using
61 a `gold linker configured with plugins enabled
62 <https://llvm.org/docs/GoldPlugin.html>`_.
64 Starting with `Xcode 8 <https://developer.apple.com/xcode/>`_.
66 Starting with r284050 for ELF, r298942 for COFF.
74 To utilize ThinLTO, simply add the -flto=thin option to compile and link. E.g.
76 .. code-block:: console
78 % clang -flto=thin -O2 file1.c file2.c -c
79 % clang -flto=thin -O2 file1.o file2.o -o a.out
81 When using lld-link, the -flto option need only be added to the compile step:
83 .. code-block:: console
85 % clang-cl -flto=thin -O2 -c file1.c file2.c
86 % lld-link /out:a.exe file1.obj file2.obj
88 As mentioned earlier, by default the linkers will launch the ThinLTO backend
89 threads in parallel, passing the resulting native object files back to the
90 linker for the final native link. As such, the usage model the same as
93 With gold, if you see an error during the link of the form:
95 .. code-block:: console
97 /usr/bin/ld: error: /path/to/clang/bin/../lib/LLVMgold.so: could not load plugin library: /path/to/clang/bin/../lib/LLVMgold.so: cannot open shared object file: No such file or directory
99 Then either gold was not configured with plugins enabled, or clang
100 was not built with ``-DLLVM_BINUTILS_INCDIR`` set properly. See
101 the instructions for the
102 `LLVM gold plugin <https://llvm.org/docs/GoldPlugin.html#how-to-build-it>`_.
104 Controlling Backend Parallelism
105 -------------------------------
108 By default, the ThinLTO link step will launch as many
109 threads in parallel as there are cores. If the number of
110 cores can't be computed for the architecture, then it will launch
111 ``std::thread::hardware_concurrency`` number of threads in parallel.
112 For machines with hyper-threading, this is the total number of
113 virtual cores. For some applications and machine configurations this
114 may be too aggressive, in which case the amount of parallelism can
115 be reduced to ``N`` via:
118 ``-Wl,-plugin-opt,jobs=N``
120 ``-Wl,-mllvm,-threads=N``
122 ``-Wl,--thinlto-jobs=N``
124 ``/opt:lldltojobs=N``
126 Other possible values for ``N`` are:
129 Use one thread per physical core (default)
131 Use a single thread only (disable multi-threading)
133 Use one thread per logical core (uses all hyper-threads)
139 ThinLTO supports fast incremental builds through the use of a cache,
140 which currently must be enabled through a linker option.
142 - gold (as of LLVM 4.0):
143 ``-Wl,-plugin-opt,cache-dir=/path/to/cache``
144 - ld64 (support in clang 3.9 and Xcode 8):
145 ``-Wl,-cache_path_lto,/path/to/cache``
146 - ELF lld (as of LLVM 5.0):
147 ``-Wl,--thinlto-cache-dir=/path/to/cache``
148 - COFF lld-link (as of LLVM 6.0):
149 ``/lldltocache:/path/to/cache``
154 To help keep the size of the cache under control, ThinLTO supports cache
155 pruning. Cache pruning is supported with gold, ld64 and ELF and COFF lld, but
156 currently only gold, ELF and COFF lld allow you to control the policy with a
157 policy string. The cache policy must be specified with a linker option.
159 - gold (as of LLVM 6.0):
160 ``-Wl,-plugin-opt,cache-policy=POLICY``
161 - ELF lld (as of LLVM 5.0):
162 ``-Wl,--thinlto-cache-policy,POLICY``
163 - COFF lld-link (as of LLVM 6.0):
164 ``/lldltocachepolicy:POLICY``
166 A policy string is a series of key-value pairs separated by ``:`` characters.
167 Possible key-value pairs are:
169 - ``cache_size=X%``: The maximum size for the cache directory is ``X`` percent
170 of the available space on the disk. Set to 100 to indicate no limit,
171 50 to indicate that the cache size will not be left over half the available
172 disk space. A value over 100 is invalid. A value of 0 disables the percentage
173 size-based pruning. The default is 75%.
