| blob | 925d9811a4ed5d52a09f6d2acf0c4ec7930fcb48 |
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>
59 method</a></li>
62 </ul></li>
64 class</a>
65 <ul>
68 </ul></li>
69 </ul>
71 <ul>
73 </ul></li>
75 <ul>
77 method</a></li>
78 <li><a href="#AU::addRequired">The <tt>AnalysisUsage::addRequired<></tt> and <tt>AnalysisUsage::addRequiredTransitive<></tt> methods</a></li>
81 <li><a href="#getAnalysis">The <tt>getAnalysis<></tt> and <tt>getAnalysisToUpdate<></tt> methods</a></li>
82 </ul></li>
84 <ul>
87 </ul></li>
90 <ul>
92 </ul></li>
94 <ul>
97 </ul></li>
99 <ul>
102 </ul></li>
104 <ul>
106 </ul></li>
107 </ol>
112 </div>
114 <!-- *********************************************************************** -->
117 </div>
118 <!-- *********************************************************************** -->
122 <p>The LLVM Pass Framework is an important part of the LLVM system, because LLVM
123 passes are where most of the interesting parts of the compiler exist. Passes
124 perform the transformations and optimizations that make up the compiler, they
125 build the analysis results that are used by these transformations, and they are,
126 above all, a structuring technique for compiler code.</p>
130 class, which implement functionality by overriding virtual methods inherited
137 more information about what your pass does, and how it can be combined with
138 other passes. One of the main features of the LLVM Pass Framework is that it
139 schedules passes to run in an efficient way based on the constraints that your
140 pass meets (which are indicated by which class they derive from).</p>
142 <p>We start by showing you how to construct a pass, everything from setting up
143 the code, to compiling, loading, and executing it. After the basics are down,
144 more advanced features are discussed.</p>
146 </div>
148 <!-- *********************************************************************** -->
151 </div>
152 <!-- *********************************************************************** -->
157 is designed to simply print out the name of non-external functions that exist in
158 the program being compiled. It does not modify the program at all, it just
159 inspects it. The source code and files for this pass are available in the LLVM
162 </div>
164 <!-- ======================================================================= -->
167 </div>
171 <p>First, you need to create a new directory somewhere in the LLVM source
172 base. For this example, we'll assume that you made
174 (Makefile) that will compile the source code for the new pass. To do this,
176 <hr/>
179 # Makefile for hello pass
181 # Path to top level of LLVM heirarchy
182 LEVEL = ../../..
184 # Name of the library to build
185 LIBRARYNAME = Hello
187 # Make the shared library become a loadable module so the tools can
188 # dlopen/dlsym on the resulting library.
189 LOADABLE_MODULE = 1
191 # Tell the build system which LLVM libraries your pass needs. You'll probably
192 # need at least LLVMSystem.a, LLVMSupport.a, LLVMCore.a but possibly several
193 # others too.
194 LLVMLIBS = LLVMCore.a LLVMSupport.a LLVMSystem.a
196 # Include the makefile implementation stuff
197 include $(LEVEL)/Makefile.common
198 </pre></div>
201 directory are to be compiled and linked together into a
204 If your operating system uses a suffix other than .so (such as windows or
205 Mac OS/X), the appropriate extension will be used.</p>
207 <p>Now that we have the build scripts set up, we just need to write the code for
208 the pass itself.</p>
210 </div>
212 <!-- ======================================================================= -->
215 </div>
219 <p>Now that we have a way to compile our new pass, we just have to write it.
220 Start out with:</p>
225 </pre></div>
229 we are operating on <tt><a
235 </pre></div>
236 <p>... which is required because the functions from the include files
237 live in the llvm namespace.
238 </p>
244 </pre></div>
246 <p>... which starts out an anonymous namespace. Anonymous namespaces are to C++
247 what the "<tt>static</tt>" keyword is to C (at global scope). It makes the
248 things declared inside of the anonymous namespace only visible to the current
249 file. If you're not familiar with them, consult a decent C++ book for more
250 information.</p>
256 </pre></div><p>
260 The different builtin pass subclasses are described in detail <a
263 time.</p>
266 static char ID;
267 Hello() : FunctionPass((intptr_t)&ID) {}
268 </pre></div><p>
270 <p> This declares pass identifier used by LLVM to identify pass. This allows LLVM to
271 avoid using expensive C++ runtime information.</p>
277 }
279 </pre></div>
282 which overloads an abstract virtual method inherited from <a
284 to do our thing, so we just print out our message with the name of each
285 function.</p>
288 char Hello::ID = 0;
289 </pre></div>
291 <p> We initialize pass ID here. LLVM uses ID's address to identify pass so
292 initialization value is not important.</p>
297 </pre></div>
300 giving it a command line
314 static char ID;
315 Hello() : FunctionPass((intptr_t)&ID) {}
320 }
321 };
324 }
325 </pre></div>
328 command in the local directory and you should get a new
329 "<tt>Debug/lib/Hello.so</tt> file. Note that everything in this file is
330 contained in an anonymous namespace: this reflects the fact that passes are self
331 contained units that do not need external interfaces (although they can have
332 them) to be useful.</p>
334 </div>
336 <!-- ======================================================================= -->
339 </div>
343 <p>Now that you have a brand new shiny shared object file, we can use the
344 <tt>opt</tt> command to run an LLVM program through your pass. Because you
345 registered your pass with the <tt>RegisterPass</tt> template, you will be able to
346 use the <tt>opt</tt> tool to access it, once loaded.</p>
348 <p>To test it, follow the example at the end of the <a
350 LLVM. We can now run the bytecode file (<tt>hello.bc</tt>) for the program
351 through our transformation like this (or course, any bytecode file will
352 work):</p>
355 $ opt -load ../../../Debug/lib/Hello.so -hello < hello.bc > /dev/null
356 Hello: __main
357 Hello: puts
358 Hello: main
359 </pre></div>
361 <p>The '<tt>-load</tt>' option specifies that '<tt>opt</tt>' should load your
362 pass as a shared object, which makes '<tt>-hello</tt>' a valid command line
364 pass</a>). Because the hello pass does not modify the program in any
365 interesting way, we just throw away the result of <tt>opt</tt> (sending it to
366 <tt>/dev/null</tt>).</p>
368 <p>To see what happened to the other string you registered, try running
369 <tt>opt</tt> with the <tt>--help</tt> option:</p>
372 $ opt -load ../../../Debug/lib/Hello.so --help
373 OVERVIEW: llvm .bc -> .bc modular optimizer
375 USAGE: opt [options] <input bytecode>
377 OPTIONS:
378 Optimizations available:
379 ...
