| blob | 7e152452740b8314f5df4d27a6231b2883bcbf09 |
3 <html>
4 <head>
8 </head>
9 <body>
12 Writing an LLVM Pass
13 </div>
15 <ol>
18 <ul>
22 </ul></li>
24 <ul>
27 <ul>
29 </ul></li>
31 <ul>
37 </ul></li>
39 <ul>
45 </ul></li>
47 <ul>
53 </ul></li>
55 <ul>
61 </ul></li>
63 <ul>
67 method</a></li>
70 </ul></li>
72 class</a>
73 <ul>
76 </ul></li>
77 </ul>
79 <ul>
81 </ul></li>
83 <ul>
85 method</a></li>
86 <li><a href="#AU::addRequired">The <tt>AnalysisUsage::addRequired<></tt> and <tt>AnalysisUsage::addRequiredTransitive<></tt> methods</a></li>
91 </ul></li>
93 <ul>
96 </ul></li>
99 <ul>
101 </ul></li>
103 <ul>
106 </ul></li>
108 <ul>
111 </ul></li>
113 <ul>
115 </ul></li>
116 </ol>
121 </div>
123 <!-- *********************************************************************** -->
126 </div>
127 <!-- *********************************************************************** -->
131 <p>The LLVM Pass Framework is an important part of the LLVM system, because LLVM
132 passes are where most of the interesting parts of the compiler exist. Passes
133 perform the transformations and optimizations that make up the compiler, they
134 build the analysis results that are used by these transformations, and they are,
135 above all, a structuring technique for compiler code.</p>
139 class, which implement functionality by overriding virtual methods inherited
147 more information about what your pass does, and how it can be combined with
148 other passes. One of the main features of the LLVM Pass Framework is that it
149 schedules passes to run in an efficient way based on the constraints that your
150 pass meets (which are indicated by which class they derive from).</p>
152 <p>We start by showing you how to construct a pass, everything from setting up
153 the code, to compiling, loading, and executing it. After the basics are down,
154 more advanced features are discussed.</p>
156 </div>
158 <!-- *********************************************************************** -->
161 </div>
162 <!-- *********************************************************************** -->
167 is designed to simply print out the name of non-external functions that exist in
168 the program being compiled. It does not modify the program at all, it just
169 inspects it. The source code and files for this pass are available in the LLVM
172 </div>
174 <!-- ======================================================================= -->
177 </div>
181 <p>First, configure and build LLVM. This needs to be done directly inside the
182 LLVM source tree rather than in a separate objects directory.
183 Next, you need to create a new directory somewhere in the LLVM source
184 base. For this example, we'll assume that you made
186 (Makefile) that will compile the source code for the new pass. To do this,
188 <hr>
191 # Makefile for hello pass
193 # Path to top level of LLVM hierarchy
194 LEVEL = ../../..
196 # Name of the library to build
197 LIBRARYNAME = Hello
199 # Make the shared library become a loadable module so the tools can
200 # dlopen/dlsym on the resulting library.
201 LOADABLE_MODULE = 1
203 # Include the makefile implementation stuff
204 include $(LEVEL)/Makefile.common
205 </pre></div>
208 directory are to be compiled and linked together into a shared object
211 If your operating system uses a suffix other than .so (such as windows or
212 Mac OS/X), the appropriate extension will be used.</p>
214 <p>If you are used CMake to build LLVM, see
217 <p>Now that we have the build scripts set up, we just need to write the code for
218 the pass itself.</p>
220 </div>
222 <!-- ======================================================================= -->
225 </div>
229 <p>Now that we have a way to compile our new pass, we just have to write it.
230 Start out with:</p>
235 <b>#include</b> "<a href="http://llvm.org/doxygen/raw__ostream_8h.html">llvm/Support/raw_ostream.h</a>"
236 </pre></div>
240 we are operating on <tt><a
242 and we will be doing some printing.</p>
247 </pre></div>
248 <p>... which is required because the functions from the include files
249 live in the llvm namespace.
250 </p>
256 </pre></div>
258 <p>... which starts out an anonymous namespace. Anonymous namespaces are to C++
259 what the "<tt>static</tt>" keyword is to C (at global scope). It makes the
260 things declared inside of the anonymous namespace only visible to the current
261 file. If you're not familiar with them, consult a decent C++ book for more
262 information.</p>
268 </pre></div><p>
272 The different builtin pass subclasses are described in detail <a
275 time.</p>
278 static char ID;
279 Hello() : FunctionPass(ID) {}
280 </pre></div><p>
282 <p> This declares pass identifier used by LLVM to identify pass. This allows LLVM to
283 avoid using expensive C++ runtime information.</p>
289 }
291 </pre></div>
294 which overloads an abstract virtual method inherited from <a
296 to do our thing, so we just print out our message with the name of each
297 function.</p>
300 char Hello::ID = 0;
301 </pre></div>
303 <p> We initialize pass ID here. LLVM uses ID's address to identify pass so
304 initialization value is not important.</p>
308 false /* Only looks at CFG */,
309 false /* Analysis Pass */);
