| blob | edc8631d0abcdc3638b157d3621fa102da02ce55 |
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>Now that we have the build scripts set up, we just need to write the code for
215 the pass itself.</p>
217 </div>
219 <!-- ======================================================================= -->
222 </div>
226 <p>Now that we have a way to compile our new pass, we just have to write it.
227 Start out with:</p>
232 <b>#include</b> "<a href="http://llvm.org/doxygen/raw__ostream_8h.html">llvm/Support/raw_ostream.h</a>"
233 </pre></div>
237 we are operating on <tt><a
239 and we will be doing some printing.</p>
244 </pre></div>
245 <p>... which is required because the functions from the include files
246 live in the llvm namespace.
247 </p>
253 </pre></div>
255 <p>... which starts out an anonymous namespace. Anonymous namespaces are to C++
256 what the "<tt>static</tt>" keyword is to C (at global scope). It makes the
257 things declared inside of the anonymous namespace only visible to the current
258 file. If you're not familiar with them, consult a decent C++ book for more
259 information.</p>
265 </pre></div><p>
269 The different builtin pass subclasses are described in detail <a
272 time.</p>
275 static char ID;
276 Hello() : FunctionPass(ID) {}
277 </pre></div><p>
279 <p> This declares pass identifier used by LLVM to identify pass. This allows LLVM to
280 avoid using expensive C++ runtime information.</p>
286 }
288 </pre></div>
291 which overloads an abstract virtual method inherited from <a
293 to do our thing, so we just print out our message with the name of each
294 function.</p>
297 char Hello::ID = 0;
298 </pre></div>
300 <p> We initialize pass ID here. LLVM uses ID's address to identify pass so
301 initialization value is not important.</p>
305 false /* Only looks at CFG */,
306 false /* Analysis Pass */);
308 </pre></div>
311 giving it a command line
313 Last two arguments describe its behavior.
314 If a pass walks CFG without modifying it then third argument is set to true.
315 If a pass is an analysis pass, for example dominator tree pass, then true
316 is supplied as fourth argument. </p>
323 <b>#include</b> "<a href="http://llvm.org/doxygen/raw__ostream_8h.html">llvm/Support/raw_ostream.h</a>"
330 static char ID;
331 Hello() : FunctionPass(ID) {}
336 }
337 };
339 char Hello::ID = 0;
341 }
343 </pre></div>
346 command in the local directory and you should get a new file
347 "<tt>Debug+Asserts/lib/Hello.so</tt>" under the top level directory of the LLVM
348 source tree (not in the local directory). Note that everything in this file is
349 contained in an anonymous namespace: this reflects the fact that passes are self
350 contained units that do not need external interfaces (although they can have
351 them) to be useful.</p>
353 </div>
355 <!-- ======================================================================= -->
358 </div>
362 <p>Now that you have a brand new shiny shared object file, we can use the
370 through our transformation like this (or course, any bitcode file will
371 work):</p>
375 Hello: __main
376 Hello: puts
377 Hello: main
378 </pre></div>
383 pass</a>). Because the hello pass does not modify the program in any
387 <p>To see what happened to the other string you registered, try running
391 $ opt -load ../../../Debug+Asserts/lib/Hello.so -help
392 OVERVIEW: llvm .bc -> .bc modular optimizer
396 OPTIONS:
397 Optimizations available:
398 ...
399 -funcresolve - Resolve Functions
400 -gcse - Global Common Subexpression Elimination
401 -globaldce - Dead Global Elimination
403 -indvars - Canonicalize Induction Variables
404 -inline - Function Integration/Inlining
405 -instcombine - Combine redundant instructions
406 ...
407 </pre></div>
409 <p>The pass name get added as the information string for your pass, giving some
411 would go ahead and make it do the cool transformations you want. Once you get
412 it all working and tested, it may become useful to find out how fast your pass
415 the execution time of your pass along with the other passes you queue up. For
416 example:</p>
419 $ opt -load ../../../Debug+Asserts/lib/Hello.so -hello -time-passes < hello.bc > /dev/null
420 Hello: __main
421 Hello: puts
422 Hello: main
423 ===============================================================================
424 ... Pass execution timing report ...