380 -funcresolve - Resolve Functions
381 -gcse - Global Common Subexpression Elimination
382 -globaldce - Dead Global Elimination
383 <b>-hello - Hello World Pass</b>
384 -indvars - Canonicalize Induction Variables
385 -inline - Function Integration/Inlining
386 -instcombine - Combine redundant instructions
387 ...
388 </pre></div>
390 <p>The pass name get added as the information string for your pass, giving some
391 documentation to users of <tt>opt</tt>. Now that you have a working pass, you
392 would go ahead and make it do the cool transformations you want. Once you get
393 it all working and tested, it may become useful to find out how fast your pass
395 line option (<tt>--time-passes</tt>) that allows you to get information about
396 the execution time of your pass along with the other passes you queue up. For
397 example:</p>
400 $ opt -load ../../../Debug/lib/Hello.so -hello -time-passes < hello.bc > /dev/null
401 Hello: __main
402 Hello: puts
403 Hello: main
404 ===============================================================================
405 ... Pass execution timing report ...
406 ===============================================================================
407 Total Execution Time: 0.02 seconds (0.0479059 wall clock)
409 ---User Time--- --System Time-- --User+System-- ---Wall Time--- --- Pass Name ---
410 0.0100 (100.0%) 0.0000 ( 0.0%) 0.0100 ( 50.0%) 0.0402 ( 84.0%) Bytecode Writer
411 0.0000 ( 0.0%) 0.0100 (100.0%) 0.0100 ( 50.0%) 0.0031 ( 6.4%) Dominator Set Construction
412 0.0000 ( 0.0%) 0.0000 ( 0.0%) 0.0000 ( 0.0%) 0.0013 ( 2.7%) Module Verifier
413 <b> 0.0000 ( 0.0%) 0.0000 ( 0.0%) 0.0000 ( 0.0%) 0.0033 ( 6.9%) Hello World Pass</b>
414 0.0100 (100.0%) 0.0100 (100.0%) 0.0200 (100.0%) 0.0479 (100.0%) TOTAL
415 </pre></div>
417 <p>As you can see, our implementation above is pretty fast :). The additional
418 passes listed are automatically inserted by the '<tt>opt</tt>' tool to verify
419 that the LLVM emitted by your pass is still valid and well formed LLVM, which
420 hasn't been broken somehow.</p>
422 <p>Now that you have seen the basics of the mechanics behind passes, we can talk
423 about some more details of how they work and how to use them.</p>
425 </div>
427 <!-- *********************************************************************** -->
430 </div>
431 <!-- *********************************************************************** -->
435 <p>One of the first things that you should do when designing a new pass is to
436 decide what class you should subclass for your pass. The <a
439 did not discuss why or when this should occur. Here we talk about the classes
440 available, from the most general to the most specific.</p>
442 <p>When choosing a superclass for your Pass, you should choose the <b>most
443 specific</b> class possible, while still being able to meet the requirements
444 listed. This gives the LLVM Pass Infrastructure information necessary to
445 optimize how passes are run, so that the resultant compiler isn't unneccesarily
446 slow.</p>
448 </div>
450 <!-- ======================================================================= -->
453 </div>
459 class. This pass type is used for passes that do not have to be run, do not
460 change state, and never need to be updated. This is not a normal type of
461 transformation or analysis, but can provide information about the current
462 compiler configuration.</p>
464 <p>Although this pass class is very infrequently used, it is important for
465 providing information about the current target machine being compiled for, and
466 other static information that can affect the various transformations.</p>
468 <p><tt>ImmutablePass</tt>es never invalidate other transformations, are never
471 </div>
473 <!-- ======================================================================= -->
476 </div>
482 class is the most general of all superclasses that you can use. Deriving from
483 <tt>ModulePass</tt> indicates that your pass uses the entire program as a unit,
484 refering to function bodies in no predictable order, or adding and removing
485 functions. Because nothing is known about the behavior of <tt>ModulePass</tt>
486 subclasses, no optimization can be done for their execution. A module pass
487 can use function level passes (e.g. dominators) using getAnalysis interface
488 <tt> getAnalysis<DominatorTree>(Function)</tt>. </p>
490 <p>To write a correct <tt>ModulePass</tt> subclass, derive from
491 <tt>ModulePass</tt> and overload the <tt>runOnModule</tt> method with the
492 following signature:</p>
494 </div>
496 <!-- _______________________________________________________________________ -->
499 </div>
504 <b>virtual bool</b> runOnModule(Module &M) = 0;
505 </pre></div>
507 <p>The <tt>runOnModule</tt> method performs the interesting work of the pass.
508 It should return true if the module was modified by the transformation and
509 false otherwise.</p>
511 </div>
513 <!-- ======================================================================= -->
516 </div>
522 is used by passes that need to traverse the program bottom-up on the call graph
523 (callees before callers). Deriving from CallGraphSCCPass provides some
524 mechanics for building and traversing the CallGraph, but also allows the system
525 to optimize execution of CallGraphSCCPass's. If your pass meets the
526 requirements outlined below, and doesn't meet the requirements of a <tt><a
529 <tt>CallGraphSCCPass</tt>.</p>
531 <p><b>TODO</b>: explain briefly what SCC, Tarjan's algo, and B-U mean.</p>
533 <p>To be explicit, <tt>CallGraphSCCPass</tt> subclasses are:</p>
535 <ol>
537 <li>... <em>not allowed</em> to modify any <tt>Function</tt>s that are not in
538 the current SCC.</li>
540 <li>... <em>allowed</em> to inspect any Function's other than those in the
541 current SCC and the direct callees of the SCC.</li>
543 <li>... <em>required</em> to preserve the current CallGraph object, updating it
544 to reflect any changes made to the program.</li>
546 <li>... <em>not allowed</em> to add or remove SCC's from the current Module,
547 though they may change the contents of an SCC.</li>
549 <li>... <em>allowed</em> to add or remove global variables from the current
550 Module.</li>
552 <li>... <em>allowed</em> to maintain state across invocations of
554 </ol>
556 <p>Implementing a <tt>CallGraphSCCPass</tt> is slightly tricky in some cases
557 because it has to handle SCCs with more than one node in it. All of the virtual
558 methods described below should return true if they modified the program, or
559 false if they didn't.</p>
561 </div>
563 <!-- _______________________________________________________________________ -->
566 method</a>
567 </div>
572 <b>virtual bool</b> doInitialization(CallGraph &CG);
573 </pre></div>
575 <p>The <tt>doIninitialize</tt> method is allowed to do most of the things that
576 <tt>CallGraphSCCPass</tt>'s are not allowed to do. They can add and remove
577 functions, get pointers to functions, etc. The <tt>doInitialization</tt> method
578 is designed to do simple initialization type of stuff that does not depend on
579 the SCCs being processed. The <tt>doInitialization</tt> method call is not
580 scheduled to overlap with any other pass executions (thus it should be very
581 fast).</p>
583 </div>
585 <!-- _______________________________________________________________________ -->
588 </div>
593 <b>virtual bool</b> runOnSCC(const std::vector<CallGraphNode *> &SCCM) = 0;
594 </pre></div>
596 <p>The <tt>runOnSCC</tt> method performs the interesting work of the pass, and
597 should return true if the module was modified by the transformation, false
598 otherwise.</p>
600 </div>
602 <!-- _______________________________________________________________________ -->
605 &)</tt> method</a>
606 </div>
611 <b>virtual bool</b> doFinalization(CallGraph &CG);
612 </pre></div>
614 <p>The <tt>doFinalization</tt> method is an infrequently used method that is
615 called when the pass framework has finished calling <a
617 program being compiled.</p>
619 </div>
621 <!-- ======================================================================= -->
624 </div>
628 <p>In contrast to <tt>ModulePass</tt> subclasses, <tt><a
630 subclasses do have a predictable, local behavior that can be expected by the
631 system. All <tt>FunctionPass</tt> execute on each function in the program
632 independent of all of the other functions in the program.
