| blob | 1f28d7c840377506741c7f09df52726286d9bb94 |
3 <html>
4 <head>
7 </head>
10 LLVM Link Time Optimization: Design and Implementation
11 </div>
13 <ul>
16 <ul>
19 </ul></li>
21 <ul>
26 </ul></li>
28 <ul>
35 </ul></li>
37 </ul>
41 </div>
43 <!-- *********************************************************************** -->
46 </div>
47 <!-- *********************************************************************** -->
50 <p>
51 LLVM features powerful intermodular optimizations which can be used at link
52 time. Link Time Optimization is another name for intermodular optimization
53 when performed during the link stage. This document describes the interface
54 and design between the LLVM intermodular optimizer and the linker.</p>
55 </div>
57 <!-- *********************************************************************** -->
60 </div>
61 <!-- *********************************************************************** -->
64 <p>
65 The LLVM Link Time Optimizer provides complete transparency, while doing
66 intermodular optimization, in the compiler tool chain. Its main goal is to let
67 the developer take advantage of intermodular optimizations without making any
68 significant changes to the developer's makefiles or build system. This is
69 achieved through tight integration with the linker. In this model, the linker
70 treates LLVM bitcode files like native object files and allows mixing and
72 loaded library, to handle LLVM bitcode files. This tight integration between
73 the linker and LLVM optimizer helps to do optimizations that are not possible
74 in other models. The linker input allows the optimizer to avoid relying on
75 conservative escape analysis.
76 </p>
77 </div>
79 <!-- ======================================================================= -->
82 </div>
85 <p>The following example illustrates the advantages of LTO's integrated
86 approach and clean interface. This example requires a system linker which
87 supports LTO through the interface described in this document. Here,
88 llvm-gcc transparently invokes system linker. </p>
89 <ul>
92 </ul>
94 --- a.h ---
95 extern int foo1(void);
96 extern void foo2(void);
97 extern void foo4(void);
98 --- a.c ---
99 #include "a.h"
101 static signed int i = 0;
103 void foo2(void) {
104 i = -1;
105 }
107 static int foo3() {
108 foo4();
109 return 10;
110 }
112 int foo1(void) {
113 int data = 0;
117 data = data + 42;
118 return data;
119 }
121 --- main.c ---
123 #include "a.h"
125 void foo4(void) {
126 printf ("Hi\n");
127 }
129 int main() {
130 return foo1();
131 }
133 --- command lines ---
134 $ llvm-gcc --emit-llvm -c a.c -o a.o # <-- a.o is LLVM bitcode file
135 $ llvm-gcc -c main.c -o main.o # <-- main.o is native object file
136 $ llvm-gcc a.o main.o -o main # <-- standard link command without any modifications
137 </pre></div>
139 externally visible symbol defined in LLVM bitcode file. This information
141 Based on this information, the linker completes its usual symbol resolution
148 advantage of tight integration with the linker. Here, the optimizer can not
150 </p>
151 </div>
153 <!-- ======================================================================= -->
156 </div>
159 <dl>
161 <dd>In this model the link time optimizer is not able to take advantage of
162 information collected during the linker's normal symbol resolution phase.
164 the linker's input because it is externally visible. This in turn prohibits
166 <dt><b>Use separate tool to collect symbol information from all object
167 files.</b></dt>
168 <dd>In this model, a new, separate, tool or library replicates the linker's
169 capability to collect information for link time optimization. Not only is
170 this code duplication difficult to justify, but it also has several other
171 disadvantages. For example, the linking semantics and the features
172 provided by the linker on various platform are not unique. This means,
173 this new tool needs to support all such features and platforms in one
174 super tool or a separate tool per platform is required. This increases
175 maintance cost for link time optimizer significantly, which is not
176 necessary. This approach also requires staying synchronized with linker
177 developements on various platforms, which is not the main focus of the link
178 time optimizer. Finally, this approach increases end user's build time due
179 to the duplication of work done by this separate tool and the linker itself.