425 ===============================================================================
428 ---User Time--- --System Time-- --User+System-- ---Wall Time--- --- Pass Name ---
434 </pre></div>
436 <p>As you can see, our implementation above is pretty fast :). The additional
438 that the LLVM emitted by your pass is still valid and well formed LLVM, which
439 hasn't been broken somehow.</p>
441 <p>Now that you have seen the basics of the mechanics behind passes, we can talk
442 about some more details of how they work and how to use them.</p>
444 </div>
446 <!-- *********************************************************************** -->
449 </div>
450 <!-- *********************************************************************** -->
454 <p>One of the first things that you should do when designing a new pass is to
455 decide what class you should subclass for your pass. The <a
458 did not discuss why or when this should occur. Here we talk about the classes
459 available, from the most general to the most specific.</p>
462 specific</b> class possible, while still being able to meet the requirements
463 listed. This gives the LLVM Pass Infrastructure information necessary to
464 optimize how passes are run, so that the resultant compiler isn't unnecessarily
465 slow.</p>
467 </div>
469 <!-- ======================================================================= -->
472 </div>
478 class. This pass type is used for passes that do not have to be run, do not
479 change state, and never need to be updated. This is not a normal type of
480 transformation or analysis, but can provide information about the current
481 compiler configuration.</p>
483 <p>Although this pass class is very infrequently used, it is important for
484 providing information about the current target machine being compiled for, and
485 other static information that can affect the various transformations.</p>
490 </div>
492 <!-- ======================================================================= -->
495 </div>
501 class is the most general of all superclasses that you can use. Deriving from
503 referring to function bodies in no predictable order, or adding and removing
505 subclasses, no optimization can be done for their execution.</p>
507 <p>A module pass can use function level passes (e.g. dominators) using
508 the getAnalysis interface
510 function to retrieve analysis result for, if the function pass does not require
511 any module or immutable passes. Note that this can only be done for functions for which the
512 analysis ran, e.g. in the case of dominators you should only ask for the
513 DominatorTree for function definitions, not declarations.</p>
517 following signature:</p>
519 </div>
521 <!-- _______________________________________________________________________ -->
524 </div>
530 </pre></div>
533 It should return true if the module was modified by the transformation and
534 false otherwise.</p>
536 </div>
538 <!-- ======================================================================= -->
541 </div>
547 is used by passes that need to traverse the program bottom-up on the call graph
548 (callees before callers). Deriving from CallGraphSCCPass provides some
549 mechanics for building and traversing the CallGraph, but also allows the system
550 to optimize execution of CallGraphSCCPass's. If your pass meets the
551 requirements outlined below, and doesn't meet the requirements of a <tt><a
560 <ol>
563 than those in the current SCC and the direct callers and direct callees of the
564 SCC.</li>
567 to reflect any changes made to the program.</li>
570 though they may change the contents of an SCC.</li>
573 Module.</li>
577 </ol>
580 because it has to handle SCCs with more than one node in it. All of the virtual
581 methods described below should return true if they modified the program, or
582 false if they didn't.</p>
584 </div>
586 <!-- _______________________________________________________________________ -->
589 method</a>
590 </div>
596 </pre></div>
601 is designed to do simple initialization type of stuff that does not depend on
603 scheduled to overlap with any other pass executions (thus it should be very
604 fast).</p>
606 </div>
608 <!-- _______________________________________________________________________ -->
611 </div>
617 </pre></div>
620 should return true if the module was modified by the transformation, false
621 otherwise.</p>
623 </div>
625 <!-- _______________________________________________________________________ -->
629 </div>
635 </pre></div>
638 called when the pass framework has finished calling <a
640 program being compiled.</p>
642 </div>
644 <!-- ======================================================================= -->
647 </div>
653 subclasses do have a predictable, local behavior that can be expected by the
655 independent of all of the other functions in the program.