633 <tt>FunctionPass</tt>'s do not require that they are executed in a particular
634 order, and <tt>FunctionPass</tt>'s do not modify external functions.</p>
636 <p>To be explicit, <tt>FunctionPass</tt> subclasses are not allowed to:</p>
638 <ol>
639 <li>Modify a Function other than the one currently being processed.</li>
640 <li>Add or remove Function's from the current Module.</li>
641 <li>Add or remove global variables from the current Module.</li>
642 <li>Maintain state across invocations of
644 </ol>
646 <p>Implementing a <tt>FunctionPass</tt> is usually straightforward (See the <a
648 may overload three virtual methods to do their work. All of these methods
649 should return true if they modified the program, or false if they didn't.</p>
651 </div>
653 <!-- _______________________________________________________________________ -->
656 method</a>
657 </div>
662 <b>virtual bool</b> doInitialization(Module &M);
663 </pre></div>
665 <p>The <tt>doIninitialize</tt> method is allowed to do most of the things that
666 <tt>FunctionPass</tt>'s are not allowed to do. They can add and remove
667 functions, get pointers to functions, etc. The <tt>doInitialization</tt> method
668 is designed to do simple initialization type of stuff that does not depend on
669 the functions being processed. The <tt>doInitialization</tt> method call is not
670 scheduled to overlap with any other pass executions (thus it should be very
671 fast).</p>
673 <p>A good example of how this method should be used is the <a
675 pass. This pass converts <tt>malloc</tt> and <tt>free</tt> instructions into
676 platform dependent <tt>malloc()</tt> and <tt>free()</tt> function calls. It
677 uses the <tt>doInitialization</tt> method to get a reference to the malloc and
678 free functions that it needs, adding prototypes to the module if necessary.</p>
680 </div>
682 <!-- _______________________________________________________________________ -->
685 </div>
690 <b>virtual bool</b> runOnFunction(Function &F) = 0;
691 </pre></div><p>
693 <p>The <tt>runOnFunction</tt> method must be implemented by your subclass to do
694 the transformation or analysis work of your pass. As usual, a true value should
695 be returned if the function is modified.</p>
697 </div>
699 <!-- _______________________________________________________________________ -->
702 &)</tt> method</a>
703 </div>
708 <b>virtual bool</b> doFinalization(Module &M);
709 </pre></div>
711 <p>The <tt>doFinalization</tt> method is an infrequently used method that is
712 called when the pass framework has finished calling <a
714 program being compiled.</p>
716 </div>
718 <!-- ======================================================================= -->
721 </div>
725 <p> All <tt>LoopPass</tt> execute on each loop in the function independent of
726 all of the other loops in the function. <tt>LoopPass</tt> processes loops in
727 loop nest order such that outer most loop is processed last. </p>
729 <p> <tt>LoopPass</tt> subclasses are allowed to update loop nest using
730 <tt>LPPassManager</tt> interface. Implementing a loop pass is usually
731 straightforward. <tt>Looppass</tt>'s may overload three virtual methods to
732 do their work. All these methods should return true if they modified the
733 program, or false if they didn't. </p>
734 </div>
736 <!-- _______________________________________________________________________ -->
739 LPPassManager &)</tt>
740 method</a>
741 </div>
746 <b>virtual bool</b> doInitialization(Loop *, LPPassManager &LPM);
747 </pre></div>
749 <p>The <tt>doInitialization</tt> method is designed to do simple initialization
750 type of stuff that does not depend on the functions being processed. The
751 <tt>doInitialization</tt> method call is not scheduled to overlap with any
752 other pass executions (thus it should be very fast). LPPassManager
753 interface should be used to access Function or Module level analysis
754 information.</p>
756 </div>
759 <!-- _______________________________________________________________________ -->
762 </div>
767 <b>virtual bool</b> runOnLoop(Loop *, LPPassManager &LPM) = 0;
768 </pre></div><p>
770 <p>The <tt>runOnLoop</tt> method must be implemented by your subclass to do
771 the transformation or analysis work of your pass. As usual, a true value should
772 be returned if the function is modified. <tt>LPPassManager</tt> interface
773 should be used to update loop nest.</p>
775 </div>
777 <!-- _______________________________________________________________________ -->
780 </div>
785 <b>virtual bool</b> doFinalization();
786 </pre></div>
788 <p>The <tt>doFinalization</tt> method is an infrequently used method that is
789 called when the pass framework has finished calling <a
791 program being compiled. </p>
793 </div>
797 <!-- ======================================================================= -->
800 </div>
804 <p><tt>BasicBlockPass</tt>'s are just like <a
806 their scope of inspection and modification to a single basic block at a time.
