1 <!DOCTYPE HTML PUBLIC "-//W3C//DTD HTML 4.01//EN"
2 "http://www.w3.org/TR/html4/strict.dtd">
5 <title>Exception Handling in LLVM</title>
6 <meta http-equiv="Content-Type" content="text/html; charset=utf-8">
7 <meta name="description"
8 content="Exception Handling in LLVM.">
9 <link rel="stylesheet" href="llvm.css" type="text/css">
14 <h1>Exception Handling in LLVM</h1>
16 <table class="layout" style="width:100%">
20 <li><a href="#introduction">Introduction</a>
22 <li><a href="#itanium">Itanium ABI Zero-cost Exception Handling</a></li>
23 <li><a href="#sjlj">Setjmp/Longjmp Exception Handling</a></li>
24 <li><a href="#overview">Overview</a></li>
26 <li><a href="#codegen">LLVM Code Generation</a>
28 <li><a href="#throw">Throw</a></li>
29 <li><a href="#try_catch">Try/Catch</a></li>
30 <li><a href="#cleanups">Cleanups</a></li>
31 <li><a href="#throw_filters">Throw Filters</a></li>
32 <li><a href="#restrictions">Restrictions</a></li>
34 <li><a href="#format_common_intrinsics">Exception Handling Intrinsics</a>
36 <li><a href="#llvm_eh_typeid_for"><tt>llvm.eh.typeid.for</tt></a></li>
37 <li><a href="#llvm_eh_sjlj_setjmp"><tt>llvm.eh.sjlj.setjmp</tt></a></li>
38 <li><a href="#llvm_eh_sjlj_longjmp"><tt>llvm.eh.sjlj.longjmp</tt></a></li>
39 <li><a href="#llvm_eh_sjlj_lsda"><tt>llvm.eh.sjlj.lsda</tt></a></li>
40 <li><a href="#llvm_eh_sjlj_callsite"><tt>llvm.eh.sjlj.callsite</tt></a></li>
41 <li><a href="#llvm_eh_sjlj_dispatchsetup"><tt>llvm.eh.sjlj.dispatchsetup</tt></a></li>
43 <li><a href="#asm">Asm Table Formats</a>
45 <li><a href="#unwind_tables">Exception Handling Frame</a></li>
46 <li><a href="#exception_tables">Exception Tables</a></li>
48 <li><a href="#todo">ToDo</a></li>
53 <div class="doc_author">
54 <p>Written by <a href="mailto:jlaskey@mac.com">Jim Laskey</a></p>
58 <!-- *********************************************************************** -->
59 <h2><a name="introduction">Introduction</a></h2>
60 <!-- *********************************************************************** -->
64 <p>This document is the central repository for all information pertaining to
65 exception handling in LLVM. It describes the format that LLVM exception
66 handling information takes, which is useful for those interested in creating
67 front-ends or dealing directly with the information. Further, this document
68 provides specific examples of what exception handling information is used for
71 <!-- ======================================================================= -->
73 <a name="itanium">Itanium ABI Zero-cost Exception Handling</a>
78 <p>Exception handling for most programming languages is designed to recover from
79 conditions that rarely occur during general use of an application. To that
80 end, exception handling should not interfere with the main flow of an
81 application's algorithm by performing checkpointing tasks, such as saving the
82 current pc or register state.</p>
84 <p>The Itanium ABI Exception Handling Specification defines a methodology for
85 providing outlying data in the form of exception tables without inlining
86 speculative exception handling code in the flow of an application's main
87 algorithm. Thus, the specification is said to add "zero-cost" to the normal
88 execution of an application.</p>
90 <p>A more complete description of the Itanium ABI exception handling runtime
91 support of can be found at
92 <a href="http://www.codesourcery.com/cxx-abi/abi-eh.html">Itanium C++ ABI:
93 Exception Handling</a>. A description of the exception frame format can be
95 <a href="http://refspecs.freestandards.org/LSB_3.0.0/LSB-Core-generic/LSB-Core-generic/ehframechpt.html">Exception
96 Frames</a>, with details of the DWARF 4 specification at
97 <a href="http://dwarfstd.org/Dwarf4Std.php">DWARF 4 Standard</a>.
