LLVM 2.6 Release Notes

This document contains the release notes for the LLVM Compiler Infrastructure, release 2.6. Here we describe the status of LLVM, including major improvements from the previous release and significant known problems. All LLVM releases may be downloaded from the LLVM releases web site.

For more information about LLVM, including information about the latest release, please check out the main LLVM web site. If you have questions or comments, the LLVM Developer's Mailing List is a good place to send them.

Note that if you are reading this file from a Subversion checkout or the main LLVM web page, this document applies to the next release, not the current one. To see the release notes for a specific release, please see the releases page.

The LLVM 2.6 distribution currently consists of code from the core LLVM repository (which roughly includes the LLVM optimizers, code generators and supporting tools), the Clang repository and the llvm-gcc repository. In addition to this code, the LLVM Project includes other sub-projects that are in development. Here we include updates on these subprojects.

The Clang project is an effort to build a set of new 'LLVM native' front-end technologies for the C family of languages. LLVM 2.6 is the first release to officially include Clang, and it provides a production quality C and Objective-C compiler. If you are interested in fast compiles and good diagnostics, we encourage you to try it out. Clang currently compiles typical Objective-C code 3x faster than GCC and compiles C code about 30% faster than GCC at -O0 -g (which is when the most pressure is on the frontend).

In addition to supporting these languages, C++ support is also well under way, and mainline Clang is able to parse the libstdc++ 4.2 headers and even codegen simple apps. If you are interested in Clang C++ support or any other Clang feature, we strongly encourage you to get involved on the Clang front-end mailing list.

In the LLVM 2.6 time-frame, the Clang team has made many improvements:

C and Objective-C support are now considered production quality.
AuroraUX, FreeBSD and OpenBSD are now supported.
Most of Objective-C 2.0 is now supported with the GNU runtime.
Many many bugs are fixed and lots of features have been added.

Previously announced in the 2.4 and 2.5 LLVM releases, the Clang project also includes an early stage static source code analysis tool for automatically finding bugs in C and Objective-C programs. The tool performs checks to find bugs that occur on a specific path within a program.

In the LLVM 2.6 time-frame, the analyzer core has undergone several important improvements and cleanups and now includes a new Checker interface that is intended to eventually serve as a basis for domain-specific checks. Further, in addition to generating HTML files for reporting analysis results, the analyzer can now also emit bug reports in a structured XML format that is intended to be easily readable by other programs.

The set of checks performed by the static analyzer continues to expand, and future plans for the tool include full source-level inter-procedural analysis and deeper checks such as buffer overrun detection. There are many opportunities to extend and enhance the static analyzer, and anyone interested in working on this project is encouraged to get involved!

The VMKit project is an implementation of a JVM and a CLI Virtual Machine (Microsoft .NET is an implementation of the CLI) using LLVM for static and just-in-time compilation.

VMKit version 0.26 builds with LLVM 2.6 and you can find it on its web page. The release includes bug fixes, cleanup and new features. The major changes are:

A new llcj tool to generate shared libraries or executables of Java files.
Cooperative garbage collection.
Fast subtype checking (paper from Click et al [JGI'02]).
Implementation of a two-word header for Java objects instead of the original three-word header.
Better Java specification-compliance: division by zero checks, stack overflow checks, finalization and references support.

The new LLVM compiler-rt project is a simple library that provides an implementation of the low-level target-specific hooks required by code generation and other runtime components. For example, when compiling for a 32-bit target, converting a double to a 64-bit unsigned integer is compiled into a runtime call to the "__fixunsdfdi" function. The compiler-rt library provides highly optimized implementations of this and other low-level routines (some are 3x faster than the equivalent libgcc routines).

All of the code in the compiler-rt project is available under the standard LLVM License, a "BSD-style" license.

The new LLVM KLEE project is a symbolic execution framework for programs in LLVM bitcode form. KLEE tries to symbolically evaluate "all" paths through the application and records state transitions that lead to fault states. This allows it to construct testcases that lead to faults and can even be used to verify algorithms. For more details, please see the OSDI 2008 paper about KLEE.

