1 ====================================================== 2 LLVM Link Time Optimization: Design and Implementation 3 ====================================================== 4 5 .. contents:: 6 :local: 7 8 Description 9 =========== 10 11 LLVM features powerful intermodular optimizations which can be used at link 12 time. Link Time Optimization (LTO) is another name for intermodular 13 optimization when performed during the link stage. This document describes the 14 interface and design between the LTO optimizer and the linker. 15 16 Design Philosophy 17 ================= 18 19 The LLVM Link Time Optimizer provides complete transparency, while doing 20 intermodular optimization, in the compiler tool chain. Its main goal is to let 21 the developer take advantage of intermodular optimizations without making any 22 significant changes to the developer's makefiles or build system. This is 23 achieved through tight integration with the linker. In this model, the linker 24 treates LLVM bitcode files like native object files and allows mixing and 25 matching among them. The linker uses `libLTO`_, a shared object, to handle LLVM 26 bitcode files. This tight integration between the linker and LLVM optimizer 27 helps to do optimizations that are not possible in other models. The linker 28 input allows the optimizer to avoid relying on conservative escape analysis. 29 30 .. _libLTO-example: 31 32 Example of link time optimization 33 --------------------------------- 34 35 The following example illustrates the advantages of LTO's integrated approach 36 and clean interface. This example requires a system linker which supports LTO 37 through the interface described in this document. Here, clang transparently 38 invokes system linker. 39 40 * Input source file ``a.c`` is compiled into LLVM bitcode form. 41 * Input source file ``main.c`` is compiled into native object code. 42 43 .. code-block:: c++ 44 45 --- a.h --- 46 extern int foo1(void); 47 extern void foo2(void); 48 extern void foo4(void); 49 50 --- a.c --- 51 #include "a.h" 52 53 static signed int i = 0; 54 55 void foo2(void) { 56 i = -1; 57 } 58 59 static int foo3() { 60 foo4(); 61 return 10; 62 } 63 64 int foo1(void) { 65 int data = 0; 66 67 if (i < 0) 68 data = foo3(); 69 70 data = data + 42; 71 return data; 72 } 73 74 --- main.c --- 75 #include <stdio.h> 76 #include "a.h" 77 78 void foo4(void) { 79 printf("Hi\n"); 80 } 81 82 int main() { 83 return foo1(); 84 } 85 86 To compile, run: 87 88 .. code-block:: console 89 90 % clang -emit-llvm -c a.c -o a.o # <-- a.o is LLVM bitcode file 91 % clang -c main.c -o main.o # <-- main.o is native object file 92 % clang a.o main.o -o main # <-- standard link command without modifications 93 94 * In this example, the linker recognizes that ``foo2()`` is an externally 95 visible symbol defined in LLVM bitcode file. The linker completes its usual 96 symbol resolution pass and finds that ``foo2()`` is not used 97 anywhere. This information is used by the LLVM optimizer and it 98 removes ``foo2()``. 99 100 * As soon as ``foo2()`` is removed, the optimizer recognizes that condition ``i 101 < 0`` is always false, which means ``foo3()`` is never used. Hence, the 102 optimizer also removes ``foo3()``. 103 104 * And this in turn, enables linker to remove ``foo4()``. 105 106 This example illustrates the advantage of tight integration with the 107 linker. Here, the optimizer can not remove ``foo3()`` without the linker's 108 input. 109 110 Alternative Approaches 111 ---------------------- 112 113 **Compiler driver invokes link time optimizer separately.** 114 In this model the link time optimizer is not able to take advantage of 115 information collected during the linker's normal symbol resolution phase. 116 In the above example, the optimizer can not remove ``foo2()`` without the 117 linker's input because it is externally visible. This in turn prohibits the 118 optimizer from removing ``foo3()``. 119 120 **Use separate tool to collect symbol information from all object files.** 121 In this model, a new, separate, tool or library replicates the linker's 122 capability to collect information for link time optimization. Not only is 123 this code duplication difficult to justify, but it also has several other 124 disadvantages. For example, the linking semantics and the features provided 125 by the linker on various platform are not unique. This means, this new tool 126 needs to support all such features and platforms in one super tool or a 127 separate tool per platform is required. This increases maintenance cost for 128 link time optimizer significantly, which is not necessary. This approach 129 also requires staying synchronized with linker developements on various 130 platforms, which is not the main focus of the link time optimizer. Finally, 131 this approach increases end user's build time due to the duplication of work 132 done by this separate tool and the linker itself. 133 134 Multi-phase communication between ``libLTO`` and linker 135 ======================================================= 136 137 The linker collects information about symbol defininitions and uses in various 138 link objects which is more accurate than any information collected by other 139 tools during typical build cycles. The linker collects this information by 140 looking at the definitions and uses of symbols in native .o files and using 141 symbol visibility information. The linker also uses user-supplied information, 142 such as a list of exported symbols. LLVM optimizer collects control flow 143 information, data flow information and knows much more about program structure 144 from the optimizer's point of view. Our goal is to take advantage of tight 145 integration between the linker and the optimizer by sharing this information 146 during various linking phases. 147 148 Phase 1 : Read LLVM Bitcode Files 149 --------------------------------- 150 151 The linker first reads all object files in natural order and collects symbol 152 information. This includes native object files as well as LLVM bitcode files. 