xref: /llvm-project/llvm/docs/ExceptionHandling.rst (revision dc04d414df9c243bb90d7cfc683a632a2c032c62)
1==========================
2Exception Handling in LLVM
3==========================
4
5.. contents::
6   :local:
7
8Introduction
9============
10
11This document is the central repository for all information pertaining to
12exception handling in LLVM.  It describes the format that LLVM exception
13handling information takes, which is useful for those interested in creating
14front-ends or dealing directly with the information.  Further, this document
15provides specific examples of what exception handling information is used for in
16C and C++.
17
18Itanium ABI Zero-cost Exception Handling
19----------------------------------------
20
21Exception handling for most programming languages is designed to recover from
22conditions that rarely occur during general use of an application.  To that end,
23exception handling should not interfere with the main flow of an application's
24algorithm by performing checkpointing tasks, such as saving the current pc or
25register state.
26
27The Itanium ABI Exception Handling Specification defines a methodology for
28providing outlying data in the form of exception tables without inlining
29speculative exception handling code in the flow of an application's main
30algorithm.  Thus, the specification is said to add "zero-cost" to the normal
31execution of an application.
32
33A more complete description of the Itanium ABI exception handling runtime
34support of can be found at `Itanium C++ ABI: Exception Handling
35<http://itanium-cxx-abi.github.io/cxx-abi/abi-eh.html>`_. A description of the
36exception frame format can be found at `Exception Frames
37<http://refspecs.linuxfoundation.org/LSB_3.0.0/LSB-Core-generic/LSB-Core-generic/ehframechpt.html>`_,
38with details of the DWARF 4 specification at `DWARF 4 Standard
39<http://dwarfstd.org/Dwarf4Std.php>`_.  A description for the C++ exception
40table formats can be found at `Exception Handling Tables
41<http://itanium-cxx-abi.github.io/cxx-abi/exceptions.pdf>`_.
42
43Setjmp/Longjmp Exception Handling
44---------------------------------
45
46Setjmp/Longjmp (SJLJ) based exception handling uses LLVM intrinsics
47`llvm.eh.sjlj.setjmp`_ and `llvm.eh.sjlj.longjmp`_ to handle control flow for
48exception handling.
49
50For each function which does exception processing --- be it ``try``/``catch``
51blocks or cleanups --- that function registers itself on a global frame
52list. When exceptions are unwinding, the runtime uses this list to identify
53which functions need processing.
54
55Landing pad selection is encoded in the call site entry of the function
56context. The runtime returns to the function via `llvm.eh.sjlj.longjmp`_, where
57a switch table transfers control to the appropriate landing pad based on the
58index stored in the function context.
59
60In contrast to DWARF exception handling, which encodes exception regions and
61frame information in out-of-line tables, SJLJ exception handling builds and
62removes the unwind frame context at runtime. This results in faster exception
63handling at the expense of slower execution when no exceptions are thrown. As
64exceptions are, by their nature, intended for uncommon code paths, DWARF
65exception handling is generally preferred to SJLJ.
66
67Windows Runtime Exception Handling
68-----------------------------------
69
70LLVM supports handling exceptions produced by the Windows runtime, but it
71requires a very different intermediate representation. It is not based on the
72":ref:`landingpad <i_landingpad>`" instruction like the other two models, and is
73described later in this document under :ref:`wineh`.
74
75Overview
76--------
77
78When an exception is thrown in LLVM code, the runtime does its best to find a
79handler suited to processing the circumstance.
80
81The runtime first attempts to find an *exception frame* corresponding to the
82function where the exception was thrown.  If the programming language supports
83exception handling (e.g. C++), the exception frame contains a reference to an
84exception table describing how to process the exception.  If the language does
85not support exception handling (e.g. C), or if the exception needs to be
86forwarded to a prior activation, the exception frame contains information about
87how to unwind the current activation and restore the state of the prior
88activation.  This process is repeated until the exception is handled. If the
89exception is not handled and no activations remain, then the application is
90terminated with an appropriate error message.
91
92Because different programming languages have different behaviors when handling
93exceptions, the exception handling ABI provides a mechanism for
94supplying *personalities*. An exception handling personality is defined by
95way of a *personality function* (e.g. ``__gxx_personality_v0`` in C++),
96which receives the context of the exception, an *exception structure*
97containing the exception object type and value, and a reference to the exception
98table for the current function.  The personality function for the current
99compile unit is specified in a *common exception frame*.