311 </pre></div>
314 giving it a command line
316 Last two arguments describe its behavior.
317 If a pass walks CFG without modifying it then third argument is set to true.
318 If a pass is an analysis pass, for example dominator tree pass, then true
319 is supplied as fourth argument. </p>
326 <b>#include</b> "<a href="http://llvm.org/doxygen/raw__ostream_8h.html">llvm/Support/raw_ostream.h</a>"
333 static char ID;
334 Hello() : FunctionPass(ID) {}
339 }
340 };
342 char Hello::ID = 0;
344 }
346 </pre></div>
349 command in the local directory and you should get a new file
350 "<tt>Debug+Asserts/lib/Hello.so</tt>" under the top level directory of the LLVM
351 source tree (not in the local directory). Note that everything in this file is
352 contained in an anonymous namespace: this reflects the fact that passes are self
353 contained units that do not need external interfaces (although they can have
354 them) to be useful.</p>
356 </div>
358 <!-- ======================================================================= -->
361 </div>
365 <p>Now that you have a brand new shiny shared object file, we can use the
373 through our transformation like this (or course, any bitcode file will
374 work):</p>
378 Hello: __main
379 Hello: puts
380 Hello: main
381 </pre></div>
386 pass</a>). Because the hello pass does not modify the program in any
390 <p>To see what happened to the other string you registered, try running
394 $ opt -load ../../../Debug+Asserts/lib/Hello.so -help
395 OVERVIEW: llvm .bc -> .bc modular optimizer
399 OPTIONS:
400 Optimizations available:
401 ...
402 -funcresolve - Resolve Functions
403 -gcse - Global Common Subexpression Elimination
404 -globaldce - Dead Global Elimination
406 -indvars - Canonicalize Induction Variables
407 -inline - Function Integration/Inlining
408 -instcombine - Combine redundant instructions
409 ...
410 </pre></div>
412 <p>The pass name get added as the information string for your pass, giving some
414 would go ahead and make it do the cool transformations you want. Once you get
415 it all working and tested, it may become useful to find out how fast your pass
418 the execution time of your pass along with the other passes you queue up. For
419 example:</p>
422 $ opt -load ../../../Debug+Asserts/lib/Hello.so -hello -time-passes < hello.bc > /dev/null
423 Hello: __main
424 Hello: puts
425 Hello: main
426 ===============================================================================
427 ... Pass execution timing report ...
428 ===============================================================================
431 ---User Time--- --System Time-- --User+System-- ---Wall Time--- --- Pass Name ---
437 </pre></div>
439 <p>As you can see, our implementation above is pretty fast :). The additional
441 that the LLVM emitted by your pass is still valid and well formed LLVM, which
442 hasn't been broken somehow.</p>
444 <p>Now that you have seen the basics of the mechanics behind passes, we can talk
445 about some more details of how they work and how to use them.</p>
447 </div>
449 <!-- *********************************************************************** -->
452 </div>
453 <!-- *********************************************************************** -->
457 <p>One of the first things that you should do when designing a new pass is to
458 decide what class you should subclass for your pass. The <a
461 did not discuss why or when this should occur. Here we talk about the classes
462 available, from the most general to the most specific.</p>
465 specific</b> class possible, while still being able to meet the requirements
466 listed. This gives the LLVM Pass Infrastructure information necessary to
467 optimize how passes are run, so that the resultant compiler isn't unnecessarily
468 slow.</p>
470 </div>
472 <!-- ======================================================================= -->
475 </div>
481 class. This pass type is used for passes that do not have to be run, do not
482 change state, and never need to be updated. This is not a normal type of
483 transformation or analysis, but can provide information about the current
484 compiler configuration.</p>
486 <p>Although this pass class is very infrequently used, it is important for
487 providing information about the current target machine being compiled for, and
488 other static information that can affect the various transformations.</p>
493 </div>
495 <!-- ======================================================================= -->
498 </div>
504 class is the most general of all superclasses that you can use. Deriving from
506 referring to function bodies in no predictable order, or adding and removing
508 subclasses, no optimization can be done for their execution.</p>
510 <p>A module pass can use function level passes (e.g. dominators) using
511 the getAnalysis interface
513 function to retrieve analysis result for, if the function pass does not require
514 any module or immutable passes. Note that this can only be done for functions for which the
515 analysis ran, e.g. in the case of dominators you should only ask for the
516 DominatorTree for function definitions, not declarations.</p>
520 following signature:</p>
522 </div>
524 <!-- _______________________________________________________________________ -->
527 </div>
533 </pre></div>
536 It should return true if the module was modified by the transformation and
537 false otherwise.</p>
539 </div>
541 <!-- ======================================================================= -->