98 A description for the C++ exception table formats can be found at
99 <a href="http://www.codesourcery.com/cxx-abi/exceptions.pdf">Exception Handling
104 <!-- ======================================================================= -->
106 <a name="sjlj">Setjmp/Longjmp Exception Handling</a>
111 <p>Setjmp/Longjmp (SJLJ) based exception handling uses LLVM intrinsics
112 <a href="#llvm_eh_sjlj_setjmp"><tt>llvm.eh.sjlj.setjmp</tt></a> and
113 <a href="#llvm_eh_sjlj_longjmp"><tt>llvm.eh.sjlj.longjmp</tt></a> to
114 handle control flow for exception handling.</p>
116 <p>For each function which does exception processing — be
117 it <tt>try</tt>/<tt>catch</tt> blocks or cleanups — that function
118 registers itself on a global frame list. When exceptions are unwinding, the
119 runtime uses this list to identify which functions need processing.<p>
121 <p>Landing pad selection is encoded in the call site entry of the function
122 context. The runtime returns to the function via
123 <a href="#llvm_eh_sjlj_longjmp"><tt>llvm.eh.sjlj.longjmp</tt></a>, where
124 a switch table transfers control to the appropriate landing pad based on
125 the index stored in the function context.</p>
127 <p>In contrast to DWARF exception handling, which encodes exception regions
128 and frame information in out-of-line tables, SJLJ exception handling
129 builds and removes the unwind frame context at runtime. This results in
130 faster exception handling at the expense of slower execution when no
131 exceptions are thrown. As exceptions are, by their nature, intended for
132 uncommon code paths, DWARF exception handling is generally preferred to
137 <!-- ======================================================================= -->
139 <a name="overview">Overview</a>
144 <p>When an exception is thrown in LLVM code, the runtime does its best to find a
145 handler suited to processing the circumstance.</p>
147 <p>The runtime first attempts to find an <i>exception frame</i> corresponding to
148 the function where the exception was thrown. If the programming language
149 supports exception handling (e.g. C++), the exception frame contains a
150 reference to an exception table describing how to process the exception. If
151 the language does not support exception handling (e.g. C), or if the
152 exception needs to be forwarded to a prior activation, the exception frame
153 contains information about how to unwind the current activation and restore
154 the state of the prior activation. This process is repeated until the
155 exception is handled. If the exception is not handled and no activations
156 remain, then the application is terminated with an appropriate error
159 <p>Because different programming languages have different behaviors when
160 handling exceptions, the exception handling ABI provides a mechanism for
161 supplying <i>personalities</i>. An exception handling personality is defined
162 by way of a <i>personality function</i> (e.g. <tt>__gxx_personality_v0</tt>
163 in C++), which receives the context of the exception, an <i>exception
164 structure</i> containing the exception object type and value, and a reference
165 to the exception table for the current function. The personality function
166 for the current compile unit is specified in a <i>common exception
169 <p>The organization of an exception table is language dependent. For C++, an
170 exception table is organized as a series of code ranges defining what to do
171 if an exception occurs in that range. Typically, the information associated
172 with a range defines which types of exception objects (using C++ <i>type
173 info</i>) that are handled in that range, and an associated action that
174 should take place. Actions typically pass control to a <i>landing
177 <p>A landing pad corresponds roughly to the code found in the <tt>catch</tt>
178 portion of a <tt>try</tt>/<tt>catch</tt> sequence. When execution resumes at
179 a landing pad, it receives an <i>exception structure</i> and a
180 <i>selector value</i> corresponding to the <i>type</i> of exception
181 thrown. The selector is then used to determine which <i>catch</i> should
182 actually process the exception.</p>
188 <!-- ======================================================================= -->
190 <a name="codegen">LLVM Code Generation</a>
195 <p>From a C++ developer's perspective, exceptions are defined in terms of the
196 <tt>throw</tt> and <tt>try</tt>/<tt>catch</tt> statements. In this section
197 we will describe the implementation of LLVM exception handling in terms of
200 <!-- ======================================================================= -->
202 <a name="throw">Throw</a>
207 <p>Languages that support exception handling typically provide a <tt>throw</tt>
208 operation to initiate the exception process. Internally, a <tt>throw</tt>
209 operation breaks down into two steps.</p>
212 <li>A request is made to allocate exception space for an exception structure.