807 As such, they are <b>not</b> allowed to do any of the following:</p>
809 <ol>
810 <li>Modify or inspect any basic blocks outside of the current one</li>
811 <li>Maintain state across invocations of
813 <li>Modify the control flow graph (by altering terminator instructions)</li>
814 <li>Any of the things forbidden for
816 </ol>
819 optimizations. They may override the same <a
822 href="#FunctionPass"><tt>FunctionPass</tt></a>'s have, but also have the following virtual methods that may also be implemented:</p>
824 </div>
826 <!-- _______________________________________________________________________ -->
829 &)</tt> method</a>
830 </div>
835 <b>virtual bool</b> doInitialization(Function &F);
836 </pre></div>
838 <p>The <tt>doIninitialize</tt> method is allowed to do most of the things that
839 <tt>BasicBlockPass</tt>'s are not allowed to do, but that
840 <tt>FunctionPass</tt>'s can. The <tt>doInitialization</tt> method is designed
841 to do simple initialization that does not depend on the
842 BasicBlocks being processed. The <tt>doInitialization</tt> method call is not
843 scheduled to overlap with any other pass executions (thus it should be very
844 fast).</p>
846 </div>
848 <!-- _______________________________________________________________________ -->
851 </div>
856 <b>virtual bool</b> runOnBasicBlock(BasicBlock &BB) = 0;
857 </pre></div>
859 <p>Override this function to do the work of the <tt>BasicBlockPass</tt>. This
860 function is not allowed to inspect or modify basic blocks other than the
861 parameter, and are not allowed to modify the CFG. A true value must be returned
862 if the basic block is modified.</p>
864 </div>
866 <!-- _______________________________________________________________________ -->
869 method</a>
870 </div>
875 <b>virtual bool</b> doFinalization(Function &F);
876 </pre></div>
878 <p>The <tt>doFinalization</tt> method is an infrequently used method that is
879 called when the pass framework has finished calling <a
881 program being compiled. This can be used to perform per-function
882 finalization.</p>
884 </div>
886 <!-- ======================================================================= -->
889 </div>
893 <p>A <tt>MachineFunctionPass</tt> is a part of the LLVM code generator that
894 executes on the machine-dependent representation of each LLVM function in the
895 program. A <tt>MachineFunctionPass</tt> is also a <tt>FunctionPass</tt>, so all
896 the restrictions that apply to a <tt>FunctionPass</tt> also apply to it.
897 <tt>MachineFunctionPass</tt>es also have additional restrictions. In particular,
898 <tt>MachineFunctionPass</tt>es are not allowed to do any of the following:</p>
900 <ol>
901 <li>Modify any LLVM Instructions, BasicBlocks or Functions.</li>
902 <li>Modify a MachineFunction other than the one currently being processed.</li>
903 <li>Add or remove MachineFunctions from the current Module.</li>
904 <li>Add or remove global variables from the current Module.</li>
905 <li>Maintain state across invocations of <a
907 data)</li>
908 </ol>
910 </div>
912 <!-- _______________________________________________________________________ -->
915 &MF)</tt> method</a>
916 </div>
921 <b>virtual bool</b> runOnMachineFunction(MachineFunction &MF) = 0;
922 </pre></div>
924 <p><tt>runOnMachineFunction</tt> can be considered the main entry point of a
925 <tt>MachineFunctionPass</tt>; that is, you should override this method to do the
926 work of your <tt>MachineFunctionPass</tt>.</p>
928 <p>The <tt>runOnMachineFunction</tt> method is called on every
929 <tt>MachineFunction</tt> in a <tt>Module</tt>, so that the
930 <tt>MachineFunctionPass</tt> may perform optimizations on the machine-dependent
931 representation of the function. If you want to get at the LLVM <tt>Function</tt>
932 for the <tt>MachineFunction</tt> you're working on, use
933 <tt>MachineFunction</tt>'s <tt>getFunction()</tt> accessor method -- but
934 remember, you may not modify the LLVM <tt>Function</tt> or its contents from a
935 <tt>MachineFunctionPass</tt>.</p>
937 </div>
939 <!-- *********************************************************************** -->
942 </div>
943 <!-- *********************************************************************** -->
948 pass registration works, and discussed some of the reasons that it is used and
949 what it does. Here we discuss how and why passes are registered.</p>
951 <p>As we saw above, passes are registered with the <b><tt>RegisterPass</tt></b>
952 template, which requires you to pass at least two
953 parameters. The first parameter is the name of the pass that is to be used on
954 the command line to specify that the pass should be added to a program (for
955 example, with <tt>opt</tt> or <tt>bugpoint</tt>). The second argument is the
956 name of the pass, which is to be used for the <tt>--help</tt> output of
957 programs, as
958 well as for debug output generated by the <tt>--debug-pass</tt> option.</p>
960 <p>If you want your pass to be easily dumpable, you should
961 implement the virtual <tt>print</tt> method:</p>
963 </div>
965 <!-- _______________________________________________________________________ -->
968 </div>
973 <b>virtual void</b> print(llvm::OStream &O, <b>const</b> Module *M) <b>const</b>;
974 </pre></div>
977 a human readable version of the analysis results. This is useful for debugging
978 an analysis itself, as well as for other people to figure out how an analysis
979 works. Use the <tt>opt -analyze</tt> argument to invoke this method.</p>
981 <p>The <tt>llvm::OStream</tt> parameter specifies the stream to write the results on,
982 and the <tt>Module</tt> parameter gives a pointer to the top level module of the
983 program that has been analyzed. Note however that this pointer may be null in
984 certain circumstances (such as calling the <tt>Pass::dump()</tt> from a
985 debugger), so it should only be used to enhance debug output, it should not be
986 depended on.</p>
988 </div>
990 <!-- *********************************************************************** -->
993 </div>
994 <!-- *********************************************************************** -->
998 <p>One of the main responsibilities of the <tt>PassManager</tt> is the make sure
999 that passes interact with each other correctly. Because <tt>PassManager</tt>
1001 know how the passes interact with each other and what dependencies exist between
1002 the various passes. To track this, each pass can declare the set of passes that
1003 are required to be executed before the current pass, and the passes which are
1004 invalidated by the current pass.</p>
1006 <p>Typically this functionality is used to require that analysis results are
1007 computed before your pass is run. Running arbitrary transformation passes can
1008 invalidate the computed analysis results, which is what the invalidation set
1009 specifies. If a pass does not implement the <tt><a
1011 having any prerequisite passes, and invalidating <b>all</b> other passes.</p>
1013 </div>
1015 <!-- _______________________________________________________________________ -->
1018 </div>
1023 <b>virtual void</b> getAnalysisUsage(AnalysisUsage &Info) <b>const</b>;
1024 </pre></div>
1026 <p>By implementing the <tt>getAnalysisUsage</tt> method, the required and
1027 invalidated sets may be specified for your transformation. The implementation
1028 should fill in the <tt><a
1030 object with information about which passes are required and not invalidated. To
1031 do this, a pass may call any of the following methods on the AnalysisUsage
1032 object:</p>
1033 </div>
1035 <!-- _______________________________________________________________________ -->
1037 <a name="AU::addRequired">The <tt>AnalysisUsage::addRequired<></tt> and <tt>AnalysisUsage::addRequiredTransitive<></tt> methods</a>
1038 </div>
1041 <p>
1042 If your pass requires a previous pass to be executed (an analysis for example),
1043 it can use one of these methods to arrange for it to be run before your pass.
1044 LLVM has many different types of analyses and passes that can be required,
1045 spanning the range from <tt>DominatorSet</tt> to <tt>BreakCriticalEdges</tt>.