544 </div>
550 is used by passes that need to traverse the program bottom-up on the call graph
551 (callees before callers). Deriving from CallGraphSCCPass provides some
552 mechanics for building and traversing the CallGraph, but also allows the system
553 to optimize execution of CallGraphSCCPass's. If your pass meets the
554 requirements outlined below, and doesn't meet the requirements of a <tt><a
563 <ol>
566 than those in the current SCC and the direct callers and direct callees of the
567 SCC.</li>
570 to reflect any changes made to the program.</li>
573 though they may change the contents of an SCC.</li>
576 Module.</li>
580 </ol>
583 because it has to handle SCCs with more than one node in it. All of the virtual
584 methods described below should return true if they modified the program, or
585 false if they didn't.</p>
587 </div>
589 <!-- _______________________________________________________________________ -->
592 method</a>
593 </div>
599 </pre></div>
604 is designed to do simple initialization type of stuff that does not depend on
606 scheduled to overlap with any other pass executions (thus it should be very
607 fast).</p>
609 </div>
611 <!-- _______________________________________________________________________ -->
614 </div>
620 </pre></div>
623 should return true if the module was modified by the transformation, false
624 otherwise.</p>
626 </div>
628 <!-- _______________________________________________________________________ -->
632 </div>
638 </pre></div>
641 called when the pass framework has finished calling <a
643 program being compiled.</p>
645 </div>
647 <!-- ======================================================================= -->
650 </div>
656 subclasses do have a predictable, local behavior that can be expected by the
658 independent of all of the other functions in the program.
664 <ol>
668 <li>Maintain state across invocations of
670 </ol>
674 may overload three virtual methods to do their work. All of these methods
675 should return true if they modified the program, or false if they didn't.</p>
677 </div>
679 <!-- _______________________________________________________________________ -->
682 method</a>
683 </div>
689 </pre></div>
694 is designed to do simple initialization type of stuff that does not depend on
696 scheduled to overlap with any other pass executions (thus it should be very
697 fast).</p>
704 free functions that it needs, adding prototypes to the module if necessary.</p>
706 </div>
708 <!-- _______________________________________________________________________ -->
711 </div>
717 </pre></div><p>
720 the transformation or analysis work of your pass. As usual, a true value should
721 be returned if the function is modified.</p>
723 </div>
725 <!-- _______________________________________________________________________ -->
729 </div>
735 </pre></div>
738 called when the pass framework has finished calling <a
740 program being compiled.</p>
742 </div>
744 <!-- ======================================================================= -->
747 </div>
753 loop nest order such that outer most loop is processed last. </p>
758 do their work. All these methods should return true if they modified the
759 program, or false if they didn't. </p>
760 </div>
762 <!-- _______________________________________________________________________ -->
766 method</a>
767 </div>
773 </pre></div>
776 type of stuff that does not depend on the functions being processed. The
778 other pass executions (thus it should be very fast). LPPassManager
779 interface should be used to access Function or Module level analysis
780 information.</p>
782 </div>
785 <!-- _______________________________________________________________________ -->
788 </div>
794 </pre></div><p>
797 the transformation or analysis work of your pass. As usual, a true value should
799 should be used to update loop nest.</p>
801 </div>
803 <!-- _______________________________________________________________________ -->
806 </div>
812 </pre></div>
815 called when the pass framework has finished calling <a
817 program being compiled. </p>
819 </div>
821 <!-- ======================================================================= -->
824 </div>
829 but executes on each single entry single exit region in the function.
831 region is processed last. </p>
836 methods should return true if they modified the program, or false if they didn not.
837 </p>
838 </div>
840 <!-- _______________________________________________________________________ -->
844 method</a>
845 </div>
851 </pre></div>
854 type of stuff that does not depend on the functions being processed. The
856 other pass executions (thus it should be very fast). RPPassManager
857 interface should be used to access Function or Module level analysis
858 information.</p>
860 </div>
863 <!-- _______________________________________________________________________ -->
866 </div>
872 </pre></div><p>
875 the transformation or analysis work of your pass. As usual, a true value should
877 should be used to update region tree.</p>
879 </div>
881 <!-- _______________________________________________________________________ -->
884 </div>
890 </pre></div>
893 called when the pass framework has finished calling <a
895 program being compiled. </p>
897 </div>
901 <!-- ======================================================================= -->
904 </div>
910 their scope of inspection and modification to a single basic block at a time.