100
101The organization of an exception table is language dependent. For C++, an
102exception table is organized as a series of code ranges defining what to do if
103an exception occurs in that range. Typically, the information associated with a
104range defines which types of exception objects (using C++ *type info*) that are
105handled in that range, and an associated action that should take place. Actions
106typically pass control to a *landing pad*.
107
108A landing pad corresponds roughly to the code found in the ``catch`` portion of
109a ``try``/``catch`` sequence. When execution resumes at a landing pad, it
110receives an *exception structure* and a *selector value* corresponding to the
111*type* of exception thrown. The selector is then used to determine which *catch*
112should actually process the exception.
113
114LLVM Code Generation
115====================
116
117From a C++ developer's perspective, exceptions are defined in terms of the
118``throw`` and ``try``/``catch`` statements. In this section we will describe the
119implementation of LLVM exception handling in terms of C++ examples.
120
121Throw
122-----
123
124Languages that support exception handling typically provide a ``throw``
125operation to initiate the exception process. Internally, a ``throw`` operation
126breaks down into two steps.
127
128#. A request is made to allocate exception space for an exception structure.
129   This structure needs to survive beyond the current activation. This structure
130   will contain the type and value of the object being thrown.
131
132#. A call is made to the runtime to raise the exception, passing the exception
133   structure as an argument.
134
135In C++, the allocation of the exception structure is done by the
136``__cxa_allocate_exception`` runtime function. The exception raising is handled
137by ``__cxa_throw``. The type of the exception is represented using a C++ RTTI
138structure.
139
140Try/Catch
141---------
142
143A call within the scope of a *try* statement can potentially raise an
144exception. In those circumstances, the LLVM C++ front-end replaces the call with
145an ``invoke`` instruction. Unlike a call, the ``invoke`` has two potential
146continuation points:
147
148#. where to continue when the call succeeds as per normal, and
149
150#. where to continue if the call raises an exception, either by a throw or the
151   unwinding of a throw
152
153The term used to define the place where an ``invoke`` continues after an
154exception is called a *landing pad*. LLVM landing pads are conceptually
155alternative function entry points where an exception structure reference and a
156type info index are passed in as arguments. The landing pad saves the exception
157structure reference and then proceeds to select the catch block that corresponds
158to the type info of the exception object.
159
160The LLVM :ref:`i_landingpad` is used to convey information about the landing
161pad to the back end. For C++, the ``landingpad`` instruction returns a pointer
162and integer pair corresponding to the pointer to the *exception structure* and
163the *selector value* respectively.
164
165The ``landingpad`` instruction looks for a reference to the personality
166function to be used for this ``try``/``catch`` sequence in the parent
167function's attribute list. The instruction contains a list of *cleanup*,
168*catch*, and *filter* clauses. The exception is tested against the clauses
169sequentially from first to last. The clauses have the following meanings:
170
171-  ``catch <type> @ExcType``
172
173   - This clause means that the landingpad block should be entered if the
174     exception being thrown is of type ``@ExcType`` or a subtype of
175     ``@ExcType``. For C++, ``@ExcType`` is a pointer to the ``std::type_info``
176     object (an RTTI object) representing the C++ exception type.
177
178   - If ``@ExcType`` is ``null``, any exception matches, so the landingpad
179     should always be entered. This is used for C++ catch-all blocks ("``catch
180     (...)``").
181
182   - When this clause is matched, the selector value will be equal to the value
183     returned by "``@llvm.eh.typeid.for(i8* @ExcType)``". This will always be a
184     positive value.
185
186-  ``filter <type> [<type> @ExcType1, ..., <type> @ExcTypeN]``
187
188   - This clause means that the landingpad should be entered if the exception
189     being thrown does *not* match any of the types in the list (which, for C++,
190     are again specified as ``std::type_info`` pointers).
191
192   - C++ front-ends use this to implement the C++ exception specifications, such as
193     "``void foo() throw (ExcType1, ..., ExcTypeN) { ... }``". (Note: this
194     functionality was deprecated in C++11 and removed in C++17.)
195
196   - When this clause is matched, the selector value will be negative.
197
198   - The array argument to ``filter`` may be empty; for example, "``[0 x i8**]
199     undef``". This means that the landingpad should always be entered. (Note
200     that such a ``filter`` would not be equivalent to "``catch i8* null``",
201     because ``filter`` and ``catch`` produce negative and positive selector
202     values respectively.)
203
204-  ``cleanup``
205
206   - This clause means that the landingpad should always be entered.
207
208   - C++ front-ends use this for calling objects' destructors.
209
210   - When this clause is matched, the selector value will be zero.
211
212   - The runtime may treat "``cleanup``" differently from "``catch <type>
213     null``".
