This document is the central repository for all information pertaining to exception handling in LLVM. It describes the format that LLVM exception handling information takes, which is useful for those interested in creating front-ends or dealing directly with the information. Further, this document provides specific examples of what exception handling information is used for in C/C++.
Exception handling for most programming languages is designed to recover from conditions that rarely occur during general use of an application. To that end, exception handling should not interfere with the main flow of an application's algorithm by performing checkpointing tasks, such as saving the current pc or register state.
The Itanium ABI Exception Handling Specification defines a methodology for providing outlying data in the form of exception tables without inlining speculative exception handling code in the flow of an application's main algorithm. Thus, the specification is said to add "zero-cost" to the normal execution of an application.
A more complete description of the Itanium ABI exception handling runtime support of can be found at Itanium C++ ABI: Exception Handling. A description of the exception frame format can be found at Exception Frames, with details of the DWARF 3 specification at DWARF 3 Standard. A description for the C++ exception table formats can be found at Exception Handling Tables.
Setjmp/Longjmp (SJLJ) based exception handling uses LLVM intrinsics llvm.eh.sjlj.setjmp and llvm.eh.sjlj.longjmp to handle control flow for exception handling.
For each function which does exception processing, be it try/catch blocks or cleanups, that function registers itself on a global frame list. When exceptions are being unwound, the runtime uses this list to identify which functions need processing.
Landing pad selection is encoded in the call site entry of the function context. The runtime returns to the function via llvm.eh.sjlj.longjmp, where a switch table transfers control to the appropriate landing pad based on the index stored in the function context.
In contrast to DWARF exception handling, which encodes exception regions and frame information in out-of-line tables, SJLJ exception handling builds and removes the unwind frame context at runtime. This results in faster exception handling at the expense of slower execution when no exceptions are thrown. As exceptions are, by their nature, intended for uncommon code paths, DWARF exception handling is generally preferred to SJLJ.
When an exception is thrown in LLVM code, the runtime does its best to find a handler suited to processing the circumstance.
The runtime first attempts to find an exception frame corresponding to the function where the exception was thrown. If the programming language (e.g. C++) supports exception handling, the exception frame contains a reference to an exception table describing how to process the exception. If the language (e.g. C) does not support exception handling, or if the exception needs to be forwarded to a prior activation, the exception frame contains information about how to unwind the current activation and restore the state of the prior activation. This process is repeated until the exception is handled. If the exception is not handled and no activations remain, then the application is terminated with an appropriate error message.
Because different programming languages have different behaviors when handling exceptions, the exception handling ABI provides a mechanism for supplying personalities. An exception handling personality is defined by way of a personality function (e.g. __gxx_personality_v0 in C++), which receives the context of the exception, an exception structure containing the exception object type and value, and a reference to the exception table for the current function. The personality function for the current compile unit is specified in a common exception frame.
The organization of an exception table is language dependent. For C++, an exception table is organized as a series of code ranges defining what to do if an exception occurs in that range. Typically, the information associated with a range defines which types of exception objects (using C++ type info) that are handled in that range, and an associated action that should take place. Actions typically pass control to a landing pad.
A landing pad corresponds to the code found in the catch portion of a try/catch sequence. When execution resumes at a landing pad, it receives the exception structure and a selector corresponding to the type of exception thrown. The selector is then used to determine which catch should actually process the exception.
At the time of this writing, only C++ exception handling support is available in LLVM. So the remainder of this document will be somewhat C++-centric.
From the C++ developers perspective, exceptions are defined in terms of the throw and try/catch statements. In this section we will describe the implementation of LLVM exception handling in terms of C++ examples.
Languages that support exception handling typically provide a throw operation to initiate the exception process. Internally, a throw operation breaks down into two steps. First, a request is made to allocate exception space for an exception structure. This structure needs to survive beyond the current activation. This structure will contain the type and value of the object being thrown. Second, a call is made to the runtime to raise the exception, passing the exception structure as an argument.
In C++, the allocation of the exception structure is done by the __cxa_allocate_exception runtime function. The exception raising is handled by __cxa_throw. The type of the exception is represented using a C++ RTTI structure.