661 <ol>
665 <li>Maintain state across invocations of
667 </ol>
671 may overload three virtual methods to do their work. All of these methods
672 should return true if they modified the program, or false if they didn't.</p>
674 </div>
676 <!-- _______________________________________________________________________ -->
679 method</a>
680 </div>
686 </pre></div>
691 is designed to do simple initialization type of stuff that does not depend on
693 scheduled to overlap with any other pass executions (thus it should be very
694 fast).</p>
701 free functions that it needs, adding prototypes to the module if necessary.</p>
703 </div>
705 <!-- _______________________________________________________________________ -->
708 </div>
714 </pre></div><p>
717 the transformation or analysis work of your pass. As usual, a true value should
718 be returned if the function is modified.</p>
720 </div>
722 <!-- _______________________________________________________________________ -->
726 </div>
732 </pre></div>
735 called when the pass framework has finished calling <a
737 program being compiled.</p>
739 </div>
741 <!-- ======================================================================= -->
744 </div>
750 loop nest order such that outer most loop is processed last. </p>
755 do their work. All these methods should return true if they modified the
756 program, or false if they didn't. </p>
757 </div>
759 <!-- _______________________________________________________________________ -->
763 method</a>
764 </div>
770 </pre></div>
773 type of stuff that does not depend on the functions being processed. The
775 other pass executions (thus it should be very fast). LPPassManager
776 interface should be used to access Function or Module level analysis
777 information.</p>
779 </div>
782 <!-- _______________________________________________________________________ -->
785 </div>
791 </pre></div><p>
794 the transformation or analysis work of your pass. As usual, a true value should
796 should be used to update loop nest.</p>
798 </div>
800 <!-- _______________________________________________________________________ -->
803 </div>
809 </pre></div>
812 called when the pass framework has finished calling <a
814 program being compiled. </p>
816 </div>
818 <!-- ======================================================================= -->
821 </div>
826 but executes on each single entry single exit region in the function.
828 region is processed last. </p>
833 methods should return true if they modified the program, or false if they didn not.
834 </p>
835 </div>
837 <!-- _______________________________________________________________________ -->
841 method</a>
842 </div>
848 </pre></div>
851 type of stuff that does not depend on the functions being processed. The
853 other pass executions (thus it should be very fast). RPPassManager
854 interface should be used to access Function or Module level analysis
855 information.</p>
857 </div>
860 <!-- _______________________________________________________________________ -->
863 </div>
869 </pre></div><p>
872 the transformation or analysis work of your pass. As usual, a true value should
874 should be used to update region tree.</p>
876 </div>
878 <!-- _______________________________________________________________________ -->
881 </div>
887 </pre></div>
890 called when the pass framework has finished calling <a
892 program being compiled. </p>
894 </div>
898 <!-- ======================================================================= -->
901 </div>
907 their scope of inspection and modification to a single basic block at a time.
910 <ol>
912 <li>Maintain state across invocations of
915 <li>Any of the things forbidden for
917 </ol>
920 optimizations. They may override the same <a
923 href="#FunctionPass"><tt>FunctionPass</tt></a>'s have, but also have the following virtual methods that may also be implemented:</p>
925 </div>
927 <!-- _______________________________________________________________________ -->
931 </div>
937 </pre></div>
942 to do simple initialization that does not depend on the
944 scheduled to overlap with any other pass executions (thus it should be very
945 fast).</p>
947 </div>
949 <!-- _______________________________________________________________________ -->
952 </div>
958 </pre></div>
961 function is not allowed to inspect or modify basic blocks other than the
962 parameter, and are not allowed to modify the CFG. A true value must be returned
963 if the basic block is modified.</p>
965 </div>
967 <!-- _______________________________________________________________________ -->
970 method</a>
971 </div>
977 </pre></div>
980 called when the pass framework has finished calling <a
982 program being compiled. This can be used to perform per-function
983 finalization.</p>
985 </div>
987 <!-- ======================================================================= -->
990 </div>
995 executes on the machine-dependent representation of each LLVM function in the
996 program.</p>
998 <p>Code generator passes are registered and initialized specially by
1001 commands.</p>
1008 <ol>
1009 <li>Modify or create any LLVM IR Instructions, BasicBlocks, Arguments,
1010 Functions, GlobalVariables, GlobalAliases, or Modules.</li>
1014 data)</li>
1015 </ol>
1017 </div>
1019 <!-- _______________________________________________________________________ -->
1023 </div>