The goal of DragonEgg is to make gcc-4.5 act like llvm-gcc without requiring any gcc modifications whatsoever. DragonEgg is a shared library (dragonegg.so) that is loaded by gcc at runtime. It uses the new gcc plugin architecture to disable the GCC optimizers and code generators, and schedule the LLVM optimizers and code generators (or direct output of LLVM IR) instead. Currently only Linux and Darwin are supported, and only on x86-32 and x86-64. It should be easy to add additional unix-like architectures and other processor families. In theory it should be possible to use DragonEgg with any language supported by gcc, however only C and Fortran work well for the moment. Ada and C++ work to some extent, while Java, Obj-C and Obj-C++ are so far entirely untested. Since gcc-4.5 has not yet been released, neither has DragonEgg. To build DragonEgg you will need to check out the development versions of gcc, llvm and DragonEgg from their respective subversion repositories, and follow the instructions in the DragonEgg README.

The LLVM Machine Code (MC) Toolkit project is a (very early) effort to build better tools for dealing with machine code, object file formats, etc. The idea is to be able to generate most of the target specific details of assemblers and disassemblers from existing LLVM target .td files (with suitable enhancements), and to build infrastructure for reading and writing common object file formats. One of the first deliverables is to build a full assembler and integrate it into the compiler, which is predicted to substantially reduce compile time in some scenarios.

In the LLVM 2.6 timeframe, the MC framework has grown to the point where it can reliably parse and pretty print (with some encoding information) a darwin/x86 .s file successfully, and has the very early phases of a Mach-O assembler in progress. Beyond the MC framework itself, major refactoring of the LLVM code generator has started. The idea is to make the code generator reason about the code it is producing in a much more semantic way, rather than a textual way. For example, the code generator now uses MCSection objects to represent section assignments, instead of text strings that print to .section directives.

MC is an early and ongoing project that will hopefully continue to lead to many improvements in the code generator and build infrastructure useful for many other situations.

An exciting aspect of LLVM is that it is used as an enabling technology for a lot of other language and tools projects. This section lists some of the projects that have already been updated to work with LLVM 2.6.

Rubinius is an environment for running Ruby code which strives to write as much of the core class implementation in Ruby as possible. Combined with a bytecode interpreting VM, it uses LLVM to optimize and compile ruby code down to machine code. Techniques such as type feedback, method inlining, and uncommon traps are all used to remove dynamism from ruby execution and increase performance.

Since LLVM 2.5, Rubinius has made several major leaps forward, implementing a counter based JIT, type feedback and speculative method inlining.

MacRuby is an implementation of Ruby on top of core Mac OS X technologies, such as the Objective-C common runtime and garbage collector and the CoreFoundation framework. It is principally developed by Apple and aims at enabling the creation of full-fledged Mac OS X applications.

MacRuby uses LLVM for optimization passes, JIT and AOT compilation of Ruby expressions. It also uses zero-cost DWARF exceptions to implement Ruby exception handling.

Pure is an algebraic/functional programming language based on term rewriting. Programs are collections of equations which are used to evaluate expressions in a symbolic fashion. Pure offers dynamic typing, eager and lazy evaluation, lexical closures, a hygienic macro system (also based on term rewriting), built-in list and matrix support (including list and matrix comprehensions) and an easy-to-use C interface. The interpreter uses LLVM as a backend to JIT-compile Pure programs to fast native code.

Pure versions 0.31 and later have been tested and are known to work with LLVM 2.6 (and continue to work with older LLVM releases >= 2.3 as well).

LDC is an implementation of the D Programming Language using the LLVM optimizer and code generator. The LDC project works great with the LLVM 2.6 release. General improvements in this cycle have included new inline asm constraint handling, better debug info support, general bug fixes and better x86-64 support. This has allowed some major improvements in LDC, getting it much closer to being as fully featured as the original DMD compiler from DigitalMars.

Roadsend PHP (rphp) is an open source implementation of the PHP programming language that uses LLVM for its optimizer, JIT and static compiler. This is a reimplementation of an earlier project that is now based on LLVM.