913 <ol>
915 <li>Maintain state across invocations of
918 <li>Any of the things forbidden for
920 </ol>
923 optimizations. They may override the same <a
926 href="#FunctionPass"><tt>FunctionPass</tt></a>'s have, but also have the following virtual methods that may also be implemented:</p>
928 </div>
930 <!-- _______________________________________________________________________ -->
934 </div>
940 </pre></div>
945 to do simple initialization that does not depend on the
947 scheduled to overlap with any other pass executions (thus it should be very
948 fast).</p>
950 </div>
952 <!-- _______________________________________________________________________ -->
955 </div>
961 </pre></div>
964 function is not allowed to inspect or modify basic blocks other than the
965 parameter, and are not allowed to modify the CFG. A true value must be returned
966 if the basic block is modified.</p>
968 </div>
970 <!-- _______________________________________________________________________ -->
973 method</a>
974 </div>
980 </pre></div>
983 called when the pass framework has finished calling <a
985 program being compiled. This can be used to perform per-function
986 finalization.</p>
988 </div>
990 <!-- ======================================================================= -->
993 </div>
998 executes on the machine-dependent representation of each LLVM function in the
999 program.</p>
1001 <p>Code generator passes are registered and initialized specially by
1004 commands.</p>
1011 <ol>
1012 <li>Modify or create any LLVM IR Instructions, BasicBlocks, Arguments,
1013 Functions, GlobalVariables, GlobalAliases, or Modules.</li>
1017 data)</li>
1018 </ol>
1020 </div>
1022 <!-- _______________________________________________________________________ -->
1026 </div>
1032 </pre></div>
1047 </div>
1049 <!-- *********************************************************************** -->
1052 </div>
1053 <!-- *********************************************************************** -->
1058 pass registration works, and discussed some of the reasons that it is used and
1059 what it does. Here we discuss how and why passes are registered.</p>
1062 template. The template parameter is the name of the pass that is to be used on
1063 the command line to specify that the pass should be added to a program (for
1066 programs, as
1069 <p>If you want your pass to be easily dumpable, you should
1072 </div>
1074 <!-- _______________________________________________________________________ -->
1077 </div>
1083 </pre></div>
1086 a human readable version of the analysis results. This is useful for debugging
1087 an analysis itself, as well as for other people to figure out how an analysis
1092 program that has been analyzed. Note however that this pointer may be null in
1094 debugger), so it should only be used to enhance debug output, it should not be
1095 depended on.</p>
1097 </div>
1099 <!-- *********************************************************************** -->
1102 </div>
1103 <!-- *********************************************************************** -->
1110 know how the passes interact with each other and what dependencies exist between
1111 the various passes. To track this, each pass can declare the set of passes that
1112 are required to be executed before the current pass, and the passes which are
1113 invalidated by the current pass.</p>
1115 <p>Typically this functionality is used to require that analysis results are
1116 computed before your pass is run. Running arbitrary transformation passes can
1117 invalidate the computed analysis results, which is what the invalidation set
1118 specifies. If a pass does not implement the <tt><a
1122 </div>
1124 <!-- _______________________________________________________________________ -->
1127 </div>
1133 </pre></div>
1136 invalidated sets may be specified for your transformation. The implementation
1137 should fill in the <tt><a
1139 object with information about which passes are required and not invalidated. To
1140 do this, a pass may call any of the following methods on the AnalysisUsage
1141 object:</p>
1142 </div>
1144 <!-- _______________________________________________________________________ -->
1146 <a name="AU::addRequired">The <tt>AnalysisUsage::addRequired<></tt> and <tt>AnalysisUsage::addRequiredTransitive<></tt> methods</a>
1147 </div>
1150 <p>
1151 If your pass requires a previous pass to be executed (an analysis for example),
1152 it can use one of these methods to arrange for it to be run before your pass.
1153 LLVM has many different types of analyses and passes that can be required,
1156 be no critical edges in the CFG when your pass has been run.
1157 </p>
1159 <p>
1160 Some analyses chain to other analyses to do their job. For example, an <a
1165 that the transitively required pass should be alive as long as the requiring
1166 pass is.
1167 </p>
1168 </div>
1170 <!-- _______________________________________________________________________ -->
1173 </div>
1176 <p>
1177 One of the jobs of the PassManager is to optimize how and when analyses are run.
1178 In particular, it attempts to avoid recomputing data unless it needs to. For
1179 this reason, passes are allowed to declare that they preserve (i.e., they don't
1180 invalidate) an existing analysis if it's available. For example, a simple
1181 constant folding pass would not modify the CFG, so it can't possibly affect the
1182 results of dominator analysis. By default, all passes are assumed to invalidate
1183 all others.