1046 Requiring <tt>BreakCriticalEdges</tt>, for example, guarantees that there will
1047 be no critical edges in the CFG when your pass has been run.
1048 </p>
1050 <p>
1051 Some analyses chain to other analyses to do their job. For example, an <a
1054 cases where analyses chain, the <tt>addRequiredTransitive</tt> method should be
1055 used instead of the <tt>addRequired</tt> method. This informs the PassManager
1056 that the transitively required pass should be alive as long as the requiring
1057 pass is.
1058 </p>
1059 </div>
1061 <!-- _______________________________________________________________________ -->
1064 </div>
1067 <p>
1068 One of the jobs of the PassManager is to optimize how and when analyses are run.
1069 In particular, it attempts to avoid recomputing data unless it needs to. For
1070 this reason, passes are allowed to declare that they preserve (i.e., they don't
1071 invalidate) an existing analysis if it's available. For example, a simple
1072 constant folding pass would not modify the CFG, so it can't possibly affect the
1073 results of dominator analysis. By default, all passes are assumed to invalidate
1074 all others.
1075 </p>
1077 <p>
1078 The <tt>AnalysisUsage</tt> class provides several methods which are useful in
1079 certain circumstances that are related to <tt>addPreserved</tt>. In particular,
1080 the <tt>setPreservesAll</tt> method can be called to indicate that the pass does
1081 not modify the LLVM program at all (which is true for analyses), and the
1082 <tt>setPreservesCFG</tt> method can be used by transformations that change
1083 instructions in the program but do not modify the CFG or terminator instructions
1084 (note that this property is implicitly set for <a
1086 </p>
1088 <p>
1089 <tt>addPreserved</tt> is particularly useful for transformations like
1090 <tt>BreakCriticalEdges</tt>. This pass knows how to update a small set of loop
1091 and dominator related analyses if they exist, so it can preserve them, despite
1092 the fact that it hacks on the CFG.
1093 </p>
1094 </div>
1096 <!-- _______________________________________________________________________ -->
1099 </div>
1104 <i>// This is an example implementation from an analysis, which does not modify
1105 // the program at all, yet has a prerequisite.</i>
1106 <b>void</b> <a href="http://llvm.org/doxygen/classllvm_1_1PostDominanceFrontier.html">PostDominanceFrontier</a>::getAnalysisUsage(AnalysisUsage &AU) <b>const</b> {
1107 AU.setPreservesAll();
1108 AU.addRequired<<a href="http://llvm.org/doxygen/classllvm_1_1PostDominatorTree.html">PostDominatorTree</a>>();
1109 }
1110 </pre></div>
1112 <p>and:</p>
1115 <i>// This example modifies the program, but does not modify the CFG</i>
1116 <b>void</b> <a href="http://llvm.org/doxygen/structLICM.html">LICM</a>::getAnalysisUsage(AnalysisUsage &AU) <b>const</b> {
1117 AU.setPreservesCFG();
1118 AU.addRequired<<a href="http://llvm.org/doxygen/classllvm_1_1LoopInfo.html">LoopInfo</a>>();
1119 }
1120 </pre></div>
1122 </div>
1124 <!-- _______________________________________________________________________ -->
1126 <a name="getAnalysis">The <tt>getAnalysis<></tt> and <tt>getAnalysisToUpdate<></tt> methods</a>
1127 </div>
1131 <p>The <tt>Pass::getAnalysis<></tt> method is automatically inherited by
1132 your class, providing you with access to the passes that you declared that you
1134 method. It takes a single template argument that specifies which pass class you
1135 want, and returns a reference to that pass. For example:</p>
1138 bool LICM::runOnFunction(Function &F) {
1139 LoopInfo &LI = getAnalysis<LoopInfo>();
1140 ...
1141 }
1142 </pre></div>
1144 <p>This method call returns a reference to the pass desired. You may get a
1145 runtime assertion failure if you attempt to get an analysis that you did not
1146 declare as required in your <a
1148 method can be called by your <tt>run*</tt> method implementation, or by any
1149 other local method invoked by your <tt>run*</tt> method.
1151 A module level pass can use function level analysis info using this interface.
1152 For example:</p>
1155 bool ModuleLevelPass::runOnModule(Module &M) {
1156 ...
1157 DominatorTree &DT = getAnalysis<DominatorTree>(Func);
1158 ...
1159 }
1160 </pre></div>
1162 <p>In above example, runOnFunction for DominatorTree is called by pass manager
1163 before returning a reference to the desired pass.</p>
1165 <p>
1166 If your pass is capable of updating analyses if they exist (e.g.,
1167 <tt>BreakCriticalEdges</tt>, as described above), you can use the
1168 <tt>getAnalysisToUpdate</tt> method, which returns a pointer to the analysis if
1169 it is active. For example:</p>
1172 ...
1173 if (DominatorSet *DS = getAnalysisToUpdate<DominatorSet>()) {
1174 <i>// A DominatorSet is active. This code will update it.</i>
1175 }
1176 ...
1177 </pre></div>
1179 </div>
1181 <!-- *********************************************************************** -->
1184 </div>
1185 <!-- *********************************************************************** -->
1189 <p>Now that we understand the basics of how passes are defined, how the are
1190 used, and how they are required from other passes, it's time to get a little bit
1191 fancier. All of the pass relationships that we have seen so far are very
1192 simple: one pass depends on one other specific pass to be run before it can run.
1193 For many applications, this is great, for others, more flexibility is
1194 required.</p>
1196 <p>In particular, some analyses are defined such that there is a single simple
1197 interface to the analysis results, but multiple ways of calculating them.
1198 Consider alias analysis for example. The most trivial alias analysis returns
1200 flow-sensitive, context-sensitive interprocedural analysis that can take a
1201 significant amount of time to execute (and obviously, there is a lot of room
1202 between these two extremes for other implementations). To cleanly support
1203 situations like this, the LLVM Pass Infrastructure supports the notion of
1204 Analysis Groups.</p>
1206 </div>
1208 <!-- _______________________________________________________________________ -->
1211 </div>
1215 <p>An Analysis Group is a single simple interface that may be implemented by
1216 multiple different passes. Analysis Groups can be given human readable names
1217 just like passes, but unlike passes, they need not derive from the <tt>Pass</tt>
1218 class. An analysis group may have one or more implementations, one of which is
1221 <p>Analysis groups are used by client passes just like other passes are: the
1222 <tt>AnalysisUsage::addRequired()</tt> and <tt>Pass::getAnalysis()</tt> methods.