213 This structure needs to survive beyond the current activation. This
214 structure will contain the type and value of the object being thrown.</li>
216 <li>A call is made to the runtime to raise the exception, passing the
217 exception structure as an argument.</li>
220 <p>In C++, the allocation of the exception structure is done by the
221 <tt>__cxa_allocate_exception</tt> runtime function. The exception raising is
222 handled by <tt>__cxa_throw</tt>. The type of the exception is represented
223 using a C++ RTTI structure.</p>
227 <!-- ======================================================================= -->
229 <a name="try_catch">Try/Catch</a>
234 <p>A call within the scope of a <i>try</i> statement can potentially raise an
235 exception. In those circumstances, the LLVM C++ front-end replaces the call
236 with an <tt>invoke</tt> instruction. Unlike a call, the <tt>invoke</tt> has
237 two potential continuation points:</p>
240 <li>where to continue when the call succeeds as per normal, and</li>
242 <li>where to continue if the call raises an exception, either by a throw or
243 the unwinding of a throw</li>
246 <p>The term used to define a the place where an <tt>invoke</tt> continues after
247 an exception is called a <i>landing pad</i>. LLVM landing pads are
248 conceptually alternative function entry points where an exception structure
249 reference and a type info index are passed in as arguments. The landing pad
250 saves the exception structure reference and then proceeds to select the catch
251 block that corresponds to the type info of the exception object.</p>
253 <p>The LLVM <a href="LangRef.html#i_landingpad"><tt>landingpad</tt>
254 instruction</a> is used to convey information about the landing pad to the
255 back end. For C++, the <tt>landingpad</tt> instruction returns a pointer and
256 integer pair corresponding to the pointer to the <i>exception structure</i>
257 and the <i>selector value</i> respectively.</p>
259 <p>The <tt>landingpad</tt> instruction takes a reference to the personality
260 function to be used for this <tt>try</tt>/<tt>catch</tt> sequence. The
261 remainder of the instruction is a list of <i>cleanup</i>, <i>catch</i>,
262 and <i>filter</i> clauses. The exception is tested against the clauses
263 sequentially from first to last. The selector value is a positive number if
264 the exception matched a type info, a negative number if it matched a filter,
265 and zero if it matched a cleanup. If nothing is matched, the behavior of the
266 program is <a href="#restrictions">undefined</a>. If a type info matched,
267 then the selector value is the index of the type info in the exception table,
268 which can be obtained using the
269 <a href="#llvm_eh_typeid_for"><tt>llvm.eh.typeid.for</tt></a> intrinsic.</p>
271 <p>Once the landing pad has the type info selector, the code branches to the
272 code for the first catch. The catch then checks the value of the type info
273 selector against the index of type info for that catch. Since the type info
274 index is not known until all the type infos have been gathered in the
275 backend, the catch code must call the
276 <a href="#llvm_eh_typeid_for"><tt>llvm.eh.typeid.for</tt></a> intrinsic to
277 determine the index for a given type info. If the catch fails to match the
278 selector then control is passed on to the next catch.</p>
280 <p><b>Note:</b> Since the landing pad will not be used if there is no match in
281 the list of type info on the call to the <tt>landingpad</tt> instruction,
282 then neither the last catch nor <i>catch all</i> need to perform the check
283 against the selector.</p>
285 <p>Finally, the entry and exit of catch code is bracketed with calls to
286 <tt>__cxa_begin_catch</tt> and <tt>__cxa_end_catch</tt>.</p>
289 <li><tt>__cxa_begin_catch</tt> takes an exception structure reference as an
290 argument and returns the value of the exception object.</li>
292 <li><tt>__cxa_end_catch</tt> takes no arguments. This function:<br><br>
294 <li>Locates the most recently caught exception and decrements its handler
296 <li>Removes the exception from the <i>caught</i> stack if the handler
297 count goes to zero, and</li>
298 <li>Destroys the exception if the handler count goes to zero and the
299 exception was not re-thrown by throw.</li>
301 <p><b>Note:</b> a rethrow from within the catch may replace this call with
302 a <tt>__cxa_rethrow</tt>.</p></li>
307 <!-- ======================================================================= -->
309 <a name="cleanups">Cleanups</a>
314 <p>A cleanup is extra code which needs to be run as part of unwinding a scope.
315 C++ destructors are a typical example, but other languages and language
316 extensions provide a variety of different kinds of cleanups. In general, a
317 landing pad may need to run arbitrary amounts of cleanup code before actually
318 entering a catch block. To indicate the presence of cleanups, a
319 <a href="LangRef.html#i_landingpad"><tt>landingpad</tt> instruction</a>
320 should have a <i>cleanup</i> clause. Otherwise, the unwinder will not stop at
321 the landing pad if there are no catches or filters that require it to.</p>
323 <p><b>Note:</b> Do not allow a new exception to propagate out of the execution
324 of a cleanup. This can corrupt the internal state of the unwinder.