1184 </p>
1186 <p>
1190 not modify the LLVM program at all (which is true for analyses), and the
1192 instructions in the program but do not modify the CFG or terminator instructions
1193 (note that this property is implicitly set for <a
1195 </p>
1197 <p>
1200 and dominator related analyses if they exist, so it can preserve them, despite
1201 the fact that it hacks on the CFG.
1202 </p>
1203 </div>
1205 <!-- _______________________________________________________________________ -->
1208 </div>
1214 <b>void</b> <a href="http://llvm.org/doxygen/structLICM.html">LICM</a>::getAnalysisUsage(AnalysisUsage &AU) <b>const</b> {
1215 AU.setPreservesCFG();
1216 AU.addRequired<<a href="http://llvm.org/doxygen/classllvm_1_1LoopInfo.html">LoopInfo</a>>();
1217 }
1218 </pre></div>
1220 </div>
1222 <!-- _______________________________________________________________________ -->
1226 </div>
1231 your class, providing you with access to the passes that you declared that you
1233 method. It takes a single template argument that specifies which pass class you
1234 want, and returns a reference to that pass. For example:</p>
1237 bool LICM::runOnFunction(Function &F) {
1239 ...
1240 }
1241 </pre></div>
1243 <p>This method call returns a reference to the pass desired. You may get a
1244 runtime assertion failure if you attempt to get an analysis that you did not
1245 declare as required in your <a
1250 A module level pass can use function level analysis info using this interface.
1251 For example:</p>
1254 bool ModuleLevelPass::runOnModule(Module &M) {
1255 ...
1257 ...
1258 }
1259 </pre></div>
1261 <p>In above example, runOnFunction for DominatorTree is called by pass manager
1262 before returning a reference to the desired pass.</p>
1264 <p>
1265 If your pass is capable of updating analyses if they exist (e.g.,
1268 if it is active. For example:</p>
1271 ...
1274 }
1275 ...
1276 </pre></div>
1278 </div>
1280 <!-- *********************************************************************** -->
1283 </div>
1284 <!-- *********************************************************************** -->
1288 <p>Now that we understand the basics of how passes are defined, how they are
1289 used, and how they are required from other passes, it's time to get a little bit
1290 fancier. All of the pass relationships that we have seen so far are very
1291 simple: one pass depends on one other specific pass to be run before it can run.
1292 For many applications, this is great, for others, more flexibility is
1293 required.</p>
1295 <p>In particular, some analyses are defined such that there is a single simple
1296 interface to the analysis results, but multiple ways of calculating them.
1297 Consider alias analysis for example. The most trivial alias analysis returns
1298 "may alias" for any alias query. The most sophisticated analysis a
1299 flow-sensitive, context-sensitive interprocedural analysis that can take a
1300 significant amount of time to execute (and obviously, there is a lot of room
1301 between these two extremes for other implementations). To cleanly support
1302 situations like this, the LLVM Pass Infrastructure supports the notion of
1303 Analysis Groups.</p>
1305 </div>
1307 <!-- _______________________________________________________________________ -->
1310 </div>
1314 <p>An Analysis Group is a single simple interface that may be implemented by
1315 multiple different passes. Analysis Groups can be given human readable names
1317 class. An analysis group may have one or more implementations, one of which is
1320 <p>Analysis groups are used by client passes just like other passes are: the
1323 scans the available passes to see if any implementations of the analysis group
1324 are available. If none is available, the default implementation is created for
1329 passes, all analysis group implementations must be registered, and must use the
1331 implementation pool. Also, a default implementation of the interface
1337 analysis group. The default implementation of the alias analysis interface (the
1338 <tt><a
1340 pass) just does a few simple checks that don't require significant analysis to
1341 compute (such as: two different globals can never alias each other, etc).
1342 Passes that use the <tt><a
1344 interface (for example the <tt><a
1346 not care which implementation of alias analysis is actually provided, they just
1347 use the designated interface.</p>
1349 <p>From the user's perspective, commands work just like normal. Issuing the
1351 instantiated and added to the pass sequence. Issuing the command '<tt>opt
1354 hypothetical example) instead.</p>
1356 </div>
1358 <!-- _______________________________________________________________________ -->
1361 </div>
1367 implementations to the analysis group. First,
1368 an analysis group should be registered, with a human readable name
1369 provided for it.
