1 <!DOCTYPE HTML PUBLIC "-//W3C//DTD HTML 4.01//EN" 2 "http://www.w3.org/TR/html4/strict.dtd"> 3 <html> 4 <head> 5 <meta http-equiv="Content-Type" content="text/html; charset=utf-8"> 6 <title>LLVM Link Time Optimization: Design and Implementation</title> 7 <link rel="stylesheet" href="llvm.css" type="text/css"> 8 </head> 9 10 <h1> 11 LLVM Link Time Optimization: Design and Implementation 12 </h1> 13 14 <ul> 15 <li><a href="#desc">Description</a></li> 16 <li><a href="#design">Design Philosophy</a> 17 <ul> 18 <li><a href="#example1">Example of link time optimization</a></li> 19 <li><a href="#alternative_approaches">Alternative Approaches</a></li> 20 </ul></li> 21 <li><a href="#multiphase">Multi-phase communication between LLVM and linker</a> 22 <ul> 23 <li><a href="#phase1">Phase 1 : Read LLVM Bitcode Files</a></li> 24 <li><a href="#phase2">Phase 2 : Symbol Resolution</a></li> 25 <li><a href="#phase3">Phase 3 : Optimize Bitcode Files</a></li> 26 <li><a href="#phase4">Phase 4 : Symbol Resolution after optimization</a></li> 27 </ul></li> 28 <li><a href="#lto">libLTO</a> 29 <ul> 30 <li><a href="#lto_module_t">lto_module_t</a></li> 31 <li><a href="#lto_code_gen_t">lto_code_gen_t</a></li> 32 </ul> 33 </ul> 34 35 <div class="doc_author"> 36 <p>Written by Devang Patel and Nick Kledzik</p> 37 </div> 38 39 <!-- *********************************************************************** --> 40 <h2> 41 <a name="desc">Description</a> 42 </h2> 43 <!-- *********************************************************************** --> 44 45 <div> 46 <p> 47 LLVM features powerful intermodular optimizations which can be used at link 48 time. Link Time Optimization (LTO) is another name for intermodular optimization 49 when performed during the link stage. This document describes the interface 50 and design between the LTO optimizer and the linker.</p> 51 </div> 52 53 <!-- *********************************************************************** --> 54 <h2> 55 <a name="design">Design Philosophy</a> 56 </h2> 57 <!-- *********************************************************************** --> 58 59 <div> 60 <p> 61 The LLVM Link Time Optimizer provides complete transparency, while doing 62 intermodular optimization, in the compiler tool chain. Its main goal is to let 63 the developer take advantage of intermodular optimizations without making any 64 significant changes to the developer's makefiles or build system. This is 65 achieved through tight integration with the linker. In this model, the linker 66 treates LLVM bitcode files like native object files and allows mixing and 67 matching among them. The linker uses <a href="#lto">libLTO</a>, a shared 68 object, to handle LLVM bitcode files. This tight integration between 69 the linker and LLVM optimizer helps to do optimizations that are not possible 70 in other models. The linker input allows the optimizer to avoid relying on 71 conservative escape analysis. 72 </p> 73 74 <!-- ======================================================================= --> 75 <h3> 76 <a name="example1">Example of link time optimization</a> 77 </h3> 78 79 <div> 80 <p>The following example illustrates the advantages of LTO's integrated 81 approach and clean interface. This example requires a system linker which 82 supports LTO through the interface described in this document. Here, 83 clang transparently invokes system linker. </p> 84 <ul> 85 <li> Input source file <tt>a.c</tt> is compiled into LLVM bitcode form. 86 <li> Input source file <tt>main.c</tt> is compiled into native object code. 87 </ul> 88 <pre class="doc_code"> 89 --- a.h --- 90 extern int foo1(void); 91 extern void foo2(void); 92 extern void foo4(void); 93 94 --- a.c --- 95 #include "a.h" 96 97 static signed int i = 0; 98 99 void foo2(void) { 100 i = -1; 101 } 102 103 static int foo3() { 104 foo4(); 105 return 10; 106 } 107 108 int foo1(void) { 109 int data = 0; 110 111 if (i < 0) 112 data = foo3(); 113 114 data = data + 42; 115 return data; 116 } 117 118 --- main.c --- 119 #include <stdio.h> 120 #include "a.h" 121 122 void foo4(void) { 123 printf("Hi\n"); 124 } 125 126 int main() { 127 return foo1(); 128 } 129 130 --- command lines --- 131 $ clang -emit-llvm -c a.c -o a.o # <-- a.o is LLVM bitcode file 132 $ clang -c main.c -o main.o # <-- main.o is native object file 133 $ clang a.o main.o -o main # <-- standard link command without any modifications 134 </pre> 135 136 <ul> 137 <li>In this example, the linker recognizes that <tt>foo2()</tt> is an 138 externally visible symbol defined in LLVM bitcode file. The linker 139 completes its usual symbol resolution pass and finds that <tt>foo2()</tt> 140 is not used anywhere. This information is used by the LLVM optimizer and 141 it removes <tt>foo2()</tt>.</li> 142 <li>As soon as <tt>foo2()</tt> is removed, the optimizer recognizes that condition 143 <tt>i < 0</tt> is always false, which means <tt>foo3()</tt> is never 144 used. Hence, the optimizer also removes <tt>foo3()</tt>.</li> 145 <li>And this in turn, enables linker to remove <tt>foo4()</tt>.</li> 146 </ul> 147 148 <p>This example illustrates the advantage of tight integration with the 149 linker. Here, the optimizer can not remove <tt>foo3()</tt> without the 150 linker's input.</p> 151 152 </div> 153 154 <!-- ======================================================================= --> 155 <h3> 156 <a name="alternative_approaches">Alternative Approaches</a> 157 </h3> 158 159 <div> 160 <dl> 161 <dt><b>Compiler driver invokes link time optimizer separately.</b></dt> 162 <dd>In this model the link time optimizer is not able to take advantage of 163 information collected during the linker's normal symbol resolution phase. 164 In the above example, the optimizer can not remove <tt>foo2()</tt> without 165 the linker's input because it is externally visible. This in turn prohibits 166 the optimizer from removing <tt>foo3()</tt>.</dd> 167 <dt><b>Use separate tool to collect symbol information from all object 168 files.</b></dt> 169 <dd>In this model, a new, separate, tool or library replicates the linker's 170 capability to collect information for link time optimization. Not only is 171 this code duplication difficult to justify, but it also has several other 172 disadvantages. For example, the linking semantics and the features 173 provided by the linker on various platform are not unique. This means, 174 this new tool needs to support all such features and platforms in one 175 super tool or a separate tool per platform is required. This increases 176 maintenance cost for link time optimizer significantly, which is not 177 necessary. This approach also requires staying synchronized with linker 178 developements on various platforms, which is not the main focus of the link 179 time optimizer. Finally, this approach increases end user's build time due 180 to the duplication of work done by this separate tool and the linker itself. 181 </dd> 182 </dl> 183 </div> 184 185 </div> 186 187 <!-- *********************************************************************** --> 188 <h2> 189 <a name="multiphase">Multi-phase communication between libLTO and linker</a> 190 </h2> 191 192 <div> 193 <p>The linker collects information about symbol defininitions and uses in 194 various link objects which is more accurate than any information collected 195 by other tools during typical build cycles. The linker collects this 196 information by looking at the definitions and uses of symbols in native .o 197 files and using symbol visibility information. The linker also uses 198 user-supplied information, such as a list of exported symbols. LLVM 199 optimizer collects control flow information, data flow information and knows 200 much more about program structure from the optimizer's point of view. 201 Our goal is to take advantage of tight integration between the linker and 202 the optimizer by sharing this information during various linking phases. 