214
215     In C++, if an unhandled exception occurs, the language runtime will call
216     ``std::terminate()``, but it is implementation-defined whether the runtime
217     unwinds the stack and calls object destructors first. For example, the GNU
218     C++ unwinder does not call object destructors when an unhandled exception
219     occurs. The reason for this is to improve debuggability: it ensures that
220     ``std::terminate()`` is called from the context of the ``throw``, so that
221     this context is not lost by unwinding the stack. A runtime will typically
222     implement this by searching for a matching non-``cleanup`` clause, and
223     aborting if it does not find one, before entering any landingpad blocks.
224
225Once the landing pad has the type info selector, the code branches to the code
226for the first catch. The catch then checks the value of the type info selector
227against the index of type info for that catch.  Since the type info index is not
228known until all the type infos have been gathered in the backend, the catch code
229must call the `llvm.eh.typeid.for`_ intrinsic to determine the index for a given
230type info. If the catch fails to match the selector then control is passed on to
231the next catch.
232
233Finally, the entry and exit of catch code is bracketed with calls to
234``__cxa_begin_catch`` and ``__cxa_end_catch``.
235
236* ``__cxa_begin_catch`` takes an exception structure reference as an argument
237  and returns the value of the exception object.
238
239* ``__cxa_end_catch`` takes no arguments. This function:
240
241  #. Locates the most recently caught exception and decrements its handler
242     count,
243
244  #. Removes the exception from the *caught* stack if the handler count goes to
245     zero, and
246
247  #. Destroys the exception if the handler count goes to zero and the exception
248     was not re-thrown by throw.
249
250  .. note::
251
252    a rethrow from within the catch may replace this call with a
253    ``__cxa_rethrow``.
254
255Cleanups
256--------
257
258A cleanup is extra code which needs to be run as part of unwinding a scope.  C++
259destructors are a typical example, but other languages and language extensions
260provide a variety of different kinds of cleanups. In general, a landing pad may
261need to run arbitrary amounts of cleanup code before actually entering a catch
262block. To indicate the presence of cleanups, a :ref:`i_landingpad` should have
263a *cleanup* clause.  Otherwise, the unwinder will not stop at the landing pad if
264there are no catches or filters that require it to.
265
266.. note::
267
268  Do not allow a new exception to propagate out of the execution of a
269  cleanup. This can corrupt the internal state of the unwinder.  Different
270  languages describe different high-level semantics for these situations: for
271  example, C++ requires that the process be terminated, whereas Ada cancels both
272  exceptions and throws a third.
273
274When all cleanups are finished, if the exception is not handled by the current
275function, resume unwinding by calling the :ref:`resume instruction <i_resume>`,
276passing in the result of the ``landingpad`` instruction for the original
277landing pad.
278
279Throw Filters
280-------------
281
282Prior to C++17, C++ allowed the specification of which exception types may be
283thrown from a function. To represent this, a top level landing pad may exist to
284filter out invalid types. To express this in LLVM code the :ref:`i_landingpad`
285will have a filter clause. The clause consists of an array of type infos.
286``landingpad`` will return a negative value
287if the exception does not match any of the type infos. If no match is found then
288a call to ``__cxa_call_unexpected`` should be made, otherwise
289``_Unwind_Resume``.  Each of these functions requires a reference to the
290exception structure.  Note that the most general form of a ``landingpad``
291instruction can have any number of catch, cleanup, and filter clauses (though
292having more than one cleanup is pointless). The LLVM C++ front-end can generate
293such ``landingpad`` instructions due to inlining creating nested exception
294handling scopes.
295
296Restrictions
297------------
298
299The unwinder delegates the decision of whether to stop in a call frame to that
300call frame's language-specific personality function. Not all unwinders guarantee
301that they will stop to perform cleanups. For example, the GNU C++ unwinder
302doesn't do so unless the exception is actually caught somewhere further up the
303stack.
304
305In order for inlining to behave correctly, landing pads must be prepared to
306handle selector results that they did not originally advertise. Suppose that a
307function catches exceptions of type ``A``, and it's inlined into a function that
308catches exceptions of type ``B``. The inliner will update the ``landingpad``
309instruction for the inlined landing pad to include the fact that ``B`` is also
310caught. If that landing pad assumes that it will only be entered to catch an
311``A``, it's in for a rude awakening.  Consequently, landing pads must test for
312the selector results they understand and then resume exception propagation with
313the `resume instruction <LangRef.html#i_resume>`_ if none of the conditions
314match.