1370 Unlike registration of passes, there is no command line argument to be specified
1374 <b>static</b> RegisterAnalysisGroup<<a href="http://llvm.org/doxygen/classllvm_1_1AliasAnalysis.html">AliasAnalysis</a>> A("<i>Alias Analysis</i>");
1375 </pre></div>
1377 <p>Once the analysis is registered, passes can declare that they are valid
1378 implementations of the interface by using the following code:</p>
1383 INITIALIZE_AG_PASS(FancyAA, <a href="http://llvm.org/doxygen/classllvm_1_1AliasAnalysis.html">AliasAnalysis</a>, "<i>somefancyaa</i>",
1384 "<i>A more complex alias analysis implementation</i>",
1388 );
1389 }
1390 </pre></div>
1394 to "join" the <tt><a href="http://llvm.org/doxygen/classllvm_1_1AliasAnalysis.html">AliasAnalysis</a></tt>
1395 analysis group. Every implementation of an analysis group should join using
1396 this macro.</p>
1401 INITIALIZE_AG_PASS(BasicAA, <a href="http://llvm.org/doxygen/classllvm_1_1AliasAnalysis.html">AliasAnalysis</a>, "<i>basicaa</i>",
1402 "<i>Basic Alias Analysis (default AA impl)</i>",
1406 );
1407 }
1408 </pre></div>
1410 <p>Here we show how the default implementation is specified (using the final
1412 one default implementation available at all times for an Analysis Group to be
1414 Here we declare that the
1415 <tt><a href="http://llvm.org/doxygen/structBasicAliasAnalysis.html">BasicAliasAnalysis</a></tt>
1416 pass is the default implementation for the interface.</p>
1418 </div>
1420 <!-- *********************************************************************** -->
1423 </div>
1424 <!-- *********************************************************************** -->
1429 class is designed to be an easy way to expose various success
1430 metrics from passes. These statistics are printed at the end of a
1431 run, when the -stats command line option is enabled on the command
1432 line. See the <a href="http://llvm.org/docs/ProgrammersManual.html#Statistic">Statistics section</a> in the Programmer's Manual for details.
1434 </div>
1437 <!-- *********************************************************************** -->
1440 </div>
1441 <!-- *********************************************************************** -->
1447 <a
1450 are set up correctly, and then schedules passes to run efficiently. All of the
1452 passes.</p>
1455 time of a series of passes:</p>
1457 <ol>
1459 recomputing analysis results as much as possible. This means keeping track of
1460 which analyses are available already, which analyses get invalidated, and which
1461 analyses are needed to be run for a pass. An important part of work is that the
1464 results as soon as they are no longer needed.</li>
1468 of a series of passes by pipelining the passes together. This means that, given
1471 the first function, then all of the <a
1473 etc... until the entire program has been run through the passes.
1475 <p>This improves the cache behavior of the compiler, because it is only touching
1476 the LLVM program representation for a single function at a time, instead of
1477 traversing the entire program. It reduces the memory consumption of compiler,
1478 because, for example, only one <a
1480 needs to be calculated at a time. This also makes it possible to implement
1481 some <a
1484 </ol>
1487 much information it has about the behaviors of the passes it is scheduling. For
1488 example, the "preserved" set is intentionally conservative in the face of an
1490 Not implementing when it should be implemented will have the effect of not
1491 allowing any analysis results to live across the execution of your pass.</p>
1494 options that is useful for debugging pass execution, seeing how things work, and
1495 diagnosing when you should be preserving more analyses than you currently are
1504 $ opt -load ../../../Debug+Asserts/lib/Hello.so -gcse -licm --debug-pass=Structure < hello.bc > /dev/null
1505 Module Pass Manager
1506 Function Pass Manager
1507 Dominator Set Construction
1508 Immediate Dominators Construction
1509 Global Common Subexpression Elimination
1510 -- Immediate Dominators Construction
1511 -- Global Common Subexpression Elimination
1512 Natural Loop Construction
1513 Loop Invariant Code Motion
1514 -- Natural Loop Construction
1515 -- Loop Invariant Code Motion
1516 Module Verifier
1517 -- Dominator Set Construction
1518 -- Module Verifier
1519 Bitcode Writer
1520 --Bitcode Writer
1521 </pre></div>
1523 <p>This output shows us when passes are constructed and when the analysis
1525 GCSE uses dominator and immediate dominator information to do its job. The LICM
1526 pass uses natural loop information, which uses dominator sets, but not immediate
1527 dominators. Because immediate dominators are no longer useful after the GCSE
1528 pass, it is immediately destroyed. The dominator sets are then reused to
1529 compute natural loop information, which is then used by the LICM pass.</p>
1531 <p>After the LICM pass, the module verifier runs (which is automatically added