203 </p> 204 205 <!-- ======================================================================= --> 206 <h3> 207 <a name="phase1">Phase 1 : Read LLVM Bitcode Files</a> 208 </h3> 209 210 <div> 211 <p>The linker first reads all object files in natural order and collects 212 symbol information. This includes native object files as well as LLVM bitcode 213 files. To minimize the cost to the linker in the case that all .o files 214 are native object files, the linker only calls <tt>lto_module_create()</tt> 215 when a supplied object file is found to not be a native object file. If 216 <tt>lto_module_create()</tt> returns that the file is an LLVM bitcode file, 217 the linker 218 then iterates over the module using <tt>lto_module_get_symbol_name()</tt> and 219 <tt>lto_module_get_symbol_attribute()</tt> to get all symbols defined and 220 referenced. 221 This information is added to the linker's global symbol table. 222 </p> 223 <p>The lto* functions are all implemented in a shared object libLTO. This 224 allows the LLVM LTO code to be updated independently of the linker tool. 225 On platforms that support it, the shared object is lazily loaded. 226 </p> 227 </div> 228 229 <!-- ======================================================================= --> 230 <h3> 231 <a name="phase2">Phase 2 : Symbol Resolution</a> 232 </h3> 233 234 <div> 235 <p>In this stage, the linker resolves symbols using global symbol table. 236 It may report undefined symbol errors, read archive members, replace 237 weak symbols, etc. The linker is able to do this seamlessly even though it 238 does not know the exact content of input LLVM bitcode files. If dead code 239 stripping is enabled then the linker collects the list of live symbols. 240 </p> 241 </div> 242 243 <!-- ======================================================================= --> 244 <h3> 245 <a name="phase3">Phase 3 : Optimize Bitcode Files</a> 246 </h3> 247 <div> 248 <p>After symbol resolution, the linker tells the LTO shared object which 249 symbols are needed by native object files. In the example above, the linker 250 reports that only <tt>foo1()</tt> is used by native object files using 251 <tt>lto_codegen_add_must_preserve_symbol()</tt>. Next the linker invokes 252 the LLVM optimizer and code generators using <tt>lto_codegen_compile()</tt> 253 which returns a native object file creating by merging the LLVM bitcode files 254 and applying various optimization passes. 255 </p> 256 </div> 257 258 <!-- ======================================================================= --> 259 <h3> 260 <a name="phase4">Phase 4 : Symbol Resolution after optimization</a> 261 </h3> 262 263 <div> 264 <p>In this phase, the linker reads optimized a native object file and 265 updates the internal global symbol table to reflect any changes. The linker 266 also collects information about any changes in use of external symbols by 267 LLVM bitcode files. In the example above, the linker notes that 268 <tt>foo4()</tt> is not used any more. If dead code stripping is enabled then 269 the linker refreshes the live symbol information appropriately and performs 270 dead code stripping.</p> 271 <p>After this phase, the linker continues linking as if it never saw LLVM 272 bitcode files.</p> 273 </div> 274 275 </div> 276 277 <!-- *********************************************************************** --> 278 <h2> 279 <a name="lto">libLTO</a> 280 </h2> 281 282 <div> 283 <p><tt>libLTO</tt> is a shared object that is part of the LLVM tools, and 284 is intended for use by a linker. <tt>libLTO</tt> provides an abstract C 285 interface to use the LLVM interprocedural optimizer without exposing details 286 of LLVM's internals. The intention is to keep the interface as stable as 287 possible even when the LLVM optimizer continues to evolve. It should even 288 be possible for a completely different compilation technology to provide 289 a different libLTO that works with their object files and the standard 290 linker tool.