Unladen Swallow is a branch of Python intended to be fully compatible and significantly faster. It uses LLVM's optimization passes and JIT compiler.

LLVM-Lua uses LLVM to add JIT and static compiling support to the Lua VM. Lua bytecode is analyzed to remove type checks, then LLVM is used to compile the bytecode down to machine code.

IcedTea provides a harness to build OpenJDK using only free software build tools and to provide replacements for the not-yet free parts of OpenJDK. One of the extensions that IcedTea provides is a new JIT compiler named Shark which uses LLVM to provide native code generation without introducing processor-dependent code.

This release includes a huge number of bug fixes, performance tweaks and minor improvements. Some of the major improvements and new features are listed in this section.

LLVM 2.6 includes several major new capabilities:

New compiler-rt, KLEE and machine code toolkit sub-projects.
Debug information now includes line numbers when optimizations are enabled. This allows statistical sampling tools like OProfile and Shark to map samples back to source lines.
LLVM now includes new experimental backends to support the MSP430, SystemZ and BlackFin architectures.
LLVM supports a new Gold Linker Plugin which enables support for transparent link-time optimization on ELF targets when used with the Gold binutils linker.
LLVM now supports doing optimization and code generation on multiple threads. Please see the LLVM Programmer's Manual for more information.
LLVM now has experimental support for embedded metadata in LLVM IR, though the implementation is not guaranteed to be final and the .bc file format may change in future releases. Debug info does not yet use this format in LLVM 2.6.

LLVM IR has several new features for better support of new targets and that expose new optimization opportunities:

The add, sub and mul instructions have been split into integer and floating point versions (like divide and remainder), introducing new fadd, fsub, and fmul instructions.
The add, sub and mul instructions now support optional "nsw" and "nuw" bits which indicate that the operation is guaranteed to not overflow (in the signed or unsigned case, respectively). This gives the optimizer more information and can be used for things like C signed integer values, which are undefined on overflow.
The sdiv instruction now supports an optional "exact" flag which indicates that the result of the division is guaranteed to have a remainder of zero. This is useful for optimizing pointer subtraction in C.
The getelementptr instruction now supports arbitrary integer index values for array/pointer indices. This allows for better code generation on 16-bit pointer targets like PIC16.
The getelementptr instruction now supports an "inbounds" optimization hint that tells the optimizer that the pointer is guaranteed to be within its allocated object.
LLVM now support a series of new linkage types for global values which allow for better optimization and new capabilities:
- linkonce_odr and weak_odr have the same linkage semantics as the non-"odr" linkage types. The difference is that these linkage types indicate that all definitions of the specified function are guaranteed to have the same semantics. This allows inlining templates functions in C++ but not inlining weak functions in C, which previously both got the same linkage type.
- available_externally is a new linkage type that gives the optimizer visibility into the definition of a function (allowing inlining and side effect analysis) but that does not cause code to be generated. This allows better optimization of "GNU inline" functions, extern templates, etc.
- linker_private is a new linkage type (which is only useful on Mac OS X) that is used for some metadata generation and other obscure things.
Finally, target-specific intrinsics can now return multiple values, which is useful for modeling target operations with multiple results.

In addition to a large array of minor performance tweaks and bug fixes, this release includes a few major enhancements and additions to the optimizers:

The Scalar Replacement of Aggregates pass has many improvements that allow it to better promote vector unions, variables which are memset, and much more strange code that can happen to do bitfield accesses to register operations. An interesting change is that it now produces "unusual" integer sizes (like i1704) in some cases and lets other optimizers clean things up.
The Loop Strength Reduction pass now promotes small integer induction variables to 64-bit on 64-bit targets, which provides a major performance boost for much numerical code. It also promotes shorts to int on 32-bit hosts, etc. LSR now also analyzes pointer expressions (e.g. getelementptrs), as well as integers.
The GVN pass now eliminates partial redundancies of loads in simple cases.
The Inliner now reuses stack space when inlining similar arrays from multiple callees into one caller.
LLVM includes a new experimental Static Single Information (SSI) construction pass.