175 - ``cache_size_bytes=X``, ``cache_size_bytes=Xk``, ``cache_size_bytes=Xm``,
176 ``cache_size_bytes=Xg``:
177 Sets the maximum size for the cache directory to ``X`` bytes (or KB, MB,
178 GB respectively). A value over the amount of available space on the disk
179 will be reduced to the amount of available space. A value of 0 disables
180 the byte size-based pruning. The default is no byte size-based pruning.
182 Note that ThinLTO will apply both size-based pruning policies simultaneously,
183 and changing one does not affect the other. For example, a policy of
184 ``cache_size_bytes=1g`` on its own will cause both the 1GB and default 75%
185 policies to be applied unless the default ``cache_size`` is overridden.
187 - ``cache_size_files=X``:
188 Set the maximum number of files in the cache directory. Set to 0 to indicate
189 no limit. The default is 1000000 files.
191 - ``prune_after=Xs``, ``prune_after=Xm``, ``prune_after=Xh``: Sets the
192 expiration time for cache files to ``X`` seconds (or minutes, hours
193 respectively). When a file hasn't been accessed for ``prune_after`` seconds,
194 it is removed from the cache. A value of 0 disables the expiration-based
195 pruning. The default is 1 week.
197 - ``prune_interval=Xs``, ``prune_interval=Xm``, ``prune_interval=Xh``:
198 Sets the pruning interval to ``X`` seconds (or minutes, hours
199 respectively). This is intended to be used to avoid scanning the directory
200 too often. It does not impact the decision of which files to prune. A
201 value of 0 forces the scan to occur. The default is every 20 minutes.
206 To `bootstrap clang/LLVM <https://llvm.org/docs/AdvancedBuilds.html#bootstrap-builds>`_
207 with ThinLTO, follow these steps:
209 1. The host compiler_ must be a version of clang that supports ThinLTO.
210 #. The host linker_ must support ThinLTO (and in the case of gold, must be
211 `configured with plugins enabled <https://llvm.org/docs/GoldPlugin.html>`_).
212 #. Use the following additional `CMake variables
213 <https://llvm.org/docs/CMake.html#options-and-variables>`_
214 when configuring the bootstrap compiler build:
216 * ``-DLLVM_ENABLE_LTO=Thin``
217 * ``-DCMAKE_C_COMPILER=/path/to/host/clang``
218 * ``-DCMAKE_CXX_COMPILER=/path/to/host/clang++``
219 * ``-DCMAKE_RANLIB=/path/to/host/llvm-ranlib``
220 * ``-DCMAKE_AR=/path/to/host/llvm-ar``
224 * ``-DLLVM_ENABLE_LTO=Thin``
225 * ``-DCMAKE_C_COMPILER=/path/to/host/clang-cl.exe``
226 * ``-DCMAKE_CXX_COMPILER=/path/to/host/clang-cl.exe``
227 * ``-DCMAKE_LINKER=/path/to/host/lld-link.exe``
228 * ``-DCMAKE_RANLIB=/path/to/host/llvm-ranlib.exe``
229 * ``-DCMAKE_AR=/path/to/host/llvm-ar.exe``
231 #. To use additional linker arguments for controlling the backend
232 parallelism_ or enabling incremental_ builds of the bootstrap compiler,
233 after configuring the build, modify the resulting CMakeCache.txt file in the
234 build directory. Specify any additional linker options after
235 ``CMAKE_EXE_LINKER_FLAGS:STRING=``. Note the configure may fail if
236 linker plugin options are instead specified directly in the previous step.
238 The ``BOOTSTRAP_LLVM_ENABLE_LTO=Thin`` will enable ThinLTO for stage 2 and
239 stage 3 in case the compiler used for stage 1 does not support the ThinLTO
245 * From LLVM project blog:
246 `ThinLTO: Scalable and Incremental LTO
247 <http://blog.llvm.org/2016/06/thinlto-scalable-and-incremental-lto.html>`_