325 Different languages describe different high-level semantics for these
326 situations: for example, C++ requires that the process be terminated, whereas
327 Ada cancels both exceptions and throws a third.</p>
329 <p>When all cleanups are finished, if the exception is not handled by the
330 current function, resume unwinding by calling the
331 <a href="LangRef.html#i_resume"><tt>resume</tt> instruction</a>, passing in
332 the result of the <tt>landingpad</tt> instruction for the original landing
337 <!-- ======================================================================= -->
339 <a name="throw_filters">Throw Filters</a>
344 <p>C++ allows the specification of which exception types may be thrown from a
345 function. To represent this, a top level landing pad may exist to filter out
346 invalid types. To express this in LLVM code the
347 <a href="LangRef.html#i_landingpad"><tt>landingpad</tt> instruction</a> will
348 have a filter clause. The clause consists of an array of type infos.
349 <tt>landingpad</tt> will return a negative value if the exception does not
350 match any of the type infos. If no match is found then a call
351 to <tt>__cxa_call_unexpected</tt> should be made, otherwise
352 <tt>_Unwind_Resume</tt>. Each of these functions requires a reference to the
353 exception structure. Note that the most general form of a
354 <a href="LangRef.html#i_landingpad"><tt>landingpad</tt> instruction</a> can
355 have any number of catch, cleanup, and filter clauses (though having more
356 than one cleanup is pointless). The LLVM C++ front-end can generate such
357 <a href="LangRef.html#i_landingpad"><tt>landingpad</tt> instructions</a> due
358 to inlining creating nested exception handling scopes.</p>
362 <!-- ======================================================================= -->
364 <a name="restrictions">Restrictions</a>
369 <p>The unwinder delegates the decision of whether to stop in a call frame to
370 that call frame's language-specific personality function. Not all personality
371 functions guarantee that they will stop to perform cleanups. For example, the
372 GNU C++ personality function doesn't do so unless the exception is actually
373 caught somewhere further up the stack. When using this personality to
374 implement EH for a language that guarantees that cleanups will always be run
375 (e.g. Ada), be sure to indicate a catch-all in the
376 <a href="LangRef.html#i_landingpad"><tt>landingpad</tt> instruction</a>
377 rather than just cleanups.</p>
379 <p>In order for inlining to behave correctly, landing pads must be prepared to
380 handle selector results that they did not originally advertise. Suppose that
381 a function catches exceptions of type <tt>A</tt>, and it's inlined into a
382 function that catches exceptions of type <tt>B</tt>. The inliner will update
383 the <tt>landingpad</tt> instruction for the inlined landing pad to include
384 the fact that <tt>B</tt> is also caught. If that landing pad assumes that it
385 will only be entered to catch an <tt>A</tt>, it's in for a rude awakening.
386 Consequently, landing pads must test for the selector results they understand
387 and then resume exception propagation with the
388 <a href="LangRef.html#i_resume"><tt>resume</tt> instruction</a> if none of
389 the conditions match.</p>
395 <!-- ======================================================================= -->
397 <a name="format_common_intrinsics">Exception Handling Intrinsics</a>
402 <p>In addition to the
403 <a href="LangRef.html#i_landingpad"><tt>landingpad</tt></a> and
404 <a href="LangRef.html#i_resume"><tt>resume</tt></a> instructions, LLVM uses
405 several intrinsic functions (name prefixed with <i><tt>llvm.eh</tt></i>) to
406 provide exception handling information at various points in generated
409 <!-- ======================================================================= -->
411 <a name="llvm_eh_typeid_for">llvm.eh.typeid.for</a>
417 i32 @llvm.eh.typeid.for(i8* %type_info)
420 <p>This intrinsic returns the type info index in the exception table of the
421 current function. This value can be used to compare against the result
422 of <a href="LangRef.html#i_landingpad"><tt>landingpad</tt> instruction</a>.