153 To minimize the cost to the linker in the case that all .o files are native 154 object files, the linker only calls ``lto_module_create()`` when a supplied 155 object file is found to not be a native object file. If ``lto_module_create()`` 156 returns that the file is an LLVM bitcode file, the linker then iterates over the 157 module using ``lto_module_get_symbol_name()`` and 158 ``lto_module_get_symbol_attribute()`` to get all symbols defined and referenced. 159 This information is added to the linker's global symbol table. 160 161 162 The lto* functions are all implemented in a shared object libLTO. This allows 163 the LLVM LTO code to be updated independently of the linker tool. On platforms 164 that support it, the shared object is lazily loaded. 165 166 Phase 2 : Symbol Resolution 167 --------------------------- 168 169 In this stage, the linker resolves symbols using global symbol table. It may 170 report undefined symbol errors, read archive members, replace weak symbols, etc. 171 The linker is able to do this seamlessly even though it does not know the exact 172 content of input LLVM bitcode files. If dead code stripping is enabled then the 173 linker collects the list of live symbols. 174 175 Phase 3 : Optimize Bitcode Files 176 -------------------------------- 177 178 After symbol resolution, the linker tells the LTO shared object which symbols 179 are needed by native object files. In the example above, the linker reports 180 that only ``foo1()`` is used by native object files using 181 ``lto_codegen_add_must_preserve_symbol()``. Next the linker invokes the LLVM 182 optimizer and code generators using ``lto_codegen_compile()`` which returns a 183 native object file creating by merging the LLVM bitcode files and applying 184 various optimization passes. 185 186 Phase 4 : Symbol Resolution after optimization 187 ---------------------------------------------- 188 189 In this phase, the linker reads optimized a native object file and updates the 190 internal global symbol table to reflect any changes. The linker also collects 191 information about any changes in use of external symbols by LLVM bitcode 192 files. In the example above, the linker notes that ``foo4()`` is not used any 193 more. If dead code stripping is enabled then the linker refreshes the live 194 symbol information appropriately and performs dead code stripping. 195 196 After this phase, the linker continues linking as if it never saw LLVM bitcode 197 files. 198 199 .. _libLTO: 200 201 ``libLTO`` 202 ========== 203 204 ``libLTO`` is a shared object that is part of the LLVM tools, and is intended 205 for use by a linker. ``libLTO`` provides an abstract C interface to use the LLVM 206 interprocedural optimizer without exposing details of LLVM's internals. The 207 intention is to keep the interface as stable as possible even when the LLVM 208 optimizer continues to evolve. It should even be possible for a completely 209 different compilation technology to provide a different libLTO that works with 210 their object files and the standard linker tool. 211 212 ``lto_module_t`` 213 ---------------- 214 215 A non-native object file is handled via an ``lto_module_t``. The following 216 functions allow the linker to check if a file (on disk or in a memory buffer) is 217 a file which libLTO can process: 218 219 .. code-block:: c 220 221 lto_module_is_object_file(const char*) 222 lto_module_is_object_file_for_target(const char*, const char*) 223 lto_module_is_object_file_in_memory(const void*, size_t) 224 lto_module_is_object_file_in_memory_for_target(const void*, size_t, const char*) 225 226 If the object file can be processed by ``libLTO``, the linker creates a 227 ``lto_module_t`` by using one of: 228 229 .. code-block:: c 230 231 lto_module_create(const char*) 232 lto_module_create_from_memory(const void*, size_t) 233 234 and when done, the handle is released via 235 236 .. code-block:: c 237 238 lto_module_dispose(lto_module_t) 239 240 241 The linker can introspect the non-native object file by getting the number of 242 symbols and getting the name and attributes of each symbol via: 243 244 .. code-block:: c 245 246 lto_module_get_num_symbols(lto_module_t) 247 lto_module_get_symbol_name(lto_module_t, unsigned int) 248 lto_module_get_symbol_attribute(lto_module_t, unsigned int) 249 250 The attributes of a symbol include the alignment, visibility, and kind. 251 252 ``lto_code_gen_t`` 253 ------------------ 254 255 Once the linker has loaded each non-native object files into an 256 ``lto_module_t``, it can request ``libLTO`` to process them all and generate a 257 native object file. This is done in a couple of steps. First, a code generator 258 is created with: 259 260 .. code-block:: c 261 262 lto_codegen_create() 263 264 Then, each non-native object file is added to the code generator with: 265 266 .. code-block:: c 267 268 lto_codegen_add_module(lto_code_gen_t, lto_module_t) 269 270 The linker then has the option of setting some codegen options. Whether or not 271 to generate DWARF debug info is set with: 272 273 .. code-block:: c 274 275 lto_codegen_set_debug_model(lto_code_gen_t) 276 277 Which kind of position independence is set with: 278 279 .. code-block:: c 280 281 lto_codegen_set_pic_model(lto_code_gen_t) 282 283 And each symbol that is referenced by a native object file or otherwise must not 284 be optimized away is set with: 285 286 .. code-block:: c 287 288 lto_codegen_add_must_preserve_symbol(lto_code_gen_t, const char*) 289 290 After all these settings are done, the linker requests that a native object file 291 be created from the modules with the settings using: 292 293 .. code-block:: c 294 295 lto_codegen_compile(lto_code_gen_t, size*) 296 297 which returns a pointer to a buffer containing the generated native object file. 298 The linker then parses that and links it with the rest of the native object 299 files. 300