A call within the scope of a try statement can potentially raise an exception. In those circumstances, the LLVM C++ front-end replaces the call with an invoke instruction. Unlike a call, the invoke has two potential continuation points: where to continue when the call succeeds as per normal; and where to continue if the call raises an exception, either by a throw or the unwinding of a throw.
The term used to define a the place where an invoke continues after an exception is called a landing pad. LLVM landing pads are conceptually alternative function entry points where an exception structure reference and a type info index are passed in as arguments. The landing pad saves the exception structure reference and then proceeds to select the catch block that corresponds to the type info of the exception object.
Two LLVM intrinsic functions are used to convey information about the landing pad to the back end.
Once the landing pad has the type info selector, the code branches to the code for the first catch. The catch then checks the value of the type info selector against the index of type info for that catch. Since the type info index is not known until all the type info have been gathered in the backend, the catch code will call the llvm.eh.typeid.for intrinsic to determine the index for a given type info. If the catch fails to match the selector then control is passed on to the next catch. Note: Since the landing pad will not be used if there is no match in the list of type info on the call to llvm.eh.selector, then neither the last catch nor catch all need to perform the check against the selector.
Finally, the entry and exit of catch code is bracketed with calls to __cxa_begin_catch and __cxa_end_catch.
Note: a rethrow from within the catch may replace this call with a __cxa_rethrow.
A cleanup is extra code which needs to be run as part of unwinding a scope. C++ destructors are a prominent example, but other languages and language extensions provide a variety of different kinds of cleanup. In general, a landing pad may need to run arbitrary amounts of cleanup code before actually entering a catch block. To indicate the presence of cleanups, a landing pad's call to llvm.eh.selector should end with the argument i32 0; otherwise, the unwinder will not stop at the landing pad if there are no catches or filters that require it to.
Do not allow a new exception to propagate out of the execution of a cleanup. This can corrupt the internal state of the unwinder. Different languages describe different high-level semantics for these situations: for example, C++ requires that the process be terminated, whereas Ada cancels both exceptions and throws a third.
When all cleanups have completed, if the exception is not handled by the current function, resume unwinding by calling the llvm.eh.resume intrinsic, passing in the results of llvm.eh.exception and llvm.eh.selector for the original landing pad.
C++ allows the specification of which exception types can be thrown from a function. To represent this a top level landing pad may exist to filter out invalid types. To express this in LLVM code the landing pad will call llvm.eh.selector. The arguments are a reference to the exception structure, a reference to the personality function, the length of the filter expression (the number of type infos plus one), followed by the type infos themselves. llvm.eh.selector will return a negative value if the exception does not match any of the type infos. If no match is found then a call to __cxa_call_unexpected should be made, otherwise _Unwind_Resume. Each of these functions requires a reference to the exception structure. Note that the most general form of an llvm.eh.selector call can contain any number of type infos, filter expressions and cleanups (though having more than one cleanup is pointless). The LLVM C++ front-end can generate such llvm.eh.selector calls due to inlining creating nested exception handling scopes.
The unwinder delegates the decision of whether to stop in a call frame to that call frame's language-specific personality function. Not all personalities functions guarantee that they will stop to perform cleanups: for example, the GNU C++ personality doesn't do so unless the exception is actually caught somewhere further up the stack. When using this personality to implement EH for a language that guarantees that cleanups will always be run, be sure to indicate a catch-all in the llvm.eh.selector call rather than just cleanups.
In order for inlining to behave correctly, landing pads must be prepared to handle selector results that they did not originally advertise. Suppose that a function catches exceptions of type A, and it's inlined into a function that catches exceptions of type B. The inliner will update the selector for the inlined landing pad to include the fact that B is caught. If that landing pad assumes that it will only be entered to catch an A, it's in for a rude surprise. Consequently, landing pads must test for the selector results they understand and then resume exception propagation with the llvm.eh.resume intrinsic if none of the conditions match.
LLVM uses several intrinsic functions (name prefixed with "llvm.eh") to provide exception handling information at various points in generated code.
i8* %llvm.eh.exception()
This intrinsic returns a pointer to the exception structure.
i32 %llvm.eh.selector(i8*, i8*, ...)