1029 </pre></div>
1044 </div>
1046 <!-- *********************************************************************** -->
1049 </div>
1050 <!-- *********************************************************************** -->
1055 pass registration works, and discussed some of the reasons that it is used and
1056 what it does. Here we discuss how and why passes are registered.</p>
1059 template. The template parameter is the name of the pass that is to be used on
1060 the command line to specify that the pass should be added to a program (for
1063 programs, as
1066 <p>If you want your pass to be easily dumpable, you should
1069 </div>
1071 <!-- _______________________________________________________________________ -->
1074 </div>
1080 </pre></div>
1083 a human readable version of the analysis results. This is useful for debugging
1084 an analysis itself, as well as for other people to figure out how an analysis
1089 program that has been analyzed. Note however that this pointer may be null in
1091 debugger), so it should only be used to enhance debug output, it should not be
1092 depended on.</p>
1094 </div>
1096 <!-- *********************************************************************** -->
1099 </div>
1100 <!-- *********************************************************************** -->
1107 know how the passes interact with each other and what dependencies exist between
1108 the various passes. To track this, each pass can declare the set of passes that
1109 are required to be executed before the current pass, and the passes which are
1110 invalidated by the current pass.</p>
1112 <p>Typically this functionality is used to require that analysis results are
1113 computed before your pass is run. Running arbitrary transformation passes can
1114 invalidate the computed analysis results, which is what the invalidation set
1115 specifies. If a pass does not implement the <tt><a
1119 </div>
1121 <!-- _______________________________________________________________________ -->
1124 </div>
1130 </pre></div>
1133 invalidated sets may be specified for your transformation. The implementation
1134 should fill in the <tt><a
1136 object with information about which passes are required and not invalidated. To
1137 do this, a pass may call any of the following methods on the AnalysisUsage
1138 object:</p>
1139 </div>
1141 <!-- _______________________________________________________________________ -->
1143 <a name="AU::addRequired">The <tt>AnalysisUsage::addRequired<></tt> and <tt>AnalysisUsage::addRequiredTransitive<></tt> methods</a>
1144 </div>
1147 <p>
1148 If your pass requires a previous pass to be executed (an analysis for example),
1149 it can use one of these methods to arrange for it to be run before your pass.
1150 LLVM has many different types of analyses and passes that can be required,
1153 be no critical edges in the CFG when your pass has been run.
1154 </p>
1156 <p>
1157 Some analyses chain to other analyses to do their job. For example, an <a
1162 that the transitively required pass should be alive as long as the requiring
1163 pass is.
1164 </p>
1165 </div>
1167 <!-- _______________________________________________________________________ -->
1170 </div>
1173 <p>
1174 One of the jobs of the PassManager is to optimize how and when analyses are run.
1175 In particular, it attempts to avoid recomputing data unless it needs to. For
1176 this reason, passes are allowed to declare that they preserve (i.e., they don't
1177 invalidate) an existing analysis if it's available. For example, a simple
1178 constant folding pass would not modify the CFG, so it can't possibly affect the
1179 results of dominator analysis. By default, all passes are assumed to invalidate
1180 all others.
1181 </p>
1183 <p>
1187 not modify the LLVM program at all (which is true for analyses), and the
1189 instructions in the program but do not modify the CFG or terminator instructions
1190 (note that this property is implicitly set for <a
1192 </p>
1194 <p>
1197 and dominator related analyses if they exist, so it can preserve them, despite
1198 the fact that it hacks on the CFG.
1199 </p>
1200 </div>
1202 <!-- _______________________________________________________________________ -->
1205 </div>
1210 <i>// This is an example implementation from an analysis, which does not modify
1211 // the program at all, yet has a prerequisite.</i>
1212 <b>void</b> <a href="http://llvm.org/doxygen/classllvm_1_1PostDominanceFrontier.html">PostDominanceFrontier</a>::getAnalysisUsage(AnalysisUsage &AU) <b>const</b> {
1213 AU.setPreservesAll();
1214 AU.addRequired<<a href="http://llvm.org/doxygen/classllvm_1_1PostDominatorTree.html">PostDominatorTree</a>>();
1215 }
1216 </pre></div>
1222 <b>void</b> <a href="http://llvm.org/doxygen/structLICM.html">LICM</a>::getAnalysisUsage(AnalysisUsage &AU) <b>const</b> {
1223 AU.setPreservesCFG();
1224 AU.addRequired<<a href="http://llvm.org/doxygen/classllvm_1_1LoopInfo.html">LoopInfo</a>>();
1225 }
1226 </pre></div>
1228 </div>
1230 <!-- _______________________________________________________________________ -->
1234 </div>
1239 your class, providing you with access to the passes that you declared that you
1241 method. It takes a single template argument that specifies which pass class you
1242 want, and returns a reference to that pass. For example:</p>
1245 bool LICM::runOnFunction(Function &F) {
1247 ...