1533 resultant LLVM code is well formed. After it finishes, the dominator set
1534 information is destroyed, after being computed once, and shared by three
1535 passes.</p>
1541 $ opt -load ../../../Debug+Asserts/lib/Hello.so -gcse -hello -licm --debug-pass=Structure < hello.bc > /dev/null
1542 Module Pass Manager
1543 Function Pass Manager
1544 Dominator Set Construction
1545 Immediate Dominators Construction
1546 Global Common Subexpression Elimination
1548 -- Immediate Dominators Construction
1549 -- Global Common Subexpression Elimination
1550 <b> Hello World Pass
1551 -- Hello World Pass
1552 Dominator Set Construction</b>
1553 Natural Loop Construction
1554 Loop Invariant Code Motion
1555 -- Natural Loop Construction
1556 -- Loop Invariant Code Motion
1557 Module Verifier
1558 -- Dominator Set Construction
1559 -- Module Verifier
1560 Bitcode Writer
1561 --Bitcode Writer
1562 Hello: __main
1563 Hello: puts
1564 Hello: main
1565 </pre></div>
1568 Dominator Set pass, even though it doesn't modify the code at all! To fix this,
1569 we need to add the following <a
1575 AU.setPreservesAll();
1576 }
1577 </pre></div>
1582 $ opt -load ../../../Debug+Asserts/lib/Hello.so -gcse -hello -licm --debug-pass=Structure < hello.bc > /dev/null
1583 Pass Arguments: -gcse -hello -licm
1584 Module Pass Manager
1585 Function Pass Manager
1586 Dominator Set Construction
1587 Immediate Dominators Construction
1588 Global Common Subexpression Elimination
1589 -- Immediate Dominators Construction
1590 -- Global Common Subexpression Elimination
1591 Hello World Pass
1592 -- Hello World Pass
1593 Natural Loop Construction
1594 Loop Invariant Code Motion
1595 -- Loop Invariant Code Motion
1596 -- Natural Loop Construction
1597 Module Verifier
1598 -- Dominator Set Construction
1599 -- Module Verifier
1600 Bitcode Writer
1601 --Bitcode Writer
1602 Hello: __main
1603 Hello: puts
1604 Hello: main
1605 </pre></div>
1607 <p>Which shows that we don't accidentally invalidate dominator information
1608 anymore, and therefore do not have to compute it twice.</p>
1610 </div>
1612 <!-- _______________________________________________________________________ -->
1615 </div>
1621 </pre></div>
1624 results, and how long to keep them around for. Because the lifetime of the pass
1625 object itself is effectively the entire duration of the compilation process, we
1626 need some way to free analysis results when they are no longer useful. The
1629 <p>If you are writing an analysis or any other pass that retains a significant
1630 amount of state (for use by another pass which "requires" your pass and uses the
1636 </div>
1638 <!-- *********************************************************************** -->
1641 </div>
1642 <!-- *********************************************************************** -->
1647 both for the purposes of distribution, and for regulating the resident code size
1648 when running on the target system. Therefore, it becomes desirable to
1649 selectively use some passes, while omitting others and maintain the flexibility
1650 to change configurations later on. You want to be able to do all this, and,
1651 provide feedback to the user. This is where pass registration comes into
1652 play.</p>
1654 <p>The fundamental mechanisms for pass registration are the
1660 communicates additions and deletions to the command line interface.</p>
1663 information provided about a particular pass. This information includes the
1664 command line name, the command help string and the address of the function used
1665 to create an instance of the pass. A global static constructor of one of these
1668 in the tool will be registered at start up. A dynamically loaded pass will
1669 register on load and unregister at unload.</p>
1671 </div>
1673 <!-- _______________________________________________________________________ -->
1676 </div>
1680 <p>There are predefined registries to track instruction scheduling
1683 allocator machine pass.</p>
1685 <p>Implement your register allocator machine pass. In your register allocator
1686 .cpp file add the following include;</p>
1689 #include "llvm/CodeGen/RegAllocRegistry.h"
1690 </pre></div>
1692 <p>Also in your register allocator .cpp file, define a creator function in the
1693 form; </p>
1696 FunctionPass *createMyRegisterAllocator() {
1697 return new MyRegisterAllocator();
1698 }
1699 </pre></div>
1701 <p>Note that the signature of this function should match the type of
1706 static RegisterRegAlloc myRegAlloc("myregalloc",
1707 " my register allocator help string",
1708 createMyRegisterAllocator);
1709 </pre></div>
1711 <p>Note the two spaces prior to the help string produces a tidy result on the
1712 -help query.</p>
1715 $ llc -help
1716 ...
1717 -regalloc - Register allocator to use (default=linearscan)
1718 =linearscan - linear scan register allocator
1719 =local - local register allocator
1720 =simple - simple register allocator
1721 =myregalloc - my register allocator help string
1722 ...