315
316Exception Handling Intrinsics
317=============================
318
319In addition to the ``landingpad`` and ``resume`` instructions, LLVM uses several
320intrinsic functions (name prefixed with ``llvm.eh``) to provide exception
321handling information at various points in generated code.
322
323.. _llvm.eh.typeid.for:
324
325``llvm.eh.typeid.for``
326----------------------
327
328.. code-block:: llvm
329
330  i32 @llvm.eh.typeid.for(i8* %type_info)
331
332
333This intrinsic returns the type info index in the exception table of the current
334function.  This value can be used to compare against the result of
335``landingpad`` instruction.  The single argument is a reference to a type info.
336
337Uses of this intrinsic are generated by the C++ front-end.
338
339.. _llvm.eh.exceptionpointer:
340
341``llvm.eh.exceptionpointer``
342----------------------------
343
344.. code-block:: text
345
346  i8 addrspace(N)* @llvm.eh.padparam.pNi8(token %catchpad)
347
348
349This intrinsic retrieves a pointer to the exception caught by the given
350``catchpad``.
351
352
353SJLJ Intrinsics
354---------------
355
356The ``llvm.eh.sjlj`` intrinsics are used internally within LLVM's
357backend.  Uses of them are generated by the backend's
358``SjLjEHPrepare`` pass.
359
360.. _llvm.eh.sjlj.setjmp:
361
362``llvm.eh.sjlj.setjmp``
363~~~~~~~~~~~~~~~~~~~~~~~
364
365.. code-block:: text
366
367  i32 @llvm.eh.sjlj.setjmp(i8* %setjmp_buf)
368
369For SJLJ based exception handling, this intrinsic forces register saving for the
370current function and stores the address of the following instruction for use as
371a destination address by `llvm.eh.sjlj.longjmp`_. The buffer format and the
372overall functioning of this intrinsic is compatible with the GCC
373``__builtin_setjmp`` implementation allowing code built with the clang and GCC
374to interoperate.
375
376The single parameter is a pointer to a five word buffer in which the calling
377context is saved. The format and contents of the buffer are target-specific.
378On certain targets (ARM, PowerPC, VE, X86), the front end places the
379frame pointer in the first word and the stack pointer in the third word,
380while the target implementation of this intrinsic fills in the remaining
381words.  On other targets (SystemZ), saving the calling context to the buffer
382is left completely to the target implementation.
383
384.. _llvm.eh.sjlj.longjmp:
385
386``llvm.eh.sjlj.longjmp``
387~~~~~~~~~~~~~~~~~~~~~~~~
388
389.. code-block:: llvm
390
391  void @llvm.eh.sjlj.longjmp(i8* %setjmp_buf)
392
393For SJLJ based exception handling, the ``llvm.eh.sjlj.longjmp`` intrinsic is
394used to implement ``__builtin_longjmp()``. The single parameter is a pointer to
395a buffer populated by `llvm.eh.sjlj.setjmp`_. The frame pointer and stack
396pointer are restored from the buffer, then control is transferred to the
397destination address.
398
399``llvm.eh.sjlj.lsda``
400~~~~~~~~~~~~~~~~~~~~~
401
402.. code-block:: llvm
403
404  i8* @llvm.eh.sjlj.lsda()
405
406For SJLJ based exception handling, the ``llvm.eh.sjlj.lsda`` intrinsic returns
407the address of the Language Specific Data Area (LSDA) for the current
408function. The SJLJ front-end code stores this address in the exception handling
409function context for use by the runtime.
410
411``llvm.eh.sjlj.callsite``
412~~~~~~~~~~~~~~~~~~~~~~~~~
413
414.. code-block:: llvm
415
416  void @llvm.eh.sjlj.callsite(i32 %call_site_num)
417
418For SJLJ based exception handling, the ``llvm.eh.sjlj.callsite`` intrinsic
419identifies the callsite value associated with the following ``invoke``
420instruction. This is used to ensure that landing pad entries in the LSDA are
421generated in matching order.
422
423Asm Table Formats
424=================
425
426There are two tables that are used by the exception handling runtime to
427determine which actions should be taken when an exception is thrown.
428
429Exception Handling Frame
430------------------------
431
432An exception handling frame ``eh_frame`` is very similar to the unwind frame
433used by DWARF debug info. The frame contains all the information necessary to
434tear down the current frame and restore the state of the prior frame. There is
435an exception handling frame for each function in a compile unit, plus a common
436exception handling frame that defines information common to all functions in the
437unit.
438
439The format of this call frame information (CFI) is often platform-dependent,
440however. ARM, for example, defines their own format. Apple has their own compact
441unwind info format.  On Windows, another format is used for all architectures
442since 32-bit x86.  LLVM will emit whatever information is required by the
443target.