LLVM has a new "EngineBuilder" class which makes it more obvious how to set up and configure an ExecutionEngine (a JIT or interpreter).
The JIT now supports generating more than 16M of code.
When configured with --with-oprofile, the JIT can now inform OProfile about JIT'd code, allowing OProfile to get line number and function name information for JIT'd functions.
When "libffi" is available, the LLVM interpreter now uses it, which supports calling almost arbitrary external (natively compiled) functions.
Clients of the JIT can now register a 'JITEventListener' object to receive callbacks when the JIT emits or frees machine code. The OProfile support uses this mechanism.

We have put a significant amount of work into the code generator infrastructure, which allows us to implement more aggressive algorithms and make it run faster:

The llc -asm-verbose option (exposed from llvm-gcc as -dA and clang as -fverbose-asm or -dA) now adds a lot of useful information in comments to the generated .s file. This information includes location information (if built with -g) and loop nest information.
The code generator now supports a new MachineVerifier pass which is useful for finding bugs in targets and codegen passes.
The Machine LICM is now enabled by default. It hoists instructions out of loops (such as constant pool loads, loads from read-only stubs, vector constant synthesization code, etc.) and is currently configured to only do so when the hoisted operation can be rematerialized.
The Machine Sinking pass is now enabled by default. This pass moves side-effect free operations down the CFG so that they are executed on fewer paths through a function.
The code generator now performs "stack slot coloring" of register spills, which allows spill slots to be reused. This leads to smaller stack frames in cases where there are lots of register spills.
The register allocator has many improvements to take better advantage of commutable operations, various spiller peephole optimizations, and can now coalesce cross-register-class copies.
Tblgen now supports multiclass inheritance and a number of new string and list operations like !(subst), !(foreach), !car, !cdr, !null, !if, !cast. These make the .td files more expressive and allow more aggressive factoring of duplication across instruction patterns.
Target-specific intrinsics can now be added without having to hack VMCore to add them. This makes it easier to maintain out-of-tree targets.
The instruction selector is better at propagating information about values (such as whether they are sign/zero extended etc.) across basic block boundaries.
The SelectionDAG datastructure has new nodes for representing buildvector and vector shuffle operations. This makes operations and pattern matching more efficient and easier to get right.
The Prolog/Epilog Insertion Pass now has experimental support for performing the "shrink wrapping" optimization, which moves spills and reloads around in the CFG to avoid doing saves on paths that don't need them.
LLVM includes new experimental support for writing ELF .o files directly from the compiler. It works well for many simple C testcases, but doesn't support exception handling, debug info, inline assembly, etc.
Targets can now specify register allocation hints through MachineRegisterInfo::setRegAllocationHint. A regalloc hint consists of hint type and physical register number. A hint type of zero specifies a register allocation preference. Other hint type values are target specific which are resolved by TargetRegisterInfo::ResolveRegAllocHint. An example is the ARM target which uses register hints to request that the register allocator provide an even / odd register pair to two virtual registers.

New features of the X86 target include:

SSE 4.2 builtins are now supported.
GCC-compatible soft float modes are now supported, which are typically used by OS kernels.
X86-64 now models implicit zero extensions better, which allows the code generator to remove a lot of redundant zexts. It also models the 8-bit "H" registers as subregs, which allows them to be used in some tricky situations.
X86-64 now supports the "local exec" and "initial exec" thread local storage model.
The vector forms of the icmp and fcmp instructions now select to efficient SSE operations.
Support for the win64 calling conventions have improved. The primary missing feature is support for varargs function definitions. It seems to work well for many win64 JIT purposes.
The X86 backend has preliminary support for mapping address spaces to segment register references. This allows you to write GS or FS relative memory accesses directly in LLVM IR for cases where you know exactly what you're doing (such as in an OS kernel). There are some known problems with this support, but it works in simple cases.
The X86 code generator has been refactored to move all global variable reference logic to one place (X86Subtarget::ClassifyGlobalReference) which makes it easier to reason about.

New features of the PIC16 target include:

Support for floating-point, indirect function calls, and passing/returning aggregate types to functions.
The code generator is able to generate debug info into output COFF files.
Support for placing an object into a specific section or at a specific address in memory.