1248 }
1249 </pre></div>
1251 <p>This method call returns a reference to the pass desired. You may get a
1252 runtime assertion failure if you attempt to get an analysis that you did not
1253 declare as required in your <a
1258 A module level pass can use function level analysis info using this interface.
1259 For example:</p>
1262 bool ModuleLevelPass::runOnModule(Module &M) {
1263 ...
1265 ...
1266 }
1267 </pre></div>
1269 <p>In above example, runOnFunction for DominatorTree is called by pass manager
1270 before returning a reference to the desired pass.</p>
1272 <p>
1273 If your pass is capable of updating analyses if they exist (e.g.,
1276 if it is active. For example:</p>
1279 ...
1282 }
1283 ...
1284 </pre></div>
1286 </div>
1288 <!-- *********************************************************************** -->
1291 </div>
1292 <!-- *********************************************************************** -->
1296 <p>Now that we understand the basics of how passes are defined, how they are
1297 used, and how they are required from other passes, it's time to get a little bit
1298 fancier. All of the pass relationships that we have seen so far are very
1299 simple: one pass depends on one other specific pass to be run before it can run.
1300 For many applications, this is great, for others, more flexibility is
1301 required.</p>
1303 <p>In particular, some analyses are defined such that there is a single simple
1304 interface to the analysis results, but multiple ways of calculating them.
1305 Consider alias analysis for example. The most trivial alias analysis returns
1306 "may alias" for any alias query. The most sophisticated analysis a
1307 flow-sensitive, context-sensitive interprocedural analysis that can take a
1308 significant amount of time to execute (and obviously, there is a lot of room
1309 between these two extremes for other implementations). To cleanly support
1310 situations like this, the LLVM Pass Infrastructure supports the notion of
1311 Analysis Groups.</p>
1313 </div>
1315 <!-- _______________________________________________________________________ -->
1318 </div>
1322 <p>An Analysis Group is a single simple interface that may be implemented by
1323 multiple different passes. Analysis Groups can be given human readable names
1325 class. An analysis group may have one or more implementations, one of which is
1328 <p>Analysis groups are used by client passes just like other passes are: the
1331 scans the available passes to see if any implementations of the analysis group
1332 are available. If none is available, the default implementation is created for
1337 passes, all analysis group implementations must be registered, and must use the
1339 implementation pool. Also, a default implementation of the interface
1345 analysis group. The default implementation of the alias analysis interface (the
1346 <tt><a
1348 pass) just does a few simple checks that don't require significant analysis to
1349 compute (such as: two different globals can never alias each other, etc).
1350 Passes that use the <tt><a
1352 interface (for example the <tt><a
1354 not care which implementation of alias analysis is actually provided, they just
1355 use the designated interface.</p>
1357 <p>From the user's perspective, commands work just like normal. Issuing the
1359 instantiated and added to the pass sequence. Issuing the command '<tt>opt
1362 hypothetical example) instead.</p>
1364 </div>
1366 <!-- _______________________________________________________________________ -->
1369 </div>
1375 implementations to the analysis group. First,
1376 an analysis group should be registered, with a human readable name
1377 provided for it.
1378 Unlike registration of passes, there is no command line argument to be specified
1382 <b>static</b> RegisterAnalysisGroup<<a href="http://llvm.org/doxygen/classllvm_1_1AliasAnalysis.html">AliasAnalysis</a>> A("<i>Alias Analysis</i>");
1383 </pre></div>
1385 <p>Once the analysis is registered, passes can declare that they are valid
1386 implementations of the interface by using the following code:</p>
1391 INITIALIZE_AG_PASS(FancyAA, <a href="http://llvm.org/doxygen/classllvm_1_1AliasAnalysis.html">AliasAnalysis</a>, "<i>somefancyaa</i>",
1392 "<i>A more complex alias analysis implementation</i>",
1396 );
1397 }
1398 </pre></div>
1402 to "join" the <tt><a href="http://llvm.org/doxygen/classllvm_1_1AliasAnalysis.html">AliasAnalysis</a></tt>
1403 analysis group. Every implementation of an analysis group should join using
1404 this macro.</p>
1409 INITIALIZE_AG_PASS(BasicAA, <a href="http://llvm.org/doxygen/classllvm_1_1AliasAnalysis.html">AliasAnalysis</a>, "<i>basicaa</i>",
1410 "<i>Basic Alias Analysis (default AA impl)</i>",
1414 );
1415 }
1416 </pre></div>
1418 <p>Here we show how the default implementation is specified (using the final
1420 one default implementation available at all times for an Analysis Group to be
1422 Here we declare that the
1423 <tt><a href="http://llvm.org/doxygen/structBasicAliasAnalysis.html">BasicAliasAnalysis</a></tt>
1424 pass is the default implementation for the interface.</p>
1426 </div>
1428 <!-- *********************************************************************** -->
1431 </div>
1432 <!-- *********************************************************************** -->
1437 class is designed to be an easy way to expose various success
1438 metrics from passes. These statistics are printed at the end of a
1439 run, when the -stats command line option is enabled on the command
1440 line. See the <a href="http://llvm.org/docs/ProgrammersManual.html#Statistic">Statistics section</a> in the Programmer's Manual for details.