444
445Exception Tables
446----------------
447
448An exception table contains information about what actions to take when an
449exception is thrown in a particular part of a function's code. This is typically
450referred to as the language-specific data area (LSDA). The format of the LSDA
451table is specific to the personality function, but the majority of personalities
452out there use a variation of the tables consumed by ``__gxx_personality_v0``.
453There is one exception table per function, except leaf functions and functions
454that have calls only to non-throwing functions. They do not need an exception
455table.
456
457.. _wineh:
458
459Exception Handling using the Windows Runtime
460=================================================
461
462Background on Windows exceptions
463---------------------------------
464
465Interacting with exceptions on Windows is significantly more complicated than
466on Itanium C++ ABI platforms. The fundamental difference between the two models
467is that Itanium EH is designed around the idea of "successive unwinding," while
468Windows EH is not.
469
470Under Itanium, throwing an exception typically involves allocating thread local
471memory to hold the exception, and calling into the EH runtime. The runtime
472identifies frames with appropriate exception handling actions, and successively
473resets the register context of the current thread to the most recently active
474frame with actions to run. In LLVM, execution resumes at a ``landingpad``
475instruction, which produces register values provided by the runtime. If a
476function is only cleaning up allocated resources, the function is responsible
477for calling ``_Unwind_Resume`` to transition to the next most recently active
478frame after it is finished cleaning up. Eventually, the frame responsible for
479handling the exception calls ``__cxa_end_catch`` to destroy the exception,
480release its memory, and resume normal control flow.
481
482The Windows EH model does not use these successive register context resets.
483Instead, the active exception is typically described by a frame on the stack.
484In the case of C++ exceptions, the exception object is allocated in stack memory
485and its address is passed to ``__CxxThrowException``. General purpose structured
486exceptions (SEH) are more analogous to Linux signals, and they are dispatched by
487userspace DLLs provided with Windows. Each frame on the stack has an assigned EH
488personality routine, which decides what actions to take to handle the exception.
489There are a few major personalities for C and C++ code: the C++ personality
490(``__CxxFrameHandler3``) and the SEH personalities (``_except_handler3``,
491``_except_handler4``, and ``__C_specific_handler``). All of them implement
492cleanups by calling back into a "funclet" contained in the parent function.
493
494Funclets, in this context, are regions of the parent function that can be called
495as though they were a function pointer with a very special calling convention.
496The frame pointer of the parent frame is passed into the funclet either using
497the standard EBP register or as the first parameter register, depending on the
498architecture. The funclet implements the EH action by accessing local variables
499in memory through the frame pointer, and returning some appropriate value,
500continuing the EH process.  No variables live in to or out of the funclet can be
501allocated in registers.
502
503The C++ personality also uses funclets to contain the code for catch blocks
504(i.e. all user code between the braces in ``catch (Type obj) { ... }``). The
505runtime must use funclets for catch bodies because the C++ exception object is
506allocated in a child stack frame of the function handling the exception. If the
507runtime rewound the stack back to frame of the catch, the memory holding the
508exception would be overwritten quickly by subsequent function calls.  The use of
509funclets also allows ``__CxxFrameHandler3`` to implement rethrow without
510resorting to TLS. Instead, the runtime throws a special exception, and then uses
511SEH (``__try / __except``) to resume execution with new information in the child
512frame.
513
514In other words, the successive unwinding approach is incompatible with Visual
515C++ exceptions and general purpose Windows exception handling. Because the C++
516exception object lives in stack memory, LLVM cannot provide a custom personality
517function that uses landingpads.  Similarly, SEH does not provide any mechanism
518to rethrow an exception or continue unwinding.  Therefore, LLVM must use the IR
519constructs described later in this document to implement compatible exception
520handling.
521
522SEH filter expressions
523-----------------------
524
525The SEH personality functions also use funclets to implement filter expressions,
526which allow executing arbitrary user code to decide which exceptions to catch.
527Filter expressions should not be confused with the ``filter`` clause of the LLVM
528``landingpad`` instruction.  Typically filter expressions are used to determine
529if the exception came from a particular DLL or code region, or if code faulted
530while accessing a particular memory address range. LLVM does not currently have
531IR to represent filter expressions because it is difficult to represent their
532control dependencies.  Filter expressions run during the first phase of EH,
533before cleanups run, making it very difficult to build a faithful control flow
534graph.  For now, the new EH instructions cannot represent SEH filter
535expressions, and frontends must outline them ahead of time. Local variables of
536the parent function can be escaped and accessed using the ``llvm.localescape``
537and ``llvm.localrecover`` intrinsics.
