3 ======================================================
4 LLVM Link Time Optimization: Design and Implementation
5 ======================================================
13 LLVM features powerful intermodular optimizations which can be used at link
14 time. Link Time Optimization (LTO) is another name for intermodular
15 optimization when performed during the link stage. This document describes the
16 interface and design between the LTO optimizer and the linker.
21 The LLVM Link Time Optimizer provides complete transparency, while doing
22 intermodular optimization, in the compiler tool chain. Its main goal is to let
23 the developer take advantage of intermodular optimizations without making any
24 significant changes to the developer's makefiles or build system. This is
25 achieved through tight integration with the linker. In this model, the linker
26 treates LLVM bitcode files like native object files and allows mixing and
27 matching among them. The linker uses `libLTO`_, a shared object, to handle LLVM
28 bitcode files. This tight integration between the linker and LLVM optimizer
29 helps to do optimizations that are not possible in other models. The linker
30 input allows the optimizer to avoid relying on conservative escape analysis.
34 Example of link time optimization
35 ---------------------------------
37 The following example illustrates the advantages of LTO's integrated approach
38 and clean interface. This example requires a system linker which supports LTO
39 through the interface described in this document. Here, clang transparently
40 invokes system linker.
42 * Input source file ``a.c`` is compiled into LLVM bitcode form.
43 * Input source file ``main.c`` is compiled into native object code.
48 extern int foo1(void);
49 extern void foo2(void);
50 extern void foo4(void);
55 static signed int i = 0;
91 % clang -emit-llvm -c a.c -o a.o # <-- a.o is LLVM bitcode file
92 % clang -c main.c -o main.o # <-- main.o is native object file
93 % clang a.o main.o -o main # <-- standard link command without modifications
95 * In this example, the linker recognizes that ``foo2()`` is an externally
96 visible symbol defined in LLVM bitcode file. The linker completes its usual
97 symbol resolution pass and finds that ``foo2()`` is not used
98 anywhere. This information is used by the LLVM optimizer and it
99 removes ``foo2()``.</li>
101 * As soon as ``foo2()`` is removed, the optimizer recognizes that condition ``i
102 < 0`` is always false, which means ``foo3()`` is never used. Hence, the
103 optimizer also removes ``foo3()``.
105 * And this in turn, enables linker to remove ``foo4()``.
107 This example illustrates the advantage of tight integration with the
108 linker. Here, the optimizer can not remove ``foo3()`` without the linker's
111 Alternative Approaches
112 ----------------------
114 **Compiler driver invokes link time optimizer separately.**
115 In this model the link time optimizer is not able to take advantage of
116 information collected during the linker's normal symbol resolution phase.
117 In the above example, the optimizer can not remove ``foo2()`` without the
118 linker's input because it is externally visible. This in turn prohibits the
119 optimizer from removing ``foo3()``.
121 **Use separate tool to collect symbol information from all object files.**
122 In this model, a new, separate, tool or library replicates the linker's
123 capability to collect information for link time optimization. Not only is
124 this code duplication difficult to justify, but it also has several other
125 disadvantages. For example, the linking semantics and the features provided
126 by the linker on various platform are not unique. This means, this new tool
127 needs to support all such features and platforms in one super tool or a
128 separate tool per platform is required. This increases maintenance cost for
129 link time optimizer significantly, which is not necessary. This approach
130 also requires staying synchronized with linker developements on various
131 platforms, which is not the main focus of the link time optimizer. Finally,
132 this approach increases end user's build time due to the duplication of work
133 done by this separate tool and the linker itself.
135 Multi-phase communication between ``libLTO`` and linker
136 =======================================================
138 The linker collects information about symbol defininitions and uses in various
139 link objects which is more accurate than any information collected by other
140 tools during typical build cycles. The linker collects this information by
141 looking at the definitions and uses of symbols in native .o files and using
142 symbol visibility information. The linker also uses user-supplied information,
143 such as a list of exported symbols. LLVM optimizer collects control flow
144 information, data flow information and knows much more about program structure
145 from the optimizer's point of view. Our goal is to take advantage of tight
146 integration between the linker and the optimizer by sharing this information
147 during various linking phases.
149 Phase 1 : Read LLVM Bitcode Files
150 ---------------------------------
152 The linker first reads all object files in natural order and collects symbol
153 information. This includes native object files as well as LLVM bitcode files.