1224 scans the available passes to see if any implementations of the analysis group
1225 are available. If none is available, the default implementation is created for
1227 between passes</a> still apply.</p>
1230 passes, all analysis group implementations must be registered, and must use the
1232 implementation pool. Also, a default implementation of the interface
1233 <b>must</b> be registered with <A
1236 <p>As a concrete example of an Analysis Group in action, consider the <a
1238 analysis group. The default implementation of the alias analysis interface (the
1239 <tt><a
1241 pass) just does a few simple checks that don't require significant analysis to
1242 compute (such as: two different globals can never alias each other, etc).
1243 Passes that use the <tt><a
1245 interface (for example the <tt><a
1247 not care which implementation of alias analysis is actually provided, they just
1248 use the designated interface.</p>
1250 <p>From the user's perspective, commands work just like normal. Issuing the
1251 command '<tt>opt -gcse ...</tt>' will cause the <tt>basicaa</tt> class to be
1252 instantiated and added to the pass sequence. Issuing the command '<tt>opt
1253 -somefancyaa -gcse ...</tt>' will cause the <tt>gcse</tt> pass to use the
1254 <tt>somefancyaa</tt> alias analysis (which doesn't actually exist, it's just a
1255 hypothetical example) instead.</p>
1257 </div>
1259 <!-- _______________________________________________________________________ -->
1262 </div>
1266 <p>The <tt>RegisterAnalysisGroup</tt> template is used to register the analysis
1267 group itself as well as add pass implementations to the analysis group. First,
1268 an analysis should be registered, with a human readable name provided for it.
1269 Unlike registration of passes, there is no command line argument to be specified
1273 <b>static</b> RegisterAnalysisGroup<<a href="http://llvm.org/doxygen/classllvm_1_1AliasAnalysis.html">AliasAnalysis</a>> A("<i>Alias Analysis</i>");
1274 </pre></div>
1276 <p>Once the analysis is registered, passes can declare that they are valid
1277 implementations of the interface by using the following code:</p>
1280 <b>namespace</b> {
1281 //<i> Analysis Group implementations <b>must</b> be registered normally...</i>
1282 RegisterPass<FancyAA>
1285 //<i> Declare that we implement the AliasAnalysis interface</i>
1286 RegisterAnalysisGroup<<a href="http://llvm.org/doxygen/classllvm_1_1AliasAnalysis.html">AliasAnalysis</a>> C(B);
1287 }
1288 </pre></div>
1290 <p>This just shows a class <tt>FancyAA</tt> that is registered normally, then
1293 analysis group. Every implementation of an analysis group should join using
1294 this template. A single pass may join multiple different analysis groups with
1295 no problem.</p>
1298 <b>namespace</b> {
1299 //<i> Analysis Group implementations <b>must</b> be registered normally...</i>
1300 RegisterPass<<a href="http://llvm.org/doxygen/structBasicAliasAnalysis.html">BasicAliasAnalysis</a>>
1303 //<i> Declare that we implement the AliasAnalysis interface</i>
1304 RegisterAnalysisGroup<<a href="http://llvm.org/doxygen/classllvm_1_1AliasAnalysis.html">AliasAnalysis</a>, <b>true</b>> E(D);
1305 }
1306 </pre></div>
1308 <p>Here we show how the default implementation is specified (using the extra
1309 argument to the <tt>RegisterAnalysisGroup</tt> template). There must be exactly
1310 one default implementation available at all times for an Analysis Group to be
1311 used. Here we declare that the <tt><a
1313 pass is the default implementation for the interface.</p>
1315 </div>
1317 <!-- *********************************************************************** -->
1320 </div>
1321 <!-- *********************************************************************** -->
1324 <p>The <a
1326 class is designed to be an easy way to expose various success
1327 metrics from passes. These statistics are printed at the end of a
1328 run, when the -stats command line option is enabled on the command
1329 line. See the <a href="http://llvm.org/docs/ProgrammersManual.html#Statistic">Statistics section</a> in the Programmer's Manual for details.
1331 </div>
1334 <!-- *********************************************************************** -->
1337 </div>
1338 <!-- *********************************************************************** -->
1342 <p>The <a
1344 <a
1347 are set up correctly, and then schedules passes to run efficiently. All of the
1348 LLVM tools that run passes use the <tt>PassManager</tt> for execution of these
1349 passes.</p>
1351 <p>The <tt>PassManager</tt> does two main things to try to reduce the execution
1352 time of a series of passes:</p>
1354 <ol>
1355 <li><b>Share analysis results</b> - The PassManager attempts to avoid
1356 recomputing analysis results as much as possible. This means keeping track of
1357 which analyses are available already, which analyses get invalidated, and which
1358 analyses are needed to be run for a pass. An important part of work is that the
1359 <tt>PassManager</tt> tracks the exact lifetime of all analysis results, allowing
1361 results as soon as they are no longer needed.</li>
1363 <li><b>Pipeline the execution of passes on the program</b> - The
1364 <tt>PassManager</tt> attempts to get better cache and memory usage behavior out
1365 of a series of passes by pipelining the passes together. This means that, given
1368 the first function, then all of the <a
1370 etc... until the entire program has been run through the passes.