Things not yet supported:

Variable arguments.
Interrupts/programs.

New features of the ARM target include:

Preliminary support for processors, such as the Cortex-A8 and Cortex-A9, that implement version v7-A of the ARM architecture. The ARM backend now supports both the Thumb2 and Advanced SIMD (Neon) instruction sets.
The AAPCS-VFP "hard float" calling conventions are also supported with the -float-abi=hard flag.
The ARM calling convention code is now tblgen generated instead of resorting to C++ code.

These features are still somewhat experimental and subject to change. The Neon intrinsics, in particular, may change in future releases of LLVM. ARMv7 support has progressed a lot on top of tree since 2.6 branched.

New features of other targets include:

Mips now supports O32 Calling Convention.
Many improvements to the 32-bit PowerPC SVR4 ABI (used on powerpc-linux) support, lots of bugs fixed.
Added support for the 64-bit PowerPC SVR4 ABI (used on powerpc64-linux). Needs more testing.

This release includes a number of new APIs that are used internally, which may also be useful for external clients.

New PrettyStackTrace class allows crashes of llvm tools (and applications that integrate them) to provide more detailed indication of what the compiler was doing at the time of the crash (e.g. running a pass). At the top level for each LLVM tool, it includes the command line arguments.
New StringRef and Twine classes make operations on character ranges and string concatenation to be more efficient. StringRef is just a const char* with a length, Twine is a light-weight rope.
LLVM has new WeakVH, AssertingVH and CallbackVH classes, which make it easier to write LLVM IR transformations. WeakVH is automatically drops to null when the referenced Value is deleted, and is updated across a replaceAllUsesWith operation. AssertingVH aborts the program if the referenced value is destroyed while it is being referenced. CallbackVH is a customizable class for handling value references. See ValueHandle.h for more information.
The new 'Triple ' class centralizes a lot of logic that reasons about target triples.
The new ' llvm_report_error()' set of APIs allows tools to embed the LLVM optimizer and backend and recover from previously unrecoverable errors.
LLVM has new abstractions for atomic operations and reader/writer locks.
LLVM has new SourceMgr and SMLoc classes which implement caret diagnostics and basic include stack processing for simple parsers. It is used by tablegen, llvm-mc, the .ll parser and FileCheck.

Other miscellaneous features include:

LLVM now includes a new internal 'FileCheck' tool which allows writing much more accurate regression tests that run faster. Please see the FileCheck section of the Testing Guide for more information.
LLVM profile information support has been significantly improved to produce correct use counts, and has support for edge profiling with reduced runtime overhead. Combined, the generated profile information is both more correct and imposes about half as much overhead (2.6. from 12% to 6% overhead on SPEC CPU2000).
The C bindings (in the llvm/include/llvm-c directory) include many newly supported APIs.
LLVM 2.6 includes a brand new experimental LLVM bindings to the Ada2005 programming language.
The LLVMC driver has several new features:
- Dynamic plugins now work on Windows.
- New option property: init. Makes possible to provide default values for options defined in plugins (interface to cl::init).
- New example: Skeleton, shows how to create a standalone LLVMC-based driver.
- New example: mcc16, a driver for the PIC16 toolchain.

If you're already an LLVM user or developer with out-of-tree changes based on LLVM 2.5, this section lists some "gotchas" that you may run into upgrading from the previous release.

The Itanium (IA64) backend has been removed. It was not actively supported and had bitrotted.
The BigBlock register allocator has been removed, it had also bitrotted.
The C Backend (-march=c) is no longer considered part of the LLVM release criteria. We still want it to work, but no one is maintaining it and it lacks support for arbitrary precision integers and other important IR features.
All LLVM tools now default to overwriting their output file, behaving more like standard unix tools. Previously, this only happened with the '-f' option.
LLVM build now builds all libraries as .a files instead of some libraries as relinked .o files. This requires some APIs like InitializeAllTargets.h.