154 To minimize the cost to the linker in the case that all .o files are native
155 object files, the linker only calls ``lto_module_create()`` when a supplied
156 object file is found to not be a native object file. If ``lto_module_create()``
157 returns that the file is an LLVM bitcode file, the linker then iterates over the
158 module using ``lto_module_get_symbol_name()`` and
159 ``lto_module_get_symbol_attribute()`` to get all symbols defined and referenced.
160 This information is added to the linker's global symbol table.
163 The lto* functions are all implemented in a shared object libLTO. This allows
164 the LLVM LTO code to be updated independently of the linker tool. On platforms
165 that support it, the shared object is lazily loaded.
167 Phase 2 : Symbol Resolution
168 ---------------------------
170 In this stage, the linker resolves symbols using global symbol table. It may
171 report undefined symbol errors, read archive members, replace weak symbols, etc.
172 The linker is able to do this seamlessly even though it does not know the exact
173 content of input LLVM bitcode files. If dead code stripping is enabled then the
174 linker collects the list of live symbols.
176 Phase 3 : Optimize Bitcode Files
177 --------------------------------
179 After symbol resolution, the linker tells the LTO shared object which symbols
180 are needed by native object files. In the example above, the linker reports
181 that only ``foo1()`` is used by native object files using
182 ``lto_codegen_add_must_preserve_symbol()``. Next the linker invokes the LLVM
183 optimizer and code generators using ``lto_codegen_compile()`` which returns a
184 native object file creating by merging the LLVM bitcode files and applying
185 various optimization passes.
187 Phase 4 : Symbol Resolution after optimization
188 ----------------------------------------------
190 In this phase, the linker reads optimized a native object file and updates the
191 internal global symbol table to reflect any changes. The linker also collects
192 information about any changes in use of external symbols by LLVM bitcode
193 files. In the example above, the linker notes that ``foo4()`` is not used any
194 more. If dead code stripping is enabled then the linker refreshes the live
195 symbol information appropriately and performs dead code stripping.
197 After this phase, the linker continues linking as if it never saw LLVM bitcode
205 ``libLTO`` is a shared object that is part of the LLVM tools, and is intended
206 for use by a linker. ``libLTO`` provides an abstract C interface to use the LLVM
207 interprocedural optimizer without exposing details of LLVM's internals. The
208 intention is to keep the interface as stable as possible even when the LLVM
209 optimizer continues to evolve. It should even be possible for a completely
210 different compilation technology to provide a different libLTO that works with
211 their object files and the standard linker tool.
216 A non-native object file is handled via an ``lto_module_t``. The following
217 functions allow the linker to check if a file (on disk or in a memory buffer) is
218 a file which libLTO can process:
222 lto_module_is_object_file(const char*)
223 lto_module_is_object_file_for_target(const char*, const char*)
224 lto_module_is_object_file_in_memory(const void*, size_t)
225 lto_module_is_object_file_in_memory_for_target(const void*, size_t, const char*)
227 If the object file can be processed by ``libLTO``, the linker creates a
228 ``lto_module_t`` by using one of:
232 lto_module_create(const char*)
233 lto_module_create_from_memory(const void*, size_t)
235 and when done, the handle is released via
239 lto_module_dispose(lto_module_t)
242 The linker can introspect the non-native object file by getting the number of
243 symbols and getting the name and attributes of each symbol via:
247 lto_module_get_num_symbols(lto_module_t)
248 lto_module_get_symbol_name(lto_module_t, unsigned int)
249 lto_module_get_symbol_attribute(lto_module_t, unsigned int)
251 The attributes of a symbol include the alignment, visibility, and kind.
256 Once the linker has loaded each non-native object files into an
257 ``lto_module_t``, it can request ``libLTO`` to process them all and generate a
258 native object file. This is done in a couple of steps. First, a code generator
265 Then, each non-native object file is added to the code generator with:
269 lto_codegen_add_module(lto_code_gen_t, lto_module_t)
271 The linker then has the option of setting some codegen options. Whether or not
272 to generate DWARF debug info is set with:
276 lto_codegen_set_debug_model(lto_code_gen_t)
278 Which kind of position independence is set with:
282 lto_codegen_set_pic_model(lto_code_gen_t)
284 And each symbol that is referenced by a native object file or otherwise must not
285 be optimized away is set with:
289 lto_codegen_add_must_preserve_symbol(lto_code_gen_t, const char*)
291 After all these settings are done, the linker requests that a native object file
292 be created from the modules with the settings using:
296 lto_codegen_compile(lto_code_gen_t, size*)
298 which returns a pointer to a buffer containing the generated native object file.
299 The linker then parses that and links it with the rest of the native object