1442 </div>
1445 <!-- *********************************************************************** -->
1448 </div>
1449 <!-- *********************************************************************** -->
1455 <a
1458 are set up correctly, and then schedules passes to run efficiently. All of the
1460 passes.</p>
1463 time of a series of passes:</p>
1465 <ol>
1467 recomputing analysis results as much as possible. This means keeping track of
1468 which analyses are available already, which analyses get invalidated, and which
1469 analyses are needed to be run for a pass. An important part of work is that the
1472 results as soon as they are no longer needed.</li>
1476 of a series of passes by pipelining the passes together. This means that, given
1479 the first function, then all of the <a
1481 etc... until the entire program has been run through the passes.
1483 <p>This improves the cache behavior of the compiler, because it is only touching
1484 the LLVM program representation for a single function at a time, instead of
1485 traversing the entire program. It reduces the memory consumption of compiler,
1486 because, for example, only one <a
1488 needs to be calculated at a time. This also makes it possible to implement
1489 some <a
1492 </ol>
1495 much information it has about the behaviors of the passes it is scheduling. For
1496 example, the "preserved" set is intentionally conservative in the face of an
1498 Not implementing when it should be implemented will have the effect of not
1499 allowing any analysis results to live across the execution of your pass.</p>
1502 options that is useful for debugging pass execution, seeing how things work, and
1503 diagnosing when you should be preserving more analyses than you currently are
1512 $ opt -load ../../../Debug+Asserts/lib/Hello.so -gcse -licm --debug-pass=Structure < hello.bc > /dev/null
1513 Module Pass Manager
1514 Function Pass Manager
1515 Dominator Set Construction
1516 Immediate Dominators Construction
1517 Global Common Subexpression Elimination
1518 -- Immediate Dominators Construction
1519 -- Global Common Subexpression Elimination
1520 Natural Loop Construction
1521 Loop Invariant Code Motion
1522 -- Natural Loop Construction
1523 -- Loop Invariant Code Motion
1524 Module Verifier
1525 -- Dominator Set Construction
1526 -- Module Verifier
1527 Bitcode Writer
1528 --Bitcode Writer
1529 </pre></div>
1531 <p>This output shows us when passes are constructed and when the analysis
1533 GCSE uses dominator and immediate dominator information to do its job. The LICM
1534 pass uses natural loop information, which uses dominator sets, but not immediate
1535 dominators. Because immediate dominators are no longer useful after the GCSE
1536 pass, it is immediately destroyed. The dominator sets are then reused to
1537 compute natural loop information, which is then used by the LICM pass.</p>
1539 <p>After the LICM pass, the module verifier runs (which is automatically added
1541 resultant LLVM code is well formed. After it finishes, the dominator set
1542 information is destroyed, after being computed once, and shared by three
1543 passes.</p>
1549 $ opt -load ../../../Debug+Asserts/lib/Hello.so -gcse -hello -licm --debug-pass=Structure < hello.bc > /dev/null
1550 Module Pass Manager
1551 Function Pass Manager
1552 Dominator Set Construction
1553 Immediate Dominators Construction
1554 Global Common Subexpression Elimination
1556 -- Immediate Dominators Construction
1557 -- Global Common Subexpression Elimination
1558 <b> Hello World Pass
1559 -- Hello World Pass
1560 Dominator Set Construction</b>
1561 Natural Loop Construction
1562 Loop Invariant Code Motion
1563 -- Natural Loop Construction
1564 -- Loop Invariant Code Motion
1565 Module Verifier
1566 -- Dominator Set Construction
1567 -- Module Verifier
1568 Bitcode Writer
1569 --Bitcode Writer
1570 Hello: __main
1571 Hello: puts
1572 Hello: main
1573 </pre></div>
1576 Dominator Set pass, even though it doesn't modify the code at all! To fix this,