538
539New exception handling instructions
540------------------------------------
541
542The primary design goal of the new EH instructions is to support funclet
543generation while preserving information about the CFG so that SSA formation
544still works.  As a secondary goal, they are designed to be generic across MSVC
545and Itanium C++ exceptions. They make very few assumptions about the data
546required by the personality, so long as it uses the familiar core EH actions:
547catch, cleanup, and terminate.  However, the new instructions are hard to modify
548without knowing details of the EH personality. While they can be used to
549represent Itanium EH, the landingpad model is strictly better for optimization
550purposes.
551
552The following new instructions are considered "exception handling pads", in that
553they must be the first non-phi instruction of a basic block that may be the
554unwind destination of an EH flow edge:
555``catchswitch``, ``catchpad``, and ``cleanuppad``.
556As with landingpads, when entering a try scope, if the
557frontend encounters a call site that may throw an exception, it should emit an
558invoke that unwinds to a ``catchswitch`` block. Similarly, inside the scope of a
559C++ object with a destructor, invokes should unwind to a ``cleanuppad``.
560
561New instructions are also used to mark the points where control is transferred
562out of a catch/cleanup handler (which will correspond to exits from the
563generated funclet).  A catch handler which reaches its end by normal execution
564executes a ``catchret`` instruction, which is a terminator indicating where in
565the function control is returned to.  A cleanup handler which reaches its end
566by normal execution executes a ``cleanupret`` instruction, which is a terminator
567indicating where the active exception will unwind to next.
568
569Each of these new EH pad instructions has a way to identify which action should
570be considered after this action. The ``catchswitch`` instruction is a terminator
571and has an unwind destination operand analogous to the unwind destination of an
572invoke.  The ``cleanuppad`` instruction is not
573a terminator, so the unwind destination is stored on the ``cleanupret``
574instruction instead. Successfully executing a catch handler should resume
575normal control flow, so neither ``catchpad`` nor ``catchret`` instructions can
576unwind. All of these "unwind edges" may refer to a basic block that contains an
577EH pad instruction, or they may unwind to the caller.  Unwinding to the caller
578has roughly the same semantics as the ``resume`` instruction in the landingpad
579model. When inlining through an invoke, instructions that unwind to the caller
580are hooked up to unwind to the unwind destination of the call site.
581
582Putting things together, here is a hypothetical lowering of some C++ that uses
583all of the new IR instructions:
584
585.. code-block:: c
586
587  struct Cleanup {
588    Cleanup();
589    ~Cleanup();
590    int m;
591  };
592  void may_throw();
593  int f() noexcept {
594    try {
595      Cleanup obj;
596      may_throw();
597    } catch (int e) {
598      may_throw();
599      return e;
600    }
601    return 0;
602  }
603
604.. code-block:: text
605
606  define i32 @f() nounwind personality ptr @__CxxFrameHandler3 {
607  entry:
608    %obj = alloca %struct.Cleanup, align 4
609    %e = alloca i32, align 4
610    %call = invoke ptr @"??0Cleanup@@QEAA@XZ"(ptr nonnull %obj)
611            to label %invoke.cont unwind label %lpad.catch
612
613  invoke.cont:                                      ; preds = %entry
614    invoke void @"?may_throw@@YAXXZ"()
615            to label %invoke.cont.2 unwind label %lpad.cleanup
616
617  invoke.cont.2:                                    ; preds = %invoke.cont
618    call void @"??_DCleanup@@QEAA@XZ"(ptr nonnull %obj) nounwind
619    br label %return
620
621  return:                                           ; preds = %invoke.cont.3, %invoke.cont.2
622    %retval.0 = phi i32 [ 0, %invoke.cont.2 ], [ %3, %invoke.cont.3 ]
623    ret i32 %retval.0
624
625  lpad.cleanup:                                     ; preds = %invoke.cont.2
626    %0 = cleanuppad within none []
627    call void @"??1Cleanup@@QEAA@XZ"(ptr nonnull %obj) nounwind
628    cleanupret from %0 unwind label %lpad.catch
629
630  lpad.catch:                                       ; preds = %lpad.cleanup, %entry
631    %1 = catchswitch within none [label %catch.body] unwind label %lpad.terminate
632
633  catch.body:                                       ; preds = %lpad.catch
634    %catch = catchpad within %1 [ptr @"??_R0H@8", i32 0, ptr %e]
635    invoke void @"?may_throw@@YAXXZ"()
636            to label %invoke.cont.3 unwind label %lpad.terminate
637
638  invoke.cont.3:                                    ; preds = %catch.body
639    %3 = load i32, ptr %e, align 4
640    catchret from %catch to label %return
641
642  lpad.terminate:                                   ; preds = %catch.body, %lpad.catch
643    cleanuppad within none []
644    call void @"?terminate@@YAXXZ"()
645    unreachable
646  }
647
648Funclet parent tokens
649-----------------------
650
651In order to produce tables for EH personalities that use funclets, it is
652necessary to recover the nesting that was present in the source. This funclet
653parent relationship is encoded in the IR using tokens produced by the new "pad"
654instructions. The token operand of a "pad" or "ret" instruction indicates which
655funclet it is in, or "none" if it is not nested within another funclet.