In addition, many APIs have changed in this release. Some of the major LLVM API changes are:

All uses of hash_set and hash_map have been removed from the LLVM tree and the wrapper headers have been removed.
The llvm/Streams.h and DOUT member of Debug.h have been removed. The llvm::Ostream class has been completely removed and replaced with uses of raw_ostream.
LLVM's global uniquing tables for Types and Constants have been privatized into members of an LLVMContext. A number of APIs now take an LLVMContext as a parameter. To smooth the transition for clients that will only ever use a single context, the new getGlobalContext() API can be used to access a default global context which can be passed in any and all cases where a context is required.
The getABITypeSize methods are now called getAllocSize.
The Add, Sub and Mul operators are no longer overloaded for floating-point types. Floating-point addition, subtraction and multiplication are now represented with new operators FAdd, FSub and FMul. In the IRBuilder API, CreateAdd, CreateSub, CreateMul and CreateNeg should only be used for integer arithmetic now; CreateFAdd, CreateFSub, CreateFMul and CreateFNeg should now be used for floating-point arithmetic.
The DynamicLibrary class can no longer be constructed, its functionality has moved to static member functions.
raw_fd_ostream's constructor for opening a given filename now takes an extra Force argument. If Force is set to false, an error will be reported if a file with the given name already exists. If Force is set to true, the file will be silently truncated (which is the behavior before this flag was added).
SCEVHandle no longer exists, because reference counting is no longer done for SCEV* objects, instead const SCEV* should be used.
Many APIs, notably llvm::Value, now use the StringRef and Twine classes instead of passing const char* or std::string, as described in the Programmer's Manual. Most clients should be unaffected by this transition, unless they are used to Value::getName() returning a string. Here are some tips on updating to 2.6:
- getNameStr() is still available, and matches the old behavior. Replacing getName() calls with this is an safe option, although more efficient alternatives are now possible.
- If you were just relying on getName() being able to be sent to a std::ostream, consider migrating to llvm::raw_ostream.
- If you were using getName().c_str() to get a const char* pointer to the name, you can use getName().data(). Note that this string (as before), may not be the entire name if the name contains embedded null characters.
- If you were using operator + on the result of getName() and treating the result as an std::string, you can either use Twine::str to get the result as an std::string, or could move to a Twine based design.
- isName() should be replaced with comparison against getName() (this is now efficient).
The registration interfaces for backend Targets has changed (what was previously TargetMachineRegistry). For backend authors, see the Writing An LLVM Backend guide. For clients, the notable API changes are:
- TargetMachineRegistry has been renamed to TargetRegistry.
- Clients should move to using the TargetRegistry::lookupTarget() function to find targets.

LLVM is known to work on the following platforms:

Intel and AMD machines (IA32, X86-64, AMD64, EMT-64) running Red Hat Linux, Fedora Core, FreeBSD and AuroraUX (and probably other unix-like systems).
PowerPC and X86-based Mac OS X systems, running 10.3 and above in 32-bit and 64-bit modes.
Intel and AMD machines running on Win32 using MinGW libraries (native).
Intel and AMD machines running on Win32 with the Cygwin libraries (limited support is available for native builds with Visual C++).
Sun x86 and AMD64 machines running Solaris 10, OpenSolaris 0906.
Alpha-based machines running Debian GNU/Linux.

The core LLVM infrastructure uses GNU autoconf to adapt itself to the machine and operating system on which it is built. However, minor porting may be required to get LLVM to work on new platforms. We welcome your portability patches and reports of successful builds or error messages.

This section contains significant known problems with the LLVM system, listed by component. If you run into a problem, please check the LLVM bug database and submit a bug if there isn't already one.

The llvm-gcc bootstrap will fail with some versions of binutils (e.g. 2.15) with a message of "Error: can not do 8 byte pc-relative relocation" when building C++ code. We intend to fix this on mainline, but a workaround for 2.6 is to upgrade to binutils 2.17 or later.
LLVM will not correctly compile on Solaris and/or OpenSolaris using the stock GCC 3.x.x series 'out the box', See: Broken versions of GCC and other tools. However, A Modern GCC Build for x86/x86-64 has been made available from the third party AuroraUX Project that has been meticulously tested for bootstrapping LLVM & Clang.