180 </dd>
181 </dl>
182 </div>
184 <!-- *********************************************************************** -->
187 </div>
190 <p>The linker collects information about symbol defininitions and uses in
191 various link objects which is more accurate than any information collected
192 by other tools during typical build cycles. The linker collects this
193 information by looking at the definitions and uses of symbols in native .o
194 files and using symbol visibility information. The linker also uses
195 user-supplied information, such as a list of exported symbols. LLVM
196 optimizer collects control flow information, data flow information and knows
197 much more about program structure from the optimizer's point of view.
198 Our goal is to take advantage of tight intergration between the linker and
199 the optimizer by sharing this information during various linking phases.
200 </p>
201 </div>
203 <!-- ======================================================================= -->
206 </div>
209 <p>The linker first reads all object files in natural order and collects
210 symbol information. This includes native object files as well as LLVM bitcode
211 files. In this phase, the linker uses
213 information from each LLVM bitcode files and updates its internal global
214 symbol table accordingly. The intent of this interface is to avoid overhead
215 in the non LLVM case, where all input object files are native object files,
216 by putting this code in the error path of the linker. When the linker sees
218 to allow changes to the LLVM LTO code without relinking the linker.
219 </p>
220 </div>
222 <!-- ======================================================================= -->
225 </div>
228 <p>In this stage, the linker resolves symbols using global symbol table
229 information to report undefined symbol errors, read archive members, resolve
230 weak symbols, etc. The linker is able to do this seamlessly even though it
231 does not know the exact content of input LLVM bitcode files because it uses
232 symbol information provided by
234 stripping is enabled then the linker collects the list of live symbols.
235 </p>
236 </div>
238 <!-- ======================================================================= -->
241 </div>
243 <p>After symbol resolution, the linker updates symbol information supplied
244 by LLVM bitcode files appropriately. For example, whether certain LLVM
245 bitcode supplied symbols are used or not. In the example above, the linker
249 and requests an optimized native object file of the LLVM portion of the
250 program.
251 </p>
252 </div>
254 <!-- ======================================================================= -->
257 </div>
260 <p>In this phase, the linker reads optimized a native object file and
261 updates the internal global symbol table to reflect any changes. The linker
262 also collects information about any changes in use of external symbols by
263 LLVM bitcode files. In the examle above, the linker notes that
265 the linker refreshes the live symbol information appropriately and performs
266 dead code stripping.</p>
267 <p>After this phase, the linker continues linking as if it never saw LLVM
268 bitcode files.</p>
269 </div>
271 <!-- *********************************************************************** -->
274 </div>
279 interface to use the LLVM interprocedural optimizer without exposing details
280 of LLVM's internals. The intention is to keep the interface as stable as
281 possible even when the LLVM optimizer continues to evolve.</p>
282 </div>
284 <!-- ======================================================================= -->
287 </div>
291 functions and global variables, defined in LLVM bitcode files, to the linker.
292 This includes symbol visibility information. This information is used by
294 defined inside an LLVM bitcode module and it is an externally visible symbol.
298 optimized native object file in
301 visibility information to notify LLVM intermodular optimizer that it is dead.
302 The LLVM intermodular optimizer takes advantage of such information to
303 generate better code.</p>
304 </div>
306 <!-- ======================================================================= -->
309 </div>
313 LLVM bitcode files and collect LLVMSymbol information. This routine also
314 supplies a list of externally defined symbols that are used by LLVM bitcode
315 files. The linker uses this symbol information to do symbol resolution.
317 memory. This function also provides a list of external references used by
318 bitcode files.</p>
319 </div>
321 <!-- ======================================================================= -->
324 </div>
328 LLVM bitcode files by applying LLVM intermodular optimization techniques.
329 This function runs the LLVM intermodular optimizer and generates native
331 linker.</p>
332 </div>
334 <!-- ======================================================================= -->
337 </div>
341 while validating LLVM bitcode file.</p>
342 </div>
344 <!-- ======================================================================= -->
347 </div>
352 modules from memory. </p>
353 </div>
355 <!-- ======================================================================= -->
358 </div>
363 </div>
365 <!-- *********************************************************************** -->
368 </div>
369 <!-- *********************************************************************** -->
375 </div>
377 <!-- *********************************************************************** -->
379 <hr>
380 <address>
386 Devang Patel<br>
388 Last modified: $Date$
389 </address>
391 </body>
392 </html>