656
657The ``catchpad`` and ``cleanuppad`` instructions establish new funclets, and
658their tokens are consumed by other "pad" instructions to establish membership.
659The ``catchswitch`` instruction does not create a funclet, but it produces a
660token that is always consumed by its immediate successor ``catchpad``
661instructions. This ensures that every catch handler modelled by a ``catchpad``
662belongs to exactly one ``catchswitch``, which models the dispatch point after a
663C++ try.
664
665Here is an example of what this nesting looks like using some hypothetical
666C++ code:
667
668.. code-block:: c
669
670  void f() {
671    try {
672      throw;
673    } catch (...) {
674      try {
675        throw;
676      } catch (...) {
677      }
678    }
679  }
680
681.. code-block:: text
682
683  define void @f() #0 personality i8* bitcast (i32 (...)* @__CxxFrameHandler3 to i8*) {
684  entry:
685    invoke void @_CxxThrowException(i8* null, %eh.ThrowInfo* null) #1
686            to label %unreachable unwind label %catch.dispatch
687
688  catch.dispatch:                                   ; preds = %entry
689    %0 = catchswitch within none [label %catch] unwind to caller
690
691  catch:                                            ; preds = %catch.dispatch
692    %1 = catchpad within %0 [i8* null, i32 64, i8* null]
693    invoke void @_CxxThrowException(i8* null, %eh.ThrowInfo* null) #1
694            to label %unreachable unwind label %catch.dispatch2
695
696  catch.dispatch2:                                  ; preds = %catch
697    %2 = catchswitch within %1 [label %catch3] unwind to caller
698
699  catch3:                                           ; preds = %catch.dispatch2
700    %3 = catchpad within %2 [i8* null, i32 64, i8* null]
701    catchret from %3 to label %try.cont
702
703  try.cont:                                         ; preds = %catch3
704    catchret from %1 to label %try.cont6
705
706  try.cont6:                                        ; preds = %try.cont
707    ret void
708
709  unreachable:                                      ; preds = %catch, %entry
710    unreachable
711  }
712
713The "inner" ``catchswitch`` consumes ``%1`` which is produced by the outer
714catchswitch.
715
716.. _wineh-constraints:
717
718Funclet transitions
719-----------------------
720
721The EH tables for personalities that use funclets make implicit use of the
722funclet nesting relationship to encode unwind destinations, and so are
723constrained in the set of funclet transitions they can represent.  The related
724LLVM IR instructions accordingly have constraints that ensure encodability of
725the EH edges in the flow graph.
726
727A ``catchswitch``, ``catchpad``, or ``cleanuppad`` is said to be "entered"
728when it executes.  It may subsequently be "exited" by any of the following
729means:
730
731* A ``catchswitch`` is immediately exited when none of its constituent
732  ``catchpad``\ s are appropriate for the in-flight exception and it unwinds
733  to its unwind destination or the caller.
734* A ``catchpad`` and its parent ``catchswitch`` are both exited when a
735  ``catchret`` from the ``catchpad`` is executed.
736* A ``cleanuppad`` is exited when a ``cleanupret`` from it is executed.
737* Any of these pads is exited when control unwinds to the function's caller,
738  either by a ``call`` which unwinds all the way to the function's caller,
739  a nested ``catchswitch`` marked "``unwinds to caller``", or a nested
740  ``cleanuppad``\ 's ``cleanupret`` marked "``unwinds to caller"``.
741* Any of these pads is exited when an unwind edge (from an ``invoke``,
742  nested ``catchswitch``, or nested ``cleanuppad``\ 's ``cleanupret``)
743  unwinds to a destination pad that is not a descendant of the given pad.
744
745Note that the ``ret`` instruction is *not* a valid way to exit a funclet pad;
746it is undefined behavior to execute a ``ret`` when a pad has been entered but
747not exited.