1723 </pre></div>
1726 an option. Registering instruction schedulers is similar except use the
1731 <p>To force the load/linking of your register allocator into the llc/lli tools,
1735 </div>
1738 <!-- _______________________________________________________________________ -->
1741 </div>
1745 <p>The easiest way to get started is to clone one of the existing registries; we
1749 <p>Then you need to declare the registry. Example: if your pass registry is
1753 MachinePassRegistry RegisterMyPasses::Registry;
1754 </pre></div>
1759 cl::opt<RegisterMyPasses::FunctionPassCtor, false,
1761 MyPassOpt("mypass",
1762 cl::init(&createDefaultMyPass),
1763 cl::desc("my pass option help"));
1764 </pre></div>
1767 creator.</p>
1769 </div>
1771 <!-- *********************************************************************** -->
1774 </div>
1775 <!-- *********************************************************************** -->
1779 <p>Unfortunately, using GDB with dynamically loaded passes is not as easy as it
1780 should be. First of all, you can't set a breakpoint in a shared object that has
1781 not been loaded yet, and second of all there are problems with inlined functions
1782 in shared objects. Here are some suggestions to debugging your pass with
1783 GDB.</p>
1785 <p>For sake of discussion, I'm going to assume that you are debugging a
1787 on that.</p>
1789 </div>
1791 <!-- _______________________________________________________________________ -->
1794 </div>
1802 GNU gdb 5.0
1803 Copyright 2000 Free Software Foundation, Inc.
1804 GDB is free software, covered by the GNU General Public License, and you are
1805 welcome to change it and/or distribute copies of it under certain conditions.
1806 Type "show copying" to see the conditions.
1807 There is absolutely no warranty for GDB. Type "show warranty" for details.
1808 This GDB was configured as "sparc-sun-solaris2.6"...
1809 (gdb)
1810 </pre></div>
1813 time to load. Be patient. Since we cannot set a breakpoint in our pass yet
1814 (the shared object isn't loaded until runtime), we must execute the process, and
1815 have it stop before it invokes our pass, but after it has loaded the shared
1816 object. The most foolproof way of doing this is to set a breakpoint in
1818 want:</p>
1824 Starting program: opt test.bc -load $(LLVMTOP)/llvm/Debug+Asserts/lib/[libname].so -[passoption]
1827 (gdb)
1828 </pre></div>
1831 now free to set breakpoints in your pass so that you can trace through execution
1832 or do other standard debugging stuff.</p>
1834 </div>
1836 <!-- _______________________________________________________________________ -->
1839 </div>
1843 <p>Once you have the basics down, there are a couple of problems that GDB has,
1844 some with solutions, some without.</p>
1846 <ul>
1847 <li>Inline functions have bogus stack information. In general, GDB does a
1848 pretty good job getting stack traces and stepping through inline functions.
1849 When a pass is dynamically loaded however, it somehow completely loses this
1850 capability. The only solution I know of is to de-inline a function (move it
1851 from the body of a class to a .cpp file).</li>
1853 <li>Restarting the program breaks breakpoints. After following the information
1854 above, you have succeeded in getting some breakpoints planted in your pass. Nex
1856 and you start getting errors about breakpoints being unsettable. The only way I
1858 already set in your pass, run the program, and re-set the breakpoints once
1861 </ul>
1863 <p>Hopefully these tips will help with common case debugging situations. If
1864 you'd like to contribute some tips of your own, just contact <a
1867 </div>
1869 <!-- *********************************************************************** -->
1872 </div>
1873 <!-- *********************************************************************** -->
1877 <p>Although the LLVM Pass Infrastructure is very capable as it stands, and does
1878 some nifty stuff, there are things we'd like to add in the future. Here is
1879 where we are going:</p>
1881 </div>
1883 <!-- _______________________________________________________________________ -->
1886 </div>
1890 <p>Multiple CPU machines are becoming more common and compilation can never be
1891 fast enough: obviously we should allow for a multithreaded compiler. Because of
1892 the semantics defined for passes above (specifically they cannot maintain state
1895 to create multiple instances of each pass object, and allow the separate
1896 instances to be hacking on different parts of the program at the same time.</p>
1898 <p>This implementation would prevent each of the passes from having to implement
1899 multithreaded constructs, requiring only the LLVM core to have locking in a few
1900 places (for global resources). Although this is a simple extension, we simply
1901 haven't had time (or multiprocessor machines, thus a reason) to implement this.
1902 Despite that, we have kept the LLVM passes SMP ready, and you should too.</p>
1904 </div>
1906 <!-- *********************************************************************** -->
1907 <hr>
1908 <address>
1916 Last modified: $Date$
1917 </address>
1919 </body>
1920 </html>