1372 <p>This improves the cache behavior of the compiler, because it is only touching
1373 the LLVM program representation for a single function at a time, instead of
1374 traversing the entire program. It reduces the memory consumption of compiler,
1375 because, for example, only one <a
1377 needs to be calculated at a time. This also makes it possible some <a
1380 </ol>
1382 <p>The effectiveness of the <tt>PassManager</tt> is influenced directly by how
1383 much information it has about the behaviors of the passes it is scheduling. For
1386 Not implementing when it should be implemented will have the effect of not
1387 allowing any analysis results to live across the execution of your pass.</p>
1389 <p>The <tt>PassManager</tt> class exposes a <tt>--debug-pass</tt> command line
1390 options that is useful for debugging pass execution, seeing how things work, and
1391 diagnosing when you should be preserving more analyses than you currently are
1392 (To get information about all of the variants of the <tt>--debug-pass</tt>
1393 option, just type '<tt>opt --help-hidden</tt>').</p>
1395 <p>By using the <tt>--debug-pass=Structure</tt> option, for example, we can see
1397 Lets try it out with the <tt>gcse</tt> and <tt>licm</tt> passes:</p>
1400 $ opt -load ../../../Debug/lib/Hello.so -gcse -licm --debug-pass=Structure < hello.bc > /dev/null
1401 Module Pass Manager
1402 Function Pass Manager
1403 Dominator Set Construction
1404 Immediate Dominators Construction
1405 Global Common Subexpression Elimination
1406 -- Immediate Dominators Construction
1407 -- Global Common Subexpression Elimination
1408 Natural Loop Construction
1409 Loop Invariant Code Motion
1410 -- Natural Loop Construction
1411 -- Loop Invariant Code Motion
1412 Module Verifier
1413 -- Dominator Set Construction
1414 -- Module Verifier
1415 Bytecode Writer
1416 --Bytecode Writer
1417 </pre></div>
1419 <p>This output shows us when passes are constructed and when the analysis
1420 results are known to be dead (prefixed with '<tt>--</tt>'). Here we see that
1421 GCSE uses dominator and immediate dominator information to do its job. The LICM
1422 pass uses natural loop information, which uses dominator sets, but not immediate
1423 dominators. Because immediate dominators are no longer useful after the GCSE
1424 pass, it is immediately destroyed. The dominator sets are then reused to
1425 compute natural loop information, which is then used by the LICM pass.</p>
1427 <p>After the LICM pass, the module verifier runs (which is automatically added
1428 by the '<tt>opt</tt>' tool), which uses the dominator set to check that the
1429 resultant LLVM code is well formed. After it finishes, the dominator set
1430 information is destroyed, after being computed once, and shared by three
1431 passes.</p>
1434 World</a> pass in between the two passes:</p>
1437 $ opt -load ../../../Debug/lib/Hello.so -gcse -hello -licm --debug-pass=Structure < hello.bc > /dev/null
1438 Module Pass Manager
1439 Function Pass Manager
1440 Dominator Set Construction
1441 Immediate Dominators Construction
1442 Global Common Subexpression Elimination
1443 <b>-- Dominator Set Construction</b>
1444 -- Immediate Dominators Construction
1445 -- Global Common Subexpression Elimination
1446 <b> Hello World Pass
1447 -- Hello World Pass
1448 Dominator Set Construction</b>
1449 Natural Loop Construction
1450 Loop Invariant Code Motion
1451 -- Natural Loop Construction
1452 -- Loop Invariant Code Motion
1453 Module Verifier
1454 -- Dominator Set Construction
1455 -- Module Verifier
1456 Bytecode Writer
1457 --Bytecode Writer
1458 Hello: __main
1459 Hello: puts
1460 Hello: main
1461 </pre></div>
1464 Dominator Set pass, even though it doesn't modify the code at all! To fix this,
1465 we need to add the following <a
1469 <i>// We don't modify the program, so we preserve all analyses</i>
1470 <b>virtual void</b> getAnalysisUsage(AnalysisUsage &AU) <b>const</b> {
1471 AU.setPreservesAll();
1472 }
1473 </pre></div>
1475 <p>Now when we run our pass, we get this output:</p>
1478 $ opt -load ../../../Debug/lib/Hello.so -gcse -hello -licm --debug-pass=Structure < hello.bc > /dev/null
1479 Pass Arguments: -gcse -hello -licm
1480 Module Pass Manager
1481 Function Pass Manager
1482 Dominator Set Construction
1483 Immediate Dominators Construction
1484 Global Common Subexpression Elimination
1485 -- Immediate Dominators Construction
1486 -- Global Common Subexpression Elimination
1487 Hello World Pass
1488 -- Hello World Pass
1489 Natural Loop Construction
1490 Loop Invariant Code Motion
1491 -- Loop Invariant Code Motion
1492 -- Natural Loop Construction
1493 Module Verifier
1494 -- Dominator Set Construction
1495 -- Module Verifier
1496 Bytecode Writer
1497 --Bytecode Writer
1498 Hello: __main
1499 Hello: puts
1500 Hello: main
1501 </pre></div>
1503 <p>Which shows that we don't accidentally invalidate dominator information
1504 anymore, and therefore do not have to compute it twice.</p>
1506 </div>
1508 <!-- _______________________________________________________________________ -->
1511 </div>
1516 <b>virtual void</b> releaseMemory();
1517 </pre></div>
1519 <p>The <tt>PassManager</tt> automatically determines when to compute analysis
1520 results, and how long to keep them around for. Because the lifetime of the pass
1521 object itself is effectively the entire duration of the compilation process, we
1522 need some way to free analysis results when they are no longer useful. The
1523 <tt>releaseMemory</tt> virtual method is the way to do this.</p>
1525 <p>If you are writing an analysis or any other pass that retains a significant
1528 <tt>releaseMEmory</tt> to, well, release the memory allocated to maintain this
1529 internal state. This method is called after the <tt>run*</tt> method for the
1530 class, before the next call of <tt>run*</tt> in your pass.</p>
1532 </div>
1534 <!-- *********************************************************************** -->
1537 </div>
1538 <!-- *********************************************************************** -->
1542 <p><i>Size matters</i> when constructing production quality tools using llvm,
1543 both for the purposes of distribution, and for regulating the resident code size
1544 when running on the target system. Therefore, it becomes desirable to
1545 selectively use some passes, while omitting others and maintain the flexibility
1546 to change configurations later on. You want to be able to do all this, and,
1547 provide feedback to the user. This is where pass registration comes into
1548 play.</p>
1550 <p>The fundamental mechanisms for pass registration are the
1551 <tt>MachinePassRegistry</tt> class and subclasses of
1552 <tt>MachinePassRegistryNode</tt>.</p>
1554 <p>An instance of <tt>MachinePassRegistry</tt> is used to maintain a list of
1555 <tt>MachinePassRegistryNode</tt> objects. This instance maintains the list and
1556 communicates additions and deletions to the command line interface.</p>
1558 <p>An instance of <tt>MachinePassRegistryNode</tt> subclass is used to maintain
1559 information provided about a particular pass. This information includes the
1560 command line name, the command help string and the address of the function used
1561 to create an instance of the pass. A global static constructor of one of these
1562 instances <i>registers</i> with a corresponding <tt>MachinePassRegistry</tt>,
1563 the static destructor <i>unregisters</i>. Thus a pass that is statically linked
1564 in the tool will be registered at start up. A dynamically loaded pass will
1565 register on load and unregister at unload.</p>
1567 </div>
1569 <!-- _______________________________________________________________________ -->
1572 </div>
1576 <p>There are predefined registries to track instruction scheduling
1577 (<tt>RegisterScheduler</tt>) and register allocation (<tt>RegisterRegAlloc</tt>)
1578 machine passes. Here we will describe how to <i>register</i> a register
1579 allocator machine pass.</p>
1581 <p>Implement your register allocator machine pass. In your register allocator
1582 .cpp file add the following include;</p>
1586 </pre></div>
1588 <p>Also in your register allocator .cpp file, define a creator function in the
1589 form; </p>
1592 FunctionPass *createMyRegisterAllocator() {
1593 return new MyRegisterAllocator();
1594 }
1595 </pre></div>
1597 <p>Note that the signature of this function should match the type of
1598 <tt>RegisterRegAlloc::FunctionPassCtor</tt>. In the same file add the
1604 createMyRegisterAllocator);
1605 </pre></div>
1607 <p>Note the two spaces prior to the help string produces a tidy result on the
1608 --help query.</p>
1611 $ llc --help
1612 ...