748
749A single unwind edge may exit any number of pads (with the restrictions that
750the edge from a ``catchswitch`` must exit at least itself, and the edge from
751a ``cleanupret`` must exit at least its ``cleanuppad``), and then must enter
752exactly one pad, which must be distinct from all the exited pads.  The parent
753of the pad that an unwind edge enters must be the most-recently-entered
754not-yet-exited pad (after exiting from any pads that the unwind edge exits),
755or "none" if there is no such pad.  This ensures that the stack of executing
756funclets at run-time always corresponds to some path in the funclet pad tree
757that the parent tokens encode.
758
759All unwind edges which exit any given funclet pad (including ``cleanupret``
760edges exiting their ``cleanuppad`` and ``catchswitch`` edges exiting their
761``catchswitch``) must share the same unwind destination.  Similarly, any
762funclet pad which may be exited by unwind to caller must not be exited by
763any exception edges which unwind anywhere other than the caller.  This
764ensures that each funclet as a whole has only one unwind destination, which
765EH tables for funclet personalities may require.  Note that any unwind edge
766which exits a ``catchpad`` also exits its parent ``catchswitch``, so this
767implies that for any given ``catchswitch``, its unwind destination must also
768be the unwind destination of any unwind edge that exits any of its constituent
769``catchpad``\s.  Because ``catchswitch`` has no ``nounwind`` variant, and
770because IR producers are not *required* to annotate calls which will not
771unwind as ``nounwind``, it is legal to nest a ``call`` or an "``unwind to
772caller``\ " ``catchswitch`` within a funclet pad that has an unwind
773destination other than caller; it is undefined behavior for such a ``call``
774or ``catchswitch`` to unwind.
775
776Finally, the funclet pads' unwind destinations cannot form a cycle.  This
777ensures that EH lowering can construct "try regions" with a tree-like
778structure, which funclet-based personalities may require.
779
780Exception Handling support on the target
781=================================================
782
783In order to support exception handling on particular target, there are a few
784items need to be implemented.
785
786* CFI directives
787
788  First, you have to assign each target register with a unique DWARF number.
789  Then in ``TargetFrameLowering``'s ``emitPrologue``, you have to emit `CFI
790  directives <https://sourceware.org/binutils/docs/as/CFI-directives.html>`_
791  to specify how to calculate the CFA (Canonical Frame Address) and how register
792  is restored from the address pointed by the CFA with an offset. The assembler
793  is instructed by CFI directives to build ``.eh_frame`` section, which is used
794  by th unwinder to unwind stack during exception handling.
795
796* ``getExceptionPointerRegister`` and ``getExceptionSelectorRegister``
797
798  ``TargetLowering`` must implement both functions. The *personality function*
799  passes the *exception structure* (a pointer) and *selector value* (an integer)
800  to the landing pad through the registers specified by ``getExceptionPointerRegister``
801  and ``getExceptionSelectorRegister`` respectively. On most platforms, they
802  will be GPRs and will be the same as the ones specified in the calling convention.
803
804* ``EH_RETURN``
805
806  The ISD node represents the undocumented GCC extension ``__builtin_eh_return (offset, handler)``,
807  which adjusts the stack by offset and then jumps to the handler. ``__builtin_eh_return``
808  is used in GCC unwinder (`libgcc <https://gcc.gnu.org/onlinedocs/gccint/Libgcc.html>`_),
809  but not in LLVM unwinder (`libunwind <https://clang.llvm.org/docs/Toolchain.html#unwind-library>`_).
810  If you are on the top of ``libgcc`` and have particular requirement on your target,
811  you have to handle ``EH_RETURN`` in ``TargetLowering``.
812
813If you don't leverage the existing runtime (``libstdc++`` and ``libgcc``),
814you have to take a look on `libc++ <https://libcxx.llvm.org/>`_ and
815`libunwind <https://clang.llvm.org/docs/Toolchain.html#unwind-library>`_
816to see what have to be done there. For ``libunwind``, you have to do the following
817
818* ``__libunwind_config.h``
819
820  Define macros for your target.
821
822* ``include/libunwind.h``
823
824  Define enum for the target registers.
825
826* ``src/Registers.hpp``
827
828  Define ``Registers`` class for your target, implement setter and getter functions.
829
830* ``src/UnwindCursor.hpp``
831
832  Define ``dwarfEncoding`` and ``stepWithCompactEncoding`` for your ``Registers``
833  class.
834
835* ``src/UnwindRegistersRestore.S``
836
837  Write an assembly function to restore all your target registers from the memory.
838
839* ``src/UnwindRegistersSave.S``
840
841  Write an assembly function to save all your target registers on the memory.
842