</p> 291 292 <!-- ======================================================================= --> 293 <h3> 294 <a name="lto_module_t">lto_module_t</a> 295 </h3> 296 297 <div> 298 299 <p>A non-native object file is handled via an <tt>lto_module_t</tt>. 300 The following functions allow the linker to check if a file (on disk 301 or in a memory buffer) is a file which libLTO can process:</p> 302 303 <pre class="doc_code"> 304 lto_module_is_object_file(const char*) 305 lto_module_is_object_file_for_target(const char*, const char*) 306 lto_module_is_object_file_in_memory(const void*, size_t) 307 lto_module_is_object_file_in_memory_for_target(const void*, size_t, const char*) 308 </pre> 309 310 <p>If the object file can be processed by libLTO, the linker creates a 311 <tt>lto_module_t</tt> by using one of</p> 312 313 <pre class="doc_code"> 314 lto_module_create(const char*) 315 lto_module_create_from_memory(const void*, size_t) 316 </pre> 317 318 <p>and when done, the handle is released via</p> 319 320 <pre class="doc_code"> 321 lto_module_dispose(lto_module_t) 322 </pre> 323 324 <p>The linker can introspect the non-native object file by getting the number of 325 symbols and getting the name and attributes of each symbol via:</p> 326 327 <pre class="doc_code"> 328 lto_module_get_num_symbols(lto_module_t) 329 lto_module_get_symbol_name(lto_module_t, unsigned int) 330 lto_module_get_symbol_attribute(lto_module_t, unsigned int) 331 </pre> 332 333 <p>The attributes of a symbol include the alignment, visibility, and kind.</p> 334 </div> 335 336 <!-- ======================================================================= --> 337 <h3> 338 <a name="lto_code_gen_t">lto_code_gen_t</a> 339 </h3> 340 341 <div> 342 343 <p>Once the linker has loaded each non-native object files into an 344 <tt>lto_module_t</tt>, it can request libLTO to process them all and 345 generate a native object file. This is done in a couple of steps. 346 First, a code generator is created with:</p> 347 348 <pre class="doc_code">lto_codegen_create()</pre> 349 350 <p>Then, each non-native object file is added to the code generator with:</p> 351 352 <pre class="doc_code"> 353 lto_codegen_add_module(lto_code_gen_t, lto_module_t) 354 </pre> 355 356 <p>The linker then has the option of setting some codegen options. Whether or 357 not to generate DWARF debug info is set with:</p> 358 359 <pre class="doc_code">lto_codegen_set_debug_model(lto_code_gen_t)</pre> 360 361 <p>Which kind of position independence is set with:</p> 362 363 <pre class="doc_code">lto_codegen_set_pic_model(lto_code_gen_t) </pre> 364 365 <p>And each symbol that is referenced by a native object file or otherwise must 366 not be optimized away is set with:</p> 367 368 <pre class="doc_code"> 369 lto_codegen_add_must_preserve_symbol(lto_code_gen_t, const char*) 370 </pre> 371 372 <p>After all these settings are done, the linker requests that a native object 373 file be created from the modules with the settings using:</p> 374 375 <pre class="doc_code">lto_codegen_compile(lto_code_gen_t, size*)</pre> 376 377 <p>which returns a pointer to a buffer containing the generated native 378 object file. The linker then parses that and links it with the rest 379 of the native object files.</p> 380 381 </div> 382 383 </div> 384 385 <!-- *********************************************************************** --> 386 387 <hr> 388 <address> 389 <a href="http://jigsaw.w3.org/css-validator/check/referer"><img 390 src="http://jigsaw.w3.org/css-validator/images/vcss-blue" alt="Valid CSS"></a> 391 <a href="http://validator.w3.org/check/referer"><img 392 src="http://www.w3.org/Icons/valid-html401-blue" alt="Valid HTML 4.01"></a> 393 394 Devang Patel and Nick Kledzik<br> 395 <a href="http://llvm.org/">LLVM Compiler Infrastructure</a><br> 396 Last modified: $Date$ 397 </address> 398 399 </body> 400 </html> 401 402