1613 -regalloc - Register allocator to use: (default = linearscan)
1614 =linearscan - linear scan register allocator
1615 =local - local register allocator
1616 =simple - simple register allocator
1617 =myregalloc - my register allocator help string
1618 ...
1619 </pre></div>
1621 <p>And that's it. The user is now free to use <tt>-regalloc=myregalloc</tt> as
1622 an option. Registering instruction schedulers is similar except use the
1623 <tt>RegisterScheduler</tt> class. Note that the
1624 <tt>RegisterScheduler::FunctionPassCtor</tt> is significantly different from
1625 <tt>RegisterRegAlloc::FunctionPassCtor</tt>.</p>
1627 <p>To force the load/linking of your register allocator into the llc/lli tools,
1629 call line to <tt>llvm/Codegen/LinkAllCodegenComponents.h</tt>.</p>
1631 </div>
1634 <!-- _______________________________________________________________________ -->
1637 </div>
1641 <p>The easiest way to get started is to clone one of the existing registries; we
1642 recommend <tt>llvm/CodeGen/RegAllocRegistry.h</tt>. The key things to modify
1643 are the class name and the <tt>FunctionPassCtor</tt> type.</p>
1645 <p>Then you need to declare the registry. Example: if your pass registry is
1646 <tt>RegisterMyPasses</tt> then define;</p>
1649 MachinePassRegistry RegisterMyPasses::Registry;
1650 </pre></div>
1652 <p>And finally, declare the command line option for your passes. Example:</p>
1655 cl::opt<RegisterMyPasses::FunctionPassCtor, false,
1656 RegisterPassParser<RegisterMyPasses> >
1658 cl::init(&createDefaultMyPass),
1660 </pre></div>
1663 creator.</p>
1665 </div>
1667 <!-- *********************************************************************** -->
1670 </div>
1671 <!-- *********************************************************************** -->
1675 <p>Unfortunately, using GDB with dynamically loaded passes is not as easy as it
1676 should be. First of all, you can't set a breakpoint in a shared object that has
1677 not been loaded yet, and second of all there are problems with inlined functions
1678 in shared objects. Here are some suggestions to debugging your pass with
1679 GDB.</p>
1681 <p>For sake of discussion, I'm going to assume that you are debugging a
1682 transformation invoked by <tt>opt</tt>, although nothing described here depends
1683 on that.</p>
1685 </div>
1687 <!-- _______________________________________________________________________ -->
1690 </div>
1694 <p>First thing you do is start <tt>gdb</tt> on the <tt>opt</tt> process:</p>
1697 $ <b>gdb opt</b>
1698 GNU gdb 5.0
1699 Copyright 2000 Free Software Foundation, Inc.
1700 GDB is free software, covered by the GNU General Public License, and you are
1701 welcome to change it and/or distribute copies of it under certain conditions.
1705 (gdb)
1706 </pre></div>
1708 <p>Note that <tt>opt</tt> has a lot of debugging information in it, so it takes
1709 time to load. Be patient. Since we cannot set a breakpoint in our pass yet
1710 (the shared object isn't loaded until runtime), we must execute the process, and
1711 have it stop before it invokes our pass, but after it has loaded the shared
1712 object. The most foolproof way of doing this is to set a breakpoint in
1713 <tt>PassManager::run</tt> and then run the process with the arguments you
1714 want:</p>
1717 (gdb) <b>break llvm::PassManager::run</b>
1718 Breakpoint 1 at 0x2413bc: file Pass.cpp, line 70.
1719 (gdb) <b>run test.bc -load $(LLVMTOP)/llvm/Debug/lib/[libname].so -[passoption]</b>
1720 Starting program: opt test.bc -load $(LLVMTOP)/llvm/Debug/lib/[libname].so -[passoption]
1721 Breakpoint 1, PassManager::run (this=0xffbef174, M=@0x70b298) at Pass.cpp:70
1722 70 bool PassManager::run(Module &M) { return PM->run(M); }
1723 (gdb)
1724 </pre></div>
1726 <p>Once the <tt>opt</tt> stops in the <tt>PassManager::run</tt> method you are
1727 now free to set breakpoints in your pass so that you can trace through execution
1728 or do other standard debugging stuff.</p>
1730 </div>
1732 <!-- _______________________________________________________________________ -->
1735 </div>
1739 <p>Once you have the basics down, there are a couple of problems that GDB has,
1740 some with solutions, some without.</p>
1742 <ul>
1743 <li>Inline functions have bogus stack information. In general, GDB does a
1744 pretty good job getting stack traces and stepping through inline functions.
1745 When a pass is dynamically loaded however, it somehow completely loses this
1746 capability. The only solution I know of is to de-inline a function (move it
1747 from the body of a class to a .cpp file).</li>
1749 <li>Restarting the program breaks breakpoints. After following the information
1750 above, you have succeeded in getting some breakpoints planted in your pass. Nex
1751 thing you know, you restart the program (i.e., you type '<tt>run</tt>' again),
1752 and you start getting errors about breakpoints being unsettable. The only way I
1754 already set in your pass, run the program, and re-set the breakpoints once
1755 execution stops in <tt>PassManager::run</tt>.</li>
1757 </ul>
1759 <p>Hopefully these tips will help with common case debugging situations. If
1760 you'd like to contribute some tips of your own, just contact <a
1763 </div>
1765 <!-- *********************************************************************** -->
1768 </div>
1769 <!-- *********************************************************************** -->
1773 <p>Although the LLVM Pass Infrastructure is very capable as it stands, and does
1774 some nifty stuff, there are things we'd like to add in the future. Here is
1775 where we are going:</p>
1777 </div>
1779 <!-- _______________________________________________________________________ -->
1782 </div>
1786 <p>Multiple CPU machines are becoming more common and compilation can never be
1787 fast enough: obviously we should allow for a multithreaded compiler. Because of
1788 the semantics defined for passes above (specifically they cannot maintain state
1789 across invocations of their <tt>run*</tt> methods), a nice clean way to
1790 implement a multithreaded compiler would be for the <tt>PassManager</tt> class
1791 to create multiple instances of each pass object, and allow the separate
1792 instances to be hacking on different parts of the program at the same time.</p>
1794 <p>This implementation would prevent each of the passes from having to implement
1795 multithreaded constructs, requiring only the LLVM core to have locking in a few
1796 places (for global resources). Although this is a simple extension, we simply
1797 haven't had time (or multiprocessor machines, thus a reason) to implement this.
1798 Despite that, we have kept the LLVM passes SMP ready, and you should too.</p>
1800 </div>
1802 <!-- *********************************************************************** -->
1803 <hr>
1804 <address>
1812 Last modified: $Date$
1813 </address>
1815 </body>
1816 </html>