1577 we need to add the following <a
1583 AU.setPreservesAll();
1584 }
1585 </pre></div>
1590 $ opt -load ../../../Debug+Asserts/lib/Hello.so -gcse -hello -licm --debug-pass=Structure < hello.bc > /dev/null
1591 Pass Arguments: -gcse -hello -licm
1592 Module Pass Manager
1593 Function Pass Manager
1594 Dominator Set Construction
1595 Immediate Dominators Construction
1596 Global Common Subexpression Elimination
1597 -- Immediate Dominators Construction
1598 -- Global Common Subexpression Elimination
1599 Hello World Pass
1600 -- Hello World Pass
1601 Natural Loop Construction
1602 Loop Invariant Code Motion
1603 -- Loop Invariant Code Motion
1604 -- Natural Loop Construction
1605 Module Verifier
1606 -- Dominator Set Construction
1607 -- Module Verifier
1608 Bitcode Writer
1609 --Bitcode Writer
1610 Hello: __main
1611 Hello: puts
1612 Hello: main
1613 </pre></div>
1615 <p>Which shows that we don't accidentally invalidate dominator information
1616 anymore, and therefore do not have to compute it twice.</p>
1618 </div>
1620 <!-- _______________________________________________________________________ -->
1623 </div>
1629 </pre></div>
1632 results, and how long to keep them around for. Because the lifetime of the pass
1633 object itself is effectively the entire duration of the compilation process, we
1634 need some way to free analysis results when they are no longer useful. The
1637 <p>If you are writing an analysis or any other pass that retains a significant
1638 amount of state (for use by another pass which "requires" your pass and uses the
1644 </div>
1646 <!-- *********************************************************************** -->
1649 </div>
1650 <!-- *********************************************************************** -->
1655 both for the purposes of distribution, and for regulating the resident code size
1656 when running on the target system. Therefore, it becomes desirable to
1657 selectively use some passes, while omitting others and maintain the flexibility
1658 to change configurations later on. You want to be able to do all this, and,
1659 provide feedback to the user. This is where pass registration comes into
1660 play.</p>
1662 <p>The fundamental mechanisms for pass registration are the
1668 communicates additions and deletions to the command line interface.</p>
1671 information provided about a particular pass. This information includes the
1672 command line name, the command help string and the address of the function used
1673 to create an instance of the pass. A global static constructor of one of these
1676 in the tool will be registered at start up. A dynamically loaded pass will
1677 register on load and unregister at unload.</p>
1679 </div>
1681 <!-- _______________________________________________________________________ -->
1684 </div>
1688 <p>There are predefined registries to track instruction scheduling
1691 allocator machine pass.</p>
1693 <p>Implement your register allocator machine pass. In your register allocator
1694 .cpp file add the following include;</p>
1697 #include "llvm/CodeGen/RegAllocRegistry.h"
1698 </pre></div>
1700 <p>Also in your register allocator .cpp file, define a creator function in the
1701 form; </p>
1704 FunctionPass *createMyRegisterAllocator() {
1705 return new MyRegisterAllocator();
1706 }
1707 </pre></div>
1709 <p>Note that the signature of this function should match the type of
1714 static RegisterRegAlloc myRegAlloc("myregalloc",
1715 " my register allocator help string",
1716 createMyRegisterAllocator);
1717 </pre></div>
1719 <p>Note the two spaces prior to the help string produces a tidy result on the
1720 -help query.</p>
1723 $ llc -help
1724 ...
1725 -regalloc - Register allocator to use (default=linearscan)
1726 =linearscan - linear scan register allocator
1727 =local - local register allocator
1728 =simple - simple register allocator
1729 =myregalloc - my register allocator help string
1730 ...