The following components of this LLVM release are either untested, known to be broken or unreliable, or are in early development. These components should not be relied on, and bugs should not be filed against them, but they may be useful to some people. In particular, if you would like to work on one of these components, please contact us on the LLVMdev list.

The MSIL, Alpha, SPU, MIPS, PIC16, Blackfin, MSP430 and SystemZ backends are experimental.
The llc "-filetype=asm" (the default) is the only supported value for this option. The ELF writer is experimental.
The implementation of Andersen's Alias Analysis has many known bugs.

The X86 backend does not yet support all inline assembly that uses the X86 floating point stack. It supports the 'f' and 't' constraints, but not 'u'.
The X86 backend generates inefficient floating point code when configured to generate code for systems that don't have SSE2.
Win64 code generation wasn't widely tested. Everything should work, but we expect small issues to happen. Also, llvm-gcc cannot build the mingw64 runtime currently due to several bugs and due to lack of support for the 'u' inline assembly constraint and for X87 floating point inline assembly.
The X86-64 backend does not yet support the LLVM IR instruction va_arg. Currently, the llvm-gcc and front-ends support variadic argument constructs on X86-64 by lowering them manually.

The Linux PPC32/ABI support needs testing for the interpreter and static compilation, and lacks support for debug information.

Support for the Advanced SIMD (Neon) instruction set is still incomplete and not well tested. Some features may not work at all, and the code quality may be poor in some cases.
Thumb mode works only on ARMv6 or higher processors. On sub-ARMv6 processors, thumb programs can crash or produce wrong results (PR1388).
Compilation for ARM Linux OABI (old ABI) is supported but not fully tested.

The SPARC backend only supports the 32-bit SPARC ABI (-m32); it does not support the 64-bit SPARC ABI (-m64).

64-bit MIPS targets are not supported yet.

On 21164s, some rare FP arithmetic sequences which may trap do not have the appropriate nops inserted to ensure restartability.

The C backend has only basic support for inline assembly code.
The C backend violates the ABI of common C++ programs, preventing intermixing between C++ compiled by the CBE and C++ code compiled with llc or native compilers.
The C backend does not support all exception handling constructs.
The C backend does not support arbitrary precision integers.

The only major language feature of GCC not supported by llvm-gcc is the __builtin_apply family of builtins. However, some extensions are only supported on some targets. For example, trampolines are only supported on some targets (these are used when you take the address of a nested function).

If you run into GCC extensions which are not supported, please let us know.

The C++ front-end is considered to be fully tested and works for a number of non-trivial programs, including LLVM itself, Qt, Mozilla, etc.

Exception handling works well on the X86 and PowerPC targets. Currently only Linux and Darwin targets are supported (both 32 and 64 bit).

Fortran support generally works, but there are still several unresolved bugs in Bugzilla. Please see the tools/gfortran component for details.

The llvm-gcc 4.2 Ada compiler works fairly well; however, this is not a mature technology, and problems should be expected.

The Ada front-end currently only builds on X86-32. This is mainly due to lack of trampoline support (pointers to nested functions) on other platforms. However, it also fails to build on X86-64 which does support trampolines.
The Ada front-end fails to bootstrap. This is due to lack of LLVM support for setjmp/longjmp style exception handling, which is used internally by the compiler. Workaround: configure with --disable-bootstrap.
The c380004, c393010 and cxg2021 ACATS tests fail (c380004 also fails with gcc-4.2 mainline). If the compiler is built with checks disabled then c393010 causes the compiler to go into an infinite loop, using up all system memory.
Some GCC specific Ada tests continue to crash the compiler.
The -E binder option (exception backtraces) does not work and will result in programs crashing if an exception is raised. Workaround: do not use -E.
Only discrete types are allowed to start or finish at a non-byte offset in a record. Workaround: do not pack records or use representation clauses that result in a field of a non-discrete type starting or finishing in the middle of a byte.
The lli interpreter considers 'main' as generated by the Ada binder to be invalid. Workaround: hand edit the file to use pointers for argv and envp rather than integers.
The -fstack-check option is ignored.