423 The single argument is a reference to a type info.</p>
427 <!-- ======================================================================= -->
429 <a name="llvm_eh_sjlj_setjmp">llvm.eh.sjlj.setjmp</a>
435 i32 @llvm.eh.sjlj.setjmp(i8* %setjmp_buf)
438 <p>For SJLJ based exception handling, this intrinsic forces register saving for
439 the current function and stores the address of the following instruction for
440 use as a destination address
441 by <a href="#llvm_eh_sjlj_longjmp"><tt>llvm.eh.sjlj.longjmp</tt></a>. The
442 buffer format and the overall functioning of this intrinsic is compatible
443 with the GCC <tt>__builtin_setjmp</tt> implementation allowing code built
444 with the clang and GCC to interoperate.</p>
446 <p>The single parameter is a pointer to a five word buffer in which the calling
447 context is saved. The front end places the frame pointer in the first word,
448 and the target implementation of this intrinsic should place the destination
450 <a href="#llvm_eh_sjlj_longjmp"><tt>llvm.eh.sjlj.longjmp</tt></a> in the
451 second word. The following three words are available for use in a
452 target-specific manner.</p>
456 <!-- ======================================================================= -->
458 <a name="llvm_eh_sjlj_longjmp">llvm.eh.sjlj.longjmp</a>
464 void @llvm.eh.sjlj.longjmp(i8* %setjmp_buf)
467 <p>For SJLJ based exception handling, the <tt>llvm.eh.sjlj.longjmp</tt>
468 intrinsic is used to implement <tt>__builtin_longjmp()</tt>. The single
469 parameter is a pointer to a buffer populated
470 by <a href="#llvm_eh_sjlj_setjmp"><tt>llvm.eh.sjlj.setjmp</tt></a>. The frame
471 pointer and stack pointer are restored from the buffer, then control is
472 transferred to the destination address.</p>
475 <!-- ======================================================================= -->
477 <a name="llvm_eh_sjlj_lsda">llvm.eh.sjlj.lsda</a>
483 i8* @llvm.eh.sjlj.lsda()
486 <p>For SJLJ based exception handling, the <tt>llvm.eh.sjlj.lsda</tt> intrinsic
487 returns the address of the Language Specific Data Area (LSDA) for the current
488 function. The SJLJ front-end code stores this address in the exception
489 handling function context for use by the runtime.</p>
493 <!-- ======================================================================= -->
495 <a name="llvm_eh_sjlj_callsite">llvm.eh.sjlj.callsite</a>
501 void @llvm.eh.sjlj.callsite(i32 %call_site_num)
504 <p>For SJLJ based exception handling, the <tt>llvm.eh.sjlj.callsite</tt>
505 intrinsic identifies the callsite value associated with the
506 following <tt>invoke</tt> instruction. This is used to ensure that landing
507 pad entries in the LSDA are generated in matching order.</p>
511 <!-- ======================================================================= -->
513 <a name="llvm_eh_sjlj_dispatchsetup">llvm.eh.sjlj.dispatchsetup</a>
519 void @llvm.eh.sjlj.dispatchsetup(i32 %dispatch_value)
522 <p>For SJLJ based exception handling, the <tt>llvm.eh.sjlj.dispatchsetup</tt>
523 intrinsic is used by targets to do any unwind edge setup they need. By
524 default, no action is taken.</p>
530 <!-- ======================================================================= -->
532 <a name="asm">Asm Table Formats</a>
537 <p>There are two tables that are used by the exception handling runtime to
538 determine which actions should be taken when an exception is thrown.</p>
540 <!-- ======================================================================= -->
542 <a name="unwind_tables">Exception Handling Frame</a>
547 <p>An exception handling frame <tt>eh_frame</tt> is very similar to the unwind
548 frame used by DWARF debug info. The frame contains all the information
549 necessary to tear down the current frame and restore the state of the prior
550 frame. There is an exception handling frame for each function in a compile
551 unit, plus a common exception handling frame that defines information common
552 to all functions in the unit.</p>
554 <!-- Todo - Table details here. -->
558 <!-- ======================================================================= -->
560 <a name="exception_tables">Exception Tables</a>
565 <p>An exception table contains information about what actions to take when an
566 exception is thrown in a particular part of a function's code. There is one
567 exception table per function, except leaf functions and functions that have
568 calls only to non-throwing functions. They do not need an exception
571 <!-- Todo - Table details here. -->
577 <!-- *********************************************************************** -->
581 <a href="http://jigsaw.w3.org/css-validator/check/referer"><img
582 src="http://jigsaw.w3.org/css-validator/images/vcss-blue" alt="Valid CSS"></a>
583 <a href="http://validator.w3.org/check/referer"><img
584 src="http://www.w3.org/Icons/valid-html401-blue" alt="Valid HTML 4.01"></a>
586 <a href="mailto:sabre@nondot.org">Chris Lattner</a><br>
587 <a href="http://llvm.org/">LLVM Compiler Infrastructure</a><br>
588 Last modified: $Date$