1731 </pre></div>
1734 an option. Registering instruction schedulers is similar except use the
1739 <p>To force the load/linking of your register allocator into the llc/lli tools,
1743 </div>
1746 <!-- _______________________________________________________________________ -->
1749 </div>
1753 <p>The easiest way to get started is to clone one of the existing registries; we
1757 <p>Then you need to declare the registry. Example: if your pass registry is
1761 MachinePassRegistry RegisterMyPasses::Registry;
1762 </pre></div>
1767 cl::opt<RegisterMyPasses::FunctionPassCtor, false,
1769 MyPassOpt("mypass",
1770 cl::init(&createDefaultMyPass),
1771 cl::desc("my pass option help"));
1772 </pre></div>
1775 creator.</p>
1777 </div>
1779 <!-- *********************************************************************** -->
1782 </div>
1783 <!-- *********************************************************************** -->
1787 <p>Unfortunately, using GDB with dynamically loaded passes is not as easy as it
1788 should be. First of all, you can't set a breakpoint in a shared object that has
1789 not been loaded yet, and second of all there are problems with inlined functions
1790 in shared objects. Here are some suggestions to debugging your pass with
1791 GDB.</p>
1793 <p>For sake of discussion, I'm going to assume that you are debugging a
1795 on that.</p>
1797 </div>
1799 <!-- _______________________________________________________________________ -->
1802 </div>
1810 GNU gdb 5.0
1811 Copyright 2000 Free Software Foundation, Inc.
1812 GDB is free software, covered by the GNU General Public License, and you are
1813 welcome to change it and/or distribute copies of it under certain conditions.
1814 Type "show copying" to see the conditions.
1815 There is absolutely no warranty for GDB. Type "show warranty" for details.
1816 This GDB was configured as "sparc-sun-solaris2.6"...
1817 (gdb)
1818 </pre></div>
1821 time to load. Be patient. Since we cannot set a breakpoint in our pass yet
1822 (the shared object isn't loaded until runtime), we must execute the process, and
1823 have it stop before it invokes our pass, but after it has loaded the shared
1824 object. The most foolproof way of doing this is to set a breakpoint in
1826 want:</p>
1832 Starting program: opt test.bc -load $(LLVMTOP)/llvm/Debug+Asserts/lib/[libname].so -[passoption]
1835 (gdb)
1836 </pre></div>
1839 now free to set breakpoints in your pass so that you can trace through execution
1840 or do other standard debugging stuff.</p>
1842 </div>
1844 <!-- _______________________________________________________________________ -->
1847 </div>
1851 <p>Once you have the basics down, there are a couple of problems that GDB has,
1852 some with solutions, some without.</p>
1854 <ul>
1855 <li>Inline functions have bogus stack information. In general, GDB does a
1856 pretty good job getting stack traces and stepping through inline functions.
1857 When a pass is dynamically loaded however, it somehow completely loses this
1858 capability. The only solution I know of is to de-inline a function (move it
1859 from the body of a class to a .cpp file).</li>
1861 <li>Restarting the program breaks breakpoints. After following the information
1862 above, you have succeeded in getting some breakpoints planted in your pass. Nex
1864 and you start getting errors about breakpoints being unsettable. The only way I
1866 already set in your pass, run the program, and re-set the breakpoints once
1869 </ul>
1871 <p>Hopefully these tips will help with common case debugging situations. If
1872 you'd like to contribute some tips of your own, just contact <a
1875 </div>
1877 <!-- *********************************************************************** -->
1880 </div>
1881 <!-- *********************************************************************** -->
1885 <p>Although the LLVM Pass Infrastructure is very capable as it stands, and does
1886 some nifty stuff, there are things we'd like to add in the future. Here is
1887 where we are going:</p>
1889 </div>
1891 <!-- _______________________________________________________________________ -->
1894 </div>
1898 <p>Multiple CPU machines are becoming more common and compilation can never be
1899 fast enough: obviously we should allow for a multithreaded compiler. Because of
1900 the semantics defined for passes above (specifically they cannot maintain state
1903 to create multiple instances of each pass object, and allow the separate
1904 instances to be hacking on different parts of the program at the same time.</p>
1906 <p>This implementation would prevent each of the passes from having to implement
1907 multithreaded constructs, requiring only the LLVM core to have locking in a few
1908 places (for global resources). Although this is a simple extension, we simply
1909 haven't had time (or multiprocessor machines, thus a reason) to implement this.
1910 Despite that, we have kept the LLVM passes SMP ready, and you should too.</p>
1912 </div>
1914 <!-- *********************************************************************** -->
1915 <hr>
1916 <address>
1924 Last modified: $Date$
1925 </address>
1927 </body>
1928 </html>