This intrinsic is used to compare the exception with the given type infos, filters and cleanups.
llvm.eh.selector takes a minimum of three arguments. The first argument is the reference to the exception structure. The second argument is a reference to the personality function to be used for this try catch sequence. Each of the remaining arguments is either a reference to the type info for a catch statement, a filter expression, or the number zero representing a cleanup. The exception is tested against the arguments sequentially from first to last. The result of the llvm.eh.selector is a positive number if the exception matched a type info, a negative number if it matched a filter, and zero if it matched a cleanup. If nothing is matched, or if only a cleanup is matched, different personality functions may or may not cause control to stop at the landing pad; see the restrictions for more information. If a type info matched then the selector value is the index of the type info in the exception table, which can be obtained using the llvm.eh.typeid.for intrinsic.
If a landing pad containing a call to llvm.eh.selector is inlined into an invoke instruction, the selector arguments for the outer landing pad are appended to those of the inlined landing pad. Consequently, landing pads must be written to ignore selector values that they did not originally advertise.
i32 %llvm.eh.typeid.for(i8*)
This intrinsic returns the type info index in the exception table of the current function. This value can be used to compare against the result of llvm.eh.selector. The single argument is a reference to a type info.
void %llvm.eh.resume(i8*, i32) noreturn
This intrinsic is used to resume propagation of an exception after landing at a landing pad. The first argument should be the result of llvm.eh.exception for that landing pad, and the second argument should be the result of llvm.eh.selector. When a call to this intrinsic is inlined into an invoke, the call is transformed into a branch to the invoke's unwind destination, using its arguments in place of the calls to llvm.eh.exception and llvm.eh.selector there.
This intrinsic is not implicitly nounwind; calls to it will always throw. It may not be invoked.
i32 %llvm.eh.sjlj.setjmp(i8*)
The SJLJ exception handling uses this intrinsic to force register saving for the current function and to store the address of the following instruction for use as a destination address by llvm.eh.sjlj.longjmp. The buffer format and the overall functioning of this intrinsic is compatible with the GCC __builtin_setjmp implementation, allowing code built with the two compilers to interoperate.
The single parameter is a pointer to a five word buffer in which the calling context is saved. The front end places the frame pointer in the first word, and the target implementation of this intrinsic should place the destination address for a llvm.eh.sjlj.longjmp in the second word. The following three words are available for use in a target-specific manner.
void %llvm.eh.sjlj.setjmp(i8*)
The llvm.eh.sjlj.longjmp intrinsic is used to implement __builtin_longjmp() for SJLJ style exception handling. The single parameter is a pointer to a buffer populated by llvm.eh.sjlj.setjmp. The frame pointer and stack pointer are restored from the buffer, then control is transferred to the destination address.
i8* %llvm.eh.sjlj.lsda()
Used for SJLJ based exception handling, the llvm.eh.sjlj.lsda intrinsic returns the address of the Language Specific Data Area (LSDA) for the current function. The SJLJ front-end code stores this address in the exception handling function context for use by the runtime.
void %llvm.eh.sjlj.callsite(i32)
For SJLJ based exception handling, the llvm.eh.sjlj.callsite intrinsic identifies the callsite value associated with the following invoke instruction. This is used to ensure that landing pad entries in the LSDA are generated in the matching order.
void %llvm.eh.sjlj.dispatchsetup(i32)
For SJLJ based exception handling, the llvm.eh.sjlj.dispatchsetup intrinsic is used by targets to do any unwind-edge setup they need. By default, no action is taken.
There are two tables that are used by the exception handling runtime to determine which actions should take place when an exception is thrown.
An exception handling frame eh_frame is very similar to the unwind frame used by dwarf debug info. The frame contains all the information necessary to tear down the current frame and restore the state of the prior frame. There is an exception handling frame for each function in a compile unit, plus a common exception handling frame that defines information common to all functions in the unit.
Todo - Table details here.
An exception table contains information about what actions to take when an exception is thrown in a particular part of a function's code. There is one exception table per function except leaf routines and functions that have only calls to non-throwing functions will not need an exception table.
Todo - Table details here.